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.1003986 | Hippocampal Remapping Is Constrained by Sparseness rather than Capacity | Grid cells in the medial entorhinal cortex encode space with firing fields that are arranged on the nodes of spatial hexagonal lattices. Potential candidates to read out the space information of this grid code and to combine it with other sensory cues are hippocampal place cells. In this paper, we investigate a population of grid cells providing feed-forward input to place cells. The capacity of the underlying synaptic transformation is determined by both spatial acuity and the number of different spatial environments that can be represented. The codes for different environments arise from phase shifts of the periodical entorhinal cortex patterns that induce a global remapping of hippocampal place fields, i.e., a new random assignment of place fields for each environment. If only a single environment is encoded, the grid code can be read out at high acuity with only few place cells. A surplus in place cells can be used to store a space code for more environments via remapping. The number of stored environments can be increased even more efficiently by stronger recurrent inhibition and by partitioning the place cell population such that learning affects only a small fraction of them in each environment. We find that the spatial decoding acuity is much more resilient to multiple remappings than the sparseness of the place code. Since the hippocampal place code is sparse, we thus conclude that the projection from grid cells to the place cells is not using its full capacity to transfer space information. Both populations may encode different aspects of space.
| The mammalian brain represents space in the population of hippocampal place cells as well as in the population of medial entorhinal cortex grid cells. Since both populations are active at the same time, space information has to be synchronized between the two. Both brain areas are reciprocally connected, and it is unclear how the two codes influence each other. In this paper, we analyze a theoretical model of how a place code processes inputs from the grid cell population. The model shows that the sparseness of the place code poses a much stronger constraint than maximal information transfer. We thus conclude that the potentially high spatial acuity of the grid code cannot be efficiently conveyed to a sparse place cell population and thus propose that sparseness and spatial acuity are two independent objectives of the neuronal place representation.
| The neuronal representation of space that is necessary for navigation and orientation has been traditionally assigned to the hippocampal place cell system [1], where cells fire only at few distinct locations and are silent elsewhere. Since the discovery of grid cells in the medial entorhinal cortex (MEC) [2], [3], which fire on a hexagonal spatial lattice, a second space representation is now known and it has become unclear what the functional differences of the two are. It is speculated that the MEC grid cells are predominantly used in path integration, whereas the place cells may connect position and context information [4]. From the coding perspective it is remarkable that the hippocampal place fields are considerably sparse, whereas the grid fields generate a much denser code with approximately one third of all grid cells active at any one time [3]. Since both networks are reciprocally connected anatomically [5], [6] and functionally [7], [8], the two space representations have to be synchronized. Understanding the interplay of both codes thus leads to the more general question of how a dense neuronal code can be efficiently transferred into a sparse code and vice versa.
In this paper, we focus on the mapping from grid to place cells. This extends previous coding approaches in so far as they studied the isolated grid cell system from a mainly information theoretic perspective [9], [10]. Here, we discuss a coding theory by including the further constraint that the grid code has to be readable by the place code at a similar and behaviorally relevant resolution, since we assume that space information is only relevant for the brain if it can be read out by other neurons. Employing two population models, for grid cells and place cells, we show that a relevant resolution of the order of centimeters can be easily transferred from a relatively small grid-cell to a relatively small place-cell population. Larger numbers (particularly of place cells) can thus be used to encode multiple environments [11] at a similar spatial resolution. Our model also shows that may interference owing to multiple environments reduces the sparseness of the hippocampal code much faster than it reduces the space information of the population patterns measured by the number of different environments that can be encoded at a given spatial resolution. These findings argue against a pure feed-forward model of place field formation from grid cells, consistent with recent experimental findings [7], [12]–[16].
Here we briefly summarize the general structure of our model, whereas a detailed account is provided in the Materials and Methods Section. A population of grid cells is connected to place cells via a feed-forward synaptic matrix. The grid cells are organized in four modules that differ in the spatial period (or grid spacing) of the periodic hexagonal firing patterns [17]. The neuronal activities of the MEC and hippocampal populations are assumed to encode either linear tracks or square boxes both of length 1 m (Figs. 1 and 2). Different environments are represented by phase shifts of the grid fields that are identical for all cells in a module [18] but random between modules [19].
The spike count of the grid cells is assumed to follow Poisson statistics. For the place cells we first define place fields that optimally cover the whole environment but are only used as teacher patterns in a training step in which we construct synaptic weights between grid cells and place cells by supervised Hebbian learning. The teacher place fields are randomly assigned in each environment (shuffling of place cells) resembling the global remapping [20] of hippocampal place fields found in experiments. For each such remapping synaptic weights are incremented according to the Hebb rule such that all shifted grid patterns activate the corresponding remapped place code.
Realizations of grid field spikes are projected via the learned feed-forward connections to the place field population that employs a soft winner-take-all mechanism (E-MAX rule) to emulate recurrent inhibition [21]. The activity from these simulations determines the actual firing fields and spike statistics of the place cells. The spatial acuity of both codes is measured by the empirical minimum mean square decoding error of single trial activity. The simulations are evaluated by a variety of measures including sparseness and the similarity between the place fields used during training and those obtained in the simulation.
The capacity of a spatial code consists of two components. First, the spatial resolution [9], or how precisely one can infer a spatial position. Second, how many different environments can be represented. Since different environments are obtained by MEC phase shifts and hippocampal remapping, all spatial information is conveyed by the same synaptic connections. Thus the multiple stored environments interfere at the cost of spatial resolution.
To assess the ground truth of our model, we first evaluate the coding capacity of the grid cell population on a one-dimensional linear track (Fig. 3). The spatial resolution (denoted as root-mean square estimation error; RMSE) non-trivially depends on the tuning width of the grid code and the number of neurons [9], [22]. Three examples of grid codes are shown in Fig. 3A–C for three different values of . Grids as usually observed in MEC are most similar to the situation in Fig. 3B, whereas Fig. 3A and C illustrate settings with extremely thin and broad tuning curves, respectively. Thus, the biological value of is about 1, which corresponds to a ratio between tuning width and spatial period of about (see Fig. S4 of [3]). However, the RMSE non-monotonically depends on [22] with a minimum at rather thin tuning curves (Fig. 3D).
The resolution (RMSE) improves with such that even for moderate cell numbers (several hundreds) it is easy to obtain spatial resolutions in the range of 1 mm and below. From a behavioral perspective, however, one may ask whether such a resolution is actually psychophysically reasonable, or even useful. We thus suggest that resolution is probably not the major objective of the grid code and test the alternative possibility that the grid code may be designed to display a reasonable spatial resolution in as many environments as possible. As a lower bound for such a reasonable resolution we postulate an RMSE of 0.5 cm (dashed line in Fig. 3D) and ask the question, which parameter setting in -space would actually result in this behaviorally relevant RMSE (Fig. 3E). The minimum scales supra-linearly with , i.e. it flattens out for smaller . We thus argue that is a good choice because it still is in the super-linear regime requiring only relatively small cell numbers and at the same time results in tuning widths that are similar to biology (like Fig. 3B). For further analysis we thus fix the grid code to and .
The spatial acuity of the population code of grid cells can only be made use of if it can be read out by downstream centers. We therefore asked under which conditions the resolution of grid cell network from the previous subsection can be preserved in the place cell network under the ideal conditions that only one environment has to be represented (number of environments ); Fig. 4.
Since the tuning curves are actually learned there exists a clear lower bound for the tuning widths that reflects the minimal width of the grid cell population (Fig. 4A–F). Narrower place fields cannot be achieved by the present model even if the fields used during training are much narrower than the smallest grid fields. Similar as for the grid cell code, a reduction in the place field width effectively improves the RMSE, however, the resolution is limited by that of the grid code (0.5 cm). Therefore an increase in the number of place cells reduces the RMSE and the performance quickly converges to the minimum for ; Fig. 4G. Only relatively few neurons are needed to achieve such a behaviorally relevant resolution, and thus we next asked how many different environments can be represented at this resolution.
Storing multiple environments generates interferences of the place codes since each remapping taxes synaptic resources. Thus the spatial resolution of the place code is getting worse when storing multiple environments (Fig. 5). However, even for 21 remappings in our parameter regime () the decoding error is still relatively low (). Also the number of remapped environments for which decoding is possible increases with the number of place cells (Fig. 6A), such that even for moderate place cell numbers many environments can be easily decoded at physiological resolution.
Although space information is retained for considerably large values of , the place code degenerates already for much smaller . This degeneration is best described by a loss of sparseness (Fig. 6B, [23]) resulting from less localized firing fields, while the average spike count remains constant (see Materials and Methods). This delocalization results in a reduction of the number of proper place cells (Fig. 6C) which exhibit an increased number of regular-sized firing fields (Fig. 6D, E) before they cease to be place cells and are active over almost the whole track as indicated by a mean population sparseness (average fraction of active cells at a position) close to 1 (Fig. 6F). Also the firing fields quickly loose their similarity to the trained firing fields (Fig. 6G). From these observations we conclude that although a large number of putative place cells allow to reliably decode a large number of environments by remapping, the place field quality (i.e. the sparseness) of the encoding neurons disappears. Thus the observation of a sparse place code in the hippocampus must result from further objectives beyond decoding quality and remapping capacity.
To test whether these observations are specific to the one-dimensional paradigm, we repeated the same simulations and analysis for a two-dimensional enclosure (see Materials and Methods and Fig. 2). As in the one-dimensional case, inspection of single examples for high numbers of remappings reveals that the place-selectivity of the readout neurons (the putative place cells) deteriorates much faster than the decoding quality (Fig. 7). Even random spatial patches (for ; Fig. 7 B) allow for almost perfect decoding (Fig. 7 E). Spatial estimation only breaks down, if hardly any space modulation is observable in the firing patterns (Fig. 7 C, F). These exemplary observations are corroborated by a systematic quantitative assessment of the code and the firing fields in Fig. 8.
In analogy to the one-dimensional case, decoding quality increases with the number of putative place cells and remains in the centimeter range for 40 and more remappings if (Fig. 8A). At the same time, the place field characteristics deteriorate with increasing as was described in the one-dimensional case (Fig. 6): sparseness decreases (Fig. 8B, F), place field number increases before no clear place fields are visible anymore (Fig. 8C, D, E), place fields loose their similarity to the trained patterns (Fig. 8G).
In the two-dimensional case for few place cells , we observe an improvement in resolution when going from one to about 10 remappings before the decoding error again increases with . Although counter-intuitive, this effect reflects that an increase in mean population sparseness at first provides a better coverage of the square box. To make the model work also for small , the number of place cells has to be large to overcome this finite size effect. It therefore imposes a constraint on a minimum number of . This effect also exemplifies that decoding RMSE depends on many different aspects and thus it is generally difficult to use it as a single measure for comparing the "quality" of a population code.
We also assessed the robustness of our findings with respect to essential model parameters. We evaluated the place code for different number of grid cells , while keeping a constant total number of input spikes and found essentially no difference (S1 Figure). Also, a mere increase in the number of place field spikes only improves the spatial resolution but does not alter any of the other place field characteristics (S2 Figure).
A substantial effect on the population code can be observed by altering the strength of feedback inhibition in the place field population by means of the E% value (Fig. 9). This parameter determines the firing threshold as the input strength E% below the maximum (see Methods and [21]). The E% value directly controls the sparseness of the code (Fig. 9B–G). For low E% values (sparse codes) and low numbers of environments, we again observe the finite size effect of high RMSE, which then improves with increasing (Fig. 9A). This initially high RMSE, however, can again be compensated for by using larger numbers of place cells (as in Fig. 8 A). As a result, the decreasing E% generally allows to store more environments, however, at the cost of high to achieve a sufficiently small RMSE for low .
If one constrains the parameter space to biologically realistic mean population sparseness values for the hippocampal place fields about to (Supporting Information of [24] and [25], see Discussion) our simulations of the standard parameter regime (Fig. 8) show that such a regular place code can only be observed for up to about ten environments. Also for increased E% value the number of sparsely encoded environments is only increased to several tens (Fig. 9). A major factor limiting the number of environments is that in our model the synapses to the place cells are updated in each remapping, i.e., the place cells experience maximal interference. One can considerably extend the number of remappings for a given sparseness if the synaptic changes from different remappings are distributed to varying subsets of place cells, thereby increasing the overall number of putative place cells (partial learning). This strategy is motivated by an experimental report showing that only a small subset of CA1 pyramidal cells shows intracellular determinants for being recruited as a place cell in a novel environment [26]. We illustrate the benefits of partial learning by a further set of simulations in which the synaptic weights to only a fraction of the place cells are updated in each individual remapping (partial learning; Fig. 10). Using mean population sparseness as a criterion for the breakdown of the place code, partial learning increases the number of possible remappings (Fig. 10A) to over a hundred. As a measure for capacity, one can define a critical number of environments at which the mean population sparseness exceeds a (biologically motivated) threshold value of (see Discussion). This critical only weakly increases with the number of place fields but strongly decreases with increasing fraction of partial learning (Fig. 10B, C).
In rat hippocampus the number of CA1 neurons is in the order of several 100 thousands and thus according to Fig. 10B, a sparse place representation may still be consistent with storing hundreds to thousands of remappings if each place cell is involved in only a small fraction of environments.
The encoding acuity (RMSE) is generally not affected by partial learning as long as is not too small (Fig. 10D). Only for very small values of , when a winner-take-all effect of the E%-MAX rule decreases sparseness for , spatial acuity deteriorates. However, this regime is biologically unrealistic, since there the number of neurons encoding an environment tends to zero.
The geometry of the spatial firing patterns (place field size and number), is virtually unaffected by (Fig. 10 D, E). The place field sizes we find in the model (up to 0.05 m2) are within the range reported in the experimental literature [25], [27], the mean number of place fields (about 3) is at the upper bound of the fields per m2 experimentally found in the hippocampus and dentate gyrus [24], [27], which indicates that the place code might in fact even be sparser than than the threshold motivated by current experimental data (see Discussion).
The hippocampal formation hosts two space representations. A sparse one in the hippocampus proper, in which the neurons have a limited number of distinct firing fields (place fields) and a dense one in the MEC, where grid cells exhibit multiple firing fields located on the nodes of a hexagonal lattice. If both brain regions encode the unique physical spatial position of the animal, the two codes have to be coherent. Anatomically both brain areas are reciprocally connected [5]–[8] and thus place cell activity will influence grid cell activity and vice versa.
In this paper, we focus on the connections from the medial entorhinal grid cells to the hippocampus, which anatomically correspond to the perforant pathway and the temporo-ammonic pathway. These pathways have initially been thought to predominantly underly the transformation from grid to place cells [19], [28]–[32]. More recently, developmental studies [12], [13] and pharmacological interventions that block grid cell firing [7], [14]–[16], have shown that place cells can also be observed independently of grid-field firing (but see [33]). Thus, while the MEC-to-hippocampus connections seem to be unnecessary to generate place fields, they are likely important in synchronizing both codes. This view is further corroborated by the observation that place cell firing is less stable if MEC input is eliminated [34].
Although it is known from information theory that capacity and sparseness cannot be maximized simultaneously [35], [36], our paper exemplifies this rule for a specific neuronal network example, in that it shows that maximization of capacity of MEC-to-hippocampal connections destroys the sparseness of the hippocampal place code.
From the theoretical perspective, if the synaptic matrix is know that transforms one code into another, reading out a dense code is more difficult than reading out a sparse code. This is because the synaptic matrix gives rise to a much noisier postsynaptic signal for dense input patterns [37]. Therefore the transformation from place cells to grid cells is less problematic than the other way round. The grid to place transformation provides an interesting test case to study information transfer between different brain areas in general.
Our model is largely based on experimental reports of grid and place cell remapping [18], [20], [38]–[40]. While place cells turn on, turn off, or show random relocation during global remapping [40], grid fields shift and rotate. In our model, we consider only shifts, since rotations were shown to be less efficient for remapping previously [19]. Although the grid modules seem to operate functionally independent [17], it is not yet clear whether the modules remap independently as proposed in [19]. A further finding from [19] was that a few () modules suffice for strong remapping and data [17] suggest that MEC has only about 5 to 9 modules. Only a part of these modules innervate any one place cell, owing to the dorso-ventrally ordered topography of the input fibers. We therefore concluded that a biologically reasonable number of modules influencing any single place cell is about 4. We further assume that the number of cells per module is constant, which is optimal from a theoretical perspective [9] but might not necessarily be the case [17].
To connect our simulations to hippocampal physiology, we assume a population sparseness value of . This value can be estimated by combining data from the supporting information (Table S1 of [24]) (mean number of place cells: 1.1/(0.8 m)2 for CA3, 2/(0.8 m)2 for DG; percentage of place fields: 62/71 for CA3, 41/44 for DG) and place field areas measured in [25] in a circular enclosure of diameter 76 cm (field area: 0.08 m2 for CA3, 0.06 m2 for DG). The estimate of the population sparseness for a 1 m2 enclosure (as in our simulations) thus follows from the product of these three values, i.e., we obtain about 0.12 for CA3 and for DG. However, in our simulations, a sparseness value of yields a number of place fields per place cell that is slightly higher than observed in experiments, and thus the above numbers may over-estimate the sparseness values in the real rodent brain.
Previous coding theories of MEC grid cells have extensively investigated spatial resolution. According to [9], [41], hierarchical grid codes outperform place codes by far in terms of their scaling behavior. A main reason is that for a constant resolution, the number of place cells scales with area, whereas for grid cells only those with larger period have to be scaled up with area for disambiguation, however, the resolution mostly rests on the smallest grid periodicity and thus the size of the population with small periodicity is independent of spatial range to be encoded. The parameter regimes in which grid codes are particularly superior to place codes provide relative root mean square errors in the range of and even far below [9]. For a one meter environment, this would correspond to (sub-)millimeter resolution which is biologically irrelevant for encoding but might be important for MEC models of path integration [42], [43] where errors can accumulate over time. In the regime used for the present model (Figs. 3 and 4), the surplus in resolution of the grid code is relatively small, consistent with a biologically relevant decoding situation of high noise and few modules [44].
A further noteworthy result of our simulations is that a population code still contains almost maximal space information (in terms of minimal RMSE), even if no clear spatial firing fields can be delineated anymore. On the one hand this shows that also brain areas like the lateral entorhinal cortex [45] and the subiculum [46] with only weakly space-modulated individual neurons can provide high-resolution space information on the population level and thus a superposition of such weakly modulated firing fields via synaptic inputs is sufficient to provide place information to any downstream structure. This means that also the hippocampus and the MEC may not generate their strongly spatially modulated firing fields de-novo but inherit them from weakly modulated populations as e.g. the lateral entorhinal cortex. On the other hand our findings show that sparseness of the hippocampal place representation is not due to coding precision requirements but must serve other purposes. Manifold advantages of sparseness have been proposed [47] including energy efficiency [48]. A further classical benefit of sparse representations arises for auto-associative memory networks, where it facilitates memory retrieval due to reduced interference [37], [49]–[52].
Although our model includes lateral inhibition via the E% rule to limit the overall network activity the network cannot enforce sparseness except for unrealistically low values of . So it is still possible that other assumptions about the recurrent connections may enforce sparseness more effectively, while allowing remappings. For example, in a model using a combination of recurrent excitation and inhibition [53], [54] place fields arise from stable attractor states, where each attractor reflects the topology of place field positions for one remapping. The capacity (number of remappings per neuron) of this autoassociator is in the range of few percent and, thus for may end up slightly above the capacity derived from our model () (for fixed realistic sparseness). So, recurrent excitatory connections between place cells can potentially help to keep the place fields compact. The disadvantage of attractor-like solutions is that they show catastrophic forgetting, whereas our model exhibits a gradual decline of the order parameters (Figs. 6, 8 and 9).
The view on how space information is communicated between the reciprocally connected brain areas hippocampus and MEC has recently undergone a dramatic change from a completely feed-forward grid-to-place dogma [19], [28]–[32] to an almost reversed place-to-grid picture [7], [12]–[16]. We started out under the assumption that the spatial precision in the hippocampus mostly relies on inputs from MEC grid cells and remapping the MEC triggers remapping on the hippocampus. If this was the only function of the MEC-to-hippocampus connections, they should be filled with as much space information as possible and the representation would no longer be sparse. Our results thus show that functionally the classical pure grid-to-place hypothesis would only suboptimally use the coding resources. The required compact place fields and the MEC-to-hippocampus synapses thus do not seem to be optimized to transfer space information.
Since new experimental data [7], [12]–[16] show that MEC is actually not essential for generating place cells, our findings suggest the possibility that hippocampal space information might actually primarily stem from other regions than the MEC. The grid field input to place fields thus likely imposes only modulatory or stabilizing effects. Conversely, no grid cells have been so far observed without place cell activity, and thus the place-to-grid hypothesis is still a possible candidate. However, it is unclear why hexagonal symmetry might emerge from the perspective of a transformation of a sparse place code to a dense code, and thus it might as well be that the two codes are generated independently for different computational purposes and the reciprocal connections are only required for synchronization and stabilization.
The grid cells are modeled as Poisson spikers with firing maps that denote the mean spike count of cell conditioned on the position on a 1 meter track. All cells have the same maximal spike count and the same field width parameter . The cells differ in their spatial periods and grid phases . The specific model for the cells' Poisson spike counts follows a von Mises function:
Each cell belongs to one of modules. Cells in a module share a spatial period . The phases in each module are chosen equidistantly such that the firing fields cover the linear track; Fig. 1A.
Though we have only one width parameter for all cells, the tuning width for the cells in one specific module scales with the period , as can be seen from expanding the cosine term in .
The spike count is adjusted such that the whole grid cell population generates a constant given number of spikes averaged over all positions and cells , i.e.,(1)
Here, the locations are discretized in bins . The value used for is 1.5 spikes per cell. Since for Poisson spikers the spike count is a product of averaging duration, firing rate and number of cells with the same rate function , the three factors cannot be distinguished. Although, for simplicity, we call the number of grid cells, it is more correctly referred to as the number of grid cell channels (different rate functions ).
The different modules are defined by their grid period . In our grid cell population, the first module is assigned the largest spatial period, which we take such that each cell in this module only has one unique firing field on the track. The smaller periods of the other modules are obtained via geometric progression, , with a period ratio , and . The period ratio is defined via the number of modules and the smallest period , which is set to 30 cm, a lower bound suggested by experiments [3], [17]. Thus the only remaining degrees of freedom for the grid code are the number of modules, the width constant and the mean spike count per length . We choose , and unless otherwise mentioned.
The synaptic weights of the feed forward connections from grid to place cells are set by Hebbian learning based on the rate maps of the grid cells from eq. (1) and the desired rate maps(2)of the place cells with width and centers that uniformly cover the interval ; Fig. 1C.
With these idealized place fields, the weights are calculated according to outer product (Hebbian) rule: using discretized locations we define(3)
The denominator ensures that connections to place cells with fields at the borders are as strong as the ones to centered place fields.
The two networks (grid and place cells) are supposed to encode environments. Each environment has a new grid code generated by shifting each module's phases by a constant . These shifts have experimentally been shown to be coherent within one module [18] and have been theoretically proposed to be uncorrelated between modules [19]. The shifted grid field patterns are denoted by . A new place code is generated by randomly choosing the place field centers . Hebbian learning as in eq. 3 is repeated times and weights are added.
The place cell spikes for cell at a position are produced by drawing Poisson spikes for the grid cells, then taking the weighted sumof those, to yield a membrane potential of the place cells. The activity is then generated following the E%-MAX rule [21], that emulates the effect of recurrent inhibition: after finding the maximum membrane potential , all are set to zero and the ones above this threshold are multiplied with a constant , and used as place cell firing rate from which spike counts are derived according to Poisson statistics.
Decoding the place code via a minimum mean square estimator [55](4)requires a statistical model of place cell firing. Since in the model the single trial spike counts are statistically independent the posterior can be obtained using Bayes' rule,
The prior is taken as constant, . The individual likelihoods are obtained by repeating the above stochastic process 800 times for each cell and each sampled position and sampling the relative frequencies of spike counts . This distribution is then fitted with a bimodal model function consisting of a probability of cell not firing, and probability of firing spikes following a normal distribution with fit parameters mean and variance :(5)
Examples for such fits are shown in Fig. 11. Again, the constant is obtained by fixing the number of spikes per centimeter per cell in an iterative fashion. The resulting value is unless otherwise mentioned.
For comparison we also implemented the model in two spatial dimensions . There, the grid cell's firing maps are set as in [31]with being a unitary vector pointing into direction . Using , and , the three spatial waves add up to a hexagonal firing pattern with spatial period , a maximum at , and orientation (Fig. 2A). The nonlinearity both adjusts the minimal firing rate to zero and matches the spatial decay of the firing rate peaks to experiments [31]. Like for the one-dimensional simulations we use four modules. Cells in one module share spatial period and orientation. The period for the first module is m (larger than the box). The smallest period is set to m. The two intermediate periods are again obtained by geometric progression . Orientation for each module is drawn at random. The "centers" are uniformly distributed over the Wigner cell of size . For all computational purposes, we used spatial bins to discretize the box.
To generate two-dimensional place fields we set feed-forward weights by Hebbian learning, using Gaussian tuning curves as firing maps for place fields as in eq. (2), but with and replaced by their two-dimensional counterparts (Fig. 2 B, C). The centers cover the box uniformly on a square grid. Centers of teacher place fields for cell exceeding the number of nodes on the square lattice were distributed randomly. Weights are then calculated using eq. (3).
The spikes are produced as in the one-dimensional case. Decoding follows eq. (4) with one-dimensional quantities replaced by their two-dimensional counterparts.
For a remapping, each grid cell module is assigned one random spatial shift vector, added to all from that module. The shift is obtained by drawing a vector from the Wigner cell of that module using a uniform distribution (Fig. 2 D). For remapping, the place cells are assigned new centers at random, which again cover the box equidistantly. Then Hebbian learning is repeated, adding to the existing weights (Fig. 2 E, F).
Partial learning as used in the simulations of Fig. 10 was implemented as follows. For each environment we selected a random set of cells such that each cell is selected approximately the same amount of times across environments. This was achieved via random permutations of the cell indices. The sets of cells were taken from such a random index sequence one after the other, and only if less than items were left in the index sequence, a new random permutation was generated.
For each set of selected cells we defined teacher place fields that cover the whole environment as uniformly as possible on a square grid with nodes (see previous section). Hebbian learning according to eq. (3) was applied to only the synapses between the grid field population and the selected set of postsynaptic cells.
By construction, some place cells will be used in more environments than others. We normalize the rows of after all environments have been learned to avoid that the cells that are involved in more environments (and thus have larger weights) are overly excited and exert too much inhibition on the remaining cells via the E-MAX rule.
According to [23], single cell sparseness is defined as , where denotes the firing rate of the specific cell as a function of position and indicates the average over space.
Population sparseness is defined as the percentage of place cells firing above a threshold of of the maximum firing rate at any position.
The number and size of place fields was found by first thresholding the rate maps, discarding all bins below of the maximal rate, and then applying the algorithm by Hoshen and Kopelman [56]. Bins were considered neighboring if they share an edge, hence diagonal bins were not neighbors. Place fields were only included in the analysis (proper place fields) if they were larger than 50 cm2 and smaller than of the total environment.
Learning of place fields was considered successful in a cell if the learned field showed sufficient similarity to the training field according to three criteria: 1) the total area above a threshold of peak rate has to be smaller than , 2) the place field center has to be detected close to the desired location, i.e., no further away than the place field radius (), and 3) the desired place field has to have an area at least twice the size of all other place fields.
|
10.1371/journal.pgen.1004637 | Dominant Sequences of Human Major Histocompatibility Complex Conserved Extended Haplotypes from HLA-DQA2 to DAXX | We resequenced and phased 27 kb of DNA within 580 kb of the MHC class II region in 158 population chromosomes, most of which were conserved extended haplotypes (CEHs) of European descent or contained their centromeric fragments. We determined the single nucleotide polymorphism and deletion-insertion polymorphism alleles of the dominant sequences from HLA-DQA2 to DAXX for these CEHs. Nine of 13 CEHs remained sufficiently intact to possess a dominant sequence extending at least to DAXX, 230 kb centromeric to HLA-DPB1. We identified the regions centromeric to HLA-DQB1 within which single instances of eight “common” European MHC haplotypes previously sequenced by the MHC Haplotype Project (MHP) were representative of those dominant CEH sequences. Only two MHP haplotypes had a dominant CEH sequence throughout the centromeric and extended class II region and one MHP haplotype did not represent a known European CEH anywhere in the region. We identified the centromeric recombination transition points of other MHP sequences from CEH representation to non-representation. Several CEH pairs or groups shared sequence identity in small blocks but had significantly different (although still conserved for each separate CEH) sequences in surrounding regions. These patterns partly explain strong calculated linkage disequilibrium over only short (tens to hundreds of kilobases) distances in the context of a finite number of observed megabase-length CEHs comprising half a population's haplotypes. Our results provide a clearer picture of European CEH class II allelic structure and population haplotype architecture, improved regional CEH markers, and raise questions concerning regional recombination hotspots.
| The human major histocompatibility complex (MHC) is a gene-dense region highly enriched in immune response genes. MHC genetic variation is among the highest in the human genome and is associated with both tissue transplant compatibility and many genetic disorders. Long-range (1–3 Mb) MHC haplotypes of essentially identical DNA sequence at relatively high (≥0.5%) population frequency (“genetic fixity”), called conserved extended haplotypes (CEHs), comprise roughly half of all European population haplotypes. We sequenced an aggregate of 27 kb over 580 kb in the MHC class II region from HLA-DQA2 to DAXX in 158 European haplotypes to quantify the breakdown of this genetic fixity in the centromeric portion of the MHC and to determine the representative nature within that region of eight previously fully or nearly fully sequenced “common” European haplotypes. We identified the dominant sequences of 13 European CEHs and determined where the “common” sequences did (or did not) represent related CEHs. We found patterns of shared sequence identity among different CEHs surrounded by fixed (for each CEH) but differing sequence. Our direct observational results for population haplotypes explain the mutual occurrence of CEHs and short (5–200 kb) blocks of fixed sequence detected by the statistical measure of linkage disequilibrium.
| The human major histocompatibility complex (MHC) is a highly polymorphic genomic region of over 3 Mb on chromosome 6p21. MHC polymorphisms include critical determinants for tissue transplantation success and show strong correlation both with many genetic diseases and with ethnic origin. Haplotype analysis of DNA sequence containing specific allele combinations of two or more nearby genetic loci was first established in the MHC. Many individuals within a human population share a small number of specific MHC haplotypes. These 1 to 3 Mb stretches of nearly identical MHC DNA sequence with high population frequency are called conserved extended haplotypes (CEHs) [1]–[4] or ancestral haplotypes [5], [6]. Virtually all MHC allele-disease associations involve marker alleles of CEHs [2]–[6].
Early work defined CEHs by their alleles at HLA-B, HLA-DR loci and at intermediate MHC genes (i.e., ‘complotypes’ [7]). Later CEH reports extended the core region from HLA-C to HLA-DQB1. With technological refinements, it became clear individual CEHs carry only one allele (or, rarely, a limited number of variants) at any given locus in this region [2]–[6]; [8]–[16] without apparent recombination. Intervening DNA sequence is therefore essentially conserved (i.e., identical) among the population haplotypes comprising each CEH, and CEHs are essentially identical by descent common population haplotypes. CEH sequence conservation has been verified whenever investigated, whether determined by microsatellite, restriction fragment length polymorphism, dense single nucleotide polymorphism (SNP) or partial resequencing analyses of multiple haplotypes from unrelated individuals [9]–[13], [15]. We have previously referred to the existence of such “fixed” CEH alleles [1]–[4], [8], [14] and the intervening sequence conservation of CEHs as “genetic fixity” [3], [4], [8], [14]–[16].
We define genetic, haplotype or sequence “fixity” to be sequence identity and conservation of a large stretch of genomic sequence, shared by a relatively large number of apparently unrelated individuals, without apparent recombination from an ancestral sequence. By “identity” we mean “essential identity,” thus allowing for minor private mutation or microvariation within individual haplotype sequences comprising a particular CEH (or CEH fragment or block). Several studies have described the extension of genetic fixity in some CEHs telomerically in the class I region to HLA-A and centromerically in the class II region to HLA-DPB1 [1]–[6], [8], [14]–[16]. However, little is known about CEH alleles and conserved dominant sequences between HLA-DQB1 and HLA-DPB1 or those centromeric to HLA-DPB1. We sought both to confirm the existence of sequence conservation in multiple examples of a wide variety of CEHs and to identify the dominant sequences centromeric to HLA-DQB1 for the core MHC CEHs most similar to the previously sequenced “common” European haplotypes [17]–[20].
Using consanguineous cell lines, the Wellcome Trust Sanger Institute (http://www.sanger.ac.uk) undertook the MHC Haplotype Project (MHP) [17]–[20] (http://www.ucl.ac.uk/cancer/medical-genomics/mhc) in which eight “common” European MHC haplotypes were fully or nearly fully sequenced over a genomic distance of up to 4.75 Mb. However, no systematic analysis has determined whether these sequences accurately represent the previously established CEHs [1]–[6]. As we argued earlier [4], the extent to which the MHP sequences could be exploited for deciphering genotype-phenotype relationships would require resequencing multiple independent population haplotypes to determine consensus sequence, microvariation, and population representation for each CEH.
Here, we sought to answer two main questions: 1. Are the centromeric portions of the MHP sequences representative of European CEHs? 2. What is the extent of and sequence of a retained dominant sequence centromeric to HLA-DQB1 for each CEH “represented” by a MHP sequence? We compare the classical HLA markers of the eight MHP sequences with the analogous markers of previously reported European CEHs [3], [4], [6], [21]. Then, we describe partial resequencing of the region centromeric to HLA-DQB1 of multiple population haplotypes for each CEH these eight sequences putatively represent. The MHC class II region centromeric to HLA-DQB1 is important both because of its strong association with many complex diseases and its known or suspected recombination hotspots [22], [23]. We document the extent of CEH dominant sequence conservation from HLA-DQA2 to DAXX, and we identify the SNP and deletion-insertion polymorphism (DIP) alleles of the dominant sequence for each CEH. Finally, we identify where MHP sequences accurately represent those dominant sequences, and we discuss several structural and conceptual issues related to local recombination hotspots and linkage disequilibrium (LD).
Five MHP haplotypes contain the markers of a previously reported CEH [3], [4], [6], [21] from HLA-C to HLA-DQB1 (what we define as the “core” or “classical” (for CEH purposes) MHC region). HLA specificities are used in Table 1 to designate each CEH name for historical reasons. The five apparent MHP CEHs are: PGF, the human reference sequence for the MHC, containing the core MHC markers of the CEH “B7,DR15”; COX, representing the CEH “B8,DR3”; QBL, representing the CEH “B18,DR3”; MANN, with the core MHC markers of two “B44,DR7” CEHs (exhibiting both HLA-C*04/C*16 variation and, independently, C4A*0/C4A*3 microvariation); and, DBB, representing the CEH “B57,DR7.”
The other three MHP cell lines do not contain a previously described MHC CEH, and it is therefore uncertain how “common” these haplotypes are in any European population. SSTO contains HLA-DQB1*03:05, a DQ8 specificity allele, but is not a reported CEH. Several reported CEHs contain a DR4,DQ8 specificity block (like SSTO) MHC haplotype, although they all have HLA-DQB1*03:02 alleles. The MCF haplotype exists in our database only once and is not a known CEH [21]. The CEH “B44,DR4,DQ7” (in which there is C4B*Q0/C4B*1 microvariation) is the reported CEH most similar to MCF in the class II region. (HLA-DQB1*03:01 is a DQ7 specificity allele.) Finally, the APD MHC haplotype (HLA-C*06:02, B*40:01, DRB1*13:01, DQB1*06:03) does not exist among our 2675 normal Boston haplotypes [21] and has not been reported to be a CEH. In the core MHC region, the CEH [HLA-C*03:04, B*40:01, SC02, DRB1*13:02, DQB1*06:04] (“B60,DR13”) is the reported CEH [3], [4], [6], [21] most similar to the APD haplotype. (HLA-B60 is the specificity of the HLA-B*40:01 allele in this CEH.)
After aligning MHP sequences [18]–[20], we chose centromeric class II amplicons that would maximize the resultant SNP-DIP haplotypic variation if the MHP sequences were representative of common population haplotypes. We supplemented our initial choices with amplicons to localize transition points where MHP sequences ceased to represent common long-range haplotypes. We designed 56 primer pairs to sequence amplicons in five regions from HLA-DQA2 to DAXX (Figure 1). We resequenced about 27 kb of this 580 kb region. HLA-DQA2 and DAXX are located approximately 80 kb and 660 kb centromeric to HLA-DQB1, respectively. The sequenced amplicons and their polymorphisms we report along with their human genome (assembly 37.p10) locations and rs number designations are given in Table S1 and Table S2. Where known polymorphisms are missing in Table S2 from genomic locations within the boundaries of amplicons described in Table S1, all of the haplotypes we report here matched the human reference sequence.
We determined the phased sequences of 158 population haplotypes primarily through segregation analysis in pedigrees. For each pedigree, we determined between two and four founder haplotypes, varying by the number of unrelated haplotypes identifiable in the subjects available for each pedigree. All haplotypes followed Mendelian inheritance patterns except for rare null alleles and intra-family crossovers. For the latter, as with all pedigrees, we only report the unrelated non-crossover founder haplotypes. We achieved 91% to 100% sequence completion in the amplicons from HLA-DQA2 through HLA-DMA and from RING1 through DAXX, 85% sequence completion for the 158 haplotypes in BRD2 and about 45% to 65% completion from HLA-DOA through HLA-DPB1. Text S1 provides details about the population haplotypes chosen for each CEH group, and Table S3 provides details of the resequencing coverage (i.e., completion) for each haplotype group and region. Table S5 provides the complete phased SNP-DIP sequence data for all 158 haplotypes and the eight MHP haplotypes for a central portion of the covered region from HLA-DOB to BRD2. Table S4 provides the MHP and dominant CEH sequences from HLA-DQA2 to DAXX, including the annotated genomic locations of the SNPs and DIPs. The dominant sequences were taken from Table S5 and analogous data from other regions. The CEH groups are organized in numerical order by their HLA-DR/DQ specificities.
Some current MHC haplotype scaffolds contain sequence gaps for and/or do not extend centromerically to several of the amplicons we sequenced. We sequenced the QBL (Figure S1), MANN (Figure S2) and DBB (Figure S3) haplotypes within those amplicons. The SNP and DIP alleles within those amplicons for the three MHP haplotypes are highlighted both within the sequences shown in Figures S1–S3 and surrounded by yellow borders in the summary Table S4 and in Table S5. These sequences have assigned GenBank accession numbers as described in the Materials and Methods and the Figures.
We compared the sequences of the five MHP haplotypes (PGF, COX, QBL, MANN and DBB) containing markers of specific CEHs in the core MHC with population haplotypes bearing the same CEH markers (Table 1). We also compared both the PGF and MANN haplotypes each with a second CEH with which they shared HLA-DRB1 and HLA-DQB1 alleles. We compared SSTO and MCF with CEHs that shared at least some class II specificity, allele or sequence identity. Because 90% of the 50 DR4,DQ8 haplotypes we sequenced shared identity to SSTO in the HLA-DQA2 and HLA-DQB2 regions, we compared SSTO with the sequences of all six resequenced DR4,DQ8 CEHs (Table 1). We compared the MCF sequence with the only known CEH with which it shares HLA-DR/DQ alleles (please see above). Finally, we compared the APD sequence with the only DR13 population haplotype (out of 13 total) we sequenced having an identical HLA-DQA2 to HLA-DOB sequence. We therefore did not compare the APD sequence with any known CEH because no known CEH shared sequence identity with APD.
The break point (or lack thereof) at which a MHP sequence no longer represented any CEH is shown in Figure 2. The APD cell line is not shown in Figure 2 because it does not represent a reported CEH. Figure 2 displays were based on analyses of sequences shown in Table S4, and those analyses were based on phased data (Table S5 and analogous data in other regions). PGF had the B7,DR15 CEH dominant sequence and QBL had the B18,DR3 CEH dominant sequence throughout the HLA-DQA2 to DAXX region (Figure 2). PGF also represented the B18,DR15 CEH up to a point within the 1 kb region from intron 8 to intron 6 of TAP2 (Table S4 and Table S5), where this CEH's dominant sequence began to differ from that of the B7,DR15 CEH.
The other five MHP sequences represented at least one CEH in class II for variable distances. We determined precisely the centromeric break point for the COX sequence. Twenty-five of the 30 B8,DR3-like haplotypes showed sequence identity to COX from HLA-DQA2 through the DOB9 amplicon within intron 6 of TAP2. However, within the same intron, at rs60045856 (SNP ID #147, Table S2), less than 600 bp centromeric to the DOB9 amplicon, the COX sequence differed from all of the previously identical 25 haplotypes (Table S4 and Table S5). The G allele at rs60045856 appeared to be a regional tag marker of the B8,DR3 CEH in that all 25 haplotypes identical to the COX sequence up to the DOB9 amplicon possessed this allele whereas all of the other 133 haplotypes reported here (as well as COX and the other seven MHP sequences) had the T allele (Table S5). COX never shared significant regional sequence with the dominant B8,DR3 sequence centromeric to TAP2.
MANN represented two related B44,DR7 CEHs through at least a region 2 kb centromeric to HLA-DOB (amplicon DOB5). Approximately 5.5 kb telomeric to TAP2, where the two B44,DR7 CEH dominant sequences diverged, MANN continued to represent the C4,B44,DR7 CEH (despite MANN carrying the HLA-C*16:01-defining allele of the C16,B44,DR7 CEH). Within intron 8 of TAP2, the C4,B44,DR7 dominant sequence split into major and minor variants, and MANN ceased to represent either B44,DR7 CEH dominant sequence, although it continued to be identical to the B49,DR4,DQ8 CEH throughout all of the TAP2 amplicons we sequenced (Table S4 and Table S5). DBB possibly represented the B57,DR7 CEH throughout the class II region in which the CEH maintained a dominant sequence. We could not pinpoint the end of representation by DBB or MCF due to incomplete sequencing (Figure 2; Text S1). However, MCF stopped representing its CEH within the 26.9 kb region at or centromeric to BRD2 (Table S4), telomeric to lost CEH fixity.
Of the seven MHP sequences in Figure 2, SSTO represented its nearest CEH group for the shortest distance. The centromeric break point for SSTO representation of the CEH HLA-B62,SC33,DR4,DQ8 was within a 13.5 kb region between HLA-DQB2 and HLA-DOB. The SSTO centromeric break point for any DR4,DQ8 CEH was narrowed to a different 11.2 kb region between HLA-DQB2 and HLA-DOB (Table S4). However, we can predict the latter break point more precisely using MHP sequence data. Telomeric to rs9276712 (SNP ID #89, amplicon DC13, Table S2), SSTO was identical to the CEHs HLA-B62,SB42,DR4,DQ8, HLA-B60,SC31,DR4,DQ8 and HLA-B38,SC21,DR4,DQ8 and was highly similar to the PGF sequence but differed significantly from the APD sequence. At and centromeric to rs9276712 (for 141 kb, until amplicon DMP1), the SSTO sequence was highly similar to APD but significantly different from PGF (which remained highly similar to the three CEHs). The recombination event that caused this SSTO switch from similarity to PGF to similarity to APD was likely between rs9276712 and rs1158783, the last SNP telomeric to rs9276712 at which the telomeric APD-PGF-SSTO pattern was clear. The distance between the two SNPs is 286 bp in the human reference sequence.
In an attempt to identify SNP/DIP markers near TAP2 differentiating the relatively similar sequences of the B18,DR3 CEH (represented by QBL) and the B44,DR7 CEHs (represented by MANN), we resequenced B44,DR7 haplotypes (including MANN) at the DOB7.5 amplicon (Table 2). Among the eight MHP sequences, only MANN had a T allele at SNP DOB7.5-2 (rs2857100; ID #2, Table 2). We confirmed this by sequencing MANN at amplicon DOB7.5. However, all nine of the 10 B44,DR7 haplotypes we sequenced (including all five essentially identical to MANN telomeric to the DOB7.5 amplicon) had the C allele at rs2857100. We concluded the MANN haplotype had a private mutation at rs2857100. If the T allele exists in other B44,DR7 haplotypes, the frequency is likely to be extremely low. This was the only private SNP allele we found in any MHP sequence within a region in which it otherwise had the dominant sequence of a CEH it represented.
From HLA-DQA2 to HLA-DQB2, all CEHs retained a dominant sequence (i.e., maintained “genetic fixity”) shared by 50 to 100% of the population haplotypes in each group (Figure 3A-C). Most CEHs also retained a dominant sequence by DAXX (660 kb centromeric to HLA-DQB1), although there was usually a gradual reduction in the number of population haplotypes sharing that sequence. CEH dominant sequence results are organized based on the MHP cell line sequences in the order in which they were published. Complete detail of fixity loss for each CEH is provided in Text S1. No results are presented for the DR13,DQ6 CEH because APD did not represent that CEH in the regions we resequenced.
Loss of sequence fixity (Figure 3A–C) is in terms of the gene and/or amplicon at which individual haplotypes stopped sharing SNP/DIP alleles of their CEH dominant sequence. Minor variations, almost always apparently unlinked and isolated private SNPs or DIPs and isolated dominant sequence microvariation, were not counted as loss of sequence fixity (e.g., Table S5). CEH fixity loss is therefore due to past recombination of the dominant sequence with other haplotype sequences. This conclusion is strengthened by our observation that sequences centromeric to the dominant sequence break point are rarely unique and are often found in other CEH groups (e.g., Table S5). We therefore make the explicit assumption that intervening unsequenced regions of population haplotypes sharing a CEH dominant sequence (i.e., telomeric to the break point for any given population haplotype) have similarly limited microvariation and are essentially identical sequences.
Although recombination is the explanation for population haplotype crossover from a CEH dominant sequence (previously identical by descent), it is not possible to calculate CEH recombination rates. The number of meioses experienced by the CEH prior to (or since) its recombination to form any particular population haplotype is unknown. To quantify and display observed population haplotype recombinants responsible for the breakdown of each CEH dominant sequence, we developed a new metric, normalized crossover frequency (NCF, see Materials and Methods ). Our sequencing amplicons and data were not distributed evenly across the region analyzed (Figure 1, Table S1, Table S2, Table S4). We therefore display NCF values, calculated for 11 separate sub-regions (see Materials and Methods and Table S2), on genomic maps drawn to scale (Figure 3D–O). The areas and locations of the bars in those figures quantify and localize the effect of recombination on the loss of CEH fixity displayed in Figure 3A–C.
Sequence recombination is difficult to localize with precision (please see below). Furthermore, the breakdown of a CEH dominant sequence likely varies in different population cohorts. Nevertheless, a few general observations are evident from the results shown in Figure 3. First, the breakdown locations and frequencies vary significantly between different CEHs. Although certain sub-regional crossover sites are more common (e.g., between HLA-DQB1 and HLA-DQA2 (sub-region 1), between HLA-DQB2 and HLA-DOB (sub-region 3), between TAP2 and HLA-DMA (sub-region 6), between BRD2 and HLA-DPB1 (sub-regions 8 and 9), and between HLA-DPB1 and VPS52 (sub-region 10)), none is common to the majority of all analyzed CEHs. Also, some dominant sequences break down gradually in many locations whereas others seem to break down in a more focused fashion. These differences may be due to different relative timelines of CEH expansion and recombination events. Finally, while specific CEHs show a range from high to no sequence conservation through DAXX, most CEHs show approximately 50% dominant sequence retention around BRD2 (in sub-region 7, between HLA-DMB and HLA-DOA). Below, we highlight the results for specific CEHs.
The precise location (and, consequently, quantitation) of historic recombination leading to the breakdown of a dominant sequence is often not definable. Perhaps the clearest example of this is the apparent strong crossover frequency for the B7,DR15 CEH in TAP2 (sub-region 5, Figure 3D). Sequence fixity was maintained for 19 of 23 (83%) B7,DR15 haplotypes from HLA-DQA2 through intron 8 of TAP2 (Figure 3A). Beginning in intron 6 of TAP2, fixity of the B7,DR15 CEH declined to 14 haplotypes (61%), and was maintained through TAP2. The detected crossover of five B7,DR15 haplotypes was to a sequence that defined the B18,DR15 CEH dominant sequence. Thus, the crossovers detected within TAP2 in these five haplotypes could have occurred anywhere within the region shared by the two CEHs. That region extends telomerically past HLA-DRB1. The B18,DR15 CEH dominant sequence from HLA-DQA2 to BRD2 (apparently identical to the B7,DR15 dominant sequence through intron 8 of TAP2) was found in 80% of the population haplotypes we resequenced (Table S4 and Table S5).
B7,DR15 haplotype resequencing centromeric to TAP2 showed a gradual loss in CEH fixity through DAXX (Figure 3A) and declined below 50% near HLA-DMB. The nine B7,DR15 haplotypes identical at HLA-DPB1 (39% of the original 23) either had the unique exon 2 sequence of or were classically typed as HLA-DPB1*04:01. The dominant B7,DR15 CEH sequence from HLA-DQA2 through DAXX was found in 26% of all resequenced B7,DR15 haplotypes.
All 30 B8,DR3 haplotypes were identical at HLA-DQA2. Sequence conservation decreased to 29 haplotypes (97%) at HLA-DQB2, to 87% at 16.4 kb centromeric to HLA-DQB2 (amplicon DC10) and to 25 haplotypes (83%) from HLA-DOB through TAP2 (Figure 3A). B8,DR3 sequence fixity decreased below 50% in the 20 kb region between amplicons DMP10 and DMP11 and declined to 40% at HLA-DOA (sub-region 8, Figure 3E). Seven of the 30 haplotypes (23%) had the dominant sequence and were essentially identical through DAXX. Six of those contained HLA-DPB1*01:01 (the seventh had HLA-DPB1*03:01). (If the haplotype with HLA-DPB1*03:01 is different from the other six (in spite of having essentially identical SNP-DIP alleles from HLA-DQA2 to DAXX), the dominant sequence at DAXX would be represented by only six of 30 (20%) of the studied haplotypes rather than 23%.)
Only 12 of the 18 (67%) B18,DR3 haplotypes showed sequence identity at HLA-DQA2 (Figure 3A). In contrast to this relatively high crossover frequency between HLA-DQB1 and HLA-DQA2 (sub-region 1; Figure 3F), 61% of all 18 haplotypes remained essentially identical to one another and QBL from HLA-DQB2 through BRD2. The only microvariation found among these 11 haplotypes was at the second BRD2 DIP (ID #204; Table S2). By DAXX, nine of the 18 (50%) sequences were still essentially identical.
All 10 B44,DR7 haplotypes shared sequence identity from HLA-DQA2 through 21 kb telomeric to HLA-DOB (amplicon DC13), and 90% of the haplotypes remained essentially identical up to 6.5 kb centromeric to HLA-DOB (Figure 3B). As outlined in Text S1, we studied six C4,B44,DR7 (Figure 3G) and four C16,B44,DR7 (Figure 3H) haplotypes. The dominant sequences of these two CEHs became different near and within TAP2 (Table S4 and Table S5). The DOB6 DIP at 5.5 kb telomeric to TAP2 and the DOB7-2 SNP 2.8 kb telomeric to TAP2 defined this split. The four C16,B44,DR7 examples we sequenced maintained sequence identity to one another for the remaining amplicons in TAP2 and three of these (75%) remained identical through the BRD2 region. The five essentially identical C4,B44,DR7 haplotypes split into two groups within intron 8 of TAP2 (at SNP DOB8.4). The dominant sequence, in three of the haplotypes (50%) appeared to be shared through at least 2 kb telomeric to BRD2. We did not sequence the haplotypes comprising the dominant sequence of either CEH sufficiently to determine the extent (or lack) of sequence identity centromeric to BRD2.
Three of the four (75%) B57,DR7-like haplotypes shared a common sequence from HLA-DQA2 through approximately 16.5 kb centromeric to HLA-DQB2. Sequence fixity declined to two of the four (50%) haplotypes between HLA-DQB2 and HLA-DOB (Figure 3B; sub-region 3, Figure 3I) and continued through BRD2. We did not sequence the two identical haplotypes from HLA-DOA (amplicon DMP11) through HLA-DPB1 (amplicon DMP17), but we began sequencing again just centromeric to RING1 (at amplicon CTB8). At amplicon CTB8, those two haplotypes remained identical to one another. However, 4.5 kb centromeric, at amplicon CTB9, the two haplotypes also differed from one another. Thus, we could only localize the B57,DR7 CEH lost fixity to a 244 kb region between amplicons DMP10 and CTB9 (Figure 2 and Figure 3B).
Although there was a significant loss of CEH sequence fixity between HLA-DQB1 and HLA-DQA2, four of the seven (57%) B44,DR4,DQ7 haplotypes retained identical sequence from within intron 1 of HLA-DQA2 through at least HLA-DOA (Figure 3B and Figure 3J). Within the first intron of HLA-DPA1 (amplicon DMP15), the number of identical CEH sequences decreased to three haplotypes (43%). The haplotype that became different was not sequenced at amplicon DMP14. The three identical haplotypes retained sequence identity to one another through DAXX. The sequence presented in Table S4 contains HLA-DPB1*04:01 and represents the dominant B44,DR4,DQ7 CEH sequence.
Although APD did not represent any known CEH in the resequenced region, APD shared sequence identity with its single represented population haplotype from at least HLA-DQA2 through TAP2, a distance of almost 100 kb. The previously identical sequence differed from the APD sequence at every SNP 2.1 kb telomeric to HLA-DMB but was otherwise identical at amplicons DMP2 through DMP6. Within and near HLA-DMB and HLA-DMA (amplicons DMP1-6), no other haplotype we sequenced had the APD sequence, and only two other haplotypes (both non-standard haplotypes from other groups) had the same sequence as the DR13,DQ6 haplotype we report in Table S4. The APD sequence in amplicons DMP7 through DMP13 has not been reported, but its sequence near and in the HLA-DP genes differed from the other DR13,DQ6 haplotype we report.
Identifying the genetic elements responsible for complex genetic diseases requires knowing the genomic haplotype architecture of the population(s) in which the diseases exist. Toward that goal, the MHP made a major advance in the early to mid-2000's by determining the sequences of eight European Caucasian MHC haplotypes [17]–[20]. However, although the MHP sequences were described as “common” European Caucasian MHC haplotypes, that remains an open question [4]. Other than a comparison of the eight sequences with 180 European Caucasian population haplotypes at 54 SNPs covering 214 kb of the MHC class II region (from HLA-DRB9 to 20 kb centromeric to HLA-DQB1 (Figure 1)) [20], no systematic study has determined the representative nature of each MHP haplotype's complete sequence.
MHC CEHs are common population haplotypes [1]–[6], [8]–[16]. Outside of their core conserved region from HLA-C to HLA-DQB1, CEHs contain dominant alleles on the telomeric class I side at HLA-E [14] and HLA-A [1]–[6], [8]–[11], which is consistent with the region between HLA-A and HLA-C being one of notably low recombination [22]. The class II region between HLA-DQB1 and DAXX is thought to be less conserved than that analogous distal class I region, and class II contains several reported recombination hotspots [22], [23] and regional LD breaks [24] (which are thought to be related). Furthermore, the best-characterized CEH, [HLA-B8, SC01, DR3], previously showed significant variability centromeric to HLA-DQB1 [9], [10], [25]. By contrast, previous reports also documented dominant HLA-DPB1 alleles for a number of European Caucasian CEHs [3], [4], [15], including the B8,DR3 (in shortened nomenclature) CEH.
We used an amplicon resequencing approach [10] to determine the dominant class II sequences centromeric to HLA-DQB1 and to delineate the breakdown of sequence conservation among multiple examples of previously identified CEHs sharing telomeric class II alleles or specificities with the eight MHP sequences. The dominant class II sequences were unique for each CEH, and we confirmed or discovered over 300 SNP and DIP markers. The phased polymorphisms of each CEH dominant sequence are shown in Table S4. Most CEHs showed significant sequence conservation (“fixity”) centromeric to HLA-DQB1, crossing multiple reported recombination hotspots [22], [23]. Although a few CEHs lost a dominant sequence by DAXX (660 kb centromeric to HLA-DQB1), most CEHs retained a dominant sequence throughout the region. Seven of the eight MHP sequences represented the dominant class II sequence of at least one CEH for varying distances.
Several general observations derive from our data. First, microvariation was low within a CEH's dominant class II sequence, even at DIPs, similar to findings within the core MHC for the two CEHs previously reported [9]–[11]. We detected only a single private mutation among the MHP sequences within regions where they otherwise represented the dominant sequences. These findings suggest CEH sequences are recent enough not to have sustained significant mutation during their expansion. Second, CEH dominant sequence conservation appears to be lost primarily due to recombination events with other relatively high frequency haplotypes because non-consensus sequences centromeric to the point of differentiation typically are identical to other common regional sequences. These are often the local fragments of other CEH dominant sequences. Third, except in a few cases where CEHs split apart from a common sequence shared by related CEHs, the dominant sequences did not usually transition to multiple examples of a conserved minor sequence past the recombination point. Although this last observation remains to be confirmed in studies of larger numbers of the same CEH, it also suggests the minor variant recombinants are relatively recent compared to the ages of the original CEHs.
A fourth conclusion is somewhat complex. Although the location of dominant sequence breakdown varied between CEHs (Figure 3D–O) and did not appear to be primarily at previously reported recombination hotspots or LD breaks [22]–[24], many CEHs showed a steady loss of fixity throughout the region. The major reported recombination hotspots in the region are [22]–[24]: a) between HLA-DQB1 and HLA-DQA2 (near MTCO3P1) [22]; b) within intron 2 of TAP2 (just centromeric to the most centromeric amplicon of TAP2 we sequenced) [22]; c) just telomeric to HLA-DMB, near the next amplicon we sequenced centromeric to TAP2 [23]; d) between BRD2 and HLA-DOA [22], [23]; e) between HLA-DOA and HLA-DPA1 [22], [24]; and, f) between HLA-DPB1 and RING1 [24]. However, losses of sequence conservation occurred in the region between HLA-DQA2 and intron 2 of TAP2, a region not previously reported to contain a recombination hotspot.
Our directly observed haplotype results reveal complexity missed by a casual analysis of LD maps. The results we present regarding CEH structure renew questions we previously raised regarding both LD and recombination hotspots [3]. However, our study was not designed to identify nor to challenge the existence of recombination hotspots in the extended class II region, and further study of this region is warranted.
An interesting feature of several CEH pairs and groups is a pattern of shared sequence identity surrounded both telomerically and centromerically by regions in which the CEHs differ significantly (Figure 4). This “block” nature of CEHs and haplotype groups sharing regional alleles has been noted previously [3]–[8], [13]–[16], [19], [22], [25]. The MHP reported [19] a 158 kb “SNP desert” from HLA-DRB1 and MTCO3P1 between the two DR3 CEHs (Figure 4A). Our study expands upon that concept and provides a richer picture of these relationships.
For example, the B7,DR15 and B18,DR15 CEHs were previously known to share alleles within the HLA-DR/DQ block [1], [3]–[6], [8], but it was unknown whether they had identical or distinct sequences centromeric to HLA-DQB1. Our results show these two CEHs share sequence identity throughout the 88 kb stretch from HLA-DQA2 through intron 8 of TAP2, centromeric to which they maintain fixed but distinctly different sequences (Figure 4B). Although the two CEHs theoretically could have different sequences between HLA-DQB1 and HLA-DQA2 and in the domains we skipped within the 88 kb region mentioned above, previously published results suggest such variation would be minimal. The MHP showed, using a set of dense SNP typings, that a set of (HLA-DRB1*15:01, -DQB1*06:02) population haplotypes were identical to one another centromeric to HLA-DQB1 until they split into primarily two subtypes in a region near or within TAP2 [19]. The sudden TAP2 transition they reported was likely both the centromeric break point of the shared sequence for the two DR15 CEHs and the continuation of the two separate but conserved CEH sequences we report here. Similarly, the two B44,DR7 CEHs [16] may share essential sequence identity for the region from HLA-B to HLA-DOB but have separate conserved sequences on either side of that larger than 1.5 Mb region. The two CEHs may have recombined in the early history of a common ancestral haplotype and expanded separately.
We observed a more complex structural pattern among the DR4,DQ8 CEHs than among the DR15 CEHs: two separate regions of shared sequence separated by a variable region of sequence divergence. Specifically, four DR4,DQ8 CEHs telomerically identical at HLA-DQB1, HLA-DQA2 and HLA-DQB2 and centromerically identical from HLA-DMB through BRD2, each had different sequences for varying distances within the 170 kb span between the two sub-regions (Figure 4C). This pattern may be analogous to the pattern within the core MHC region exhibited by the related CEHs [HLA-B62, SB42, DR4, DQ8] and [HLA-B62, SC33, DR4, DQ8] (which, interestingly, are the most divergent of the four DR4,DQ8 CEHs within the 170 kb mentioned above). These patterns of alternating blocks of shared and divergent sequence/alleles may be a type of CEH supergroup microvariation created by early differentiation from a common ancestral sequence due to recombination or, perhaps more likely, localized hypermutation followed by expansion of separate but related CEHs.
Although our dense sequencing results raise questions specific to the class II region, the main issue is essentially the same question we and others have asked about CEHs generally: How can long-range conserved sequences comprise up to half a population's haplotypes crossing numerous putative recombination hotspots or regions of LD breakdown? For example, one of the strongest reported MHC recombination hotspots is located in the TNF-LTA region [22], yet that region is located within the core MHC, the only human genomic region well-documented to contain CEHs.
We conclude CEHs are recent expansions of separate ancestral progenitors. Thus, multiple population examples of each CEH are essentially identical by descent but have spread through the population into pedigrees that are not now highly related. The few mutations within a stretch of conserved sequence can be used to calculate the age of the long-range haplotype [10], [26]. However, plausible values for the variables in such calculations are often difficult to verify.
We also conclude LD values are not particularly useful indicators of population haplotype architecture [3], [4]. LD variation is likely useful to demarcate localized changes in the relationships between individual haplotypes, but LD is all too often simplistically and incorrectly interpreted to suggest the population haplotype architectual dominance of short blocks of conserved sequence separated by narrow regions of relatively frequent randomized sorting. It is likely not coincidental that the MHC is both the region most often studied by segregation analysis in pedigrees and the only well-documented region to contain megabase-length CEHs.
Haplotype sequence and population haplotype architecture accuracy requires both direct observation and the consideration of long-range sequence fixity. Whole genome sequencing will soon allow direct determination of full haplotype sequences if analyzed appropriately [27]. This requires either sequencing individual chromosomes after physical isolation [28] or sequencing moderate to large pedigrees to phase pedigree data directly [29], [30]. The latter allows both sequence integrity crosschecking and directly observed recombination. Samples homozygous for a particular long-range haplotype are useful for identifying putative CEH alleles [10], [17]–[20], [31], [32], but such cell lines are rare. Direct haplotype determination and counting [2] is the only method capable of revealing the details of haplotype structure and population haplotype architecture essential for disease gene localization [4]. Computational phasing to “impute” haplotype structure in unrelated subjects has been advocated for monetary or feasibility reasons, but this does not usually provide accurate haplotype structure [33].
Reports over 30 years show that MHC CEHs are high population frequency (“common”) megabase-length conserved sequences [1], [3]–[6], [8]–[16], [23], [24]. The evidence for CEH sequence conservation (with minor microvariation) increased whenever loci were defined at higher resolution or at intervening locations. We update and improve the definition of the centromeric points up to which the published reference MHC sequences essentially represent CEH dominant sequences. The dominant class II CEH sequences we provide (far from a complete list) should be useful for future European Caucasian haplotype comparisons. More complete resequencing of larger numbers of pedigree-determined haplotypes is required to determine population haplotype architecture both within the MHC and throughout the genome. Furthermore, non-European CEHs [34] must be studied in a similar manner. Finally, an appreciation of long-range haplotype sequence conservation throughout the genome is required to localize efficiently the genomic structural elements responsible for complex genetic traits (including disease susceptibility).
All participants gave informed consent in accordance with Institutional Review Board (IRB)-approved protocols. All work was conducted under IRB protocols approved by the Immune Disease Institute (or its predecessors) and/or Boston Children's Hospital IRBs.
Blood samples were provided by 180 individuals in 43 unrelated families and by 10 unrelated subjects (the latter homozygous for portions of the MHC), mostly from the Boston metropolitan area. We obtained extensive demographic and personal health information (including family histories) from all subjects. The relatively diverse European Caucasian population in Boston and our recruitment methods make it highly unlikely any of the pedigrees or unrelated subjects are directly related to one another. We also obtained B-lymphocytic cell lines of 15 individuals in four families from the Human Biological Data Interchange (HBDI; Philadelphia, PA). International Histocompatibility Workshop (IHW) homozygous cell lines (n = 12), including three of the MHP (DBB, MANN and QBL), were used for a limited number of haplotypes. All samples had been typed at classical markers within the MHC prior to selection, although typing was conducted at various resolutions (from serological to high resolution DNA typing). Pedigrees were chosen to obtain multiple examples of a wide variety of MHC CEHs or at least the HLA-DR/DQ fragments of CEHs putatively represented by MHP haplotypes. HLA-DPB1 typing was not considered during subject and haplotype selection so that the degree of fixity in the centromeric class II region was random.
DNA was extracted from EDTA-treated blood, peripheral blood mononuclear cells or B-lymphocytic cell lines. Genomic DNA was isolated using the QIAmp DNA mini kit (Qiagen, Valencia, CA).
Molecular MHC allele typing was performed by PCR and sequence-specific oligonucleotide probes (in-house or Lifecodes) or by sequence-specific primer kits (Invitrogen) at low to high resolution. Some HLA types were identified serologically [35]. CFB (previously known as BF) and C4 allele typing was by agarose gel electrophoresis and immunofixation with specific antisera; C2 alleles were determined by isoelectric focusing of serum in polyacrylamide gels and a C2-sensitive hemolytic overlay [36]. MHC complement gene haplotypes or complotypes are designated by their CFB, C2, C4A, and C4B alleles, in that arbitrary order [7]. Null or Q0 alleles are simply designated 0. Thus, FC31 indicates the complotype CFB*F, C2*C, C4A*3, C4B*1. Complotypes for some IHW cell lines were described previously [30], [31]. IHW cell typing was known (http://www.ebi.ac.uk/ipd/imgt/hla/cell_query.html) and/or verified as described above.
We analyzed eight different MHC class II and extended class II sequences determined by the Sanger Institute [17]–[20] for distinguishing SNPs and deletion/insertion polymorphisms (DIPs). Currently available MHC sequence data for these cell lines may be found via: http://www.ucl.ac.uk/cancer/medical-genomics/mhc or http://www.ensembl.org/index.html or at the URL listed under “MHC Typing.” MHP haplotypes represent the human reference sequence (PGF) as well as the following alternative sequences for the human MHC: ALT_REF_LOCI_1 (APD), ALT_REF_LOCI_2 (COX), ALT_REF_LOCI_3 (DBB), ALT_REF_LOCI_4 (MANN), ALT_REF_LOCI_5 (MCF), ALT_REF_LOCI_6 (QBL), and ALT_REF_LOCI_7 (SSTO).
We used an amplicon-based resequencing approach [10] to distinguish the dominant sequences of CEHs in the class II region. CLC Combined Workbench software program (CLCBio LLC, Cambridge, MA) was used to align these sequences for the region from MTCO3P1 to DAXX (Figure 1). After aligning all MHP sequence data available for the eight haplotypes, we analyzed the entire region from MTCO3P1 to DAXX to find an optimal distribution of amplicons that balanced the needs for relatively even coverage and for maximizing differences between the sequences. After preliminary resequencing and localization of regions in which some of the MHP haplotypes appeared to cease representing many population haplotypes or in which we had poor sequencing results, we added or substituted amplicons. Finally, in some cases, we skipped relatively large regions having known low polymorphism.
We designed primers, using a version of Primer 3 software (http://frodo.wi.mit.edu), at monomorphic (or near-monomorphic) positions in regions near or within genes that would likely offer maximal differentiation of the various MHP haplotypes. The primer sequences we used and the amplicons we resequenced are shown in Table S1. We sequenced a total of approximately 27 kb using 56 sets of primers covering five separate regions spanning a total distance of approximately 580 kb of genomic DNA. In some cases where sequence phase could be determined without some members of a particular pedigree, the DNA of those members was not sequenced, but we often sequenced all members of a pedigree to confirm results for a particular haplotype in multiple carriers of that haplotype. We also sequenced portions of three MHP cell lines in data gaps of current scaffolds, and we report and have provided that information to GenBank.
PCR products were excised from agarose gels and purified using the QIAEX II gel extraction kit (Qiagen) or were drawn out of recovery wells directly (Lonza, Inc.) and sequenced by dideoxy sequencing using Big Dye Terminator V3.0 chemistry (Genewiz, Inc., South Plainfield, NJ and/or Davis Sequencing, Inc., Davis, CA). All sequences were analyzed and compared using both alignment software and direct visual inspection of chromatograms. DIP sizes in heterozygotes were usually decipherable and deducible in both directions based on the known sequence surrounding the DIP. At least two individuals inspected visually and agreed upon the sequence of each chromatogram used to determine sequence.
Excluding private mutations, we identified 274 SNPs and DIPs in the 342 kb region from HLA-DQA2 to HLA-DPB1 and 34 SNPs and DIPs in the 103 kb region from centromeric to RING1 to DAXX (Table S2). We defined the centromeric point where a particular MHP sequence no longer represented a haplotype group as the location at which the dominant sequence shared by those haplotypes began to contain SNP and DIP alleles not in the MHP sequence.
The vast majority of haplotypes (n = 132; 83.5%) were phased by segregation analysis in pedigrees, showed Mendelian inheritance patterns (except in rare cases of null alleles or detected crossovers) and were assigned unique identifiers as unrelated founder chromosomes. Six unrelated subjects or IHW cell lines each known not to be consanguineous were homozygous for specific haplotypes throughout the region analyzed and provided 12 additional unrelated chromosomes (7.6%). Six IHW cell lines either known to be consanguineous or of unknown status provided six additional unrelated chromosomes (3.8%). Finally, four unrelated subjects known not to be consanguineous who were heterozygous for at least some portion of the region studied provided the final eight unrelated chromosomes (5.1%).
Haplotype phasing in the classical CEH region (between HLA-C and HLA-DQB1) was established for 95% of all haplotypes or was inferred from known CEH allele combinations. Over 96% of SNP and DIP alleles were unambiguously phased: a) by segregation analysis in pedigrees [1], [2]; or, b) using IHW or locally-identified MHC homozygous samples. Such cell lines were assumed to be of consanguineous origin unless known not to be and received only one haplotype assignment. The remaining alleles (<4% overall and <4% in all regions except for HLA-DPB1, where the percentage of inferred phasing was 10.8%) were assigned to haplotypes by inference as follows. In a family in which all subjects were heterozygous identical at a locus or in a heterozygous individual without relatives, one of the alleles was arbitrarily assigned to one of the haplotypes to be consistent with its surrounding (unambiguous) markers, defined by the unambiguous haplotypes in the group to which it belonged or, if the haplotype was no longer representative, by all unambiguous haplotypes. Phasing of the remaining pedigree haplotype(s) was/were thus established. We report here on 158 haplotypes (and an additional seven at HLA-DQA2) that fell into one of the eight MHP groups.
Physical distances between MHC genes, locations and amplicons were found at the NCBI website (http://www.ncbi.nlm.nih.gov/projects/genome/guide/human/index.shtml). We used the human reference sequence NC_000006.11 from Genome Reference Consortium assembly GRCh37.p10 and reference sequence (rs) numbers are from dbSNP build 138. All novel SNP (n = 1) and DIP (n = 7) variations (shown in Table S2) were submitted to dbSNP (http://www.ncbi.nlm.nih.gov/snp/) using the handle CAALPER. All novel DNA sequences for the three MHP cell lines have GenBank accession numbers (http://www.ncbi.nlm.nih.gov/genbank/) KF880997-KF880999 (for QBL), KF881000-KF881006 (for MANN) and KF881007-KF881009 (for DBB) (Figures S1–S3). To determine sequence fixity, we assumed sequence identity within intervening regions we did not resequence among the population haplotypes bearing the genotypic markers and/or dominant sequence of a CEH (except for rare private mutations and infrequent microvariations).
To quantify and represent crossover events leading to the breakdown of CEH dominant sequences, we define a new metric: normalized crossover frequency (NCF). NCF is the fraction of remaining dominant sequences of a single CEH that begin to differ from the dominant sequence due to apparent recombination within a defined region, normalized over a unit (1 Mb) distance. Our data were not distributed evenly across the region we studied, and we therefore calculated our data over sub-regions of varying size (Table S2). Thus, we required normalization to a unit distance to compare the sequence breakdown by separate crossovers. We displayed these data in a bar graph format in which the abscissa is drawn to genomic scale. Therefore, the areas (not the heights) of the bars representing NCFs are compared to determine the relative contribution of regional recombinants to the breakdown of CEH sequence conservation. NCF was calculated using the equation:
NCF = (crossovers/total remaining haplotypes) × (1 Mb/distance covered)
where:
a) “crossovers” are the number of haplotypes that lost the CEH dominant sequence centromeric to the prior (telomeric) analyzed region due to recombination events (as opposed to minor microvariation in an otherwise identical sequence), and include both any crossovers directly observed in the currently analyzed region and any deduced to have occurred between the currently and prior analyzed regions.
b) the “total remaining haplotypes” are the number of remaining population haplotypes having the dominant sequence of a given CEH throughout the region immediately telomeric to the analyzed region. For the first region (HLA-DQA2), it was assumed all population haplotypes of a given CEH had the dominant sequence at the centromeric end of HLA-DQB1.
c) the “distance covered” is measured by subtracting the genomic position of the most centromeric point of the prior region (the region immediately telomeric to the currently analyzed region) from the genomic position of the most centromeric point of the currently analyzed region.
As an example, if 16 population haplotypes of a single CEH had the dominant sequence through HLA-DQA2 and 3 of these crossed over to a non-dominant sequence by the centromeric end of the HLA-DQB2 region (in which the distance from the most centromeric polymorphism analyzed in the HLA-DQA2 region through the most centromeric polymorphism analyzed in the HLA-DQB2 region is 21,096 bases), the NCF would be:
(3/16) × (1,000,000/21,096) = 8.9
and the “total remaining haplotypes” for the next region (between HLA-DQB2 and HLA-DOB, covering 25,927 bases) would be 13 (i.e., 16–3).
|
10.1371/journal.pntd.0000948 | Assessment of the Anthelmintic Efficacy of Albendazole in School
Children in Seven Countries Where Soil-Transmitted Helminths Are
Endemic | The three major soil-transmitted helminths (STH) Ascaris
lumbricoides, Trichuris trichiura and
Necator americanus/Ancylostoma duodenale are among the
most widespread parasites worldwide. Despite the global expansion of
preventive anthelmintic treatment, standard operating procedures to monitor
anthelmintic drug efficacy are lacking. The objective of this study,
therefore, was to define the efficacy of a single 400 milligram dose of
albendazole (ALB) against these three STH using a standardized protocol.
Seven trials were undertaken among school children in Brazil, Cameroon,
Cambodia, Ethiopia, India, Tanzania and Vietnam. Efficacy was assessed by
the Cure Rate (CR) and the Fecal Egg Count Reduction (FECR) using the
McMaster egg counting technique to determine fecal egg counts (FEC).
Overall, the highest CRs were observed for A. lumbricoides
(98.2%) followed by hookworms (87.8%) and T.
trichiura (46.6%). There was considerable variation in
the CR for the three parasites across trials (country), by age or the
pre-intervention FEC (pre-treatment). The latter is probably the most
important as it had a considerable effect on the CR of all three STH.
Therapeutic efficacies, as reflected by the FECRs, were very high for
A. lumbricoides (99.5%) and hookworms
(94.8%) but significantly lower for T. trichiura
(50.8%), and were affected to different extents among the 3 species
by the pre-intervention FEC counts and trial (country), but not by sex or
age.
Our findings suggest that a FECR (based on arithmetic means) of
>95% for A. lumbricoides and >90% for
hookworms should be the expected minimum in all future surveys, and that
therapeutic efficacy below this level following a single dose of ALB should
be viewed with concern in light of potential drug resistance. A standard
threshold for efficacy against T. trichiura has yet to be
established, as a single-dose of ALB is unlikely to be satisfactory for this
parasite.
ClinicalTrials.gov NCT01087099
| Soil-transmitted helminths (roundworms, whipworms and hookworms) infect millions
of children in (sub)tropical countries, resulting in malnutrition, growth
stunting, intellectual retardation and cognitive deficits. Currently, there is a
need to closely monitor anthelmintic drug efficacy and to develop standard
operating procedures, as highlighted in a World Health Organization–World
Bank meeting on “Monitoring of Drug Efficacy in Large Scale Treatment
Programs for Human Helminthiasis” in Washington DC at the end of 2007.
Therefore, we have evaluated the efficacy of a commonly used treatment against
these parasitic infections in school children in Africa, Asia and South-America
using a standardized protocol. In addition, different statistical approaches to
analyzing the data were evaluated in order to develop standardized procedures
for data analysis. The results demonstrate that the applied treatment was highly
efficacious against round- and hookworms, but not against whipworms. However,
there was large variation in efficacy across the different trials which warrants
further attention. This study also provides new insights into the statistical
analysis of efficacy data, which should be considered in future monitoring and
evaluation studies of large scale anthelmintic treatment programs. Finally, our
findings emphasize the need to update the World Health Organization recommended
efficacy threshold for the treatment of STH.
| The three major Soil-Transmitted Helminths (STH), Ascaris
lumbricoides (roundworm), Trichuris trichiura
(whipworm) and Necator americanus/Ancylostoma duodenale (the
hookworms) are amongst the most widespread parasites worldwide. An estimated 4.5
billion individuals are at risk of STH infection and more than one billion
individuals are thought to be infected, of whom 450 million suffer morbidity from
their infection, the majority of who are children. An additional 44 million infected
pregnant women suffer significant morbidity and mortality due to hookworm-associated
anemia. Approximately 135,000 deaths occur per year, mainly due to infections with
hookworms or A. lumbricoides
[1].
The most widely implemented method of controlling STH infections is through periodic
administration of anthelmintics. Rather than aiming to achieve eradication, current
control programs are focused on reducing infection intensity and transmission
potential, primarily to reduce morbidity and avoid mortality associated with the
disease [2]. The
benzimidazole (BZ) drugs, i.e. albendazole (ALB) and mebendazole, are the most
widely used drugs for the control of STH. While both show broad-spectrum
anthelmintic activity, for hookworms a single dose of ALB is more effective than
mebendazole [3].
The scale up of chemotherapy programs that is underway in various parts of Africa,
Asia and South America, particularly targeting school children, is likely to exert
increasing drug pressure on parasite populations, a circumstance that is likely to
favor parasite genotypes that can resist anthelmintic drugs. Given the paucity of
suitable alternative anthelmintics it is imperative that monitoring programs are
introduced, both to assess progress and to detect any changes in therapeutic
efficacy that may arise from the selection of worms carrying genes responsible for
drug resistance. The well documented occurrence of resistance to anthelmintics in
nematode populations of livestock [4], highlights the potential
for frequent treatments used in chemotherapy programs to select drug resistant
worms. Such an eventuality threatens the success of treatment programs in humans,
both at individual and community levels [5]. Although some small scale
studies [6], [7], have
suggested emerging drug resistance in human STH, these studies should be interpreted
with some caution, since suboptimal efficacy could have been due to factors other
than drug resistance. Moreover, although for the BZ drugs there are many published
studies reporting the Cure Rate (CR) and the Fecal Egg Count Reduction (FECR), the
two most widely used indicators for assessing the efficacy of an anthelmintic in
human medicine, comparison of such studies is difficult, largely because there is no
widely accepted standard operating procedure for undertaking such trials [8]. Published
studies are confounded by methodological variations including treatment regimens,
poor quality of drugs, differing statistical analyses used to calculate therapeutic
efficacy, as well as a range of other problems in study design, such as small sample
size, diagnostic methods, variation in pre-intervention infection intensities and
confounding factors related to geographical locations. Such variation among studies
greatly complicates direct comparison [3]. A World Health
Organization-World Bank (WHO-WB) meeting on “Monitoring of Drug Efficacy in
Large Scale Treatment Programs for Human Helminthiasis”, held in Washington DC
at the end of 2007, highlighted the need to closely monitor anthelmintic drug
efficacy and to develop standard operating procedures for this purpose. In a
systematic meta-analysis of published single-dose studies, Keiser and Utzinger [8], confirmed that
there was a paucity of high quality trials, and that the majority of trials were
carried out more than 20 years ago. They recommended that well-designed, adequately
powered, and rigorously implemented trials should be undertaken to provide current
data regarding the efficacy of anthelmintics against the main species of STH. These
should be designed to establish benchmarks (including standard operating procedures)
for subsequent monitoring of drug resistance.
The objective of the present work was to validate a standard protocol that has been
developed for monitoring efficacy of anthelmintics against STH. To give the study
wide relevance, we conducted the trial in seven populations in different geographic
locations in Brazil, Cameroon, Cambodia, Ethiopia, India, Tanzania and Vietnam. In
each of the study sites, different epidemiologic patterns of infection prevail,
including different combinations of STH. We assessed the efficacy of a single dose
(400 mg) of ALB in terms of the CR and the FECR in school children between 14 and 30
days following treatment. The McMaster egg counting technique was used in a
standardized fashion, with rigorous quality control. Levecke et al. [9] reported that
the McMaster holds promise as a standardized method on account of its applicability
for quantitative screening of large numbers of subjects. This method is the
recommended method for measuring fecal egg counts (FEC) when performing FECR for the
detection of anthelmintic resistance in veterinary medicine [10], [11].
This study was carried out in seven different countries covering Africa
(Cameroon, Ethiopia and Tanzania), Asia (Cambodia, India and Vietnam) and
South-America (Brazil). However, it is important to note, that while we refer to
individual countries to identify results from particular trials, we do not make
any conclusions about any country as such. Here, names of countries are used
only to distinguish between 7 separate trials that were conducted in 7
geographically disparate regions of the world. In total ten study sites with
varying STH and treatment history were included. These seven STH endemic
countries were selected because of the presence of investigator groups with
previous extensive experience in the diagnosis and control of STH. Table 1 provides their
specific locations (district/province/state) and treatment history. Both species
of hookworms (N. americanus and A. duodenale)
were present in all study sites in varying degree with the exception of Brazil
where only N. americanus was present.
During the pre-intervention survey, school children aged 4 to 18 years at the
different study sites were asked to provide a stool sample. For the initial
sampling the aim was to enroll at least 250 infected children with a minimum of
150 eggs per gram of feces (EPG) for at least one of the STH. This sample size
was selected based on statistical analysis of study power, using random
simulations of correlated over-dispersed FEC data reflecting the
variance-covariance structure in a selection of real FEC data sets. This
analysis suggested that a sample size of up to 200 individuals
(α = 0.05, power = 80%) was
required to detect a 10 percentage point drop from a null efficacy of ∼
80% (mean percentage FEC Δ per individual) over a wide range of
infection scenarios. Standard power analyses for proportions also indicated that
the detection of a ∼10 percentage point drop from a null cure rate required
sample sizes up to 200 (the largest samples being required to detect departures
from null efficacies of around 50%). Given an anticipated non-compliance
rate of 25%, a sample of 250 individuals with >150 EPG pre-treatment
was therefore considered necessary at each study site.
Fecal samples were processed using the McMaster technique (analytic sensitivity
of 50 EPG) for the detection and the enumeration of infections with A.
lumbricoides, T. trichiura and hookworms [9]. None of
the samples were preserved. Samples which could not be processed within 24 hours
were kept at 4°C. A single dose of 400 mg ALB (Zentel) from the same
manufacturer (GlaxoSmithKline Pharmaceuticals Limited, India) and same lot
(batch number: B.N°: L298) was used at all trial sites. No placebo control
subjects were included in the trial for ethical and operational reasons. Between
14 to 30 days after the pre-intervention survey, stool samples were collected
from the treated subjects and processed by the McMaster technique. All of the
trials were carried out in a single calendar year (2009). Subjects who were
unable to provide a stool sample at follow-up, or who were experiencing a severe
concurrent medical condition or had diarrhea at time of the first sampling, were
excluded from the study. The participation, the occurrence of STH and sample
submission compliance for pre- and post-intervention surveys are summarized in
Figure 1.
The McMaster counting technique (McMaster) was based on the modified McMaster
described by the Ministry of Agriculture, Fisheries and Food (UK; 1986) [12]. Two
grams of fresh stool samples were suspended in 30 ml of saturated salt solution
(density = 1.2). The suspension was poured three times
through a wire mesh to remove large debris. Then 0.15 ml aliquots were added to
each of the 2 chambers of a McMaster slide. Both chambers were examined under a
light microscope using a 100x magnification and the FEC for each helminth
species was obtained by multiplying the total number of eggs by 50.
The efficacy of the treatment for each of the three STH was evaluated
qualitatively based on the reduction in infected children (CR) and
quantitatively based on the reduction in fecal egg counts (FECR). The outcome of
the FECR was calculated using three formulae. The first two formulae were based
on the mean (arithmetic/geometric) of the pre- and post-intervention fecal egg
count (FEC) ignoring the individual variability, whereas the third formula
represented the mean of the reduction in the FEC per subject. The latter is the
only quantitative indicator of efficacy for which the importance of confounding
factors can be assessed by statistical analysis.
The CR and the FECR
(1-3) outputs were calculated for the different trials, both sexes, age classes
(A: 4–8 years; B: 9–13 years and C: 14–18 years) and for the
level of egg excretion intensity at the pre-intervention survey. These levels
corresponded to the low, moderate and high intensities of infection as described
Montresor et al. [13] For A. lumbricoides these were
1–4,999 EPG, 5,000–49,999 EPG and >49,999 EPG; for T.
trichiura these levels were 1–999 EPG, 1000–9,999 EPG
and >9,999 EPG; and for hookworms these were 1–1,999 EPG,
2,000–3,999 EPG and >3,999 EPG, respectively.
In addition, the robustness of the three FECR formulae was explored by comparing
the FEC reduction rate obtained from all samples containing STH and those
obtained from samples containing more than 150 EPG as recommended in the
anthelmintic resistance guidelines of the World Association for the Advancement
of Veterinary Parasitology [9]. Finally, putative factors affecting the CR and the
FECR (3) were evaluated. For the CR, generalized linear models (binomial error)
were built with the test result (infected /uninfected) as the outcome,
‘trial’ (7 levels: trials in Brazil, Cambodia, Cameroon, Ethiopia,
India, Tanzania and Vietnam) and ‘sex’ (2 levels: female and male)
as factors, and ‘age’ and the log transformed pre-intervention FEC
as covariates. Full factorial models were evaluated by the backward selection
procedure using the likelihood ratio test of χ2. Finally, the CR
for each of the observed values of the covariate and factor was calculated based
on these models (The R Foundation for Statistical Computing, version 2.10.0
[14]).
For analysis of the data from FECR (3), non-parametric methods were used,
because models based on parametric statistics, even with negative binomial error
structures, or based on transformed data would not converge satisfactorily as a
consequence of the high proportion of FEC with zero EPG. Hence, the impact of
the factors ‘trial’ and ‘sex’ were assessed by the
Kruskal-Wallis test (for more than 2 group comparisons) and the Mann-Whitney U
test, respectively. The correlation between the outputs of FECR (3) and the
covariates (age and pre-intervention FEC) was estimated by the Spearman rank
order correlation coefficient (SAS 9.1.3, SAS Institute Inc.; Cary, NC,
USA).
The overall protocol of the study was approved by the Ethics committee of the
Faculty of Medicine, Ghent University (Nr B67020084254) and was followed by a
separate local ethical approval for each study site. For Brazil, approval was
obtained from the Institutional Review Board from Centro de Pesquisas
René Rachou (Nr 21/2008), for Cambodia from the National Ethic Commitee
for Health Research, for Cameroon from the National Ethics Committee (Nr
072/CNE/DNM08), for Ethiopia from the Ethical Review Board of Jimma University,
for India from the Institutional Review Board of the Christian Medical College
(Nr 6541), for Tanzania (Nr 20) from the Zanzibar Health Research Council and
the Ministry of Health and Social Welfare, for Vietnam by the Ministry of Health
of Vietnam. An informed consent form was signed by the parents of all subjects
included in the study. This clinical trial was registered under the
ClinicalTrials.gov Identifier NCT01087099.
Overall, the highest CRs were observed for A. lumbricoides
(98.2%), followed by hookworm (87.8%) and T.
trichiura (46.6%). However, as shown in Table 2, the CRs varied
across the different trials, age classes and pre-intervention FEC levels. The
differences in CRs between trials were most pronounced for T.
trichiura, ranging from 21.0 (Tanzania) to 88.9% (India).
The T. trichiura CRs of 100% for the trials in Brazil
and Cambodia are not considered here as they were based on only 1 and 2
individuals, respectively. For hookworms and A. lumbricoides,
the CRs varied from 74.7 (India) to 100% (Vietnam) and from 96.4
(Tanzania) to 99.3% (Ethiopia and Cameroon), respectively. The CRs for
A. lumbricoides in Cambodia (100%) and India
(95.2%) are not considered here as they were based on fewer than 50
individuals. The CRs increased over the three age classes (A.
lumbricoides: 95.8 to 100%; T. trichiura:
44.7 to 54.1%), except for hookworms where the CRs ranged from 86.1 to
88.3, and then to 87.5%. For each of the three STH, there was a decline
in the CR with increasing levels of infection intensities at the
pre-intervention survey. The largest drop was observed for T.
trichiura, which decreased from 53.9 to 12.5%. For the two
other STH, the drop in the CR was less pronounced, ranging from 88.6 to
76.9% for hookworms and only from 98.3 to 95% for A.
lumbricoides. The observed differences between sexes were
negligible for all three STH.
Differences in CR by trial, age and pre-intervention FEC are illustrated in Figure 2. The variability in
the CR of the three parasites was significantly associated with these three
factors (predictive value >75%). The pre-intervention FEC was probably
the most important as it had a considerable effect on the CR of A.
lumbricoides
(χ21 = 4.14,
p<0.05), T. trichiura
(χ21 = 66.3,
p<0.0001) and hookworms
(χ21 = 11.9,
p<0.001). Age only contributed to variation in the CR of
A. lumbricoides
(χ21 = 6.8,
p<0.01). Differences among the trials (countries) in the CR
were observed for T. trichiura
(χ23 = 33.8,
p<0.0001) and hookworms
(χ26 = 35.1,
p<0.0001), but not for A. lumbricoides. In
addition, there was an interaction between the pre-intervention FEC for
A. lumbricoides
(χ21 = 4.7,
p<0.05) and for T. trichiura
(χ23 = 18.4,
p<0.0005) with age and trial (country) respectively
(lines cross one another). The impact of pre-intervention FEC on the CR of
A. lumbricoides was more pronounced for older individuals
than younger ones. For T. trichura the effect of
pre-intervention FEC varied considerably across the trials conducted in the
different countries, particularly for the trial in Ethiopia where the CR dropped
from almost 100 to nearly 0% as the pre-intervention FEC increased.
The pre-intervention FEC for the different STH ranged from 50 to 170,500 EPG for
A. lumbricoides (arithmetic
mean = 6877 EPG), from 50 to 23,200 EPG for T.
trichiura (arithmetic mean = 824 EPG) and from
50 to 13,800 EPG for hookworm (arithmetic mean = 650 EPG).
The data in Table 3 show
that there was considerable variation in the arithmetic means of the FEC from
the trial groups in the 7 participating countries for each of the three STH
species. As illustrated in Figure
3, pre-intervention FEC were highly aggregated among the subjects,
and high FEC were only observed in relatively few subjects.
The FEC reduction rate calculated using all three formulae (based on FECR 1-3) in
turn for A. lumbricoides, T. trichiura and
hookworms across the 7 trials (countries), age classes, sexes and
pre-intervention infection intensities are summarized in Table 4. Overall, the FEC reduction rate for
FECR(1) was the highest for A. lumbricoides (99.5%),
followed by hookworm (94.8%) and T. trichiura
(50.8%). However, there was considerable variation in the FEC reduction
rate among the 7 trials, age classes and infection intensities at
pre-intervention survey. For A. lumbricoides, the FEC reduction
rate remained roughly unchanged over these variables, only ranging from 97.8 to
100%. This contrasts with T. trichiura, for which the
FEC reduction rate differed between the trials (from 39.2 [Cameroon]
to 92.4% [Ethiopia]), age classes (from 45.4 [B] to
62.7% [A]) and pre-intervention infection intensity (from 40.0
[high] to 58.7% [moderate]). There was no difference
in the FEC reduction rate between the sexes. For hookworms, only small
differences in the FEC reduction rate were observed between the trials, ranging
from 87.1 [India] to 100% [Vietnam]. However, there
were only negligible differences between the age classes (from 94.7
[B] to 96.4% [C]).
Compared to the results of FECR (1), the outputs of FECR (2) resulted in higher
values for all three STH, except for A. lumbricoides where the
FEC reduction rate already showed a ceiling effect (100%). Considerable
variation in the FEC reduction rate (FECR (2)) occurred with T.
trichiura among the trials (from 82.6 [Tanzania] to
99.1% [Ethiopia]) and pre-intervention infection intensity
(from 88.6 [high] to 94.3% [low]). For hookworms, the
differences between the trials were virtually negligible, all indicating a
potent effect just short of the maximum 100% (FECR (2)
>99.3%).
The results of FECR (3) mostly yielded comparable or lower values than those from
FECR (1). The low values (sometimes negative) can be explained by subjects for
whom the post-intervention FEC exceeded the pre-intervention FEC. These subjects
contributed to a negative FEC reduction rate which had a significant impact on
the final FEC reduction rate calculated with FECR (3). This became apparent in
the FEC reduction rate for A. lumbricoides, where a Cameroonian
male subject of 7 years with a pre-intervention FEC of 100 and a
post-intervention FEC of 22,050 EPG, contributed markedly to lowering the
overall values for the data-set from the trial in Cameroon (FECR (1):
99.2%; FECR (3): 26.0%). This lowering of FECR (3) compared to
FECR (1) for A. lumbricoides also occurred with age class A
(FECR (1): 98.9%; FECR (3): −2.7%) and the low
pre-intervention infection intensity level (FECR (1): 97.8%; FECR (3):
66.6%), but not for the remaining variables. The number of negative
individual FEC reduction rates, and the magnitude of the difference between pre-
and post-intervention FEC, both contributed to the discrepancies found for
T. trichiura (176 subjects) and hookworms (10
subjects).
Table 5 summarizes the FEC
reduction rates restricted to samples of more than 150 EPG indicating that the
results of FECR (1) and FECR (2) remained roughly unchanged. The values from
FECR (3) increased and were mostly comparable with those obtained by FECR (1).
This change in the results of FECR (3) is due to the exclusion of negative
individual FEC reduction rates which mostly occurred among the subjects with low
pre-intervention FEC (see also Table 4). Differences of more than 5% between the results of
FECR (3) and FECR (1) were limited to T. trichiura (country:
Cameroon, India, Tanzania and Vietnam; age class: A and C).
The assessment of putative factors affecting the results from FECR (3) was
restricted to samples containing more than 150 EPG. Due to the limited variation
in the FEC reduction rates (FECR (3)) of A. lumbricoides across
the different variables, this species was not analyzed further. Also, because of
the limited number of infected subjects (<50), the trials in Brazil, Cambodia
and India were excluded from analyses of T. trichiura. For
hookworms, and for the same reasons, subjects from the trials in Brazil and
Vietnam were not included. Significant differences in the FEC reduction rates
between the trials were found for both T. trichiura
(χ23 = 117.3,
p<0.0001) and hookworms
(χ24 = 20.2,
p = 0.0005). High pre-intervention FEC of
T. trichiura yielded lower FEC reduction rates (3)
(Rs = −0.18,
n = 701, p<0.0001), but this was not found for hookworm
(Rs = −0.04,
n = 601, p = 0.34). In addition, there
was an interaction between the pre-intervention FEC of T.
trichiura and trial (country), reflected in the negative
correlations in the trials in Cameroon
(Rs = −0.28, n = 233,
p<0.0001), and Ethiopia, (Rs = −0.34,
n = 72, p = 0.0034), but a positive
correlation for the trial in Tanzania
(Rs = +0.28, n = 325,
p<0.0001) and a non-significant correlation for the trial in Vietnam
(Rs = −0.07,
n = 71, p = 0.58). Host sex and age
did not contribute significantly to variation of the results of (FECR (3)) in
any of the STH examined.
To our knowledge, the present study is the first to evaluate drug efficacy for STH in
school children across different endemic regions using a protocol which was
standardized in terms of the treatment (a single-oral 400 mg dose of ALB originating
from the same batch), the follow up (between 14 and 30 days after) and the detection
technique (the McMaster counting technique). Moreover, efficacy was evaluated by
both the CR and the FECR, and compared statistically between the seven trials which
took place in geographically disparate parts of the world.
Overall, this study supports previous reports that indicated that single dose ALB
treatment is most effective for infection with A. lumbricoides,
followed by hookworm, but is relatively ineffective for T.
trichiura, confirming the efficacy studies reviewed by Bennet and
Guyatt [3], and
by Keiser and Utzinger [8]. The low efficacy observed for T.
trichiura compared to the two other STH, is in keeping with previous
studies, where a 3-day dose schedule of ALB has been shown to be necessary to
achieve acceptable therapeutic efficacy [3].
At present, the most commonly reported indicator of drug efficacy in this field is
the CR [3]. Our
results support the view that the CR should not be the recommended parameter, as it
is sensitive to variation in the intensity of infection before treatment. The CRs
declined in all three STH with increasing intensity of infection (FEC) at the
pre-intervention survey. Hence, comparison between populations (countries, villages,
schools, etc.) differing in pre-intervention FEC are guaranteed to arrive at
different conclusions about drug efficacy. Differences in the outputs of
calculations based on processing quantitative data in different ways also showed
variation that requires careful review if standard operating procedures for data
processing are to be adopted. The observation that therapeutic efficacies based on
arithmetic means were mostly lower than those based on geometric means is in
agreement with other studies [15], and arises because the arithmetic means captures the
variation more effectively, while the geometric means compress the data such that
efficacies are highly overestimated. Our exploratory analysis of different
statistical approaches for analyzing data also indicates that FECR based on
individuals was highly affected by excluding subjects with pre-intervention FEC
below 150 EPG. Therefore, we conclude that the group based formula using an
arithmetic mean is the best summary statistic to employ in analysis of therapeutic
efficacy in future large scale drug administration trials, since it represents a
robust indicator that is sensitive to changes in drug efficacy.
The efficacy (CR and FEC reduction rate) varied widely across the trials, except for
A. lumbricoides. Possible explanations for the observed
differences include (1) treatment history, (2) geographic differences within STH
species, (3) fecal consistency and (4) diet. It is therefore pertinent to comment on
each. Although the lowest efficacies for T. trichiura (Cameroon and
Tanzania) and hookworms (India) were obtained in countries with a treatment history,
the observed low efficacies are not likely to be attributable to large scale
anthelmintic treatment in Cameroon and India. In these countries, a comparison
between different study sites with a history of large scale anthelmintic treatment
(Cameroon: Loum; India: Vellore) and without such a history (Cameroon: Yoyo; India:
Thiruvanamalai) indicated that these large scale programs did not result in a
reduced efficacy compared to sites were they were absent (data not shown and to be
published separately). For Tanzania, the impact of large scale anthelmintic
treatment programs could be ruled out, as studies before and during these
interventions have shown similar drug efficacy figures for T.
trichuria
[16], [17].
Current molecular studies indicate that geographical differences exist within STH
species [18], [19]. For T.
trichiura varying anthelmintic efficacy has been suggested to be
attributable to the presence/absence of the β-tubulin codon 200 polymorphism
that has been linked to BZ resistance [20]. Strain differences have been
demonstrated in some species with different drug tolerance as assessed both by
efficacy and molecular studies [20], [21]. Nevertheless, the exact impact of genetic differences
within the 3 STH in this study on the efficacy of specific anthelmintics remains
speculative. Of note, even at a higher taxonomic level, information on the relative
therapeutic efficacy of a single dose ALB on N. americanus and
A. duodenale is scarce, this despite the distinct and well
known biological differences between these hookworms [22]-[24]. FEC was calculated in the
current study without compensation for fecal consistency. It is well recognized that
well-formed stools can concentrate helminth eggs, compared to looser or diarrheic
feces where they are diluted [25], thus confounding assessment of drug efficacy. Finally,
the diet of subjects varied considerably across the seven participating countries.
Differences in the quality of food consumed would have created differences in fat
and roughage content and/or increased the rate of passage of substances through the
gastrointestinal tract. This may have reduced the period over which ALB could have
acted on the parasites, thereby reducing efficacy [26]-[28].
Kopp et al. [29]
demonstrated that a reduction in adult canine hookworm (A. caninum)
counts following chemotherapy did not always yield a reduction in FEC, due to an
increase in fecundity among the small residual worm population that survived the
anthelmintic treatment (i/e., density dependent fecundity), consequently confounding
the FECR. As described by Kotze and Kopp [30], density dependent effects could
be manifested in a FECR as a reduced drug efficacy for subjects with higher
pre-intervention FEC. However, this did not occur in the present study for
A. lumbricoides and hookworm. For T.
trichiura, the efficacy did decrease with increasing pre-intervention FEC,
but this should be interpreted with some caution. This effect was not consistent
across the different trials (e.g., no correlation in Vietnam but a positive
correlation in Tanzania), suggesting that other factors as discussed above may have
confounded this result. It is also possible that increases in FEC may have arisen
because of the inability of ALB to cure infections during the pre-patent period
(with an onset of patency after the pre-intervention egg count time point). This is
a complication that cannot be avoided in studies taking place in endemic areas where
transmission occurs daily because of soil and food contaminated with infective
stages of the parasites, and is not interrupted in the population during the period
of study. Finally, a negative correlation between the FEC and efficacy is expected,
as the probability of having a FEC of zero after treatment in the follow-up survey,
consequently a FECR of 100%, will be higher for low FEC than for high FEC
before the administration of the drug.
Our findings emphasize a need to adhere to strict standard operating procedures and
methodologies, and to change the WHO recommended threshold levels for the efficacy
of ALB [31], where
a FEC reduction rate below 70% in the case of A.
lumbricoides or below 50% for the hookworms are the currently
accepted thresholds. We recommend that in future monitoring of single-dose
ALB-dependent control programs a minimum FEC reduction rate (based on arithmetic
means) of >95% for A. lumbricoides and >90% for
hookworms are appropriate thresholds, and that efficacy levels below this should
raise concern. The great variability of the FECR for T. trichiura
and the relatively low efficacy of ALB, confirmed in this present study, indicate
that it is not possible to propose an efficacy threshold for this parasite based on
our data.
In conclusion, the present study is the first to evaluate drug efficacy of a
single-oral dose of ALB on such a scale and across three continents. The results
confirm the therapeutic efficacy of this treatment against A.
lumbricoides and hookworms, and the low efficacy against T.
trichiura. Efficacy varied widely across the seven different trials,
particularly in the case of T. trichiura and it remains unclear
which factors were principally responsible for this variation, although
pre-intervention FEC and age played clear roles in this respect. The FEC reduction
rate based on arithmetic means is the best available indicator of drug efficacy, and
should be adopted in future monitoring and evaluation studies of large scale
anthelmintic treatment programs. Finally, our findings emphasize the need to revise
the WHO recommended efficacy threshold for single dose ALB treatments.
|
10.1371/journal.pbio.0060160 | Human and Drosophila Cryptochromes Are Light Activated by Flavin Photoreduction in Living Cells | Cryptochromes are a class of flavoprotein blue-light signaling receptors found in plants, animals, and humans that control plant development and the entrainment of circadian rhythms. In plant cryptochromes, light activation is proposed to result from photoreduction of a protein-bound flavin chromophore through intramolecular electron transfer. However, although similar in structure to plant cryptochromes, the light-response mechanism of animal cryptochromes remains entirely unknown. To complicate matters further, there is currently a debate on whether mammalian cryptochromes respond to light at all or are instead activated by non–light-dependent mechanisms. To resolve these questions, we have expressed both human and Drosophila cryptochrome proteins to high levels in living Sf21 insect cells using a baculovirus-derived expression system. Intact cells are irradiated with blue light, and the resulting cryptochrome photoconversion is monitored by fluorescence and electron paramagnetic resonance spectroscopic techniques. We demonstrate that light induces a change in the redox state of flavin bound to the receptor in both human and Drosophila cryptochromes. Photoreduction from oxidized flavin and subsequent accumulation of a semiquinone intermediate signaling state occurs by a conserved mechanism that has been previously identified for plant cryptochromes. These results provide the first evidence of how animal-type cryptochromes are activated by light in living cells. Furthermore, human cryptochrome is also shown to undergo this light response. Therefore, human cryptochromes in exposed peripheral and/or visual tissues may have novel light-sensing roles that remain to be elucidated.
| Vision in animals is generally associated with light-sensitive rhodopsin pigments located in the eyes. However, animals ranging from flies to humans also possess ancient visual receptors known as cryptochromes in multiple cell types. In this work, we study the mechanism of light sensing in two representative animal cryptochromes: a light-sensitive Drosophila cryptochrome (Dmcry) and a presumed light-insensitive mammalian cryptochrome from humans (Hscry1). We expressed recombinant cryptochromes to high levels in living cells, irradiated the cells with blue light, and analyzed the proteins' response to irradiation with electron paramagnetic resonance and fluorescence spectroscopic techniques. Photoreduction of protein-bound oxidized FAD cofactor to its radical form emerged as the primary cryptochrome photoreaction in living cells, and was correlated with a light-sensitive biological response in whole organisms. These results indicate that both Dmcry and Hscry1 are capable of undergoing similar light-driven reactions and suggest the possibility of an as-yet unknown photo-perception role for human cryptochromes in tissues exposed to light.
| Cryptochromes are blue-light–absorbing photoreceptors found throughout the biological kingdom, involved in diverse and important signaling roles [1–3]. Cryptochromes were first identified in plants from a mutant of Arabidopsis thaliana (A. thaliana), hy4, which failed to show normal plant growth and developmental responses to blue light [4]. The N-terminal region of the HY4 encoding protein, renamed cryptochrome or A. thaliana cry1 (Atcry1) was found to be highly homologous to a previously characterized class of enzymes, DNA photolyases, which utilizes blue light as a source of energy for the repair of UV-light–generated DNA lesions [2,5]. However, cryptochrome did not repair DNA but instead, participated in numerous blue-light–dependent plant growth responses, including early seedling development, leaf and stem expansion, initiation of flowering, and gene regulation [6,7]. The defining characteristic of a cryptochrome-type photoreceptor is therefore a light receptor molecule that is structurally highly similar to DNA photolyases, but has lost DNA repair activity and acquired a novel role in signaling.
Subsequent to their discovery in plants, cryptochromes were identified in animal (human and mouse) systems by isolation of homologous cDNAs whose encoded proteins were likewise not functional in DNA repair [8]. Interestingly, these animal-type cryptochromes were more similar to a type of DNA photolyase that repairs 6–4 photoproducts, than to the type I cyclobutane pyrimidine dimer (CPD)-repairing DNA photolyases to which the plant cryptochromes are most closely related. Therefore, animal-type cryptochromes are thought to have evolved independently from different photolyase ancestors than plant cryptochromes [9]. A signaling role for animal-type cryptochromes was first identified in insects, Drosophila melanogaster (D. melanogaster), through isolation of a mutation in the D. melanogaster cryptochrome (Dmcry) resulting in failure to properly entrain the peripheral circadian clock [10,11]. The role of Dmcry as a light-sensing input to the circadian clock is now well established, occurring by interaction of Dmcry with known clock proteins such as timeless or period [12,13]. Thus, although derived from different evolutionary photolyase ancestors, both plant Atcry1 and Dmcry act as signaling molecules that undergo light-sensitive interactions with partners to initiate signaling reactions. A schematic of the basic structural characteristics and evolutionary relation of cryptochromes and photolyases is presented in Figure 1A.
Despite rapid advances in identifying the signaling pathways and molecular targets of cryptochromes in various organisms, the primary light-driven reactions that initiate the signaling process have remained elusive until very recently. Like DNA photolyases, cryptochromes bind flavin adenine dinucleotide (FAD) as a blue-light–absorbing cofactor [1,2]. However, the resting state of flavin in Atcry appears to be the fully oxidized redox form rather than the reduced form as found to be catalytically active in DNA photolyases [14]. Through a combination of in vitro studies with purified proteins and whole-cell in vivo spectroscopic techniques, it was deduced that protein-bound flavin both in Atcry1 and Atcry2 is reduced by blue light through intraprotein electron transfer resulting in accumulation of a relatively long-lived semiquinone intermediate form which is believed to represent the active signaling state [15–18]. This photoreaction is thought to be the basis for conformational changes that occur in the protein to initiate signaling [19,20]. When returned to darkness, plant cryptochromes slowly reoxidize back to the fully oxidized state of the flavin chromophore. In DNA photolyases, a similar light-induced photoreduction, known as photoactivation, can also occur but is relatively unimportant to biological activity since generally the FAD cofactor of DNA photolyase is fully reduced and mostly remains in this redox state independent of light conditions in vivo [21]. Therefore, plant cryptochromes have apparently transformed a minor photoreaction intrinsic to DNA photolyases to form the basis for a novel function as photoreceptor; a summary of the plant cryptochrome photocycle is presented in Figure 1B.
The mechanism of light activation of animal-type cryptochromes is currently unknown. Although structurally similar to plant cryptochromes, the animal proteins apparently stem from 6–4 photolyase-type ancestors and not from CPD photolyases (which are related to plant cryptochromes) [9] and so could have evolved a different photocycle. Nevertheless, purified preparations of Dmcry have been recently shown to undergo a photoreduction reaction similar to plant cryptochromes in vitro, to a relatively stable radical intermediate [22]. Furthermore, studies of wavelength sensitivity show little Dmcry activity above 500 nm (green/red) light [23,24]. These characteristics are consistent with the known mechanism of activation of plant cryptochromes. On the other hand, mutation of conserved tryptophan residues that play a role in photoreduction [21] to redox inactive Phe have not been reported to affect biological activity of either mouse cry [25] or a recently characterized insect cryptochrome from monarch butterfly (Danaus plexippus) (Dpcry1) and from other insects [26,27].
A further complication arose with the identification of a light-independent function for mammalian cryptochromes. Transgenic mice in which both existing cryptochrome alleles (mcry1 and mcry2) had been knocked out showed a complete absence of rhythmic activity. This led to the conclusion that mammalian cryptochrome functions as a central component of the circadian oscillator [28,29]. Like insect cryptochromes, mcry1 and mcry2 were shown to interact with known components of the mammalian clock and thereby obtain a novel biological role. However, these interactions were entirely independent of light [30], and moreover, the clock phenotype of mcry1 and mcry2 occurs in continuous darkness without light interruption over several days. Mammalian cryptochromes are similar to insect cryptochromes and apparently stem from the same 6–4 photolyase ancestor. Moreover, they are flavoproteins and show conservation in amino acids required for light activation in photolyases and other cryptochromes. Despite these similarities, the light-independent nature of mammalian cryptochrome response leads to the question of whether these signaling molecules retain the ability to respond to light at all.
The aim of the present study is to resolve the question of whether and how light activates animal-type cryptochromes. We have employed in vivo spectroscopic techniques including a novel application of electron paramagnetic resonance (EPR) to detect photoconversion of flavin and accumulation of radical in living whole cells. We have examined both Dmcry and Homo sapiens cryptochrome-1 (Hscry1) as representative light-sensitive and light-insensitive cryptochromes, respectively. These experiments showed that light activation of animal cryptochromes occurs by photoreduction and accumulation of a radical signaling intermediate, similar to plant cryptochromes and unlike photolyases. Furthermore, because Hscry1 undergoes the same photoreactions, mammalian cry is demonstrated to have the capacity to function as a light sensor.
In plant cryptochromes cry1 and cry2, the dark state of the flavin is in the oxidized form in vivo (Figure 1B). To determine the nature of the dark state of flavin in animal-type cryptochromes, we measured a classic action spectrum for Dmcry activity in living flies. An action spectrum is a dose-response curve for photoreceptor sensitivity in which the response of an organism is determined at multiple wavelengths of light and at multiple light intensities at each wavelength. In this way, the response will depend on how well the photoreceptor absorbs light at the given wavelength. The wavelength at which peak activity can be observed in the living organism indicates the absorption maximum (in which the light is absorbed at highest efficiency) of the responsible photoreceptor. If performed to sufficient resolution, such action spectra can be compared to the absorption spectrum of a purified pigment or photoreceptor and in this way identify the nature of the photoactive pigment implicated in a given biological response [6].
As a possible assay for Dmcry function, we investigated its characteristic property of degradation that followed upon photoreceptor activation. Levels of Dmcry protein decrease rapidly in flies subsequent to blue-light irradiation, likely due to conformational change in the photoreceptor leading to targeting to the proteasome [11,23,31]. This degradation can be quantitatively monitored by western blot analysis with Dmcry antibody. However, in order to be useful for action spectroscopy, the amplitude of the response (decline in Dmcry concentration) must be proportional to the number of photons of light energy absorbed by the photoreceptor, and not simply a delayed response with little direct relation to the light input signal. To test this property, we irradiated living flies for fixed time intervals with blue light (450 nm) and observed decrease in levels of Dmcry protein over time as previously described [11,23,31] (Figure 2A). Importantly, when blue-light irradiation was performed at different blue-light intensities, the time required to reach a given decline in Dmcry protein levels was proportional to the given light intensity. For example, to obtain a decrease to 50% of the original Dmcry protein concentration requires 6.2 min at irradiance of 200 μmol m−2 sec−1, 9.3 min at 150 μmol m−2 sec−1, and 12.5 min at 100 μmol m−2 sec−1 (from Figure 2A), respectively. Therefore, Dmcry degradation obeys the Bunsen-Roscoe law of reciprocity, indicating that it is a response to the total number of photons, independent of irradiance time, and so represents an accurate measure of photoreceptor light responsivity [32].
For generation of this Dmcry action spectrum, living flies were first dark adapted to accumulate maximum levels of dark state cryptochrome. Flies were then subjected to continuous irradiation at a set photon fluence rate of 17 μmol m−2 s−1 at wavelengths between 380 and 502 nm. Shorter wavelengths are impractical because of the increasing absorption from cell components and therefore increasing errors. Levels of Dmcry protein were monitored by western blot analysis. Illumination time was varied to provide the different irradiance, as permitted by reciprocity. Shorter or longer wavelengths of light proved to be ineffective at eliciting significant response (unpublished data). Dose-response plots of the time course of Dmcry protein degradation at different wavelengths of light showed linear decay as a function of the log irradiation time for points between 20% and 90% of dark levels of protein accumulation (Figure S1). The action spectrum is plotted from these dose-response curves using the total irradiance required to reduce Dmcry protein levels to 50% of dark controls at each wavelength (Figure 2B). The curve is inverted to give a visual image whereby the peak efficiency (the wavelength that required the shortest time to elicit 50% Dmcry degradation) represents the absorption maximum of the responsible photoreceptor. Peak wavelength sensitivity was at 450 nm, with defined shoulders around 420 and 480 nm, matching the spectrum of protein-bound oxidized flavin [33,34]. Oxidized flavin is therefore the likely photoactive pigment of Dmcry in living whole flies, similar to plant cryptochromes [14] and in marked contrast to DNA photolyases in which flavin is fully reduced [2].
We next determined the nature of the chemical reaction induced by light in animal cryptochromes in living cells. We performed baculovirus-driven expression of Dmcry and Hscry1 cryptochromes in Sf21 insect cells, where photoreceptor protein accumulates to sufficiently high levels for direct application of spectroscopic and biophysical techniques in vivo [18]. To verify whether cryptochrome-bound flavin can be directly observed, expressing whole Sf21 cells were harvested and placed intact inside a fluorimeter. Fluorescence emission was measured at 525 nm (characteristic of oxidized flavin) over an excitation range of 400–500 nm. Despite the substantial scatter due to measurements of these living intact cells, there was clearly observable signal increase peaking for excitation at 450 nm in cells overexpressing both Dmcry and Hscry1 as compared to uninfected control cells. These results are consistent with oxidized flavin bound to the dark state of the photoreceptors (Figure S2). These data showing increased oxidized flavin in cryptochrome-expressing cells are in agreement with the resting state of Dmcry determined from action spectroscopy (Figure 2B). Interestingly, mammalian cryptochrome also accumulates in the oxidized form and thereby shows functional similarity to Dmcry and not to DNA photolyases. Similar results have been previously obtained for Atcry1 [18].
To initiate the photochemical reaction, Dmcry- or Hscry1-expressing cells were irradiated with blue light and returned to the fluorimeter at intervals for measurement of excitation spectra. This assay detects change in levels of oxidized flavin in these living cells. For both Dmcry and Hscry1, peak excitation at 450 nm showed a progressive decrease over time that matches the spectra for photoreduction of oxidized flavin (Figure 2C and 2D). This decrease was not due to protein degradation since both Dmcry and Hscry1 protein levels remain stable throughout the time course of illumination in Sf21 cells (Figure S3). Furthermore, flavin reoxidation is observed when illuminated cell cultures are returned to darkness (unpublished data), indicating that no cryptochrome degradation has occurred. Therefore, both tested cryptochromes had undergone a photoreaction in vivo, leading to change in redox state of protein-bound flavin (Figure 2D).
A similar reaction, known as photoactivation, occurs in DNA photolyases, wherein the flavin chromophore is converted to the fully reduced form by an electron transfer reaction ultimately fed by an extrinsic reductant. An intraprotein electron transfer pathway from the protein surface to the buried flavin has been derived for this light-driven reaction in Escherichia coli DNA photolyase (EcPl) based on crystallographic structural information and on a combination of site-directed mutagenesis and spectroscopy [35–38]. This pathway comprises a chain of three tryptophan residues (W382–W359–W306) that are highly conserved throughout the photolyase/cryptochrome family. Recently, a study with purified Atcry1 has demonstrated the functional relevance of this reaction to cryptochrome photoreceptor activity [16] by substitution of redox-inactive phenylalanines for two tryptophan residues, W400 and W324, which are found in the Atcry1 sequence and crystal structure [39] at the homologous positions to W382 and W306 of EcPl, respectively. These mutant proteins (W400F and W324F) lack the predicted electron donor proximal to the flavin (W400) or exposed to the protein surface (W324). Both proteins were found to have impaired electron transfer activity in vitro and reduced biological activity in living plants. To determine whether flavin photoreduction may occur by a similar pathway of intermolecular electron transfer in animal-type cryptochromes, we have made point mutations in Dmcry of two conserved tryptophan residues. One mutation is distal to the flavin in this pathway (W342F), which corresponds to W306 in EcPl and W328F in Dpcry1 [26], respectively. Second, we have introduced a substitution into the middle member of the electron transfer chain of Dmcry, corresponding to W359 of EcPl. The mutant Dmcry proteins were expressed in Sf21 insect cells to high levels and subjected to in vivo fluorescence spectroscopy to follow flavin photoreduction.
For determination of in vivo photoreduction, cryptochrome-expressing Sf21 cells were irradiated with blue light and returned to the fluorimeter periodically to determine remaining levels of oxidized flavins. Unexpectedly, both wild-type and mutant Dmcry proteins showed similar rates of photoreduction in these living cells at high light intensity (150 μmol m−2 sec−1 white light), as did the W400F mutant of Atcry1 (Figure 3A). This result is surprising since purified preparations of Atcry1 W400F protein and of the W328F Dpcry1 (homolog to Dmcry W342F) showed greatly reduced photoreduction in vitro at even higher light intensities, and there was no significant radical accumulation after this time period [16,26,27]. Therefore, the efficiency of cryptochrome photoconversion in vivo is much higher than that of the purified, isolated protein in vitro, perhaps due to a more conducive redox environment and the presence of relevant electron donors/acceptors in vivo. Nevertheless, at lower photon fluence (10 μmol m−2 sec−1 blue light), a significant decline in the rate of photoreduction is observed in the phenylalanine mutants of both Dmcry and Atcry1 as compared to wild-type proteins (Figure 3B—see also Figure 4SA and 4SB for experiments performed with further reduced photon fluence). This phenomenon provides a consistent explanation for the observed biological activity of amino acid–substitution mutants in Dpcry and Atcry1. In Arabidopsis, biological activity of the mutant proteins (W400F and W324F) was determined at only low blue-light intensity and found impaired at this irradiance in vivo.
To determine the state of the photoreceptor (radical or fully reduced) in the activated cryptochrome, fluorescence emission techniques in whole cells are not sufficient as they can identify only the oxidized form of the flavin chromophore. It cannot, therefore, be concluded from the above studies whether photoreduction in vivo leads to accumulation of a semiquinone intermediate as for plant cryptochromes [16–18], or whether the fully reduced form of flavin accumulates in animal cryptochromes as for DNA photolyases [1,2]. To directly monitor for radical accumulation in response to light in vivo, whole-cell EPR spectra were recorded as previously described [17,18]. Intact Sf21 insect cells with overexpressed Dmcry or Hscry1 protein were irradiated in parallel with nonexpressing control cells at the identical intensities of blue light and rapidly frozen for EPR analysis. A paramagnetic species that does not accumulate in control cells was induced by blue-light irradiation of Dmcry- (Figure 4A, traces B and C) and Hscry1-expressing cells (Figure 4A, traces E and F). This species appears with similar kinetics to both plant cryptochromes [17,18]. Interestingly, there was detectable amount of a radical present even in the dark in some samples (see trace D), although not in all trials, perhaps due to concentrations below our level of detection. This result is in marked contrast to previous experiments, for which radical accumulation was never observed in unilluminated cells [17,18]. Finally, we have examined the Drosophila mutant proteins W397F and W342F for radical accumulation in vivo (Figure 4A, traces H–K). Saturating illumination (40 μmol m−2 sec−1 blue light) leads to accumulation of a radical intermediate form.
To further characterize these signals, X-band–pulsed ENDOR spectroscopy was applied to illuminated whole cells expressing Dmcry (Figure 4B). The observed spectrum in the expressing cells (trace A) is very similar to that obtained from the purified Dmcry protein (trace B). Both spectra differ from those of neutral flavin radicals as seen in plant cryptochromes and corroborate the assignment to an anionic radical species as given previously for the purified protein [22]. Taken together, the in vivo spectroscopic data conclusively indicate that the photocycle for both Dmcry and Hscry1 involves light-dependent flavin reduction and accumulation of the radical state.
Finally, it is necessary to establish the biological relevance of the observed in vivo photoconversion of animal cryptochromes. In plant cryptochromes, the radical state has been demonstrated to be the biologically active signaling state for both Atcry1 and Atcry2 [17,18]. This conclusion resulted from the observation that green light reversed the effect of blue light in the course of cryptochrome activation, due to photoconversion of the active, radical form to the fully reduced, inactive species [17,18,40] (see also Figure 1A). A simple means to determine whether light-induced radical accumulation also has biological relevance for animal cryptochromes in vivo is therefore to measure whether green light (above 525 nm) affects both Dmcry protein accumulation and the kinetics of cryptochrome photoreduction.
To test this prediction, we performed bichromatic irradiation of flies simultaneously with blue and green light (B+G) and compared the response to that obtained with the identical intensity of blue light by itself (B) (Figure 5A). Green-light irradiation by itself resulted in no change in Dmcry protein levels (unpublished data). In each of three independent trials, we observed more rapid decline in Dmcry protein levels in blue light (B) as compared to coirradiation with blue and green light (B+G). This antagonistic effect can only be explained by photoconversion of the (green-light absorbing) radical signaling state to an inactive redox form. In the case of Dmcry-expressing cell cultures, an effect of green light on cryptochrome photoreduction was directly monitored. Cell cultures irradiated with blue and green light (B+G) show accelerated photoreduction of Dmcry as compared to blue light (B) alone (Figure 5B). Although the accumulation of fully reduced flavin can not be directly monitored by this technique, these data are consistent with a shift in the overall flavin photoequilibrium subsequent to formation of the radical, and thereby consistent with the effect of green light on biological activity observed in living flies.
Although the present study so far has shown that mammalian cryptochromes undergo similar photoreactions to those of insect and plant, a functional role for light in biological activation remains to be demonstrated. To address this question, we have assayed for a form of activation of Hscry1 in response to light in living flies, where endogenous cryptochrome (Dmcry) is known to undergo light-dependent changes resulting in proteolysis (Figure 2). Transgenic flies expressing full-length Hscry1 under the control of the UAS promoter element were obtained by established procedures (see Materials and Methods). Expression of the recombinant Hscry1 was verified by western blot analysis in two independent transformed lines (A and B). Expressing flies were then dark adapted to accumulate maximal cryptochrome protein and subsequently irradiated with white light. Levels of Hscry1 were assayed during the course of the irradiation. Interestingly, as is the case for Dmcry, significant decrease in Hscry1 protein levels were observed shortly after transfer to white light (Figure 6). These results indicate that Hscry1 undergoes light-dependent proteolysis as does Dmcry in living flies. Since degradation of Dmcry correlates with activation by light and biologically relevant radical formation, a similar mechanism of biological activation is also likely for Hscry1.
In this work, we provide, for the first time, evidence for a photocycle of animal-type cryptochromes such as found in insects and mammals. Cryptochrome-bound flavin is found in an oxidized redox state in vivo, and light activation results in flavin photoreduction to a radical intermediate that represents the likely signaling state. The biological significance of this reaction is supported by the observation of antagonistic effects of green light on Dmcry function, which reduces levels of radical intermediate [17,18,40]. This mechanism contrasts with that of DNA photolyases in which flavin is fully reduced for catalytic activity. Most importantly, Hscry1 from a cryptochrome subfamily with no established light response also has the capacity to undergo this photoreaction in living cells, suggesting the possibility of novel light-sensing capabilities in humans.
A number of studies have indicated that Dmcry responsivity occurs primarily at wavelengths below 500 nm [23,24]. The current study extends these prior observations by providing sufficient fine structure to identify oxidized flavin, with peak of activity at 450 nm and defined shoulders around 420 and 480 nm, as the likely responsible photopigment in the visible range. Further corroboration for the assignment of oxidized flavin as the ground state for animal-type cryptochromes is provided by a classic action spectrum of phase shift in pupal emergence of Drosophila [41], a response involving phase shift of the circadian clock which is now known to be under the control of cryptochrome [11]. Consistent with the current work, peak activity was at 450 nm, and the spectrum matches that of protein-bound oxidized flavin.
Interestingly, Dmcry degradation in Schneider cells has been reported to have a peak of activity in the near-UV spectral region [24], whereas in living flies and Dmcry-expressing Sf21 cells, the peak of activity is at 450 nm (see Figure 2B). Like DNA photolyases, cryptochromes are proposed to bind folate derivatives as cofactors in addition to flavin [8]. In DNA photolyases, a folate derivative absorbs light primarily in the near-UV spectral region (370–400 nm) and transfers energy to the flavin chromophore [2]. Recently, a similar role for folate has been postulated in plant cryptochromes [42] whereby light energy for photoreduction is transferred to flavin through a UV antenna pigment. It is therefore likely that the reported near-UV responsivity of Dmcry in Schneider cells also results from light absorption by a folate (or another, yet unspecified) antenna pigment. In that case, absence of near-UV responsivity in Dmcry extracted from whole flies (Figure 2B) suggests that the second chromophore of animal-type cryptochromes may not be available in the majority of insect cell types. This is in line with older experiments done with DNA photolyases, where a low binding constant of the folate chromophore and a therefore heterogeneous folate concentration was concluded.
In plant (Atcry1 and Atcry2) cryptochromes, flavin photoreduction leading to a meta-stable neutral radical accumulation can be observed in in vitro experiments. This property of purified plant cryptochrome contrasts published DNA photolyase data, in which oxidized flavin is rapidly converted to the fully reduced redox state (necessary for DNA repair). Recently, photoreduction experiments were performed with purified preparations of several insect cryptochromes in vitro resulting in similar photoreactions (accumulation of radical and not fully reduced flavin), although the anionic radical, and not the neutral radical, was accumulated [22,26,27]. These results are consistent with the presently derived in vivo photocycle for animal cry activation.
Results from recent studies performed with various insect cryptochromes (fruitfly [Dmcry], butterfly [Dpcry1], mosquito [Agcry1], and silk moth [Apcry1]) [26,27] have called into question the assignment of oxidized flavin as the ground state for animal cryptochromes and argued against a photocycle involving flavin photoreduction. Their interpretation focused on the observation that substitution of amino acids necessary for flavin photoreduction in vitro does not abolish biological activity of these proteins in vivo. This apparent discrepancy between the absence of photoreduction in vitro yet significant biological activity in vivo is resolved by the observation that amino acid substitutions abolishing in vitro photoreduction of purified Dmcry does not, in fact, abolish photoreduction activity in vivo. Photoreduction of oxidized flavin measured by fluorescence techniques (Figure 3A and 3B) in these substitution mutants correlates with concomitant appearance of anionic radical as determined by EPR spectroscopic techniques (Figure 4). The same is true for the W400F mutant of Atcry1, which shows normal rates of photoreduction in vivo under high light intensities even though flavin photoreduction in vitro is virtually zero under these conditions [16]. Cryptochrome photoreduction, therefore, occurs far more efficiently, and by additional alternate pathways, in vivo than is observed for the purified protein in vitro. A similar discrepancy between the light required to activate a purified photoreceptor protein in vitro as compared to activation in vivo has been noted for other photoreceptors, for instance, the class of phototropins [34,43,44] in which blue-light–dependent autophosphorylation requires a much higher irradiance in vitro to obtain a similar effect than is required in vivo.
Results from recent studies showing reduced biological activity at lower light intensity in the W342F mutant of Dmcry [27] are consistent with our observed reduced rates of photoreduction at low photon fluence (Figure 3B). Although quantitation was not formally performed, a prior study in which function of amino acid substitutions in the Trp triad of Dmcry was analyzed [25] also showed apparent reduced biological activity in W-F substitution mutants. In this study, the authors proposed that F can function as electron donor similarly to W, which, however, is not observed [16,26,27]; nevertheless, their data are entirely consistent with the present work. Finally, observations from point mutations that reduce the rate of radical formation in Apcry1 (C402A) abolish protein function in vivo [27] are entirely consistent with our assignment of the radical as the signaling state of the receptor [16]. A proposed mechanism whereby the anionic radical may be the resting state [27] is unlikely given that peak activity is not observed at 470 nm either in the present (Figure 2) or former studies of wavelength sensitivity for this receptor [24,41].
The derived photocycle of animal cryptochromes is therefore similar to the reaction mechanism of plant cryptochromes (Figure 1B). Both photocycles involve reduction of flavin leading to cycling between radical (active) and oxidized (inactive) redox forms. Since these different families of cryptochromes evolved independently from unrelated DNA repair enzyme ancestors, there must be a latent property of DNA photolyases that lends itself to development of photoreceptor properties. The likeliest possibility is that the flavin semiquinone form confers some conformational change on the protein, which can be recognized by yet-unidentified signaling partners and thereby be readily adapted to a role in a signaling pathway. In addition to the classic plant and animal-type cryptochromes, a third family of cryptochromes (cryDASH) has been identified in Synechocystis and many other organisms [45]. CryDASH cryptochromes are structurally similar to DNA photolyases, but do not efficiently repair DNA. They evolved from a 6–4 photolyase ancestor but are apparently unrelated to either the plant (Atcry1 and Atcry2) or animal-type cryptochromes described in this study. Although no information as yet exists on the in vivo redox state and photocycle of cryDASH, the purified proteins are converted by light to the fully reduced flavin in vitro [45,46] and retain some single-strand DNA repair activity [47], in this respect appearing more similar to DNA photolyases. It is therefore possible that an additional, entirely unrelated photocycle has evolved for cryDASH-type cryptochromes that is not similar to plant or animal-type cryptochromes.
Perhaps the most intriguing finding of the present study is that mammalian cryptochromes, in particular Hscry1, are responsive to light in vivo. Mammals are generally large, dense (and also often nocturnal) animals, and it makes sense that a molecule such as cryptochrome, which is an essential component of the circadian clock, should be regulated by other means than by direct absorption of light. In fact, with the exception of an isolated report of cryptochrome effect on pupil dilation in mice [48], there has to date been no definite report of light-responsive phenotypes attributed to mammalian cryptochromes at all. Nevertheless, as the present study has shown, Hscry1 can undergo photoconversion in living organisms via a mechanism conserved with that of light-responsive cryptochromes. Hscry1 further undergoes light-dependent proteolysis in living flies, similar to the response mediated by appropriate E3 ubiquitin ligases such as COP1 in the case of Arabidopsis cry2 [49], Therefore, the observed light sensitivity of Hscry1 in Drosophila is likely to result from surface changes leading to fortuitous recognition of the activated form by a fly E3 ligase. Since Hscry1 is widely distributed in many tissue types of humans, it could be activated by light in the retina or in locations close to the surface, providing the basis for novel biological signaling functions that remain to be determined.
As a light source, white light was produced by slide projectors placed before interference filters of 8–15-nm half-band width (Schott Glaswerke). Filters used in the irradiations for the action spectrum were 380 ± 10 nm, 402 ± 12 nm, 418 ± 13 nm, 428 ± 10 nm, 438 ± 11 nm, 445 ± 10 nm, 466 ± 11 nm, 471 ± 16 nm, 492 ± 15 nm, 502 ± 15 nm, 515 ± 15 nm, and 525 ± 12 nm.
Preparation of fly extracts and detection of dcry was performed essentially as described [50]. Two-week-old adult Drosophila (after eclosion) were adapted to dark for 3 d prior to an experiment. Between ten and 15 flies were placed under the indicated wavelengths of light during their subjective night phase (between circadian time [CT] 20–22) and irradiated for the indicated times. Unirradiated control flies did not show variation in Dmcry levels during the time period that illuminations were performed on the test flies. Whole flies were harvested into liquid nitrogen, and extracted proteins prepared as described [50,51]. A total of 20 μg of protein was loaded per lane of an SDS polyacrylamide gel and transferred to PVDF membrane. Protein load was verified prior to load by Bio-Rad Bradford assay and subsequently by Coomassie staining of the blotted gels. Western blot analysis was performed by established methods with affinity-purified anti-Dmcry antibody [51] or Hscry1 antibody to a peptide fragment comprising amino acids 178–194 of the coding sequence (Neosystems). Quantitation of resolved bands was performed digitally using Quantity One imaging software from Bio-Rad.
cDNA for Dmcry and Hscry1 were cloned into a baculovirus transfer vector (pBakPak9; Clontech) by established protocols (Clontech). A histidine tag was introduced upstream of the initial ATG in each construct to allow fast purification. For protein expression, amplified viral supernatant of the recombinant clones were mixed with cell culture and incubated as recommended (Clontech). Protein expression was verified by western blot analysis with appropriate antibody of both whole-cell extract and of proteins purified by metal-affinity chromatography. Presence of flavin was verified by absorption and fluorescence spectra of the purified proteins. For construction of Dmcry mutants W397F and W342F, side-directed mutagenesis was performed by the recommended protocol using Altered Sites II in vitro mutagenesis kit (Promega). The primers designed for mutagenesis are as follows: W397F, GTGCTGCAGTCCATGCTCGAAGCTCTGCCACAA; W342F, CTCGTTCGGCTTAGCGAACGGGATGCTCAGGCA. All clones were sequenced in entirety prior to protein expression.
Whole-cell fluorescence experiments of Sf21 insect cells expressing recombinant cryptochrome photoreceptors were performed essentially as described [18,42]. Living Sf21 insect cells expressing cryptochrome protein or uninfected controls were centrifuged from culture medium, resuspended in PBS buffer (0.02 M sodium phosphate [pH 7.4], 0.15 M sodium chloride), and placed directly into cuvettes at 10 °C for measurement of fluorescence spectra. Fluorescence emission at 525 nm was monitored in a Varian fluorescence spectrophotometer over a range of excitation wavelengths or at a single designated wavelength as indicated (see Figure 3 legend). Excitation and/or emission spectra were always determined in parallel, both for infected (cryptochrome-expressing) and uninfected cell cultures at identical cell density. For light treatments, samples were removed from the spectrophotometer and placed on ice. Illumination was carried out for the indicated times, using the designated interference filters. Samples were then returned to the fluorescence spectrophotometer to monitor differences in excitation and emission spectra. All experiments were repeated for a minimum of three independent trials.
Sf21 insect cells expressing both Dmcry and Hscry1, control Sf21 cells, and purified Dmcry from Sf21 insect cells were resuspended in phosphate-buffered saline supplemented with 50% (v/v) glycerol in the dark. Aliquots were transferred into EPR quartz tubes (3 mm i.d.) and illuminated for different times at 290 K with blue light (Halolux 100HL; Streppel) using a 420–470-nm band filter (Schott). Samples were then frozen rapidly under illumination in liquid nitrogen and stored therein.
X-band continuous-wave (cw) EPR spectra were recorded using a pulsed EPR spectrometer (Bruker Elexsys E580) with a cw resonator (Bruker ER 4122SHQE) immersed in a helium-gas flow cryostat (Oxford CF-935). X-band–pulsed ENDOR spectra were recorded on the same spectrometer using an ENDOR accessory (Bruker E560-DP), an RF amplifier (Amplifier Research 250A250A), and employing a dielectric-ring ENDOR resonator (Bruker EN4118X-MD-4W1). The temperature was regulated to ±0.1 K by a temperature controller (Oxford ITC-503S). The cw-EPR spectra were recorded at 120 K with a microwave power of 3.0 μW at 9.7 GHz microwave frequency with field modulation amplitude of 0.3 mT (at 100 kHz modulation frequency). For Davies-type ENDOR spectroscopy, a microwave pulse sequence π–T–π/2–τ–π with 64- and 128-ns π/2- and π-pulses, respectively, and a RF pulse of 10-μs duration starting 1 μs after the first microwave pulse were used. The pulse separations T and τ between the microwave pulses were selected to be 13 μs and 500 ns, respectively. To avoid saturation effects due to long relaxation times, the entire pulse pattern was repeated with a low repetition frequency of 200 Hz. Spectra were taken at a magnetic field of 345.7 mT and a microwave frequency of 9.71 GHz.
The entire coding sequence from the 5′ ATG onwards of HSCRY1 was introduced behind the upstream UAS promoter element of the pP(UAST) vector via PCR amplification and verified by sequencing; constructs were subsequently introduced into flies by standard methods [51]. Two high-expressing transformed lines with insertions on chromosome II and III, respectively, were selected for further study. Expression from the UAS upstream promoter element was induced by crossing flies to homozygous tim-GAL4 lines expressing the gal4 transcription activator driven by the TIMELESS (tim) promoter as described [52]. Parental lines used for crosses were: yw;tim-GAL4 [53] and Hscry1 insertion lines w;31.2A and w;26.9A. Expression was verified in F1 heterozygote progeny by western blot analysis with anti-Hscry1 antibody (monoclonal antibody ref: BIN165979; http://Antikoerper-online.de).
|
10.1371/journal.pntd.0002528 | Regulatory T Cells in Peripheral Blood and Cerebrospinal Fluid of Syphilis Patients with and without Neurological Involvement | Syphilis, a sexually transmitted disease caused by spirochetal bacterium Treponema pallidum, can progress to affect the central nervous system, causing neurosyphilis. Accumulating evidence suggest that regulatory T cells (Tregs) may play an important role in the pathogenesis of syphilis. However, little is known about Treg response in neurosyphilis.
We analyzed Treg frequencies and Transforming Growth Factor-β (TGF-β) levels in the blood and CSF of 431 syphilis patients without neurological involvement, 100 neurosyphilis patients and 100 healthy donors. Suppressive function of Tregs in peripheral blood was also assessed. Among syphilis patients without neurological involvement, we found that secondary and serofast patients had increased Treg percentages, suppressive function and TGF-β levels in peripheral blood compared to healthy donors. Serum Rapid Plasma Reagin (RPR) titers were positively correlated with Treg numbers in these patients. Compared to these syphilis patients without neurological involvement, neurosyphilis patients had higher Treg frequency in peripheral blood. In the central nervous system, neurosyphilis patients had higher numbers of leukocytes in CSF compared to syphilis patients without neurological involvement. CD4+ T cells were the predominant cell type in the inflammatory infiltrates in CSF of neurosyphilis patients. Interestingly, among these neurosyphilis patients, a significant decrease in CSF CD4+ CD25high Treg percentage and number was observed in symptomatic neurosyphilis patients compared to those of asymptomatic neurosyphilis patients, which may be associated with low CSF TGF-β levels.
Our findings suggest that Tregs might play an important role in both bacterial persistence and neurologic compromise in the pathogenesis of syphilis.
| 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 affected neurosyphilis patients may not have any symptoms, some of the patients will develop severe symptoms that can be life-threatening. Regulatory T cells (Tregs) are a subpopulation of CD4+ T cells functioning in suppression of immune-mediated bacterial clearance and tissue damage. In this study, we conduct a comparative analysis of regulatory T cells (Tregs) in the blood and cerebrospinal fluid (CSF) of syphilis patients without neurological abnormalities, and neurosyphilis patients with or without symptoms. Our results show that neurosyphilis patients had higher Treg percentage in blood than syphilis patients without neurological involvement, suggesting that neurological progression in syphilis patients is associated with an increase in blood Treg numbers. Strikingly, a decrease in Treg percentage and numbers in CSF of symptomatic neurosyphilis patients was observed compared to asymptomatic neurosyphilis patients. These results may implicate reduced CNS Treg response as a factor underlying the development of symptoms in some neurosyphilis patients. Our findings highlight a discordant Treg response in blood and CSF in symptomatic neurosyphilis patients and further underscore the fascinating complexity of immune response in syphilis.
| China has experienced an expanding epidemic of syphilis infection in the last 10 years [1], [2]. In 2011, the national incidence rate was 32.04 per 100,000 population and 429,677 new cases were reported [3]. This sexually transmitted disease has reemerged as a significant public health issue in China due to its serious, irreversible sequelae [4] and its strong association with HIV infection [5]. The rapid rise in syphilis rates in China highlight the importance of understanding of the pathogenesis of syphilis and its complications.
The spirochetal bacterium, Treponema pallidum (T. pallidum), is the etiologic agent of syphilis [6], [7]. After T. pallidum infection, mammalian hosts mount robust humoral and cellular immune responses aimed at spirochetal clearance [8], [9], [10], [11]. However, T. pallidum has the ability to escape the host immune response and establish persistent infection. There are several strategies used by the spirochete to resist host immune effector mechanisms including poor antigenicity [12], [13], antigenic variation of membrane proteins [14], [15], [16], and impaired antibody-mediated opsonization [17]. Interestingly, several studies have demonstrated that T. pallidum may also actively harness host immune suppression mechanisms to facilitate persistence and dissemination [18], [19]. A recent study has demonstrated that T. pallidum antigen TpF1 could promote development of regulatory T cells (Tregs) in the patients with secondary syphilis [18]. Tregs represent a unique population of CD4+ T cells with potent immune suppressive activity [20], [21]. This regulatory CD4+ T cell population is classically defined by high expression of CD25 (IL-2 receptor α-chain) [22]. The forkhead family transcription factor Foxp3, the most definitive signature, is critical for Treg development and function [23]. Emerging evidence from human patients and animal models has demonstrated that Tregs contribute to impaired immune responses and chronic infection with diverse organisms [24], including mycobacterium tuberculosis [25], helicobacter pylori [26], hepatitis B virus [27], [28], HIV [29], and plasmodium falciparum [30]. The enhanced Treg response in early syphilis patients may down-regulate immune effector function to allow survival of T. pallidum within the host.
T. pallidum infection can infect many organs, including central nervous system (CNS). This form of syphilis is termed neurosyphilis. Neurosyphilis may affect the meninges or brain or spinal cord parenchyma and may be asymptomatic or symptomatic [4], [31]. Meningeal neurosyphilis usually appears during the first few years of T. pallidum infection. Patients with meningeal neurosyphilis may be manifested by meningitis (headache, stiff neck, and cranial nerve abnormalities) or meningovasculitis (focal CNS ischemia or stroke). Parenchymal neurosyphilis, presenting as general paresis and tabes dorsalis, occur in the later course of the disease, often decades after the primary infection [4], [32]. The mechanisms underlying the development of symptomatic neurosyphilis in some patients are largely unknown.
Previous studies have extensively characterized immune cell infiltrates of early syphilis lesions [8], [9], [10] and indicated that the clinical manifestations of early syphilis result from collateral tissue damage caused by host immunity to T. pallidum [6], [33]. However, little is known about the immune response in neurosyphilis patients. In the present study, we performed a comparative analysis of Tregs in peripheral blood and cerebrospinal fluid (CSF) from neurosyphilis patients and syphilis patients without neurological involvement. We found that symptomatic neurosyphilis patients had lower Treg frequencies and numbers in CSF compared to asymptomatic neurosyphilis patients, indicating that an immunopathological mechanism might be present in the onset of neurological symptoms.
This study was performed at the Shanghai Skin Disease Hospital between June 2009 and Jan 2012. The hospital is located in central Shanghai, where the syphilis prevalence is highest in China [1]. The Sexually Transmitted Diseases (STD) center in this hospital is the major STD clinic in Shanghai, which provides screening, diagnosis and treatment for most sexually transmitted diseases, including syphilis. As one of the biggest STD centers in China, more than 300 patients are served in this clinic per day. Although most of our patients are walk-in, some are referred to our clinic by their doctors at other hospitals across the country. This study was approved by the Ethics Committee of the Shanghai Skin Disease Hospital. Written informed consent was obtained from all participants.
Syphilis was determined based on medical history, physical, neurological and psychiatric symptoms and signs, and the results of nontreponemal and treponemal serological tests. The excluded criteria include HIV; prior syphilis or syphilis treatment (except in the serofast syphilis group); history of systemic inflammatory, autoimmune disease, other underlying acute or chronic disease, were receiving anti-inflammatory medications, were immunocompromised, or use of antibiotics or immunosuppressive medications in the last four weeks. Peripheral blood was collected from all healthy donors and syphilis patients. Lumbar punctures were encouraged to be performed if i) patients had neurological or psychiatric signs or symptoms, ii) patients whose serum RPR≥1∶32, regardless of stage or presentation, iii) patients whose serofast state was more than 2 years and who are anxious regarding their serofast state. 100 healthy donors, who visited Shanghai Skin Disease Hospital voluntarily for STD prevention and a medical check-up, were recruited to the study. All healthy control subjects were negative for HIV and serological tests for syphilis.
Primary syphilis: i) Chancres or ulcers; and/or ii) detection of spirochetes in a dark-field microscopy examination; and iii) positive RPR confirmed by Treponema pallidum particle agglutination assay (TPPA); and iv) absence of other causes of genital ulcers, including herpes simplex virus (HSV) infections. Secondary syphilis: i) positive RPR confirmed by TPPA; and ii) skin or mucocutaneous lesions; Latent syphilis: i) positive RPR confirmed by TPPA; and ii) without skin or mucocutaneous lesions or any symptoms of syphilis; Serofast syphilis: i) previously treated syphilis of any stage; ii) an appropriate 4-fold decline in serum RPR titer at 6 months after treatment (Benzathine penicillin 2.4 MU/qw im for 2 or 3 weeks or procaine penicillin 0.8 MU/day im for 15 days in most cases, if patient allergic to penicillin ceftriaxone 250 mg/day im for 10 days would be as an alternative); iii) persistently reactive serum RPR two or more years after treatment; iv) no evidence of reinfection. The clinical and laboratory characteristics of 71 patients with primary syphilis, 136 patients with secondary syphilis, 127 patients with latent syphilis, and 97 patients with serofast syphilis were shown in Table 1.
All neurosyphilis patients have positive serum RPR and TPPA tests. The diagnosis of confirmed neurosyphilis also 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 nonreactive CSF-VDRL but reactive CSF-TPPA with either or both of the following: i) CSF protein concentration >45 mg/dL or CSF white blood cell (WBC) count ≥8/µL in the absence of other known causes for these abnormalities; ii) neurological or psychiatric manifestations consistent with neurosyphilis without other known causes for these abnormalities. Fourteen patients with presumptive neurosyphilis were also included in the study and the data of these patients were combined with those of confirmed neurosyphilis patients for analysis. In the case of presumptive neurosyphilis, the patient has a nonreactive CSF-VDRL test plus a reactive CSF-TPPA along with either or both of the following: (i) elevated CSF proteins (normal: 15–45 mg/dL) or elevated CSF white blood cell (WBC) count (normal: <8/µL) in the absence of other known causes of the abnormalities; (ii) clinical neurological or psychiatric manifestations without other known causes of these clinical abnormalities.
Neurosyphilis is categorized as asymptomatic, meningeal (meningitis and meningovasculitis) and parenchymal (general paresis and tabes dorsalis). Asymptomatic neurosyphilis is defined by the presence of CSF abnormalities consistent with neurosyphilis and the absence of neurological and psychiatric signs or symptoms. Meningitis is diagnosed by CSF abnormalities and headache, stiff neck, nausea, or cranial neuropathies. Meningovasculitis is defined by clinical features of meningitis and stoke with or without neuroradiological confirmation. General paresis is characterized by personality changes, dementia and psychiatric symptoms including mania or psychosis. Tabes dorsalis is characterized by sensory loss, ataxia, lancinating pains, and bowel and bladder dysfunction. All patients diagnosed with neurosyphilis should have no other known causes for these clinical abnormalities. The features of 100 neurosyphilis patients are shown in Table 2. These patients are mutually exclusive of those in Table 1.
Peripheral blood mononuclear cells (PBMC) were isolated from whole blood from syphilis, neurosyphilis patients and healthy donors via density centrifugation over Lymphoprep (Axis-Shield). CSF was centrifuged and stained immediately at 4°C after spinal tap. The volume was 5 mL. Multicolor fluorescence activated cell sorting (FACS) analysis was performed using the following antibodies: PE-, FITC-, PerCP, or PE-Cy5-conjugated antibodies against human CD45 (Biolegend), CD3 (Biolegend), CD4 (Biolegend), CD25 (Biolegend). For Foxp3 staining, cells were stained using One Step Staining Human Treg Flow Kit (Biolegend) according to the manufacturer's protocols. Cells were assessed with FACScalibur (Becton Dickinson) or Epics XL (Beckman Coulter) cytometers as previously described [34]. For CSF samples, acquisition of ≥5,000 events for gated CD45+ cells was performed. The CSF Treg number was defined as the total number of CSF cells multiplied by the percentage of Tregs identified by flow cytometry. Data were analyzed using FlowJo software (Tree Star).
Treg suppression assay was performed as described [35], [36]. Briefly, PBMC were used for CD4+ CD25+ and CD4+ CD25− T cell isolation using a Regulatory T Cell Isolation Kit according to the manufacturer's instruction (Miltenyi Biotec). Purity of the cell fractions as determined by flow cytometry was >90%. Purified CD4+ CD25− T responder cells (5×104 cells/well) were incubated in RPMI 1640 medium with 10% FBS in 96-well U-bottom plates precoated with anti-CD3 antibody (1 µg/mL; eBioscience). To assess suppressive ability, purified autologous CD4+ CD25+ T cells were added, at a CD25+/CD25− ratio of 1∶1, 1∶2, 1∶4, or 1∶8. All cells were cultured in a final volume of 200 µl in the presence of 2×104 irradiated allogeneic PBMC/well. After 4 days of culture, [3H] thymidine (Amersham) was added for an additional 18 h to each well. [3H] thymidine incorporation was measured using a liquid scintillation counter. Percent inhibition of proliferation was determined as (1- [3H] thymidine incorporation of CD25+ and CD25− T cells coculture/[3H] thymidine incorporation of CD25− T cells alone)×100.
Serum and CSF TGF-β1 levels were determined using Human TGF-β1 ELISA kit from eBioscience.
We performed statistical analysis using GraphPad Prism version 5.01 (GraphPad Software). All datasets were first assessed for deviation from a normal distribution using the D'Agostino-Pearson omnibus normality test. Non-normally distributed variables were compared between groups using the nonparametric Kruskal–Wallis test followed by Dunn's multiple comparison tests. If the variables were approximately normally distributed, differences between experimental groups were analyzed using one-way ANOVA followed by Bonferroni test for the selected pairs. Pearson correlation analysis was used to determine the relationship between the frequency of CD4+ CD25high Treg and other parameters. A value of P<0.05 was considered significant.
Human Tregs were identified as CD4+CD25high or CD4+Foxp3+ T cells [20], [21]. The representative gating strategy for CD4+ CD25high and CD4+ Foxp3+ T cells are depicted in Figure 1A. The majority of Foxp3+ T cells co-expressed high levels of CD25 (Figure 1A). The baseline frequency of CD25high Tregs among CD4+ T cells in PBMCs from healthy individuals was 2.7%±0.1% (Figure 1B). A comparison between syphilis patients and healthy individuals revealed a 1.3-fold increase in mean frequency of CD4+ CD25high T cells in primary syphilis patients (3.6%±0.2%, p<0.01), 1.7-fold increase in secondary syphilis patients (4.5%±0.2%, p<0.001), 1.5-fold increase in early latent syphilis patients (4.1%±0.2%, p<0.001), and 1.7-fold increase in serofast syphilis patients (4.7%±0.3%, p<0.001) (Figure 1B). Consistently with CD25 expression, the highest percentage of Foxp3+ Tregs among CD4+ T cells were observed in patients with secondary syphilis (4.3%±0.4%, p<0.001) and serofast syphilis (4.3%±0.3%, p<0.001) patients, followed by latent syphilis (3.9%±0.4%, p<0.01) and primary syphilis patients (3.6%±0.4%, p<0.05), which were all significantly higher than healthy donors (2.3%±0.1%) (Figure 1B).
We next investigate the suppressive function of Tregs from syphilis patients on T cell proliferation. CD4+ CD25+ suppressor T cells were cocultured with autologous CD4+ CD25− T responder cells at different ratios (suppressor/responder ratios: 1∶1, 1∶2, 1∶4, and 1∶8). We found that blood CD4+ CD25+ Tregs isolated from secondary syphilis (84.0%±1.4%, P<0.05) and serofast syphilis (84.3%±3.0%, P<0.01) but not primary syphilis (74.5%±1.1%, P>0.05) and latent syphilis (73.8%±1.1%, P>0.05) patients exhibited significantly higher suppressive activity than healthy controls (66.3%±1.1%) at a 1∶1 (suppressor: responder) ratio (Figure 1C). Significant increases in suppressive effect of CD4+ CD25+ Tregs were also observed at ratios of 1∶2 and 1∶4 in secondary and serofast syphilis patients compared with healthy donors (Figure 1C). These data indicated that CD4+ CD25+ Tregs derived from secondary and serofast syphilis patients display enhanced suppressive function.
Since Transforming Growth Factor-β (TGF-β) is critical to Treg differentiation and suppressive function [37], [38], [39], we determined whether higher Treg frequency and function in syphilis patients were associated with serum TGF-β levels. It was shown that serum concentrations of TGF-β were significantly increased in patients with secondary (5.4±0.7 ng/ml, P<0.001) and, to a lesser extent, in primary syphilis patients (4.4±0.9 ng/ml, P<0.05), latent patients (4.6±0.5 ng/ml, P<0.01) and serofast patients (4.4±0.6 ng/ml, P<0.01) compared with healthy controls (1.1±0.2 ng/ml) (Figure 1D). There was a positive correlation between the percentage of circulating CD4+ CD25high Tregs and serum TGF-β levels in these syphilis patients (r = 0.20, P<0.05, Figure 1E).
Nontreponemal test antibody titers usually correlate with disease activity [40]. We thus assessed whether serum RPR titers were associated with circulating Treg percentage in these syphilis patients. Pearson correlation analysis showed that there was a positive correlation between the percentage of circulating CD4+ CD25high Tregs and serum RPR titer in secondary syphilis (r = 0.27, P<0.01, Figure 2B), latent syphilis (r = 0.27, P<0.05, Figure 2C) and serofast (r = 0.44, P<0.01, Figure 2D) syphilis patients, but no correlation in primary syphilis patients (r = 0.10, P = 0.44, Figure 2A).
If untreated or treated improperly, some syphilis patients will progress to neurosyphilis. To investigate whether Tregs are associated with the progression of neurosyphilis, we analyzed Treg numbers in the peripheral blood of 49 asymptomatic and 41 symptomatic neurosyphilis patients. As shown in Figure 3A and 3B, syphilis patients with neurological involvement (including both asymptomatic and symptomatic syphilis patients) had higher percentage of CD4+ CD25high Tregs (4.7%±0.2%, P<0.001) and CD4+ Foxp3+ Tregs (5.0%±0.4%, P<0.001) in peripheral blood compared with healthy individuals (2.7%±0.1% and 2.4%±0.1%, respectively). Compared to syphilis patients without neurological involvement (including primary, secondary, latent and serofast syphilis patients), there was a slight but not significant increase in CD4+ CD25high Treg frequency in peripheral blood of neurosyphilis patients (P = 0.06) (Figure 3A), but the percentage of CD4+ Foxp3+ Treg were significantly higher (P<0.05) (Figure 3B). Among syphilis individuals with neurological involvement, there was no significant difference in CD4+ CD25high Treg frequency (P>0.05) (Figure 3C) and CD4+ Foxp3+ Treg frequency (P>0.05) (Figure 3D) in peripheral blood among asymptomatic, meningeal, and parenchymal neurosyphilis patients.
CSF mononuclear pleocytosis is one of diagnostic criteria for neurosyphilis [41], [42]. As expected, higher numbers of leukocytes were observed in asymptomatic (14±3 cells/µL), meningeal (35±15 cells/µL) and parenchymal (16±4 cells/µL) neurosyphilis patients compared to those from syphilis patients without neurological involvement (4±1 cells/µL) (P<0.001, P<0.001, and P<0.001, respectively) (Table 3). Among the CSF leukocytes, higher percentage of CD4+ T cells were found in patients with asymptomatic (41.8%±2.3%, P<0.01) and parenchymal (46.4%±2.1%, P<0.001) neurosyphilis compared with syphilis patients without neurological involvement (29.7%±2.4%) (Table 3). There was no significant difference in CSF CD4+ T cell frequency (P>0.05) among different types of neurosyphilis patients (Table 3).
The average percentage of CD25high Tregs in the CD4 compartment was 22.0%±1.0% for the patients without neurological involvement and did not differ from those with asymptomatic neurosyphilis (20.0%±1.1%, P>0.05). Both meningeal (12.5%±1.4%) and parenchymal (12.0%±1.2%) neurosyphilis patients showed pronounced decreases in CD4+CD25high Treg percentage compared to syphilis patients without neurological involvement (P<0.05, P<0.001, respectively) and asymptomatic neurosyphilis patients (P<0.05, P<0.001, respectively) (Table 3). Due to preferential accumulation of CD4+ T cells in the CSF of neurosyphilis patients, both asymptomatic and symptomatic neurosyphilis patients have higher numbers of CD4+CD25high Tregs than syphilis patients without neurological involvement. Interestingly, lower number of Tregs was observed in meningeal (0.9±0.3 cells/µL) and parenchymal (0.5±0.1 cells/µL) neurosyphilis patients than asymptomatic neurosyphilis patients (1.2±0.2 cells/µL). In addition, meningeal (3.4±0.9 ng/ml) and parenchymal (2.8±0.5 ng/ml) neurosyphilis patients had significantly lower CSF TGF-β levels than asymptomatic neurosyphilis (10.7±2.0 ng/ml) and syphilis patients without neurological involvement (8.2±1.7 ng/ml), indicating that decreased CD4+ CD25high Treg frequencies in CSF of symptomatic neurosyphilis patients may be associated with low CSF TGF-β concentration.
Syphilis is a multistage chronic disease, which can cause damage to diverse tissues and organs. An influx of immune cells to skin lesions of early syphilis patients not only mediates bacterial clearance but also lead to tissue damage and clinical symptoms [9], [10], [43], [44]. Our prior study has shown that immune cells can also infiltrate into the CSF of syphilis patients [45]. However, this study was limited because of a small number of patients (n = 32), selected patient populations (latent syphilis and neurosyphilis) and lack of characterization of neurosyphilis patients [45]. In the present study, a total of 431 syphilis patients without neurological involvement (including 20 latent syphilis patients in the previous report) and 100 neurosyphilis patients (including 12 patients in the previous report) were included. This larger number of syphilis patients enables further stratification according to stage and symptoms.
Interestingly, we observed an accumulation of CD4+ T cells in the CSF of both asymptomatic and symptomatic neurosyphilis patients, which were consistent with several previous reports showing that CD4+ T cells were the primary responders to T. pallidum in syphilis lesions [8], [9], [46]. CD4+ T cells can be divided into a variety of effector subsets, including classical Th1 cells and Th2 cells, the more recently defined Th17 cells, follicular helper T cells, and regulatory T cells [47]. Though we did not elucidate the precise identity of the CD4+ T cell subset, we observed a decreased frequency of CD4+ CD25high Tregs in the CSF of symptomatic neurosyphilis patients compared with those of non-neurosyphilis and asymptomatic neurosyphilis patients. Given the important role of Tregs in controlling immune-mediated tissue damage, our results suggest that the CNS damage in neurosyphilis patients may be due to an uncontrolled host immune response. A local decrease in Tregs may facilitate CNS injury in neurosyphilis patients. A similar scenario has been observed in other CNS disorders [48], [49].
T. pallidum can establish persistent infection by promoting Treg response in early stage of syphilis. In marked contrast to reduced local Treg response in symptomatic neurosyphilis, we found that Treg numbers in circulation of neurosyphilis patients were even higher than early syphilis patients without neurological involvement. This finding suggests that suppression of the systemic immune response against T. pallidum may favor neurological progression. Consistent with this notion, studies have found that HIV-positive people infected with T. pallidum are more likely to develop neurosyphilis, even during the early stages of infection [5], [50].
The mechanisms underlying Treg differences among syphilis patients are poorly understood. Given that TGF-β was implicated in modulating Treg differentiation and activity [37], [38]; we investigated whether the frequency and functional status of Tregs were associated with this cytokine. We confirmed that serum from the patients with secondary and serofast syphilis did express significantly higher levels of TGF-β than those of healthy control subjects, which may be related to the increased frequency and enhanced function of Tregs in these patients. Lower TGF-β levels were observed in CSF of symptomatic neurosyphilis patients than asymptomatic neurosyphilis patients, which may be associated with a decrease in CSF Treg numbers.
We propose a model to summarize the role of T cell subsets in the pathogenesis of syphilis in Figure 4. T. pallidum penetrates through abraded skin where antigen presenting cells (APC), such as dendritic cells (DC), process the bacteria and then migrate to the subcutaneous lymph nodes. These activated APC [51] may present T. pallidum-derived antigens to naïve T cells and induce production of Th1 [9] and Treg [18], which enter the peripheral blood and circulate widely throughout the body. T. pallidum has the ability to preferentially enhance the generation of Tregs through TGF-β [18], which may impair Th1 function to favor bacterial persistence in the circulation and skin. Antigenic variation and poor antigenicity also enable T. pallidum to evade cell mediated immune response [13], [15], [16]. However, a defective accumulation of Tregs in the CNS (Table 3) may fail to suppress T cell-mediated inflammation and tissue damage in the meninges and parenchyma of brain and spinal cord, resulting in neurological symptoms and signs.
In our study cohort, there are differences in sex distribution among syphilis patients of different stages: 78.0% neurosyphilis patients (78/100) were male, while only 35.1% serofast syphilis patients (34/97) were male. However, there was no significant difference in blood and CSF CD4+ CD25high Tregs between males and females in each group (data not shown), which indicating that the Treg differences between stages were not due to gender preference. Serofast status represents a clinical challenge for treatment of syphilis. There is no universally accepted definition of “serofast”. The definition of “serofast” in this manuscript is “having had an appropriate 4-fold titer decline after treatment, but not reverting to seronegative”. Although these syphilis patients meet criteria for being adequately treated, we and others have shown that such “serofast” patients can progress to neurosyphilis [52], [53], suggesting that they still harbor T. pallidum. The immune status of serofast patients is unclear. A recent study reported that HIV-infected patients are at increased risk for serofast state after treatment [54]. Our results showed that these patients had enhanced circulating Treg numbers and suppressive function, also suggesting serofast status may be associated with a systemic immune suppression.
There are several limitations in the analysis of Treg activity in this study. First, future studies should examine Foxp3 expression and define the functional status of CD4+ CD25high Tregs in CSF in neurosyphilis patients. We were not able to conduct such studies because of the limited availability of CSF T cells. In addition, studies of Treg loss-of-function and gain-of-function are needed to further explore their role in syphilis, but these experiments have been hampered by inherent difficulty in conducting immunologic studies of syphilis in experimental animal models [9].
In conclusion, our findings demonstrate for the first time that neurological progression in syphilis patients is associated with increased circulating Tregs and CSF CD4+ T cells and reduced local Treg response is implicated in the development of symptoms in neurosyphilis patients.
|
10.1371/journal.ppat.1003263 | Strongyloidiasis and Infective Dermatitis Alter Human T Lymphotropic Virus-1 Clonality in vivo | Human T-lymphotropic Virus-1 (HTLV-1) is a retrovirus that persists lifelong by driving clonal proliferation of infected T-cells. HTLV-1 causes a neuroinflammatory disease and adult T-cell leukemia/lymphoma. Strongyloidiasis, a gastrointestinal infection by the helminth Strongyloides stercoralis, and Infective Dermatitis associated with HTLV-1 (IDH), appear to be risk factors for the development of HTLV-1 related diseases. We used high-throughput sequencing to map and quantify the insertion sites of the provirus in order to monitor the clonality of the HTLV-1-infected T-cell population (i.e. the number of distinct clones and abundance of each clone). A newly developed biodiversity estimator called “DivE” was used to estimate the total number of clones in the blood. We found that the major determinant of proviral load in all subjects without leukemia/lymphoma was the total number of HTLV-1-infected clones. Nevertheless, the significantly higher proviral load in patients with strongyloidiasis or IDH was due to an increase in the mean clone abundance, not to an increase in the number of infected clones. These patients appear to be less capable of restricting clone abundance than those with HTLV-1 alone. In patients co-infected with Strongyloides there was an increased degree of oligoclonal expansion and a higher rate of turnover (i.e. appearance and disappearance) of HTLV-1-infected clones. In Strongyloides co-infected patients and those with IDH, proliferation of the most abundant HTLV-1+ T-cell clones is independent of the genomic environment of the provirus, in sharp contrast to patients with HTLV-1 infection alone. This implies that new selection forces are driving oligoclonal proliferation in Strongyloides co-infection and IDH. We conclude that strongyloidiasis and IDH increase the risk of development of HTLV-1-associated diseases by increasing the rate of infection of new clones and the abundance of existing HTLV-1+ clones.
| HTLV-1 is a human retrovirus estimated to infect 20 million people world-wide and is causing in a small proportion of the infected individuals an inflammatory disease or a leukemia/lymphoma. HTLV-1 persists lifelong by driving clonal proliferation of infected T-cells. Strongyloidiasis, a gastrointestinal infection by an helminth (Strongyloides stercoralis) and Infective Dermatitis associated with HTLV-1 (IDH), a skin inflammation with bacterial infection, appear to increase the risk of developing HTLV-1-related diseases. It is well known that the chance of developing HTLV-1-related diseases increases with the number of cells infected by the virus (also called proviral load). It is also known that HTLV-1-infected individuals co-infected by Strongyloides or affected by IDH have a higher proviral load, but the mechanism is still unclear. Consequently, the aim of this study was to test if co-infection increases the total number and/or the abundance (or size) of HTLV-1-infected T-cell clones. We have shown that the significantly increased proviral load in HTLV-1-infected individuals with IDH or strongyloidiasis is due to an increase in the mean clone abundance (bigger clones), not to an increase in the number of infected clones. These patients appear to be less capable of restricting clone abundance than those with HTLV-1 alone.
| Human T-lymphotropic virus type 1 (HTLV-1) causes adult T-cell leukemia-lymphoma (ATLL) and HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP). It has been estimated that 10 to 20 million persons live with HTLV-1 infection worldwide. A small proportion (up to 7%, depending on the area) of HTLV-1-infected individuals develop disease, while the majority remain asymptomatic carriers (ACs). Infection occurs via breastfeeding, transfusion of infected cellular blood products, or sexual intercourse. Symptoms usually appear after a long period (years or decades) of clinical latency [1]. The HTLV-1 proviral load remains stable within each infected person and correlates with the outcome of infection. However, the proviral load varies widely among infected people, even within a particular diagnostic group [2], [3], [4]. Infectious transmission of HTLV-1 across the virological synapse [5] is believed to be important early in infection, whereas mitotic replication is thought to be mainly responsible for maintaining proviral load once a persistent infection has been established and has reached equilibrium with the immune response [6]. We recently showed that the abundance of each established HTLV-1 clones is determined by genomic features of the host DNA flanking the provirus. HTLV-1 clonal expansion in vivo is enhanced by proviral integration in an actively transcribed area of the genome [7].
The helminth Strongyloides stercoralis (St) is estimated to infect 50-100 million individuals, mainly in the tropics and subtropics. Most people with strongyloidiasis have mild diarrhea, vague abdominal complaints, pruritus, perianal rash or simply remain asymptomatic. The Strongyloides stercoralis larvae migrate to a range of sites like the lung, liver, kidney, and central nervous system. The larvae can carry bacteria from the colon and cause fatal sepsis and meningitis. A severe form of the disease named strongyloides hyperinfection syndrome, characterized by abundant disseminated parasites, has been described in patients with corticosteroid therapy, severe malnutrition, transplantation, haematological malignancies (especially lymphoma) and HTLV-1 infection [8]. Epidemiological evidence shows that HTLV-1 is associated with a high frequency of Strongyloides stercoralis infection, a high risk of the strongyloides hyperinfection syndrome, and with relapse after treatment with ivermectin, thiabendazole, or albendazole [9], [10], [11], [12], [13], [14]. Patients with HTLV-1 and Strongyloides stercoralis co-infection had a higher HTLV-1 proviral load and a higher Strongyloides stercoralis burden than patients with either infection alone [13], [15]. The high proviral load measured in Strongyloides stercoralis co-infected patients has been linked with oligoclonal expansion of HTLV-1 infected T-cells [16]. Strongyloides stercoralis co-infection is suspected to be a risk factor for the development of ATLL, but the mechanism is still unclear [17], [18], [19], [20], [21], [22], [23], [24].
Infective dermatitis (IDH) is a severe, chronic, relapsing dermatitis associated with HTLV-1. IDH has been reported in several populations with endemic HTLV-1 infection, including South Africa, Jamaica, Trinidad, Brazil, Colombia, Peru and Japan. Staphylococcus aureus and/or beta-hemolytic Streptococci are commonly cultured from the skin lesions. The average age at disease onset is 2 years. The disease decreases in severity with age and rarely continues until adulthood [25]. IDH patients typically have a high HTLV-1 proviral load, comparable to HAM/TSP patients [26]. IDH is suspected to increase the risk of HAM/TSP or ATLL development [26], [27], [28], [29], but the evidence is not conclusive.
The aim of this study was to identify and quantify the impact of co-infection on HTLV-1 clonality. Because HTLV-1 proviral load and oligoclonality are closely correlated with disease status, we aimed to test the hypothesis that each co-infection increases the risk of HTLV-1-associated diseases by increasing the number or the abundance of HTLV-1-infected T-cell clones. We used a newly developed method to map and quantify thousands of HTLV-1 proviral insertion sites. We demonstrate that co-infections significantly alter the HTLV-1 clonality. Patients with strongyloidiasis or IDH had a higher proviral load and a higher average clone abundance than did asymptomatic HTLV-1 carriers even though the major determinant of proviral load was still the number of clones. The degree of oligoclonality of HTLV-1 was higher, and less stable over time, in patients with strongyloidiasis than in patients with neither co-infection nor ATLL, and there was a higher rate of turnover (i.e. appearance or disappearance) of HTLV-1-infected clones in the co-infected patients. Finally, we show that co-infections drive the proliferation of HTLV-1+ T-cell clones regardless of the genomic environment of the provirus, in contrast to infection with HTLV-1 alone, in which selective clonal expansion is favored by specific features of the proviral integration site in that clone.
The proviral load in patients with IDH and in patients co-infected with Strongyloides was significantly higher than the proviral load in asymptomatic carriers (Figure 1A, median proviral load = 0.3*105 proviral copies per 106 PBMCs for Asymptomatic Carriers, 1.1*105 for patients with Infective Dermatitis associated with HTLV-1 and 1.0*105 for patients co-infected with Strongyloides, Mann Whitney, p = 0.0001 and p = 0.002 respectively for IDH and Strongyloides co-infected patients vs Asymptomatic Carriers). These observations are in accordance with data from previous reports [26], [28], [30]. We estimated the total number of clones present in the blood using a newly developed method (DL, BA, CRMB, submitted). We found that the total number of clones in the blood of IDH and Strongyloides co-infected patients was comparable to those of asymptomatic carriers (Figure 1B, Mann Whitney, respectively p = 0.27 and p = 0.81). We also observed that HAM/TSP patients had a higher number of clones in the blood than asymptomatic carriers (Figure 1B, Mann Whitney, p = 0.05). The average clone abundance (expressed as the number of cells in a given clone per 106 PBMCs) was higher in patients with IDH and Strongyloides co-infection than in asymptomatic carriers, but no significant difference was observed between asymptomatic carriers and HAM/TSP patients (Figure 1C). The mean clone abundance, expressed as number of cells per 106 PBMCs, was 1.4 in Asymptomatic Carriers, 1.2 (HAM/TSP), 3.6 (IDH patients) and 3.0 (Strongyloides co-infected patients) (Figure 1C, Mann Whitney, respectively for HAM/TSP, IDH and Strongyloides vs Asymptomatic Carriers, p = 0.97, p = 0.02 and p = 0.02). The distribution of clone abundance is depicted in Figure S1A in Text S1. Each slice in the pie charts represents a single clone; the size of the slice is proportional to the relative abundance of that clone. The 3 most abundant clones are represented by the colored slices. The clonality of HTLV-1 in the blood of the representative patient with IDH was relatively uniform (the slices in the pie charts are of similar size) and accordingly the oligoclonality index was low in this representative subject (Figure S1A in Text S1). The HTLV-1 clone distributions in two different patients co-infected with Strongyloides stercoralis are shown to illustrate the wide variation in clonality observed in this group of patients despite similar proviral load. The oligoclonality index of the first patient was low and in the range observed in Asymptomatic Carriers and patients with HAM/TSP. The oligoclonality index of the second patient with Strongyloides was greater because of the presence of 3 relatively abundant clones (Figure S1A in Text S1, red slices made up nearly half of the proviral load) and lay in the range of patients with ATLL (Figure 1E, red dotted circle). The oligoclonality index in patients with Strongyloides co-infection was greater than the oligoclonality index in Asymptomatic Carriers (Figure 1D, Mann Whitney, p = 0.01) confirming the previous observation of oligoclonal expansion of HTLV-1 infected T-cells [16]. There was no significant difference in oligoclonality index between Asymptomatic Carriers and IDH patients (Figure 1D, Mann Whitney, p = 0.10). Two patients out of the 14 co-infected with Strongyloides had a very high oligoclonality index (0.76 and 0.78 respectively) due to oligoclonal expansion of infected clones (Figure 1E, dotted red circle). Their proviral load and clonal distribution were in the range of patients with ATLL. In patients without malignant disease, proviral load did not correlate with oligoclonality index (Figure 1E). However, proviral load was positively correlated with the total number of clones in each cohort (Figure 1F) and by contrast no correlation was observed between the number of clones and oligoclonality index in any cohort (Figure 1G). The mean oligoclonality index in patients with Strongyloides did not vary significantly after anti-helminth treatment (Figure S1B and Table S2 in Text S1), although there was a large decrease in oligoclonality index after Strongyloides clearance in the patient with the most oligoclonal distribution in the cohort (Figure S1B in Text S1, patient St6).
These data show that, in individuals without ATLL, strongyloidiasis or IDH, the proviral load of HTLV-1, which correlates with the risk of inflammatory and malignant diseases, is determined mainly by the number of infected T-cell clones and not, as previously believed, by the amount of oligoclonal proliferation. The significantly increased proviral load in HTLV-1-infected individuals with IDH or co-infected with Strongyloides is due to an increase in the mean clone abundance, not to a further increase in the number of infected clones. Moreover, the degree of oligoclonal expansion in patients with Strongyloides was significantly higher than that in asymptomatic carriers, whereas no significant difference in oligoclonality index was observed between IDH patients and asymptomatic carriers.
Figure 2A shows the evolution of proviral load with time in Strongyloides co-infected patients. Figure 2B shows the evolution of oligoclonality index with time in Strongyloides co-infected patients. The data show that the oligoclonality index in Strongyloides co-infected patients was more variable over time than in patients with HTLV-1 infection alone as described previously [7]. This conclusion was confirmed by the data shown in Figure 2C, which depicts the absolute variation in oligoclonality index per year. The oligoclonality index varied on average by 0.01 per year in patients with HTLV-1 only and by 0.04 per year in Strongyloides co-infected patients (Mann Whitney, p<0.0001). The similarity between the populations of HTLV-1-infected T-cell clones at two consecutive time points is shown in Figure 2D and 2E. To compare these two populations we used the SØrensen similarity indices, which were developed to assess the similarity between two ecosystems in term of species shared. A clone is here the equivalent of a species. The indices range from 0 to 1, with 0 indicating that no clones are shared between the two time points and 1 indicating complete identity (see Materials and Methods). Figure S2 in Text S1 shows the incidence- and abundance-based similarity indices from biological replicates (i.e. clonality analyses made in triplicate from the same blood sample of patients with HTLV-1 only with non-malignant infection). The results show that HTLV-1 clone populations in two consecutive blood samples in Strongyloides co-infected patients differed more in identity and abundance than the clone populations from two consecutive blood samples in patients with HTLV-1 only (Figure 2D and 2E, Mann Whitney; p = 0.007 and p = 0.007 respectively for incidence- and abundance-based similarity index).
We conclude that HTLV-1 clonality was less stable in Strongyloides co-infected patients, in whom there was a higher rate of turnover of clones. This observation raised the question: does co-infection with Strongyloides alter the selection forces that favor selective expansion of HTLV-1+ clones?
In patients with HTLV-1 alone, we previously reported [7] a positive correlation between clone abundance and proximity to CpG islands and host genes, a positive correlation between clone abundance and markers of active transcription, and a negative correlation with repressive epigenetic marks. We concluded from these data [7] that transcriptional activity of the flanking host genome favors selective expansion of the HTLV-1+ T-cell clone. In contrast, in patients with Strongyloides co-infection or IDH, we did not observe these trends linking clone abundance and a particular genomic environment of the proviral integration site (Figure 3, see black arrows). Specifically, the most abundant clones (clone having more than 103 cells per 106 PBMCs) had a genomic environment of the provirus similar to the random distribution; i.e. the environment of the provirus does not determine the abundance of the major clones in co-infected patients.
We conclude that the abundance of the largest clones in patients with Strongyloides co-infection or IDH is independent of the genomic environment of the proviral insertion site.
We quantified HTLV-1 clonality in two types of sample in which the infected T-cells present are believed to play a direct role in the pathogenesis of the respective inflammatory disease: CSF from patients with HAM/TSP and skin lesions from patients with IDH. The rationale was two-fold: first, over-representation of a few infected clones in these tissues may indicate immune activation of antigen-specific T-cells. Second, we wished to test whether the selectively expanded clones in these tissues are also abundant in the bloodstream.
Figure 4A illustrates the overlap between the HTLV-1 infected cell populations from blood and from the skin lesion. The pink slices denote clones present in both blood and skin lesion, the grey and black slices denote clones detected only in blood or skin. The results show that a high proportion of the observed clones were present in both skin and blood samples. The infected cell population in the skin did not differ from the blood populations in the identity of the clones (incidence-based similarity index, paired t-test, p = 0.646), but differed by the relative abundance of the common clones (abundance-based similarity index, paired t-test, p = 0.003). The mean oligoclonality index in the infected cell population in the skin lesion was significantly lower than the oligoclonality index in the corresponding blood sample (paired t-test, p = 0.015). That is, all HTLV-1-infected T-cell clones present in the skin lesion had approximately the same relative abundance, and there was no evidence of selective expansion of a specific subset of infected cells.
Figure 4B illustrates the small overlap observed between blood and CSF HTLV-1 infected cell populations. As above the yellow slices show the clones found in both blood and CSF. The largest clone in the CSF contained a provirus inserted in chromosome 11 (coordinate 41125786): this clone was also detected in blood but at a much lower relative abundance. The infected clone population in the CSF differed in both identity and abundance from that in blood (incidence and abundance-based similarity index, paired t-test, respectively p = 0.090 and p = 0.014). In other words, the clone population in the CSF did not appear to be a random subsample of the blood population but diverged significantly, with the presence of other clones and a significant variation in the relative abundance of common clones. We conclude that there was selective migration or proliferation of infected T-cells in the CSF.
The data show that HTLV-1 clonality is extensively altered in vivo by Strongyloides co-infection and IDH. By comparison with patients infected with HTLV-1 only, Strongyloidiasis stercoralis co-infected patients showed five main differences: i. a higher mean proviral load, ii. a higher mean clone abundance, iii. a higher mean oligoclonality index, iv. a higher clone turnover rate (i.e. a higher rate of appearance and disappearance of clones), and v. a proliferation of the largest clones independent of the genomic environment of the provirus. IDH patients had also a higher mean proviral load, a higher mean clone abundance and a change in the selection forces that favor expansion of the largest HTLV-1 clones. The median oligoclonality index in the IDH group was higher than that in Strongyloides co-infected patients (respectively 0.464 vs 0.460), but the difference from asymptomatic carriers did not reach statistical significance. We emphasize that the mean age of the IDH cohort (14 years) was the lowest in the study (because the disease typically manifests during childhood); the mean age of Strongyloides co-infected patients was 44 years and that of asymptomatic carriers was 55 years. It would be interesting to follow the oligoclonality index over time in these IDH patients. The cohorts of co-infected patients in the present study also differed from the non co-infected individuals by geographical origin and ethnicity. It is possible that variation in host or viral genotype also influences HTLV-1 clonality. However, HTLV-1 genetic variation is limited, and the Cosmopolitan subtype 1a (Transcontinental subgroup) dominates in the Caribbean, South Africa, Peru and Brazil [31]). Moreover, the observed effects of host genotype account for only 5–10% of the observed variation of proviral load between individuals [32], [33], that is at least an order of magnitude less than the observed variation in proviral load within a given disease group.
The increased divergence between successive time-points in the HTLV-1 clonal composition observed in Strongyloides co-infected patients suggests a higher rate of persistent infectious spread of the virus, increasing the total number of clones generated during the HTLV-1 lifelong infection (as distinct from the total number of clones measured at a given time point). To test this hypothesis, further work is needed to quantify the rate of infectious spread within a given patient.
The observation that a major determinant of proviral load (in patients without ATLL) was the number of clones has important implications for the understanding of the development of HTLV-1-associated diseases. Why do some patients carry more clones than others? It is likely that the efficiency of the host's immune response to HTLV-1, especially the quality of the HTLV-1-specific CTLs [34], plays a major part in determining the total number of HTLV-1-infected clones. The initial dose of infection may also be a significant determinant. We also found that both IDH and Strongyloides co-infected patients had on average more abundant clones (i.e. the mean number of cells per clone was higher). These patients seem to be less capable of restricting clone abundance than those with HTLV-1 alone. Consistent with this conclusion, we observed a higher oligoclonality index in Strongyloides co-infected patients. Our observation of oligoclonal expansion in Strongyloides co-infected patients confirms a previous report by Gabet et al 2000 [16] but contrasts with the conclusion of polyclonal expansion made by Satoh et al 2002 [30]. The apparent discrepancy may be due to the lower sensitivity of the inverse long PCR technique used by Satoh et al [30] to quantify clonal distribution. Gabet et al also reported [35] oligoclonal expansion of HTLV-1+ T-cells in a patient with both IDH and Strongyloides co-infection, and the authors concluded that IDH might be a co-factor for ATLL. From our observations, it appears that Strongyloides co-infection could be a co-factor for oligoclonal expansion. However, in the samples from the IDH patients in this cohort we did not observe any case with a degree of oligoclonal proliferation (oligoclonality index value) within the ATLL range; and the oligoclonality index of HTLV-1-infected cells in the skin in IDH was significantly lower than that in the blood (see below).
Based on observations in a mouse model, it has been shown that immune activation of HTLV-1-infected CD4+ T-cells induces HTLV-1 Tax expression, T-cell proliferation, and may culminate in the development of ATLL [36]. Ratner et al [37] reported a case of a patient with HTLV-1-associated chronic ATLL and Strongyloides infection, in whom active HTLV-1 transcription resolved with anti-helminthic therapy [37]. This observation supports the idea that Strongyloides co-infection can induce HTLV-1 transcription via T-cell immune activation. Nevertheless, little is known about the nature of the hypothetical activating signals induced by the co-infection. We considered two possible explanations, which are not mutually exclusive: i. co-infection induces immune activation of the few infected clones that are specific to the co-infecting pathogen (such as Strongyloides stercoralis, Staphylococcus or Streptococcus); ii. co-infection favours expansion of all HTLV-1 infected clones (either by inducing non specific immune activation of the HTLV-1 infected clones or by impairing immune surveillance against the HTLV-1 infected cells). To answer this question we compared the HTLV-1 infected T-cells present in the skin lesion from IDH patients with the infected cells from the corresponding blood sample. As a reference point, we also compared the HTLV-1 infected T-cells present in the CSF in HAM/TSP patients with the infected cells from the corresponding blood sample. We observed that the major clone in the CSF can be rare or undetectable in the bloodstream (Figure S4 in Text S1). This supports the idea that certain infected clones present in the CSF can expand in the central nervous system (CNS), perhaps through antigenic stimulation; the relatively inefficient immune surveillance (the ‘immune privilege’) in the CNS may also allow unrestricted clonal expansion. By contrast, the presence of numerous HTLV-1 clones with an approximately equal abundance in the skin lesions of IDH patients suggests that the dermatitis does not involve the selective proliferation of HTLV-1 infected T-cells specific to Staphylococcus (or Streptococcus) antigens, but rather the non-specific proliferation of the entire population of infiltrating T-cells. HTLV-1 infection in endemic regions frequently occurs during breast-feeding and so predates infection with Strongyloides or Streptococcus/Staphylococcus. Consequently, it is unlikely that HTLV-1 infection is biased towards T-cells specific to Strongyloides or Streptococcus. It remains possible that chronic antigen stimulation favours the expansion of such T-cells in co-infected subjects. However, since the degree of oligoclonality of infected T-cells observed in patients with infective dermatitis was lower in the skin lesion than in blood, we infer that antigen specificity was a minor contributor to selective clonal expansion in these co-infected individuals. It remains possible that stimulation by Strongyloides antigens contributes to clonal expansion in individuals with Strongyloides co-infection. Our observation that the abundance of the largest clones in patients with Strongyloides co-infection or IDH is independent of the proviral insertion site environment supports the idea that Strongyloides co-infection and IDH change the selection forces that favour expansion of HTLV-1-infected clones.
HTLV-1 infection causes activation and proliferation of the infected T-cells. The HTLV-1 Tax protein activates the canonical NF-κB pathway (review by Qu et al [38]), upregulates expression of the interleukin-2 receptor alpha (IL-2Rα) [39] and promotes cell proliferation [40]. Additionally, the frequency of CD4+FoxP3+ regulatory T-cells is abnormally high in HTLV-1 patients [41] and the rate of CTL-mediated lysis of autologous HTLV-1-infected cells is negatively correlated with the frequency of CD4+FoxP3+ T-cells [41]. The frequency of CD4+FoxP3+ regulatory T-cells is increased further in Strongyloides stercoralis co-infected patients [15]. These observations suggest the existence of a vicious circle in which each pathogen favors the other. We suggest that two principal factors contribute to the increased clonal proliferation of HTLV-1+ T-cells observed in Strongyloides co-infection. First, an autocrine IL2/IL-2R loop was reported by Satoh et al [30] in patients with this co-infection. Second, the high frequency of CD4+FoxP3+ T-cells may impair the host T-cell response to HTLV-1 infection.
In summary, co-infection with Strongyloides is associated with an increase in the rate of formation of new HTLV-1-infected T-cell clones, oligoclonal proliferation of certain HTLV-1+ clones, and a higher mean clone abundance. IDH, similarly, is associated with an increase in the mean abundance of HTLV-1+ T-cells in the circulation and a change of the selection forces that favour expansion of HTLV-1+ clones. We propose that repeated activation of a large number of HTLV-1-infected T-cell clones causes abundant proviral expression, resulting in both infectious spread (infection of new T-cell clones) and mitotic spread (proliferation of existing infected T-cell clones) thereby increasing the risk of both inflammatory disease and malignant transformation.
We studied 61 individuals (75% of Afrocaribbean ethnicity, 18% of Caucasian and 7% of Asian ethnicity) infected with HTLV-1 alone (14 asymptomatic HTLV-1 carriers, mean age 55 years; 1 patient with uveitis; 26 patients with HAM/TSP, mean age 62 years; 20 patients with ATLL, mean age 53 years). All individuals attended the clinic at the National Centre for Human Retrovirology (Imperial College Healthcare NHS Trust, St. Mary's Hospital, London, UK), and donated blood samples. Four HAM/TSP patients (2 from the UK, 2 from Brazil) also donated samples of cerebrospinal fluid (CSF). Fourteen individuals (11 from Peru, 3 Caribbean, mean age 44 years) infected with HTLV-1 and affected by strongyloidiasis donated blood samples. These patients had microbiology confirmed stool positive samples for Strongyloides stercoralis and confirmed negative post anti-helminth treatment. Seventeen individuals with IDH (10 from South Africa and 7 from Brazil, mean age 14 years) donated blood and skin lesion samples. These patients had active disease at time of blood sampling and biopsy. Table S1 in Text S1 details the microbiological isolates from a skin swab of each patient. PBMCs were isolated using Histopaque-1077 (Sigma-Aldrich). Cells were washed and cryopreserved in fetal calf serum (Invitrogen) with 10% DMSO (Sigma-Aldrich). Skin lesion samples were frozen directly after sampling in liquid nitrogen. DNA was extracted from PBMCs, skin tissue or CSF using DNeasy Blood and Tissue kit (Qiagen).
All subjects gave fully informed, written consent and all clinical investigations have been conducted according to the principles expressed in the Declaration of Helsinki. This study was approved by the UK National Research Ethics Service (NRES reference 09/H0606/106).
DNA was amplified for HTLV-1 DNA (using the Tax sequence-specific primers SK43 and SK44) and for β-actin (as a measure of genomic DNA using ActFw and ActRev primers) (SK43:5′CGGATACCCAGTCTACGTGT, SK44:5′GAGCCGATAACGCGTCCATCG, ActFw:5′TCACCCACACTGTGCCCATCTATGA, ActRev:5′CATCGGAACCGCTCATTGCCGATAG). Three dilutions of DNA were amplified by real time quantitative PCR in a Roche light cycler using SYBR Green 1 Dye incorporation (Roche Applied Science). Standard curves were generated using the rat cell line TARL2 which contains 1 copy per cell of the HTLV-1 provirus [42]. The sample copy number was estimated by interpolation from the standard curve, calculated as an average of the 3 dilutions, and expressed as the number of copies per 106 PBMCs.
We used a newly developed protocol to map and quantify thousands of HTLV-1 proviral insertion sites, as previously described [7], [43]. We define an HTLV-1 clone as a population of cells that carry an integrated HTLV-1 provirus in a particular insertion site in the host genome. We have demonstrated that there is a single proviral copy per cell in non-transformed cells naturally infected with HTLV-1 [44], and leukemic clones typically carry one (complete or defective) provirus per cell [45] [46]. DNA was extracted from uncultured PBMCs, skin lesion or CSF of HTLV-1-infected individuals and sheared by sonication. A partially double-stranded DNA linker containing a 6 nt index tag was ligated to the sheared DNA and nested PCR was performed between the HTLV-1 LTR and the linker. Nested PCR products were pooled to construct the library of DNA for high-throughput sequencing. Fifty-nucleotide paired-end reads (read 1 and read 2) and a 6 nucleotide index tag read were acquired on an Illumina Genome Analyzer II or an Illumina HiSeq. Read 1 and read 2 were mapped against the human genome (build hg18) and the proviral insertion site and the shear site were deduced. For each unique insertion site, we counted the number of amplicons of different length (i.e. different shear sites) to enumerate the number of sister cells of that infected T-cell clone. The absolute abundance of a given clone i (number of cells per 106 PBMCs) was calculated from the number of sister cells and the measurement of the proviral load as follow:where Xi is the number of sister cells of the ith clone, D the number of observed clones and PVL the proviral load.
The relative abundance of a given clone i (in percent of the proviral load) was expressed as follow:
To measure the clonality of the infected cell population, i.e. the non-uniformity of the clone abundance distribution, we used the oligoclonality index [7], based on the Gini coefficient [47].
Oligoclonality index = 1 indicates perfect monoclonality (only one clone constitutes the total proviral load), while oligoclonality index = 0 indicates perfect polyclonality (all clones have the same abundance).
Quantitative measures of similarity (or overlap) between two populations play an important role in statistical ecology. The first similarity indices developed were based on the presence or absence of species between the two populations. The widely used SØrensen incidence-based similarity index ranges from 0 to 1, with 0 indicating that no clones are shared between the two populations and 1 indicating complete identity (all the clones present in population 1 were also present in population 2 and vice versa). The former index was subsequently improved to take into consideration the abundance of each clone and named SØrensen abundance-based similarity index. Because this index takes clone abundance into account, populations that contain the same clones but have different clone abundance will have an index value of less than 1. Details of the calculations are given in supplemental data.
In Figure 2, the similarity index was calculated in each case by comparing two samples of equal numbers of sister cells, to preclude a bias toward the clone distribution in one sample. When samples from 3 time points for a given patient were analysed, we calculated the similarity index twice, between time point 1 and time point 2 and between time point 2 and time point 3 respectively. In Figure 4, to compare blood and skin clone populations or blood and CSF clone populations from the same patient, we first created subsamples of clones from the blood clonality analysis by randomly drawing 10% of the sister cells detected in the blood. We calculated the similarity index by comparing 2 different subsamples from the blood population (BLD vs BLD values). We then created a subsample of clones from the blood by randomly drawing the same number of sister cells detected in the corresponding skin lesion (or CSF) and calculated the similarity index between the blood subsample and the skin (or CSF) (BLD vs SKN or BLD vs CSF respectively).
We used the Integration Site Pipeline and Database (INSIPID) from the Bushman laboratory (Department of Microbiology, University of Pennsylvania School of Medicine, Pennsylvania, Philadelphia, United States of America) (http://microb215.med.upenn.edu/insipid/). This web-based tool houses sequences of newly inserted elements in vertebrate genomes, together with specific genomic annotations, to facilitate analysis of the environment of the genomic insertion site: see Figure S3 in Text S1.
The diversity estimation approach (DivE) (Daniel Laydon, Charles Bangham, Becca Asquith, submitted) used to estimate the total number of clones (observed and unobserved) involves fitting many mathematical models to species-accumulation data, and to successively smaller nested subsamples thereof. Novel criteria are used to score models in how consistently they can reproduce existing observations from incomplete data. The estimates from the best performing models are aggregated (using the geometric mean) to estimate the number of clones in the circulation. We have shown that, when applied to HTLV-1 infection and other microbiological populations, DivE significantly outperforms several classical ecological estimators of unseen species (namely the Chao, Bootstrap and Good-Turing estimators, and species-accumulation curves). Let be the total number of clones observed, and let be an estimate of from a subsample of data. We define the accuracy of a given estimator as the percentage error between and (i.e. ). When applied to HTLV-1 infection, the mean accuracy of DivE was 3.5%, compared to accuracies of 61.5%, 35.3%, 35.0%, and 29.1% for the Chao, Bootstrap and Good-Turing estimators, and the species-accumulation curves respectively (using two-tailed paired Mann-Whitney tests, p<0.0001 for all comparisons with DivE). DivE was optimized for clonal distribution of patients with non-malignant HTLV-1 infection and further work will be necessary to estimate with the same confidence the total number of clones in ATLL patients. Therefore, within this paper we do not estimate the total number of clones in patients with ATLL.
The average clone abundance was calculated from the proviral load divided by the estimated total number of clones in the blood and expressed as the number of cells per 106 PBMCs.where PVL is the proviral load and S the estimated total number of clones.
Statistical tests were performed using GraphPad Prism and R softwares and were two tailed when possible. The symbol *** was used when p<0.001, ** when p<0.01, * when p<0.05, NS (Non Significant) when p>0.05.
CXCR4, NG_011587.1; CCR5, NG_012637.1; IL-2Rα, NG_007403.1; IFN-γ, NG_015840.1; IL-10, NG_012088.1; TGF-β, NG_013364.1; HTLV-1 Tax, NC_001436.1.
|
10.1371/journal.pntd.0000361 | Prevalence Distribution and Risk Factors for Schistosoma hematobium Infection among School Children in Blantyre, Malawi | Schistosomiasis is a public health problem in Malawi but estimates of its prevalence vary widely. There is need for updated information on the extent of disease burden, communities at risk and factors associated with infection at the district and sub-district level to facilitate effective prioritization and monitoring while ensuring ownership and sustainability of prevention and control programs at the local level.
We conducted a cross-sectional study between May and July 2006 among pupils in Blantyre district from a stratified random sample of 23 primary schools. Information on socio-demographic factors, schistosomiasis symptoms and other risk factors was obtained using questionnaires. Urine samples were examined for Schistosoma hematobium ova using filtration method. Bivariate and multiple logistic regressions with robust estimates were used to assess risk factors for S. hematobium. One thousand one hundred and fifty (1,150) pupils were enrolled with a mean age of 10.5 years and 51.5% of them were boys. One thousand one hundred and thirty-nine (1,139) pupils submitted urine and S. hematobium ova were detected in 10.4% (95%CI 5.43–15.41%). Male gender (OR 1.81; 95% CI 1.06–3.07), child's knowledge of an existing open water source (includes river, dam, springs, lake, etc.) in the area (OR 1.90; 95% CI 1.14–3.46), history of urinary schistosomiasis in the past month (OR 3.65; 95% CI 2.22–6.00), distance of less than 1 km from school to the nearest open water source (OR 5.39; 95% CI 1.67–17.42) and age 8–10 years (OR 4.55; 95% CI 1.53–13.50) compared to those 14 years or older were associated with infection. Using urine microscopy as a gold standard, the sensitivity and specificity of self-reported hematuria was 68.3% and 73.6%, respectively. However, the positive predictive value was low at 23.9% and was associated with age.
The study provides an important update on the status of infection in this part of sub-Saharan Africa and exemplifies the success of deliberate national efforts to advance active participation in schistosomiasis prevention and control activities at the sub-national or sub-district levels. In this population, children who attend schools close to open water sources are at an increased risk of infection and self-reported hematuria may still be useful in older children in this region.
| Schistosoma hematobium infection is a parasitic infection endemic in Malawi. Schistosomiasis usually shows a focal distribution of infection and it is important to identify communities at high risk of infection and assess effectiveness of control programs. We conducted a survey in one district in Malawi to determine prevalence and factors associated with S. hematobium infection among primary school pupils. Using a questionnaire, information on history of passing bloody urine and known risk factors associated with infection was collected. Urine samples were collected and examined for S. hematobium eggs. One thousand one hundred and fifty (1,150) pupils were interviewed, and out of 1,139 pupils who submitted urine samples, 10.4% were infected. Our data showed that male gender, child's knowledge of an existing open water source (includes river, dam, springs, lake, etc.) in the area, history of urinary schistosomiasis in the past month, distance of less than 1 km from school to nearest open water source and age 8–10 years compared to those 14 years and older were independently associated with infection. These findings suggest that children attending schools in close proximity to open water sources are at increased risk of infection.
| Schistosomiasis remains an important public health problem globally with an estimated 200 million cases reported each year [1]. However, 85% of the cases reported annually occur in sub-Saharan Africa and over 150,000 deaths are attributable to chronic infection with S. haematobium in this region [2],[3]. The eggs of S. haematobium provoke granulomatous inflammation, ulceration, and pseudo-polyposis of the vesical and ureteral walls. Hematuria is a very common sign of infection but other signs include dysuria, pollakisuria, and proteinuria. Kidney failure deaths due to urinary tract scarring, deformity of ureters and the bladder caused by S. haematobium infection have become less common due to modern drugs [4],[5]. Subtle and indirect morbidities such as fatigue, physical or cognitive impairment and effects of co-infections with other infectious diseases like HIV, malaria have received more attention recently [4]. New evidence from a recent review of these studies suggests a causative link between schistosome infection, anti-parasite inflammation, and risk for anaemia, growth stunting and under-nutrition, as well as exacerbation of co-infections and impairment of cognitive development and physiological capacities among infected individuals [6]. The causal relationship between anaemia and schistosomiasis exists even after controlling for other co-infections and dietary factors among pregnant women and children [6]–[9]. The underlying mechanisms proposed range from social determinants to complex immune interactions.
In Malawi, schistosomiasis is endemic with S. haematobium being highly prevalent in the southern region while S. mansoni predominates on the central plain and the northern regions [10]. The national schistosomiasis control program estimates that between 40% and 50% of the total Malawian population is infected with schistosomiasis (National Plan of Action for the Control of Schistosomiasis and Soil transmitted Helminthes five-year plan 2004–2008). Other studies have suggested that these national estimates may have been derived from studies conducted years ago that had some selection bias for high risk schools [11]. A national survey conducted in 2002 among primary school pupils found the prevalence of S. hematobium and S. mansoni infection among school children to be 6.9% and 0.4% using filtration method of urine and Kato-Katz method for stools respectively [11]. These findings were much lower than expected and it was concluded that schistosomiasis is highly localized in Malawi. This implies that local estimates would be more useful than national estimates in guiding the selection of control strategies to be implemented at district or sub-district level which depends on disease prevalence rate in a community [12]. However, most districts in Malawi do not have local estimates to guide planning and implementation of control interventions at that level. The Ministry of Health through the national schistosomiasis control program is currently undertaking a deliberate effort to have schistosomiasis prevention and control efforts integrated within district plans to encourage ownership and improve sustainability by engaging district teams (National Plan of Action for the Control of Schistosomiasis and Soil transmitted Helminthes five-year plan 2004–2008).
Blantyre district health team initiated and conducted this study aimed at determining the prevalence of schistosomiasis as part of this effort. The findings are expected to be used as baseline for future evaluation and monitoring of control activities. A representative cross-sectional study of urinary schistosomiasis infection among school children in Blantyre district was therefore conducted. To establish the need for intervention in a community, we conducted our baseline survey among children sampled from grade three in elementary schools as it has been shown that this population is intensely affected in endemic communities [13]. The prevalence distribution, factors associated with S. hematobium infection and the reliability of self reported hematuria compared to the “gold standard” parasitological examination among school children are described.
This cross-sectional study was carried out in Blantyre district located in southern Malawi. The district has a land area of 2,012 km2 and the altitude above sea level ranges from 300 m to 1000 m. According to the national statistical office, it has an estimated population of one million of whom 61% reside in the urban or peri-urban areas.
From north-west to south-west boundary of the district runs the Shire River which is the main outlet of Lake Malawi (Figure 1). The average annual temperature is 27°C whilst the average rainfall is 871 mm. The rainy season is from November to April and the dry season from May to October [14].
All standard three pupils who were attending primary schools in Blantyre district during the time of data collection between May and July 2006 were eligible to participate in our study. Our study sample was selected using a stratified 2-stage probability sampling technique. First, we stratified the district was into 3 ecological risk areas for urinary schistosomiasis which was based on altitude above sea level as follows; high (<500 m), moderate (500–1000 m) and low (>1000 m). Some previous studies have shown that altitude above sea level is known to be associated with schistosomiasis [15],[16]. We used STATA software v10.1 (StataCorp Ltd, Texas, USA) to estimate sample size for comparison of proportions between strata. The sample size of 1,128 was estimated using a prevalence estimate of 40.0% with 80.0% power to detect a relative difference of 10.0% between strata at a 0.05 level of significance. Second we obtained from the district education offices the number of schools and pupils' population estimates in each stratum. Based on the pupils' population distribution within the strata and our prior decision to randomly select 50 standard three pupils from each selected school, we calculated that 23 schools would be required. Studies have shown that Lot-Quality Assurance Scheme approaches provide the ability to identify communities with a high prevalence of schistosomiasis with high levels of sensitivity and specificity, even at very small maximum sample sizes [17]. Finer classification of schools according to categories of prevalence are achieved with moderate sample sizes of ≥15 and have been found to be more accurate with extremely low probabilities of making gross classification errors. Thirdly, sampling was conducted in 2 stages. In the first stage, the primary sampling units were schools that were selected with a probability proportional to number of schools in the strata. A list of schools in each stratum was compiled and using computer generated random numbers schools in each stratum were selected as follows; six schools from the high risk stratum, eight from the moderate risk stratum and nine from the low risk stratum. In the second stage, a random sample of 50 grade three pupils in each selected school was obtained.
Children were interviewed by trained Health Surveillance Assistants (HSAs) and community health nurses using a questionnaire that was adapted from the 2002 national prevalence survey [11]. To reduce bias and improve the performance of the questionnaires, questions about schistosomiasis were disguised among other health related questions. The questionnaire was pre-tested and modifications were made after discussions with HSAs, teachers and district health office staff. Risk factors included main household water source, child's knowledge of nearby open water sources (open water source defined as any open water body including lakes, springs, rivers, streams, ponds, swamps and dams), frequency of contact with open water sources, dysuria or passing blood in urine within the past month, history of S. hematobium infection and S. hematobium treatment. Other factors included urban or rural location, proximity of school to an open water source and household socio-economic status (SES).
We used a method developed by the Centre for Social Research in Malawi (Kadzandira et al. unpublished report) and similar methods have been used in other studies to estimate SES [18]. The method combined six variables to formulate the complex indicator of SES which includes housing structures, main occupation of heads of households, and possession of selected assets such as radios, telephones, televisions, etc. For each variable, households were assigned a weight ranging from zero to two. For example, households with a grass-thatched roof were given a weight of zero while households with a plastic paper roof were given a 0.3 and households with iron-sheet or tiled roof were given a 1.0 (Table S1). The sum of weights from the six variables determined the SES score of each household. Households were then classified as low SES for those with scores less than 4.0, moderate SES for 4.0–6.0 and high SES for more than 6.0. There were no significant differences when we analyzed our data using the year 2000 World Bank asset scoring system for wealth (data not shown) [19].
In addition to the questionnaire interview, pupils were asked to submit urine in supplied 20 ml screw top plastic containers between 0900 hrs and 1400 hrs. The samples were immediately transported to the University of Malawi College of Medicine microbiology laboratory. Our primary outcome was S. hematobium infection as diagnosed by urinary microscopy examination.
Samples were tested for micro-hematuria using urine reagent strips (Uripath, Plasmatec Laboratory, UK) and results were scored as negative, +, ++ or +++ as per manufacturer recommendations. 10 mls of urine was filtered using paper filters (Gelman Sciences, Michigan USA) and the egg count was recorded per 10 mls of urine. In close collaboration with the district health office, all children who were found to be positive either through microscopic examination or urine reagent strips were referred for treatment.
Data were entered into Microsoft Access 2003. The data were then converted into SAS version 9.1 (SAS Institute, Cary, North Carolina, USA) which was used for all analyses. Based on the study design of the survey; weighting, stratification, and clustering were taken into account in all statistical analyses using Survey procedures in SAS. The survey analysis procedures use the Taylor series expansion method to estimate sampling errors of estimators based on sample designs [20]. The procedures estimate the variance from the variation among the schools and pool stratum variance estimates to compute the overall variance estimate. A weighting factor was used in the analysis to reflect the likelihood of sampling each student in a stratum. The weight used for estimation is given by the following formula:Where W1 = the inverse of the probability of selecting the school in a stratum and, W2 = the inverse of the probability of selecting a child from the classroom within the school. Bivariate analyses were used to estimate the crude odds ratios and identify variables to be included in the initial multivariate logistic regression model. Chi-square test was used to assess the association between categorical variables and S. hematobium infection in bivariate analysis. Variables that had a p value <0.20 were included in the initial multivariate logistic regression model. Using backward elimination method, variables that showed independent association with S. hematobium infection at a significance level of p value <0.05 were retained in the model. A significance level of p value <0.05 was used in all other analyses. Pearson correlation coefficient was used to test for correlation between continuous variables.
The study was conducted using protocols approved by the institutional ethics Review Boards of the University of North Carolina at Chapel Hill, USA and the University of Malawi, College of Medicine.
Prior to conducting the study, aims and procedures to be used to collect data were explained to parents or guardians and community leaders including school committees during meetings. Written consent was obtained from the local leaders, children's parents or guardians and assent was subsequently obtained from the children.
From a total of 195 schools, 23 schools were selected. A total of 1,150 pupils were interviewed and 1,139 pupils submitted urine specimen which was examined for S. hematobium ova. 4 pupils (0.35%) refused to submit urine specimen and 7 pupils (0.60%) reported that either they did not have the urge to produce urine during the time the study team was at the school or they submitted urine specimen of less than 10 mls.
The mean age of the study population was 10.5 years and most of the children (86.2%) were aged 8 to 13 years with no significant difference by gender. About 44.2% were from poor SES while 21.9% of the children were from high socio-economic households. Characteristics of the study sample are shown in Table 1.
S. hematobium eggs were found in 10.4% (95% confidence interval (CI): 5.4–15.4%) of the pupils who submitted urine specimen. Prevalence in schools ranged from 0.0% to 46.0% and 39.1% (9/23) of the schools had infection prevalence of ten percent or more (Figure 1). Prevalence varied by altitude above sea level; 17.8% of children from <500 m stratum had positive results for S. hematobium ova, 12.4% in the 500–1000 m stratum and 3.60% in the >1000 m stratum (p = 0.007).
Infection was higher in rural areas at 14.4% (95% CI: 6.6–22.2%) compared to the urban areas 3.6% (95% CI: 1.4–5.8%). 13.2% of the boys in the study population were infected compared to 7.4% among the girls. Across all age ranges, boys had higher prevalence compared to girls (Figure 2). The prevalence showed an increasing trend with increasing age and peaked around 11–13 years then started to decline.
Boys were more likely to be knowledgeable of existence of open water source in their area compared to girls, 61.8% and 55.0% respectively (p = 0.03). Among pupils who were knowledgeable of an open water source in their area, children aged 14 years or older were less likely to report playing or swimming in water compared to those less than 14 years old, 79.1% and 90.8% respectively (p = 0.003). There was no significant difference in the proportions that reported daily/ weekly water contact frequency among boys and girls who reported playing or bathing in open water sources, 85.6% and 88.1% respectively (p = 0.4). Children who reported previous history of urinary schistosomiasis infection were more likely to report ever being treated for urinary schistosomiasis in the past compared to children who reported no previous history of schistosomiasis (27.1% and 3.7%; p<0.001).
The overall ova density (eggs/10 ml of urine) among the infected in the study population was 10.1 (range 1–80). The prevalence of heavy infection (≥50 eggs/10 mls) among those infected was 2.4%. The average ova density was calculated for each school and was compared with school prevalence. There was a strong correlation between prevalence of S. hematobium infection and mean ova density in schools (Pearson correlation coefficient r = 0.65 p = 0.005). There was no correlation between age and ova density among those infected. There was no difference in the mean ova density between boys and girls (mean; 9.9 and 10.4 eggs/10 mls urine p = 0.8).
Of the 1,150 pupils interviewed, 353 (31.6%, 95%CI: 23.1–40.1%) reported passing blood in urine (hematuria) over the past month. Proportion of pupils reporting passing blood in urine ranged from 8% to 70% in schools. Using urine microscopic examination as a gold standard, the sensitivity and specificity of self-reported hematuria was 68.3% and 73.6% respectively. The negative predictive value (NPV) was 95.0% while positive predictive (PPV) value was very low at 23.9%. PPV of self-reported hematuria was higher among boys compared to girls (27.3% versus 18.4%) and was higher among pupils aged 9 years or more compared to those aged 8 years old or less (17.0% versus 24.8%).
Table 2 shows the association between risk factors and S. hematobium infection in bivariate analyses. The following factors were significantly associated with infection in bivariate analysis; age 11–13 years, socio-economic status, gender, location, household water source, pupil's knowledge of existence of an open water source, previous history of urinary schistosomiasis and proximity of school to an open water source. Distance from home to an open water source, previous history of schistosomiasis treatment, frequency of water contact and dysuria were not significantly associated with infection. In multivariate analysis, factors that remained significantly associated with S. hematobium infection were male gender (OR 1.81; 95% CI 1.06–3.07), child's knowledge of an open water source in the area (OR 1.90; 95% CI 1.14–3.46), previous history of urinary schistosomiasis (OR 3.65; 95% CI 2.22–6.00) distance of less than 1 km from school to nearest open water source (OR 5.32; 95% CI 1.66–17.07) and age 8–13 compared to those age 14 and older (Table 3).
To our knowledge this is the first large-scale survey that has been conducted in Blantyre district to determine prevalence and risk factors associated with S. hematobium infection. The prevalence of S. hematobium infection was 10.4% and ranged from 0.0% to 46% in schools in our study. Male gender, age, child's knowledge of an open water source in the area, previous history of schistosomiasis and proximity of school to an open water source were independently associated with infection in our study.
The last published S. hematobium prevalence estimates for the district were from studies conducted between 1979 and 1981 among pupils from three schools and individuals from two villages where prevalence ranged from 22% to 72% using centrifugation and reagent dipsticks respectively [10]. Our study prevalence estimate is lower than the national estimates of between 40% and 50% currently in use. Our prevalence findings are similar to the national schistosomiasis survey results [11]. The observed differences between the national estimates in current use and the recent observed findings have been attributed to effective schistosomiasis control and availability of anti-schistosomiasis drugs [11]. In addition the differences in size and methodology of studies could partly explain the observed findings. The more recent large-scale surveys are likely to be more accurate compared to previous studies that may have had some selection bias for high risk schools [11]. Also, the wide range of infection prevalence rates among schools in our study illustrates the focal distribution characteristic of schistosomiasis.
The association between male gender and S. hematobium found in our study is similar to findings from other studies [21]–[23]. Boys were more likely to be infected and be knowledgeable of the existence of an open water source in their area compared to girls. In other studies, boys had more water-contact compared to girls [23]. This could be partly explained by the fact that boys are usually more adventurous, they are more likely to be knowledgeable of their environments including water bodies and therefore be more likely to play in them compared to girls. The association between gender and S. hematobium infection varies in different communities. Some studies have reported no association between S. hematobium infection and gender [22],[24] while other studies have reported association with female gender [25].
Our study showed that there was an increasing trend of infection among children from six years to thirteen years with a decline from 14 years. Also, children aged 14 years or more were less likely to report playing or swimming in water in our study. This suggests that children 14 years and older have lower risk of being infected as they are less likely to be engaged in recreational water-contact behaviors compared to younger children. Studies conducted elsewhere have reported similar results, for example, throughout the nine year schistosomiasis study in Kenya, schistosomiasis infection was lower in older children [22],[26]. Other studies have reported that age-acquired immunity to re-infection contributes to the declining trend in prevalence among children aged 15 years and older [27]. However, we cannot conclude whether age-acquired immunity contributed to our findings since we did not find a negative association between ova load and age.
School proximity to an open water source showed a very strong association with infection. Proximity to water sources has been consistently shown to be associated with schistosomiasis infection in other studies [23],[28],[29]. In contrast, other studies have found no association or variable influence on schistosomiasis infection and it has been suggested that multiple uses of various water bodies with different transmission levels could explain these findings [30],[31]. Interestingly, we did not find a significant association between proximity of household to an open water source in our study. This could be partly explained by the following reasons; first, 95.0% of pupils who were infected and reported that their homes were far or very far from an open water source attended schools that were less than 1 km from an open water source. This means that while children may not engage in water contact recreational activities close to their homes, they might have been exposed when going or coming from school or other places away from home. Secondly, our questionnaire method of estimating distance from home to nearest water source was more subjective and prone to misclassification. Hence our estimate for the effect of household proximity to water source on infection could have been biased.
Communities with high infection rates are usually clustered around contaminated water sources [32],[33]. This could partly explain why we observed an association between S. hematobium infection and previous history of urinary schistosomiasis since children who reported previous history of schistosomiasis were more likely to report ever being treated for urinary schistosomiasis. We did not find a significant association with reported water contact frequency. Studies have found variable influence of water contact frequency on schistosomiasis infection. This has been attributed to variability in amount of body exposed to water and others have suggested that other factors play major role [22],[34].
Socioeconomic status of the household was not an independent factor associated with infection in our study population. Conflicting results have been reported in previous studies [35]–[37]. Based on our findings, possibly improving socioeconomic status alone may not significantly reduce the rate of infection among this population.
The performance of self reported hematuria in our study is similar findings from other endemic countries with sensitivity and specificity range of 50–100% and 58–96% respectively in moderate and high transmission areas [38]–[40]. Our positive predicted value of self-reported hematuria was low and could be due to several reasons. First, our study was limited since only one urine specimen was collected and we did not encourage the children to conduct exercises prior to urine collection. Studies have shown that repeated examination of urine specimen over consecutive days and exercises prior to urine collection improve egg detection [41],[42]. It is possible to have an infected child reporting hematuria with no ova detected one urine sample. Also, this could have been exacerbated by the fact that our study population had low infection intensity (10.1 eggs/10 mls urine). Our results probably could have been different if reported hematuria was compared to Circulating Anodic Antigen (CAA) from S. haematobium. Secondly, children that either reported passing blood in urine or tested positive for S. hematobium infection were referred for treatment and it might be possible that some children reported hematuria to be referred to health facilities and hence affecting accuracy.
This study was initiated and implemented by the district health office. It demonstrates the success of the persistent and deliberate national efforts advocating for active participation at the sub-national or sub-district level in schistosomiasis prevention and control activities in countries where schistosomiasis is endemic. This is crucial for long-term ownership and sustainability of schistosomiasis control efforts. S. hematobium infection was found to be localized in Blantyre district. The district health office in collaboration with the district education offices may consider targeted treatment every two years in schools and surrounding communities with prevalence of between 10% but less than 50% as recommended by WHO in addition to passive treatment in the health facilities and snail control [12]. Even though the predictive value for the self-reported hematuria was low, the sensitivity and specificity was comparable to other previous studies and suggests that the use of questionnaires in this part of Sub-Saharan Africa may still valuable in older children [43]. Further intervention studies to determine the best and cost-effective strategies to provide treatment to children and communities in the affected areas are required. Ecological studies are also needed to identify transmission foci to facilitate implementation of ecologically targeted control measures.
|
10.1371/journal.pntd.0005627 | Expression of inhibitory receptors and polyfunctional responses of T cells are linked to the risk of congenital transmission of T. cruzi | Congenital T. cruzi infections involve multiple factors in which complex interactions between the parasite and the immune system of pregnant women play important roles. In this study, we used an experimental murine model of chronic infection with T. cruzi to evaluate the changes in the expression of inhibitory receptors and the polyfunctionality of T cells during gestation and their association with congenital transmission rate of T. cruzi infection. The results showed that pregnant naïve mice had a higher percentage of CD4+ and CD8+ T cells that expressed inhibitory receptors than cells from non-pregnant naïve mice. However, in mice chronically infected with T. cruzi, gestation induced a significant decrease in the frequency of T cells that expressed or co-expressed inhibitory receptors, as well as an increase in the frequency of polyfunctional CD4+ and CD8+ T cells. This different behavior may be due to the breakdown in the infected mice of the gestation-induced immune homeostasis, probably to control the parasite load. Remarkably, it was observed that the mothers that transmitted the parasite had a higher frequency of T cells that expressed and co-expressed inhibitory receptors as well as a lower frequency of polyfunctional parasite-specific T cells than those that did not transmit it, even though the parasitemia load was similar in both groups. All together these data suggest that the maternal immune profile of the CD4+ and CD8+ T cells could be a determining factor in the congenital transmission of T. cruzi.
| Chagas disease or American trypanosomiasis is a complex parasitic disease caused by the protozoan Trypanosoma cruzi. This disease that affects approximately 10 million people worldwide may be mother-to-child transmitted which is an important public health problem with great relevance in endemic and non-endemic areas and regions where the vector transmission has been controlled. During gestation, the maternal immune system must defend both the mother and the fetus from infections, while, at the same time, it must tolerate a semiallogenic fetus. This immune homeostasis is characterized by a natural process of immunosuppression that in T. cruzi-infected pregnant women can lead to an increase in the mother´s parasite load that favors the risk of congenital transmission. In this study, it was determined the immunological modifications induced by gestation in an experimental model of T. cruzi chronic infection and their influence in the parasite congenital transmission. The results indicate that a T. cruzi infection induces a reversion of the pregnancy-associated homeostasis. Furthermore, it is shown that females who are not able to reverse the biological profile induced by pregnancy are those that transmit the parasite.
| Chagas disease, which is caused by the protozoan Trypanosoma cruzi, is a tropical parasitic disease that affects approximately 10 million people worldwide. The disease is primarily transmitted by contact with the infected feces of triatomine bugs [1]. This parasite can also be transmitted by non-vector mechanisms, including blood transfusions, organ transplants, congenital infections, and oral transmission through food contaminated with insect feces [1, 2]. In a T. cruzi infection, as in other chronic infectious diseases, it has been shown that antigen persistence is associated with a process of a gradual dysfunction of CD8+ T cells [3]. The development of the dysfunction in the T cell response is linked to the constitutive expression of several inhibitory receptors that might negatively regulate the function of antigen-specific T cells, thus compromising pathogen control [4, 5]. In Chagas’ disease patients with increased heart involvement, an active silencing of the CD8+ T cell response is characterized by an impaired ability to simultaneously produce multiple cytokines (polyfunctionality) and an increase in the frequency of CD8+ T cells that co-express inhibitory receptors [3].
Vertical transmission of T. cruzi is a worldwide problem with great relevance in endemic and non-endemic areas [6]. In this regard, some authors consider the congenital T. cruzi infection to be an ecological model that involves multiple and complex interactions between the parasite and (i) the immune system of pregnant women in whom the responses depend on genetic and environmental factors, (ii) the placenta, which possesses its own defense mechanisms, and (iii) the immune system of the fetus [7]. In fact, in a case of twins congenitally infected with T. cruzi, it was reported that a lower level of pro-inflammatory cytokine production was associated with a poorer clinical prognosis in one of the brothers [8].
During gestation, the maternal immune system must defend both the mother and the fetus from infections, while, at the same time, it must tolerate a semiallogenic fetus [9]. This immune homeostasis is mediated by cellular and molecular mechanisms that include the immunosuppressive effects of regulatory T cells and the immunoregulatory capacity of molecules that are involved in the inhibition of specific immune responses [9, 10]. This natural process of immunosuppression in pregnant women infected with T. cruzi can lead to an increase in the mother´s parasite load, especially during the third trimester of pregnancy, and this process increases the risk of congenital transmission during this time [11, 12]. Taking into consideration that the pregnancy-associated homeostasis can increase the parasitemia and that the lack of a polyfunctional response correlates with a high expression of inhibitory receptors and, in turn, with the progression of the Chagas disease [3], we considered important to assess the immunological changes regarding the polyfunctionality of T cells and the expression of inhibitory receptors during gestation in a murine experimental model of T. cruzi chronic infection. The obtained results show that the balance between the expression of the inhibitory receptors and the functionality of CD4+ and CD8+ T cells plays a decisive role in the process of congenital transmission.
The Ethic Committees from the Instituto de Parasitología y Biomedicina López Neyra (IPBLN) and from Consejo Superior de Investigaciones Científicas (CSIC) reviewed and approved the animal care and protocols used in this study (registry number MCL.2/14) as well as the project entitled “Immunological and molecular approaches for the Chagas’ disease control” (identification number CEEA/2014/MCL/1, n° 13-03-14-48). The experiments were performed in the Animal Experimentation Unit from IPBLN (registry number ES-180210000022) and conducted according to the Ethics Guidelines of the Animal Care Unit Committee of IPBLN-CSIC provided by the European Union Directive 2010/63 and Spanish Legislative Decree 53/2013.
Young (6–10 weeks old) female BALB/c mice were purchased from Charles Rivers Laboratories International, Inc. (Paris, France) and housed at the animal facilities of the Institute of Parasitology and Biomedicine “López-Neyra” (Granada, Spain). The mice were maintained in polyethylene cages with food and water ad libitum, on a 12 h light/dark cycle at 20–22°C and 40–60% humidity. The animals were divided into the following four groups: uninfected non-pregnant (n = 10), uninfected pregnant (n = 11), infected non-pregnant (n = 11) and infected pregnant (n = 22).
The infections were performed by intraperitoneal (i.p.) inoculation with 104 trypomastigotes of DA (MHOM/CO/01/DA; DTU I) or SOL (MHOM/ES/2008/SOL; DTU V) strains of T. cruzi in 0.1 ml of sterile phosphate-buffered saline (PBS). DA and SOL strains were, respectively, obtained from an infected human donor in the town of Sutatenza, Boyacá, Colombia [13] and an infant infected by congenital transmission in Spain, Infant 1 referred in [14]. Before infection, the parasites were grown and purified from the monolayers of LLC-MK2 monkey kidney fibroblast line, provided by Dra. N. Andrews (New York University) [15]. The control mice received the same volume of sterile PBS.
Healthy (uninfected) mice, 140 and 240 days old, and infected mice in the chronic phase of Chagas disease (70 and 170 days post-infection) were mated. All mice presented T. cruzi-specific IgG antibodies two months post-infection (S1 Table) measured using an enzyme-linked immunosorbent assay (ELISA) as previously described [16]. The mating was permitted for 5 days by placing one uninfected male and two females in the same cage. The gravid females were euthanized on day 18th of gestation using carbon dioxide inhalation. Blood was collected by cardiac puncture. The fetuses were extracted by caesarean section, washed and stored at -80°C until DNA extraction. The spleens were aseptically removed after dissection and maintained in RPMI-1640 culture medium (Gibco, Grand Island, NY) supplemented with 10% FBS (Gibco), 100 U/ml penicillin, and 10 μg/ml streptomycin (Gibco) until processing.
Genomic DNA (gDNA) extraction from whole blood for conventional PCR and qPCR was carried out using Chelex-100 resin (Bio-Rad Laboratories, Hercules, CA) [17]. The DNA extraction from 250 mg of the fetuses’ tissues from the abdominal-thoracic area was performed using a commercial kit for tissue DNA purification (Qiagen, Valencia, CA) according to the manufacturer´s instructions. For parasite detection, conventional PCR was carried out employing gDNA from whole blood from the mothers and from fetuses tissues using the primers S35 (5’-AAATAATGTACGGG(T/G)GAGATGCATGA-3’) and 122 (5’-GGTTCGATTGGGGTTGGTGTAATATA-3’), which are based on the conserved regions of minicircles from T. cruzi kinetoplast DNA [18, 19], and the primers TcZ1 (5´-CGAGCTCTTGCCCACACGGGTGCT-3´) and TcZ2 (5´-CCTCCAAGCAGCGGATAGTTCAGG-3´), which are based on the satellite DNA of T. cruzi [20]. The quantitative PCR performed for parasite load detection was made with gDNA extracted from peripheral blood from the infected mothers as template and the previously mentioned TcZ1 and TcZ2 primers. The parasite load was estimated through extrapolation of the values obtained in a standard curve generated using DNA purified from the blood of non-infected mice reconstituted with serial dilutions of T. cruzi DNA (DA and SOL strains) ranging from 1 to 0.0001 ng.
Soluble total protein extract (STcA) from amastigote/trypomastigote forms of T. cruzi was obtained as previously described [21]. Briefly, semiconfluent monolayers of monkey kidney fibroblast cells (LLC-MK2) were infected with tripomastigotes form of the Y strain of T. cruzi (MHOM/BR/1950/Y human isolate) at parasite:cell ratio 4:1 for 12 hours. At 96-120h post-infection, a mixture of amastigotes and trypomastigotes forms were recovered from infected-culture supernatants and washed in PBS. Subsequently, the parasites were resuspended in lysis buffer (50 mM Tris-HCl at pH 7.4, 0.05% Nonidet P-40, 50 mM NaCl, 1 mM phenylmethylsulfonyl fluoride (PMSF), 1 μg/mL leupeptin) and sonicated. Soluble protein extracts were obtained after centrifugation at 10,000 rpm for 20 min at 4°C.
The following Abs were used for cell surface staining: anti-CD3 PerCP-Cy5.5 (clone 17A2), anti-CD8a allophycocyanin-H7 (clone 53–6.7), anti-CD4 Alexa Flour 700 (clone GK1.5), anti-PD-1 allophycocyanin (clone J43), anti-2B4 FITC (clone 2B4), and anti-CD160 PE-CF594 (clone CNX46-3) (BD Biosciences, San Jose, CA). The abs for intracellular staining included anti-IFN-γ PE-CF594 (clone XMG1.2), anti-TNF-α PE-Cy7 (clone MP6-XT22), anti-IL-2 Brilliant Violet 421 (clone JES6-5H4), anti-CTLA-4 PE (clone UC10-4F10-11) (BD Biosciences), anti-perforin allophycocyanin (clone eBioOMAK-D) and anti-granzyme B PE (clone NGZB) (eBiosciences, San Diego, CA). A LIVE/DEAD Fixable Aqua Dead Cell Stain Kit (Invitrogen Molecular Probes, Eugene, OR) was used for dead cell exclusion. All conjugated Abs were titrated as previously reported [22].
A total of 1 x 106 splenocytes/ml of medium were cultured with STcA (1 μg/ml) for 12 h at 37°C and 5% CO2. During the last 11 h of culture, Brefeldin A (1 μg/ml) and monensin (0.7 μg/ml) (BD Pharmingen) were present. The splenocytes were adjusted to 1 x 106 cells/tube and labeled with LIVE/DEAD Fixable Aqua for 20 min in darkness at room temperature. To determine the T cell function, the cells were subsequently stained with anti-CD3, anti-CD8 and anti-CD4 mAbs, followed by fixation and permeation for intracellular staining with anti-IFN-γ, anti-TNF-α, anti-IL-2, anti-perforin and anti-granzyme B for 30 min at 4°C. To determine the frequency of the T cells that expressed the inhibitory receptors, the cells were stained with the anti-CD3, anti-CD8, anti-CD4, anti-PD-1, anti-2B4, and anti-CD160 mAbs, followed by fixation and permeation for intracellular staining with anti-CTLA-4 for 30 min at 4°C. In each experiment, non-stimulated cells were included as a negative control. At least 50,000 events, gated on live CD3+ cells, were acquired through flow cytometry using a FACSAria III flow cytometer (BD Immunocytometry Systems, San Jose, CA), and the results were subsequently analyzed using FlowJo 9.3.2 software (Tree Star, Ashland, OR). The gates for positivity in the multicolor panels were determined using fluorescence-minus-one control staining as recommended [23]. The analyses of the co-expression of the inhibitory receptors and polyfunctional responses were performed using a Boolean gating strategy. The data were analyzed and visualized using Pestle version 1.7 and SPICE version 5.3 software (the National Institutes of Health, Bethesda, MD) [24].
The Mann–Whitney U test was used to evaluate the differences between two groups. The differences were considered statistically significant when p < 0.05. GraphPad Prism version 6.0 for Mac OS X statistics software (GraphPad Software, San Diego, CA) was used for the statistical analyses. Co-expression pie charts were compared using 10,000 permutations calculated with the SPICE software.
To determine the effect of gestation on the regulation of adaptive immune-cell function, we evaluated the expression of inhibitory receptors (PD-1, 2B4, CD160, and CTLA-4) on the CD4+ and CD8+ T cells from non-pregnant and pregnant BALB/c naive mice. The results showed that the pregnant mice had significantly higher percentages of CD4+ T cells that expressed PD-1, 2B4 or CD160 and CD8+ T cells that expressed CD160 than those observed in the non-pregnant mice (Fig 1A and 1C) (p<0.01 or p<0.05).
With the aim of determining whether the immune regulation mechanism that prevents rejection of the fetus in naive mice was modified in the context of chronic infection by T. cruzi, the frequencies of the CD4+ and CD8+ T cells that expressed or co-expressed the inhibitory receptors in the uninfected and infected pregnant mice were compared. Significantly lower frequencies of CD4+ T cells that expressed 2B4, CD160 or CTLA-4 were observed in the group of T. cruzi-infected pregnant mice than in the uninfected (healthy) pregnant mice (Fig 1A) (p<0.0001). Among the infected mice, a statistically significant increase in the frequency of CD4+ T cells expressing CTLA-4 was observed in infected non-pregnant regarding that of the infected pregnant mice (Fig 1A) (p<0.001). The evaluation of the co-expression of PD-1, 2B4, CD160, and CTLA-4 showed that significantly lower frequencies of CD4+ T cells that co-expressed two or three inhibitory receptors were found in T. cruzi-infected pregnant mice than in the uninfected pregnant mice (Fig 1B) (p<0.0001). Among the infected mice, a statistically significantly increase in the frequency of CD4+ T cells that co-expressed two inhibitory receptors was observed in infected non-pregnant mice compared to the infected pregnant mice (Fig 1B) (p<0.001). Furthermore, evaluation of the expression of the inhibitory receptors in the CD8+ T cells showed significantly lower frequencies of CD8+ cells that expressed PD-1, 2B4, CD160 or CTLA-4 in the infected pregnant mice than in the uninfected pregnant mice (Fig 1C) (p<0.001 or p<0.0001). Such significant decrease in the frequency of CD8+ T cells expressing CD160 or CTLA-4 was significantly reversed in infected non-pregnant mice (Fig 1C) in regards to the infected pregnant mice (p<0.001 or p<0.0001). Finally, evaluation of the co-expression of these inhibitory receptors showed a significantly lower frequency of CD8+ T cells that co-expressed two or three molecules in the infected pregnant mice group than in the uninfected pregnant mice (Fig 1D) (p<0.0001), whereas in infected non-pregnant mice there was a significant increase in the frequency of CD8+ T cells that co-expressed two and those that expressed one inhibitory receptor with regard to the infected pregnant mice group (Fig 1D) (p<0.001). Taken together, these results indicate that the regulation of the maternal immune system induced during gestation can be modulated by external factors such as T. cruzi infection.
To determine whether the observed patterns of expression and co-expression of the inhibitory receptors was associated with the T cell functions, we evaluated the responses of the T. cruzi-specific CD4+ and CD8+ T cells from the pregnant and non-pregnant mice with chronic T. cruzi infections. Specifically, the production of the intracellular cytokines (IL-2, IFN-γ and TNF-α) and cytotoxic molecules (granzyme B and perforin) after TcSA stimulation was determined. The obtained results showed that in the infected pregnant mice compared to infected non-pregnant mice there was not a significantly higher percentage of CD4+ T cells expressing IL-2+, perforin+ and granzyme B+ (Fig 2A1) but there was a statistically significant higher percentage of CD8+ T cells expressing IFN-γ+ and perforin+ (Fig 2B1) (p<0.05). In order to evaluate the polyfunctional response the frequency of the 16 most prevalent combinations of these functional markers in the T. cruzi-specific CD4+ and CD8+ T cells from the infected non-pregnant and pregnant mice are shown in Fig 2A2 and 2B2. The results showed that in the infected pregnant mice, the frequency of polyfunctional CD4+ T cells tended to be greater than in the non-pregnant mice although these differences were not statistically significant. In fact, CD4+ T cells that demonstrated five (granzyme B+, IFN-γ+, IL-2+, perforin+, TNF-α+) or four (granzyme B+, IFN-γ+, IL-2+, TNF-α+) functions were detected only in the pregnant mice (Fig 2A2). Similarly, a 4% increase in the frequency of the CD4+ T cells that expressed three markers (mainly granzyme B+, IFN-γ+, TNF-α+) and a10% increase in the frequency of CD4+ T cells that expressed two markers (mainly IFN-γ+, TNF-α+) were observed in the pregnant mice compared with the non-pregnant mice (Fig 2A2). Regarding the frequency of the polyfunctional CD8+ T cells, and although these differences were not statistically significant, a 3% decrease in the percentage of CD8+ T cells that expressed three markers (granzyme B+, IFN-γ+, TNF-α+) was observed in the pregnant mice compared with the non-pregnant mice. However, a 7% increase in the frequency of CD8+ T cells that expressed two markers (mainly granzyme B+, perforin+) was observed in the pregnant group of mice compared with the non-pregnant mice (Fig 2B2).
To evaluate the possible changes in the parasite load induced by the gestation-associated homeostasis, the presence of parasite DNA was evaluated in peripheral blood from mothers, before and after pregnancy. The presence of the parasite DNA was detected in 5 out of 22 infected mice before pregnancy, whereas after pregnancy, it was detected in 18 out of 22 mice (Table 1). However, within the group of the 5 mice in which the PCR was positive before pregnancy, only three of the mice were positive by PCR after becoming pregnant. Moreover, the presence of the parasite DNA after pregnancy was detected in 15 out of the 17 mice for which the PCR result before pregnancy was negative. The evaluation of the parasites in the fetal tissue showed that the T. cruzi DNA was detected in 7 out of 18 fetuses from mothers that demonstrated positive PCR results after pregnancy (Table 1). Conversely, in mothers with negative PCR results after pregnancy, congenital transmission was not detected. In addition, it was found that the parasitemia levels were moderately high and similar between the mice that transmitted the parasite (ct mean = 27.7) and the mice that did not transmit it (ct mean = 27.6) (S1 Table).
On the basis that gestation modulated the inhibitory receptor expression and the polyfunctionality of the CD4+ and CD8+ T cells in T. cruzi-infected mice, the inhibitory receptor expression and polyfunctionality profiles of T cells were evaluated in infected pregnant mice that transmitted or did not transmit the parasite to their progeny. A significantly higher frequency of CD4+ cells that expressed CTLA-4 was observed in the group of mothers that transmitted the parasite than in the mothers that did not transmit it p<0.05 (Fig 3A1). When the co-expression of inhibitory receptors was evaluated, significantly higher frequencies of CD4+ T cells that co-expressed two and those that expressed one inhibitory receptor were observed in the mice that transmitted the parasite p<0.05 (Fig 3A2). When the expression of individual inhibitory receptors was evaluated in the CD8+ T cells, significantly higher frequencies of CD8+ T cells that expressed CD160 or CTLA-4 were found in the group of mothers that transmitted T. cruzi compared to the mothers that did not transmit it p<0.05 (Fig 3B1). Similarly, evaluation of the co-expression of inhibitory receptors showed that a significantly higher frequency of CD8+ T cells that co-expressed two and those that expressed one molecule were present in the group of transmitting mice p<0.05 (Fig 3B2). On the basis that the quality of the T cells response is critical for the control of chronic infections such as Chagas disease, we also evaluated the polyfunctional profiles of the CD4+ and CD8+ T cells following stimulation with STcA. A decrease in the frequency of polyfunctional CD4+ (p<0.0001) and CD8+ T cells (Fig 4A1 and 4B1, respectively) and an increase in the frequency of monofunctional CD8+ T cells were observed in the group of transmitting mothers. The frequency of the 16 most prevalent combinations of functional markers in the parasite-specific CD4+ and CD8+ T cells from the infected pregnant mice with positive PCR results are shown in Fig 4B1 and 4B2. In the group of transmitting mice, the most prevalent population was the monofunctional CD4+ T cells producing TNF-α (p<0.001) (Fig 4B1). Furthermore, the CD8+ T cells that exhibited four functions (granzyme B+, IFN-γ+, IL-2+, perforin+) were only detected in the mothers who did not transmit the parasite although it had no statistically significance (Fig 4B2). Likewise, the mothers that did not transmit the parasite showed a higher frequency of CD8+ T cells that produced granzyme B, IFN-γ, and TNF-α compared with the mothers that transmitted the parasite (Fig 4B2).
To prevent rejection of the fetus, a transient suppression of maternal cell-mediated immunity is induced during gestation. This suppression is characterized by decreased effector functions of CD8+ T cells and a lower frequency of CD4+ T cells [25]. Thus, in pregnancy, the immune system has the duties of defending the mother and the fetus from foreign pathogens while simultaneously tolerating the paternal alloantigens expressed by the fetus [26]. The regulation of the immune effector functions is mediated by cellular and molecular interactions with numerous cellular components of the immune system and also by a vast amount of molecules with immunoregulatory capacities, such as inhibitory receptors [10]. However, although these immune mechanisms are needed for fetal protection, they can be a risk factor for the congenital transmission of infectious diseases. In this manner, the immunoregulation during gestation may be involved in mother-to-child transmission of T. cruzi, which has become important in endemic and non-endemic countries. Thus, the aim of this study was to determine the immunological modifications induced by gestation in an experimental model of chronic T. cruzi infection with regard to the expression of the inhibitory receptors that modulate T-cell functions and influence this immunoregulation of parasite congenital transmission. In spite of the DTU of the infecting parasite is most often unknown together to the fact that mixed infections have been described, herein two T. cruzi strains (DTUI and DTU V) were used for the experimental chronic infection in order to evaluate the influence of the parasite genetic variability in the mechanisms of immunoregulation during pregnancy and the parasite congenital transmission. However, given that no differences in the pattern of functional response and inhibitory receptor expression were observed between mice infected with DA or SOL strains (S1 Fig), the obtained results were shown together. In spite of the human monocytes and whole blood cells isolated from healthy donor which were in vitro infected with culture-derived trypomastigotes showed slight differences in the monocyte activation and cytokine production depending on the infecting strain [27], a similar expression pattern of cytokines and cytotoxic molecules was observed by the parasite-specific CD8+ and CD4+ T cells from mice infected with SOL and DA strains following stimulation with total antigens from different T. cruzi strains (S2 Fig). Consequently, in the present manuscript solely T. cruzi soluble antigens (STcA) from Y strain were used for in vitro stimulation.
Several immunomodulatory mechanisms appear to be crucial in regulating the maternal immune cell responses during pregnancy, including molecules that generate negative signals that regulate the activation of CD4+ and CD8+ T cells [28]. In particular, PD-1 has been described as playing an important role in maternal-fetal tolerance by inducing apoptosis of paternal antigen-specific T cells during pregnancy [29]. In fact, PD-1 ligand blockage in vivo resulted in increased abortion rates and reduced litter sizes [30]. Previous studies have also shown the involvement of CTLA-4 in the suppression of activated T cells [31]. Nevertheless, during pregnancy, CTLA-4 blockage did not interfere with the protective effect of the Treg cells [32]. Others molecules including CD160 and 2B4, which regulate cytotoxic activity and cytokine production [3], might play critical roles in the delicate balance between effective immunity and maternal tolerance to the fetal allograft. Similarly, our results showed that during gestation in uninfected (healthy) mice, an up-regulation of inhibitory receptors was induced, perhaps as a mechanism for the maintenance of maternal-fetal tolerance. Interestingly, although the frequency of the CD4+ and CD8+ T cells that expressed inhibitory receptors tended to increase during gestation in healthy mice, the percentage of CD4+ and CD8+ T cells that co-expressed these molecules was similar in pregnant and non-pregnant healthy mice. This finding suggests that this up-regulation was a physiological mechanism induced by the pregnancy but was not a sign of T cell exhaustion.
In this study, gestation in a murine model was found to induce a differential immunological pattern depending on the status of the mice (healthy or T. cruzi infected). Specifically, in healthy mice, gestation led to significant increases in the individual expression of the inhibitory receptors on the CD4+ and CD8+ T cells, perhaps as a mechanism involved in the feto-maternal tolerance. However, in mice chronically infected with T. cruzi, gestation induced a significant decrease in the frequency of T cells that expressed or co-expressed inhibitory receptors, an increase in the frequency of polyfunctional CD4+ and CD8+ T cells and an increase in the parasite load. This different immunological profile induced by the infection during pregnancy may be due to the breakdown of the gestation-induced immune homeostasis, probably to control the parasite load. Interestingly, this pattern of reversing the process of immune homeostasis and increased polyfunctionality in CD4+ and CD8+ cells appeared to be directly related with no T. cruzi congenital transmission. Thus, our results show that the mice that transmitted the parasite have an increase in the frequency of both the CD4+ and CD8+ cells that co-expressed the inhibitory receptors and a decrease in the polyfunctionality of these T cells. These results suggest that females who are not able to reverse the biological profile induced by pregnancy are those that transmit the parasite, which demonstrates the importance of the balance in the immune response to keep the fetus alive and to prevent the parasite transmission.
Consistent with previous reports [12, 14, 33], our results showed that there was a detectable increase in the parasitemia level (measured as a positive conventional PCR results after gestation) during gestation in the T. cruzi-infected mice and that those mice who transmitted the infection had a positive PCR during gestation. However, when the parasitemia was assessed by quantitative PCR in mice with a positive PCR result (transmitting and non-transmitting the infection) similar levels of parasitemia were detected in all the mice, those that transmitted and those who did not transmit the infection. These data indicate that although the parasitemia in pregnant mice seems to be an important factor that contributes to the congenital transmission of the parasite, it is not a determining factor. Thus, our data indicate that an effective response of the maternal T cells is critical for reducing the chances of parasite transmission from the mother to the fetus. In agreement with this, other authors have suggested that monocytes from parasite-transmitting women are less activated than those from non-transmitting women [34]. In fact, chronically infected non-pregnant women have increased levels of circulating TNF-α. In women that do not transmit the parasite, these levels remain elevated during gestation. In contrast, pregnant women that transmit the parasite showed a down-regulation in the secretion of the IFN-γ and TNF-α [35, 36]. Together, and based on the fact that there were no differences in the parasitemia levels among mice that transmitted and those who did not transmit the infection, these results show that the quality and/or capacity of the T cell response can be considered as a determining factor in the congenital transmission of T. cruzi. This is consistent with the ecological model of multiple and complex interactions among parasites, mothers, placentae and fetuses [7]. Thus, the results shown in this manuscript suggest that the immunological profile of a mother during pregnancy could be an useful tool for determining the risk factor of vertical transmission of the T. cruzi parasite. Evaluation of these profiles would facilitate the subsequent clinical follow-up of the mothers and the children.
|
10.1371/journal.ppat.1003440 | The Interactomes of Influenza Virus NS1 and NS2 Proteins Identify New Host Factors and Provide Insights for ADAR1 Playing a Supportive Role in Virus Replication | Influenza A NS1 and NS2 proteins are encoded by the RNA segment 8 of the viral genome. NS1 is a multifunctional protein and a virulence factor while NS2 is involved in nuclear export of viral ribonucleoprotein complexes. A yeast two-hybrid screening strategy was used to identify host factors supporting NS1 and NS2 functions. More than 560 interactions between 79 cellular proteins and NS1 and NS2 proteins from 9 different influenza virus strains have been identified. These interacting proteins are potentially involved in each step of the infectious process and their contribution to viral replication was tested by RNA interference. Validation of the relevance of these host cell proteins for the viral replication cycle revealed that 7 of the 79 NS1 and/or NS2-interacting proteins positively or negatively controlled virus replication. One of the main factors targeted by NS1 of all virus strains was double-stranded RNA binding domain protein family. In particular, adenosine deaminase acting on RNA 1 (ADAR1) appeared as a pro-viral host factor whose expression is necessary for optimal viral protein synthesis and replication. Surprisingly, ADAR1 also appeared as a pro-viral host factor for dengue virus replication and directly interacted with the viral NS3 protein. ADAR1 editing activity was enhanced by both viruses through dengue virus NS3 and influenza virus NS1 proteins, suggesting a similar virus-host co-evolution.
| Viruses are obligate intracellular parasites that rely on cellular functions for efficient replication. As most biological processes are sustained by protein-protein interactions, the identification of interactions between viral and host proteins can provide a global overview about the cellular functions engaged during viral replication. Influenza viruses express 13 viral proteins, including NS1 and NS2, which are translated from an alternatively spliced RNA derived from the same genome segment. We present here a comprehensive overview of possible interactions of cellular proteins with NS1 and NS2 from 9 viral strains. Seventy nine cellular proteins were identified to interact with NS1, NS2 or both NS1 and NS2. These interacting host cell proteins are potentially involved in many steps of the virus life cycle and 7 can directly control the viral replication. Most of the cellular targets are shared by the majority of the virus strains, especially the double-stranded RNA binding domain protein family that is strikingly targeted by NS1. One of its members, ADAR1, is essential for influenza virus replication. ADAR1 colocalizes with NS1 in nuclear structures and its editing activity is enhanced by NS1 expressed on its own and during virus infection. A similar phenomenon is observed for dengue virus whose NS3 protein also interacts with ADAR1, suggesting a parallel virus-host co-evolution.
| Influenza A viruses are the causative agents of seasonal and pandemic infections and are responsible for the death of at least half a million people worldwide each year. The genome of influenza A viruses is composed of eight negative-sense single-stranded RNAs encoding 13 proteins. NS1 and NS2 are derived from alternatively spliced RNAs that are transcribed from the eighth RNA segment. The segments are encapsidated by binding to nucleoproteins (NP) and the polymerase complex (PA, PB1 and PB2) forming the viral ribonucleoproteins (vRNPs). The viral particle contains eight vRNPs, the surface glycoproteins haemagglutinin (HA) and neuraminidase (NA), the matrix proteins (M1 and M2) and the NS2 protein. Some strains express the pro-apoptotic PB1-F2 protein and two additional virulence factors, PB1-N40 and PA-X, have been recently identified [1]–[3].
The NS1 protein is not incorporated in the virus. It exerts a large spectrum of functions through interactions with a variety of cellular components residing either in the cytoplasm or in the nucleus. NS1 is a pleiotropic virulence factor repressing innate antiviral mechanisms e.g. by interfering with the type I interferon system through direct interaction with PKR and TRIM25, or through the sequestration of double-stranded RNA [4]–[8]. NS1 is also known to perturb the mRNA processing by interacting with CPSF4 and PABPN1 to inhibit nuclear export of cellular mRNA [9] and is suspected to hijack the RNA translation machinery in favor of translation of viral protein e.g. by interacting with STAU1 [10]–[11].
In contrast to NS1, NS2 protein is a structural component of the viral particle and it associates with the viral matrix M1 protein [12]. NS2 mediates the export of vRNPs from the nucleus to the cytoplasm through export signal [13] via its interaction with XPO1 [14]. In addition, NS2 interacts with nucleoporins and was suggested to serve as an adaptor between vRNPs and the nuclear pore complex [13]. A role of NS2 in the regulation of influenza virus transcription and replication has also been proposed [15]. However, many functions of NS2, in particular its transit through the cytoplasm and its incorporation into the viral particle, are not understood.
Several screens have been performed to identify host factors involved in the influenza virus replication cycle, mainly focusing on interactors of vRNPs or of the polymerase by using affinity purification or yeast two-hybrid techniques [16]–[18]. A proteome-wide screen of virus-host protein-protein interactions has provided an important resource of 135 interactions [19]. However, the weak overlap of the public datasets suggests that they are far from being complete.
The impact of cellular proteins on the influenza virus replication has been extensively studied using RNAi screens [19]–[24]. Although poorly overlapping at the gene level, these screens better converge at the level of biological processes [25]–[27]. Hence, more than the identification of host factors, these studies highlighted major cellular functions that are essential for the virus replication. However, for the majority of identified host factors, the mode of action remains to be determined. Furthermore, comparisons of strain-specific virus-host interactomes are clearly missing, which is required to reveal general principles governing infection mechanisms and to identify common therapeutic targets as well as broad-spectrum antivirals.
In the present study we conducted stringent yeast two-hybrid screens to identify human proteins interacting with NS1 and NS2 from 9 influenza A virus strains representative of the variability in nature. The functional impact of all NS1 and NS2 interactors on viral replication was systematically addressed by RNA interference. In combination with published datasets, our new results offer a comprehensive view of NS1 and NS2 interactomes and corresponding targeted cellular functions. The global analysis of the NS1 and NS2 host cell targets reveals an enrichment of double-stranded RNA binding domain (DRBD) containing proteins for the 9 tested influenza virus strains. A focus was put on ADAR1 since this protein is critical for the replication of other viruses [28], is highly expressed in human lung cells [29], is induced by type I interferon [30], is interfering with interferon signalling production [31] and is interacting with all tested NS1 proteins. In addition, we also observed in another screen that ADAR1 interacts with the dengue virus NS3 protein which is a bifunctional enzyme containing protease and helicase activity [32]. We show that ADAR1 is a pro-viral host factor favoring replication of influenza virus and dengue virus and that these viral proteins can control ADAR1 editing activity.
To identify all cellular proteins interacting with influenza virus NS1 and/or NS2 proteins, yeast two-hybrid screens (Y2H) were carried out using NS1 and NS2 proteins from 9 different virus strains as baits (Table S1) and three cDNA libraries (from human spleen, fetal brain and respiratory epithelium). Key features of the virus strains are provided in Text S1. NS1 and NS2 proteins selected for this study are representative of the natural diversity since they are distributed all along the phylogenetic trees of known NS1 and NS2 sequences (Text S1, Figures S1 and S2 in Text S1, Alignments of NS1 and NS2 protein sequences are presented in Figures S3 and S4 in Text S1). Seventy nine non-redundant cellular proteins were identified to interact with NS1, NS2 or both and were individually retested in a pairwise array (Figure 1A). From a total of 1422 possible interactions tested (79 cellular proteins tested against 9 NS1 and 9 NS2 proteins), 562 tested positive. In this way, we identified 33 cellular proteins interacting exclusively with NS1, 28 exclusively with NS2, and 18 with both NS1 and NS2.
The vast majority (97.5%) of the NS1 and NS2 interactors are known to be expressed in the respiratory epithelium (Table S2). Twelve out of the 79 host interactors have already been reported (AIMP2, SCRIB, CPSF4, the kinases PIK3R1, PIK3R2, MAPK9, CRK and proteins with a double-stranded RNA-binding domain STAU1, PRKRA, ADAR1, TARBP2, ILF3) [9], [11], [19], [33]–[38]. 21.5% of host interactors are targeted by all virus strains (Figure 1B) and 5% appear to be strain specific. 80% of the cellular interactors bind to more than 50% of the tested NS1 and NS2 proteins indicating that the dataset is more appropriate to the identification of common rather than differential interaction profiles. Together with previously published data available in the VirHostNet database [39], we now provide a list of 111 non-redundant cellular proteins interacting exclusively with NS1, 32 exclusively with NS2 and 18 with both proteins (a complete list of influenza virus interactors is given in Table S3).
Consistent with observations from previous virus-host interactome studies, NS1 and NS2 proteins tend to interact with highly central proteins in the human interactome [40]–[43]. Indeed, the degree distribution of targeted human proteins was significantly higher than the degree distribution in the human interactome (U-test, p-value<2.2×10−16) (Figure 1C). Similarly, the betweenness distribution of targeted human proteins was significantly higher than the betweenness distribution in the human interactome (U-test, p-value<2.2×10−16) (Figure 1D). This suggests that influenza NS1 and NS2 proteins preferentially target pleiotropic cellular proteins [44].
Finally, an assessment of Gene Ontology categories revealed a significant enrichment (p-value = 3.3×10−14) for DRBD-containing proteins (DRBPs) in the interaction dataset. Strikingly, DRBPs were exclusively targeted by NS1 proteins. All virus strains interacted with most of the DRBPs suggesting that the direct targeting of DRBDs is of special importance for influenza A viruses.
Among the 79 NS1 and NS2 interactors identified here, 12 have been previously identified in recent genome-wide siRNA screens as modulators of viral replication - ATP6V1G1, RPL13A [21], [23], GMEB1, PIK3R2 [19], SON, EEF1A1 [23], CHCHD5, RPL23A [24] and NUP214 [22]. However, since these genome-wide siRNA screens are weakly overlapping, it is very likely that numerous modulators of viral replication have been missed or remain to be confirmed [25]–[27].
We have therefore performed a systematic siRNA-based screen in A549 human lung epithelial cells to explore the functional contribution of the 79 cellular NS1 and NS2 interactors to virus replication. The silencing phenotype was first tested by measuring replication of the A/H1N1/Puerto Rico/8/34 virus strain which was used in the yeast two- hybrid screen. The complete replication cycle was first probed by measuring the neuraminidase activity in the supernatant 48 h post-infection. The assay was calibrated by using siRNAs against ATP6V1G1 and CSNK2B that have been previously described as pro-viral host factor and anti-viral host factor respectively. ATP6V1G1 is a subunit of the vacuolar ATPase proton pump required for influenza A virus replication [21] while CSNK2B gene silencing increases virus replication in A549 infected cells [45]. As expected, siRNAs targeting ATP6V1G1 and CSNK2B respectively reduced and increased the neuraminidase activity in the supernatant, thus validating the assay (Figure 2A and 2B). By comparison with these controls, virus replication should be altered by at least 35% according to the threshold defined by König et al. in their genome-wide siRNA screen [22]. This threshold together with a silencing efficiency greater than 60% for each siRNAs without detectable cytotoxicity were used for a stringent selection of the pro-viral and anti-viral host factors (Table S4). These criteria are in the range applied in earlier siRNA-based screens [23], [46], [47] (information on individual silencing efficiency is also provided in Table S4).
In this way, we identified the two pro-viral host factors, ADAR1 and RPSA, and confirmed ATP6V1G1, RPL13A, EEF1A1 and SON (Figure 2A). In addition, one new anti-viral host factor, N-PAC, was identified (Figure 2A). These results were confirmed by using plaque assays, revealing a broader range of inhibition or activation of virus replication (Table S4). In conclusion, out of the 79 cellular interactors of NS1 and NS2 identified in this study, 7 were identified as possible direct modulators of A/H1N1/Puerto Rico/8/34 virus replication. Importantly, these results were confirmed with the A/H1N1/New Caledonia/2006 influenza virus strain that was not used in the yeast two-hybrid screen (Figure 2B), although RPL13A only scored positive 72 h post infection (Figure S5 in Text S1).
As it is a critical component of virus-host interaction, production of type I interferon was quantified in the supernatant of infected cells transfected with siRNAs (Table S4). Silencing of ADAR1, ATP6V1G1, BCLAF1, RPL13A and SON increased interferon production in infected cells. These proteins interact with NS1 and, except for BCLAF1, are pro-viral host factors for the virus. Therefore, 4 of the 6 pro-viral host factors are implicated in interferon production. This is consistent with the role of NS1 in interfering with the type I interferon system.
The function and subcellular localization of cellular interactors identified in the present and published studies indicate that both NS1 and NS2 are pleiotropic proteins required for several essential steps of the viral life cycle (Figure 3). Although NS2 function in the cytoplasm remains elusive, it is shown here that NS2 mostly targets proteins of the cytoskeleton and involved in intracellular transport (Figure 3). Given that NS2 also interacts with the vRNPs, it might also mediate their transport to the plasma membrane or to the nucleus. In case of the latter, NS2 is implicated in the export of vRNPs, consistent with the observed interaction of NS2 with the NPC [13]. The targeting of transcription-regulating proteins by NS2 is much less documented. A role of NS2 in regulating influenza virus RNA genome transcription via its interaction with vRNPs has been previously proposed [15], [48]. Although such a direct interaction with the components of the vRNPs is not ruled out, our data indicate a direct targeting of the cellular transcriptional machinery by NS2 (Figure 3, box regulation of transcription). Interestingly, since NS1 also targets this process, a potential cooperation between NS1 and NS2 for the control of the cellular transcription machinery can be speculated.
NS1 proteins target several DRBPs either localized in the nucleus or in the cytoplasm [49]. These proteins are critical transcriptional or translational checkpoints. NS1 is known to inhibit the activation of PKR, one of the major interferon-inducible antiviral effectors, through direct interaction [6]. More recently, SON has been described to be important for the trafficking of influenza virions [23]. Here, we confirmed that SON is essential for viral replication and suggest that this activity could be related to the NS1 protein. ADAR1 and PKR have an opposite effect on virus replication although they are both induced by type I interferon.
ADAR1 is a type I interferon-induced protein that is expressed in human lung [29], [30] and interacts with all tested NS1 proteins. The 150 kDa interferon-inducible ADAR1 isoform is expressed in A549 cells upon influenza A virus infection and by type I interferon. The constitutive 110 kDa ADAR1 isoform was only induced upon infection indicating that ADAR1 expression can also be controlled by an interferon-independent mechanism, at least in the setting of an influenza A virus infection (Figure 4A). ADAR1-specific siRNAs efficiently reduced the expression of ADAR1 isoforms and blocked their induction upon infection (Figure 4B). The silencing of ADAR1 inhibited virus release from 15% at 8 h post-infection to 90% at 48 h post infection (Figure 4C). Expression of viral proteins (here HA, NP, M1 and NS1) was also significantly reduced as early as 8 h post infection. NS1, NP and M1 expression was delayed while HA expression remained very low until 24 h post infection (Figure 4B). Thus, ADAR1 is a pro-viral host factor for virus protein expression and virus production.
Immunofluorescence revealed that ADAR1 is diffusely distributed in the nucleus and relocalized in nuclear structures in influenza virus-infected cells (Figure 5A). In these structures ADAR1 colocalized with NS1 but not with HA for which no interaction with ADAR1 could be detected. As NS1 interacts with several DRBD-containing proteins, the NS1 binding site in ADAR1 could be a DRBD. Amino acid sequence alignment of DRBDs revealed a conserved region of 47 amino acid residues within the two firsts DRBD of ADAR1 (Figure S6 in Text S1). A set of 4 ADAR1 deletion mutants, differing in their number of DRBDs, and a plasmid encoding the 47 amino acid residues of the first DRBD were constructed (Figure 5B). In a yeast two-hybrid array, ADAR1 interacted with NS1 even in the absence of its first DRBD while interaction was completely abrogated when the first two DRBDs were deleted. The peptide of 47 amino acid residues also interacted with NS1 (Figure 5C) in the array and in GST pull-down assays (Figure 5D). Thus, ADAR1 displays two potential NS1 interaction sites located on the first two double-stranded RNA-binding domains.
To validate these results the NS1 RNA-binding domain (RBD) and effector domain fused to GST were used in pull-down experiments for the mapping of NS1 interaction with 3×Flag tagged ADAR1 after co-expression in HEK293T cells (Figure 5E). Full-length NS1 and NS1 RBD domain efficiently co-precipitated ADAR1 but not the effector domain (Figure 5F) indicating that NS1 interacts with ADAR1 through its RBD. GST pull-down and RNAse A treatment showed that RNA is marginally involved in the NS1-ADAR-1 interaction (Figure S7 in Text S1). A mutant of NS1 that lacks double-stranded RNA-binding activity still interacts with ADAR1, albeit with reduced efficiency (Figure S8 in Text S1) confirming that RNA is not strictly required for NS1 interaction with ADAR1.
To evaluate the functional impact of ADAR1-NS1 interaction on the catalytic activity of the enzyme, an original editing reporter system was constructed. This reporter system consists of a 24 nucleotide-long minimal ADAR1 substrate derived from the sequence of the antigenome of the hepatitis delta virus that is edited by this enzyme [50]. In this sequence, ADAR1 editing activity changes a stop codon into a tryptophane codon (Figure 6A) [51]. The reporter plasmid contains the ADAR1 substrate sequence inserted in frame in-between the Renilla and the Firefly luciferase genes (Figure 6A). In this configuration, the Firefly luciferase activity reflects the extend of editing and thus ADAR1 activity, leading to the conversion of the stop codon into the tryptophane codon. ADAR1 was co-expressed in HEK293T cells with NS1 or its RBD and with the editing reporter construct. The NS1 effector domain or DLG4, which does not bind to ADAR1 (data not shown), was used as negative control in analogous co-transfection experiments. NS1 RBD and full-length NS1 increased the Firefly luciferase signal by 30% and 60% respectively (Figure 6B) suggesting that NS1 can cooperatively interact with ADAR1 via its RNA-binding domain to promote ADAR1 editing activity (Figure 6B). Editing activity was also analyzed in the context of influenza virus infection after expression of the editing reporter construct (Figure 6C). H1N1 influenza virus infection increased the editing activity of ADAR1 by 70% and this was completely reversed when ADAR1 expression was silenced by RNA interference.
To validate these observations, a catalytically inactive ADAR1 (E912A) mutant was constructed [52]. Unfortunately, A549 cells became refractory to plasmid DNA transfection after siRNA transfection, thus precluding functional tests of the mutant in this cell line (not shown). As an alternative, we tested a potential transdominant negative effect of the ADAR1 mutant on influenza virus growth. The catalytically inactive ADAR1 (E912A) mutant construct was therefore transfected into A549 cells and virus growth in these cells was compared to the one achieved with mock-transfected cells or in wild type ADAR1-transfected cells. Viral protein expression was reduced in A549 cells expressing the ADAR1 mutant compared to control cells (Figure 6D). Consistent with this result, neuraminidase activity in the supernatant was also significantly reduced (Figure 6E). Importantly, since influenza A virus infection induces endogenous ADAR1 expression, the impact of the ADAR1 mutant is most likely underestimated in this experimental system. We therefore concluded that the RNA editing function is required for the pro-viral activity of ADAR1.
During the course of a systematic screening for virus-host protein-protein interactions with a yeast two-hybrid system, we also identified ADAR1 as an interactant of the NS3 protein of dengue virus type 2. This interaction was confirmed in a yeast two-hybrid pairwise array (Figure 7A). As for NS1 of influenza A virus, interaction between NS3 and ADAR1 was validated by GST pull-down experiments (Figure 7B) and also in this case, RNA contributed to this interaction only to a very minor extent (Figure S9 in Text S1). Both ADAR1 isoforms were induced upon dengue virus infection as well as upon type I interferon treatment of Huh-7 cells (Figure 7C). Silencing of ADAR1 expression by RNA interference (Figure S10 in Text S1) resulted in a strong decrease of dengue virus replication (Figure 7D). This result was confirmed with a subgenomic dengue virus replicon stably replicating in Huh-7 cells, indicating that ADAR1 acts at a post-entry step in the dengue virus life cycle (Figure S11 in Text S1). Likewise, as observed for influenza virus, dengue virus infection strongly increased the editing activity of ADAR1 (Figure 7E). In fact, full-length NS3 and the helicase domain increased the Firefly signal by 24% and 44% respectively, suggesting that NS3 cooperatively interacts with ADAR1 to enhance its editing activity (Figure 7F).
In conclusion, both influenza virus and dengue virus (i) induce over-expression of ADAR1, (ii) interact with ADAR1 through the RNA-binding domain of influenza virus NS1 and the helicase domain of dengue virus NS3, (iii) enhance the editing activity of ADAR1 and (iv) are dependent on ADAR1 expression for efficient virus replication.
This study describes an exhaustive interaction profile for NS1 and NS2 proteins of 9 influenza virus strains. More than 560 interactions between 79 cellular proteins and NS1 and NS2 were identified. Thirty-three cellular proteins interacted exclusively with NS1, 28 exclusively with NS2, and 18 with both NS1 and NS2. Since NS1 and NS2 are the products of alternatively spliced RNAs, shared interactions may reflect binding to the common N-terminal 10 amino-acid residues long sequence. This result suggests that influenza viruses have evolved two proteins to interact with cellular proteins that are potentially essential for them.
Twelve out of the 79 NS1 and NS2 cellular interactors have already been reported in the literature, demonstrating the reliability and robustness of our screening approach. For NS1, there is a strong overlap with hits published by others (11 of the 51 interactors identified in the present study, which is well above the average overlap) [41], [53], suggesting that the NS1 interactome dataset is now close to completion. In case of NS2, only 4 cellular interactors have been published and one of them, AIMP2, has been confirmed in our screens. Although 46 new NS2 interactors have been identified, it is difficult at this stage to estimate the completion level of the NS2 interactome due to the lack of published interaction data.
Overall, most of the cellular targets interacted with the majority of NS1 or NS2 proteins of the different influenza viruses arguing that we have identified highly relevant and evolutionary conserved interactions. Interestingly, a significant proportion of these proteins is also targeted by other viruses (44.7%, exact Fisher test, p-value<2.2×10−16) indicating that these cellular proteins are likely to be involved in a generic process of viral infection [39].
Our interaction dataset indicates that NS1 and NS2 proteins are likely to be involved in multiple steps of the viral replication cycle, paving the way for challenging functional explorations. This was largely unexpected for NS2, which is known to be involved in the nuclear export of the vRNPs. Its interaction with the cytoskeleton appears particularly interesting for further studies. Although the pleiotropic nature of NS1 is well established [54], our study provides new insights into the breadth of interactions and activities of this regulatory protein. In addition to the 67 new interactors, the current dataset also provides additional information on previously known interactors and related targeted functions. For instance, the CPSF4 interaction with NS1 has been described as a potential therapeutic target [55] and is confirmed in our study. Three NS1 proteins also interacted with CPSF3L, a protein participating in the endonuclease activity of CPSF, suggesting that the corresponding viruses evolved alternative strategies to interfere with the cellular 3′end mRNA processing [56].
The phenotypic analysis of the cellular targets of NS1 and NS2 by RNA interference revealed an enrichment in modulators of influenza virus replication, further validating the interaction dataset. Indeed, out of the 79 cellular interactors of NS1 and NS2 identified in this study, 7 revealed to control positively or negatively the replication of two influenza virus strains. Interaction profiles suggest that the data could be extrapolated to other strains with the noticeable exception of RPL13A, an exclusive target of A/H1N1/Puerto Rico/8/34 NS1. The validation rate of cellular interactors by RNAi reached about 9% (15.2% when data from the literature are included) and is similar to that of Shapira et al. [19] while the validation rate of virus replication modulators identified from genome-wide siRNA screens ranges from 0.75 to 1.5%. Therefore, combining interactomic screens with genetic screens drastically enhances the rate of functional validation, providing lists of cellular proteins strongly enriched in pro- and anti-viral host factors (exact Fisher test, p-value = <2.1 10−4, Text S1).
Interaction of NS1 with some members of the DRBD protein family have been sporadically documented [6], [10], [19]. Here we observed a massive enrichment of the DRBD protein family in our NS1 interactome for which we used 9 different influenza virus strains. One hundred and sixty five independent screens have been performed with other viral baits using the same cDNA libraries (45 with the fetal brain cDNA library, 31 with the respiratory epithelium library and 89 with the spleen library). The GO term “Double-stranded RNA-binding domain (DRBD) containing proteins” has never been enriched in any of these screens while it was enriched for the 9 tested influenza virus strains. Reciprocally, a large diversity of other GO terms was enriched in these different screens and in screens performed by other laboratories using the same libraries. Therefore, we could be confident that the DRBD containing proteins enrichment reflects a real propensity of NS1 to interact with this protein family. This is most likely reflecting the ability of NS1 to interact with the double-stranded RNA-binding domain of cellular partners through its own RNA-binding domain. Two DRBD-containing proteins, SON and ADAR1, were found to be essential for virus replication. Conflicting results on the role of ADAR1 for virus replication have been published. Initially suspected to have an antiviral activity because of its induction by interferon, ADAR1 appears to promote the replication of several viruses (measles virus, vesicular stomatitis virus, hepatitis delta virus, human immunodeficiency virus type 1 and Kaposi's sarcoma-associated virus). In contrast ADAR1 was reported to display an antiviral activity against hepatitis C virus and lymphocytic choriomeningitis virus [28], [57]. Concerning influenza A virus, two studies provided evidence for an antiviral role of ADAR1. Mice lacking IKKε become highly susceptible to influenza virus infection, express ADAR1 only to low amounts and show a reduced editing of matrix M1 mRNA isolated from infected lung. However, since IKKε knock-out also strongly affects the expression of other type I interferon-stimulated genes, the susceptibility of these mice to infection could not be attributed to a unique defect in ADAR1 activity [58]. An increased cytopathic effect of influenza A virus has been observed in mouse cells derived from non-viable embryos unable to express the p150 isoform of ADAR1. However, this effect was not correlated to an increased virus replication [59]. In the present study, we show that inhibiting ADAR1 expression by RNA interference reduced viral protein expression and drastically impaired virus replication. Thus, ADAR1 appeared as an important host dependency factor for influenza viruses.
Several studies have demonstrated a role of ADAR1 in modulating interferon signaling. Inducible ADAR1 disruption in mice causes a global interferon response [31]. Mutations in ADAR1 responsible for Aicardi-Goutières syndrome in humans are associated with upregulation of interferon-stimulated genes [60]. ADAR1 also suppresses measles virus-induced production of interferon-β mRNA [61]. Here, we show that interferon-β is enhanced in ADAR1-deficient cells after infection with influenza A virus. NS1 is a well-known antagonist of the antiviral response. Its mode of action is pleiotropic including interference with signaling induced by RIG-I like receptors (RLRs) [62]. A combined action of ADAR1 and NS1 protein is suggested by our results. The double-strand RNA editing activity of ADAR1 produces double-strand RNA with I:U pairs instead of A:U pairs. Interestingly, I:U-containing double-strand RNA can suppress the induction of interferon-stimulated genes [63]. Conceivably NS1 might potentiate the hyperediting of an as yet unknown double-strand RNA substrate and thus interfere with interferon induction. A similar mechanism can be expected for dengue virus NS3 protein. Both NS1 and ADAR1 also interfere with PKR activity [6], [64]. ADAR1 and PKR are recognized by non overlapping domains of NS1 (respectively the RNA binding domain and the effector domain [7]). Thus, both NS1 and ADAR1 could sequester double-strand RNA or could form inactive complexes, suppressing PKR-mediated proapoptotic and interferon-mediated amplification activities.
Influenza A NS1 protein is considered as a valid target for the development of antiviral drugs. The druggability of NS1 has been demonstrated in a proof-of-concept study with an inhibitory peptide derived from CPSF30, a cellular protein that interacts and interferes with the effector domain of NS1 [55]. Such a strategy can be extended to other NS1 interactors once the interacting sequences have been mapped and the 3D structure is solved. The interacting sequences, e.g. the ADAR1-derived 47 amino acid peptide, could then be used for the design of low molecular weight compounds. In this respect, the systematic screening for protein-protein interactions between a virus and its host cell identifies cellular proteins promoting or restricting virus replication. Interference with these interactions may offer new alternatives to enlarge the diversity of potential therapeutic targets and prevent the emergence of resistance caused by rapid viral adaptation. Small molecules targeting these host interaction surfaces and developed for other therapeutic purposes could now be tested for their ability to control virus replication. Concerning ADAR1, new inhibitors of the RNA editing activity are being screened and could be tested for their capacity to block the replication of influenza A virus or anti-dengue virus [65]–[67]. The dual luciferase editing reporter described in this study is well suited for screening RNA editing inhibitors at a high throughput level.
Human HEK293T and human lung adenocarcinoma A549 cells were maintained in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% heat-inactivated fetal bovine serum (FBS), 50 IU/ml penicillin G, 50 µg/ml streptomycin, at 37°C under 5% CO2. Huh-7 cells were grown in DMEM supplemented with 2 mM L-glutamine, non-essential amino acids, 100 IU/ml of penicillin, 100 µg/ml of streptomycin and 10% fetal calf serum.
Influenza ORFs (Text S1) were transferred from pDONR207 into bait vector (pPC97, Lifetechnologies) to be expressed as Gal4-DB fusions in yeast. Bait vectors were transformed into AH109 (bait strain, Clontech [68]), and human spleen, fetal brain and respiratory epithelium Gal4-AD-cDNA libraries (each containing more than 106 primary clones) were transformed into Y187 (prey strain, Clontech). Single bait strains were mated with prey strains and diploids were plated on SD-W-L-H+ 10 mM 3-AT medium. Each screen has covered more than one time the libraries. Positive clones were maintained onto this selective medium for 15 days to eliminate any contaminant AD-cDNA plasmid. AD-cDNAs were PCR-amplified, sequenced and analyzed using pISTil [69]. Cellular ORFs (interacting domains found in Y2H screens) were amplified from a pool of human cDNA libraries or from a plasmid encoding the corresponding cDNA from the MGC collection (IMAGE consortium) using KOD polymerase (Toyobo) and cloned by recombinational cloning into pDONR207 (Invitrogen). Primers contained the attB1.1 and attB2.1 gateway recombination sites. All entry clones were sequence-verified and individually transferred by recombinational cloning into a prey vector (pPC86, Invitrogen) to be expressed as Gal4-AD (activating domain) fusion in yeast. Pairwise yeast two-hybrid interaction analyses were also performed by yeast mating using Y187 and AH109 yeasts strains (Clontech [68]), as described in [70]. Bait and prey strains were mated in an all-against-all array (together with negative controls, either empty bait vector or empty prey vector) and plated on a selective medium lacking histidine and supplemented with increasing concentrations of 3-AT (0, 5, 10, 15 mM) to test the interaction-dependent transactivation of HIS3 reporter gene. Interactions were scored as positive if observed in at least 2 out of 3 independent arrays. When yeasts containing an empty bait vector and a prey vector were still able to grow, the corresponding proteins were rejected as being auto-activators and thus false positives.
The R statistical environment was used to perform statistical analysis and the igraph R package to compute network topology measures [71].
Protein-protein interaction networks are formed by a set of N nodes (or vertices) representing proteins connected by E edges representing physical interactions between these proteins. The topology of protein-protein interaction networks can be described by a set of measures:
The degree or connectivity (k) of a node v in a graph is a local centrality measure which summarizes the number of edges that are incident to this node v.
The betweenness (b) of a node v in a graph is a global centrality measure which can be defined by the number of shortest paths going through this node v and is normalized by twice the total number of protein pairs in the graph . The equation used to compute betweenness centrality, b(v), for a node v is:where gij is the number of shortest paths going from node i to j, i and j ∈ V and gij(v) the number of shortest paths from i to j that pass through the node v.
DAVID database was used for functional annotation [72]. DAVID functional annotation chart tool was used to perform Gene Ontology categories analysis. Gene Ontology terms with a Benjamini-Hochberg corrected p-value smaller than 5.102 were considered as significantly overrepresented.
5 pmoles of each siRNA (stealth select RNAi, Invitrogen) were arrayed in 96 plates in 10 µl of OptiMEM (2 siRNAs per gene). After 20 minutes of room temperature incubation with a transfection agent (0.2 µl of lipofectamine RNAiMAX in 10 µl of OptiMEM), siRNA-transfection agent mix was added to 3.104 A549 suspension cells. Cells were incubated for 48 hours at 37°C and 5% CO2 before influenza A virus infection at MOI 0.5. At indicated time post-infection, supernatants were titered.
siRNA-transfected cells were washed twice with DMEM and infected with the A/H1N1/Puerto Rico/8/34 strain or the A/H1N1/New Caledonia/2006 strain at indicated MOI in infection medium (DMEM supplemented with 0.2 µg.ml−1 TPCK-trypsin (Sigma)). After 1 h at 37°C, the inoculum was discarded and cells were washed again and incubated in infection medium at 37°C and 5% CO2.
Standard fluorimetric assay was used to measure influenza virus neuraminidase activity [73]. Influenza virus neuraminidase is able to cleave the methyl-umbelliferyl-N-acetylneuraminic acid (4-MUNANA, Sigma) yielding a fluorescent product that can be quantified. In 96-black plate, 25 µl infection supernatants were diluted in 25 µl D-PBS containing calcium and magnesium and the reaction was started with 50 µl of 20 µM 4-MUNANA. After 1 h incubation at 37°C, the reaction was terminated by adding 100 µl of glycine 0.1 M, 25% ethanol pH 10.7. Fluorescence was recorded with TECAN infinite M1000 instrument at 365 nm excitation and 450 nm emission wavelengths.
ADAR1 was transferred from pDONR207 to pCIneo3×Flag (kind gift of Dr Y. Jacob, Pasteur Institute, Paris, France). NS1 and NS3 constructs were transferred in pDEST27 (Invitrogen). Plasmids coding for mutant NS1 (pCAGGS-NS1-R38AK41A) and control NS1 (pCAGGS-NS1) are kind gifts from A. Garcia-Sastre (Mount Sinai School of Medicine, New York). HEK293T cells were transfected in 6-well plates using JetPEI transfection reagent (Polyplus Transfection). 48 h post-transfection, cells were lysed in a cold extract buffer (20 mM Tris-HCl pH 8.0, 150 mM NaCl, 1 mM EDTA, 0.5% Igepal and a protease inhibitor cocktail (Roche)). Protein extracts (300 µg) were incubated overnight with Glutathione Sepharose 4B beads (GE Healthcare) at 4°C. Beads were then extensively washed with the cold extract buffer, proteins were separated by SDS-PAGE and transferred to a nitrocellulose membrane. GST-tagged viral proteins and 3×FLAG-tagged cellular proteins were detected using standard immunoblotting techniques with a mouse peroxidase-conjugated anti-GST monoclonal antibody (Sigma) or a mouse peroxidase-conjugated anti-FLAG M2 monoclonal antibody (Sigma). When indicated, pull-downs were treated with 2 µg of RNAse A (Invitrogen) in a buffer containing 100 mM NaCl for 30 min at 4°C. Proteins bound and released in the supernatants were then detected by immunoblotting using anti-ADAR (Sigma), anti-influenza A virus (Chemicon) and anti-NS1 antibodies (Abcam). Anti-actin antibody was purchased from Sigma.
HEK293T were transfected in 24-well plates with a total of 1 µg plasmid DNAs (editing reporter plasmid, 3XF-ADAR1 and plasmids coding for indicated viral proteins) using the JetPEI. 24 h post-transfection, cells were seeded in 96-well plates in DMEM and incubated for 24 h. The Dual-Glo Luciferase Assay System (Promega) was then added to measure both Firefly and Renilla luminescence activities using the TECAN infinite M1000 instrument. Relative Light Unit (RLU) is the ratio of luminescence from FLUC to luminescence from RLUC.
For influenza virus, HEK293T cells were seeded at 20,000 cells/well in 96-well plate and were transfected or not with anti-ADAR1 or control siRNAs, 24 h prior transfection with the editing reporter plasmid. 24 h post transfection, cells were infected influenza A virus at MOI 10 in DMEM supplemented with 10% FCS. 24 h post-infection cells were subjected to the procedure of editing assay described above. For dengue virus, Huh-7 cells were seeded at 3.105 cell/well in 12-well plates and were transfected or not with anti-ADAR1 or control siRNAs, 24 h day prior transfection with the editing reporter. 24 h later, cells were infected with dengue virus type 2 with an MOI of 20. Luciferase values were measured as described above.
A549 cells were infected with influenza A virus at MOI 3 in DMEM supplemented with 50 IU/ml penicillin, 50 µg/ml streptomycin and 0.25 µg/ml TPCK-trypsin. Eight hours post-infection cells were fixed with 4% formaldehyde for 30 min and permeabilized with 0.5% Triton X100. Double staining were performed by incubation with mouse monoclonal antibodies anti-NS1 (clone 1A7, kindly provided by Robert G. Webster) or anti-HA (Abcam) and rabbit anti-ADAR (Sigma) in combination with Alexa 488-labeled anti-rabbit F(ab)′2 fragment and Alexa 546-labeled anti-mouse F(ab)′2 fragment (Molecular Probes). Analyzes were performed with a laser-scanning confocal microscope (Axioplan LSM510 v3.2 (Zeiss)) and images were processed using LSM Image Browser (Zeiss).
|
10.1371/journal.pcbi.1003415 | Population Decoding in Rat Barrel Cortex: Optimizing the Linear Readout of Correlated Population Responses | Sensory information is encoded in the response of neuronal populations. How might this information be decoded by downstream neurons? Here we analyzed the responses of simultaneously recorded barrel cortex neurons to sinusoidal vibrations of varying amplitudes preceded by three adapting stimuli of 0, 6 and 12 µm in amplitude. Using the framework of signal detection theory, we quantified the performance of a linear decoder which sums the responses of neurons after applying an optimum set of weights. Optimum weights were found by the analytical solution that maximized the average signal-to-noise ratio based on Fisher linear discriminant analysis. This provided a biologically plausible decoder that took into account the neuronal variability, covariability, and signal correlations. The optimal decoder achieved consistent improvement in discrimination performance over simple pooling. Decorrelating neuronal responses by trial shuffling revealed that, unlike pooling, the performance of the optimal decoder was minimally affected by noise correlation. In the non-adapted state, noise correlation enhanced the performance of the optimal decoder for some populations. Under adaptation, however, noise correlation always degraded the performance of the optimal decoder. Nonetheless, sensory adaptation improved the performance of the optimal decoder mainly by increasing signal correlation more than noise correlation. Adaptation induced little systematic change in the relative direction of signal and noise. Thus, a decoder which was optimized under the non-adapted state generalized well across states of adaptation.
| In the natural environment, animals are constantly exposed to sensory stimulation. A key question in systems neuroscience is how attributes of a sensory stimulus can be “read out” from the activity of a population of brain cells. We chose to investigate this question in the whisker-mediated touch system of rats because of its well-established anatomy and exquisite functionality. The whisker system is one of the major channels through which rodents acquire sensory information about their surrounding environment. The response properties of brain cells dynamically adjust to the prevailing diet of sensory stimulation, a process termed sensory adaptation. Here, we applied a biologically plausible scheme whereby different brain cells contribute to sensory readout with different weights. We established the set of weights that provide the optimal readout under different states of adaptation. The results yield an upper bound for the efficiency of coding sensory information. We found that the ability to decode sensory information improves with adaptation. However, a readout mechanism that does not adjust to the state of adaptation can still perform remarkably well.
| A goal of systems neuroscience is to achieve a quantitative understanding of how cortical neurons report sensory events in their population activity. The interlaced synaptic architecture of neuronal networks provides anatomical evidence for population decoding by downstream neuronal structures. Such a synaptic organization allows an integration model in which the activity of neurons in the relevant population is summed with different weights. Under this model, discrimination of different stimuli can be formalized in terms of a linear classification of the neuronal responses. Here, we use a biologically plausible method of decoding: the model downstream neuron (the decoder) assigns a weight to each neuron before integrating the population activity (Figure 1A). The weight coefficient represents the synaptic strength between the input neuron and the decoder. This allows us to define an optimal linear decoder and establish its dependence on the adapted state of the network and its tolerance to correlated trial-to-trial covariability across neurons (noise correlation [1]–[4]).
In a recent study, we found that sensory adaptation improves coding efficiency of single neurons and the summed activity across neurons [5]. The present paper reanalyzes the same dataset with a focus on decoding. Investigating the behavior of the system under different adaptation states allows us to compare the performance of a non-adaptive decoder, which is optimal only under the non-adapted state, and an adaptive decoder, which adjusts to network dynamics and is thus optimal for any state of adaptation. In addition, by decoding simultaneously recorded single neurons, we quantify the influence of signal and noise correlations on the information available to downstream neurons.
All components of the experiment were conducted in accordance with international guidelines and were approved by the Animal Care and Ethics Committee at the University of New South Wales (ACEC 08/77B and 10/47B).
For the present study we reanalyzed the recorded neuronal data in [5]. A brief description of the recording method follows. Six adult male Wistar rats were used for acute recordings. Anesthesia was induced by intra-peritoneal administration of Urethane (1.5 gr/kg body weight). Neuronal activity was acquired using a 32-channel 4-shank multi-electrode probe (NeuroNexus Technologies, Ann Arbor, MI) from the barrel cortex. The stimulus train was composed of a 250 ms adaptation stimulus of 80 Hz sinusoidal vibration followed by a half-cycle (6.25 ms) pause and a single-cycle sinusoidal test stimulus (frequency of 80 Hz, 12.5 ms). We used 10 blocks at each of 3 adaptation amplitudes (0, 6 and 12 µm). Each block contained twelve test stimuli (amplitudes of 0 to 33 µm with equal increment steps of 3 µm) presented in a random order. Throughout a recording session each test stimulus was repeated 100 times under every adaptation state. Neuronal response to different stimulus amplitudes was characterized by counting the number of spikes generated in each trial over a 50 ms window post stimulus onset. Previous recordings from barrel cortex have revealed that most of the information about vibration stimuli is transmitted within this time window [6], [7]. In 6 male rats, a total of 73 single units and 86 multi-unit clusters were recorded across a total of 16 sessions (see Table 1 in [5]). Each session contained a distinct set of simultaneously recorded neurons that were isolated using an online amplitude threshold and an offline template-matching procedure.
To explore population decoding, we quantify the discriminability obtained from (i) the pooled activity of simultaneously recorded neurons (i.e. all spike counts summed together), and (ii) the population activity of neurons when they are integrated after applying an optimum set of weights. For a population of N neurons, the spike counts are represented as a data point in an N-dimensional space where every dimension corresponds to a neuron in the population. Each data point is then projected onto the given weight vector. Pooling gives equal weights to all neurons such that the weight vector lies along the identity line (Figure 1B, left panel). An optimum linear decoder assigns different weights to neurons based on an algorithm (detailed below) to provide maximal separation between the response distributions (Figure 1B, right panel). Once the weight vector is determined, population response histograms are calculated from the projection of data points onto the weight vector. The overlap between the two histograms is quantified by applying an ROC analysis considering all possible values of the decision criterion, ranging from the minimum to the maximum observed projection values (see left panel in Figure 1B). Each criterion yields a hit rate and false-alarm rate; plotting the hit rates versus the false alarm rates leads to an ROC curve (insets in Figure 1B). Here we use the area (denoted by A) between the ROC and the identity (non-discriminant) line. The area A is calculated by approximating the missing parts of the ROC curve between two consecutive criteria by a trapezoid. The value of A falls within the range of 0 to 0.5; A = 0 indicates that the hit rate is equal to the false alarm rate, reflecting complete overlap between two histograms, thus no discriminability. A = 0.5, on the other hand, indicates no overlap between the two histograms and thus perfect discriminability. The value of A takes into account the trial-by-trial variability in response and characterizes discrimination performance supported by the neuronal population. For the whole stimulus set, the overall discriminability was defined as the average value of A across all possible pairwise comparisons of stimuli (n = 66).
In order to identify the optimum weight vector for population decoding, we applied Fisher linear discriminant analysis [8]–[10] on the neuronal spike counts. For a population of N neurons, let the N×100 matrix denote the neuronal responses to stimulus s across 100 trials, and the N×N matrix denote the neuronal response covariance matrix for stimulus s. Let the N×1 vector be the average population responses to stimulus s across 100 trials. Here we calculate the optimal weight vector that yields maximum discrimination between stimuli. The N elements of the vector represent the weights applied to the response of individual neurons in the population. The optimal solution for the weight vector is obtained by maximizing the signal-to-noise ratio:(1)where represents the N×N signal covariance matrix, T denotes the transpose operator, the N×1 vector represents the average population responses across all stimuli (n = 12), and represents the overall neuronal trial-by-trial covariability. In Equation 1, the numerator is proportional to the population signal strength along the vector , while the denominator is proportional to the noise along the vector . The signal to noise ratio calculated in this way is invariant under scaling . Thus we can always find an optimal weight vector such that . The maximization problem in Equation 1 is a quasi-convex optimization problem [11] with the following Lagrangian function:(2)Applying the Karush–Kuhn–Tucker conditions [11] yields:(3)where represents the Lagrange multiplier corresponding to the equality constraint. Assuming is invertible, Equation 3 can be restated as(4)which is equivalent to eigenvalue decomposition of , where the optimal weight vector is along the eigenvector corresponding to the largest eigenvalue of [8], [11].
An upper bound on the performance of the linear discrimination can be achieved by finding the optimum set of weights for every pairwise stimulus discrimination. In this condition, for the particular stimulus pair and with average neuronal population responses and , and covariance matrices and , the signal-to-noise ratio along weight vector can be simplified to the following formula:(5)Solving for the optimal weight which maximizes the above equation by applying the same approach as in the problem formulated in Equation 1 yields [8]. This solution is identical to linear least square error estimation of the two classes [8], [9]. The overall discriminability (A) for the whole stimulus set was defined as the mean value of A across all possible stimulus pairs (n = 66). Throughout the paper, we refer to this upper bound as the pairwise-optimal decoder.
is not invertible when at least one of the recorded neurons does not fire any spikes in response to any stimuli. Calculation of the optimal weight vector is generalized to conditions when is singular, simply by removing the neurons with zero average spike count and then setting their corresponding weight to zero.
According to Equation 4, the solution for the optimal weight vector is the generalized eigenvalue decomposition of the signal covariance matrix , and the noise covariance matrix . The problem can be transformed into a subspace where is invertible, and hence the optimal weight vector is the first eigenvector of . However, as is not a symmetric matrix, other eigenvectors are not orthonormal. To quantify the level of tolerance of the decoder to changes in the weight vector direction, we need a symmetric representation of the effect of rotation in the space of neuronal activity with respect to the optimal direction. Thus we transpose the eigenvectors of to an orthogonal basis by rotating the eigenvectors according to the Gram–Schmidt procedure.
To characterize signal correlation in a population of more than two neurons, we applied principal component analysis (PCA) [12] on the z-scored neuronal average spike counts, similar to the quantification of noise correlation employed in [5]. For a population of N neurons, let the N×12 matrix denote the z-scored average neuronal responses to stimulus set averaged across 100 trials. Below we show that signal correlation can be represented by the largest normalized eigenvalue of the neuronal response correlation matrix . The strength of the signal correlation is proportional to the amount of stretch in the joint distribution of the average population responses. The first eigenvalue of the signal correlation matrix – denoted by – normalized to the sum of all eigenvalues specifies the maximum covariation in the average z-scored population responses relative to all dimensions forming the space of population activity. Thus normalized represents the stretch or skewness in the joint distribution of population responses, and hence identifies the signal correlation.
As the sum of all eigenvalues equals the sum of all diagonal elements of the signal correlation matrix , which is equal to N, the normalized can be re-expressed as:However, normalized has a positive constant bias which depends on the number of neurons in the population and the number of stimuli: When population size, N, is less than 12 (the number of stimuli), the maximum number of non-zero eigenvalues of signal correlation matrix is N, and hence the minimum value of normalized is . When population size is 12 or more the rank of signal correlation matrix is limited to 11. Thus the maximum number of non-zero eigenvalues of signal correlation matrix is 11, and hence the minimum value of normalized is . In order to provide a measure of signal correlation which is independent of population size or number of stimuli, we subtracted this bias from normalized and rescaled the result such that it falls between 0 and 1. We define this measure as the signal correlation index, denoted by SCI:(6)where . The signal correlation index depends solely on the correlation between the average responses of neurons. Similarly, the noise correlation index, denoted by NCI, is defined as [5]:(7)where is the ith greatest eigenvalue of the average noise correlation matrix across stimuli. For the special case of two neurons, signal and noise correlation indices are identical to the absolute value of the correlation coefficient between neuronal responses averaged across stimuli, and the correlation coefficient of trial-by-trial response variability, respectively [5].
The information that can be inferred from neuronal populations depends on the ‘readout mechanism’. A biologically plausible method of decoding applies a weight to each input neuron before integrating their response (Figure 1A). The weight coefficient represents the synaptic strength between the input neuron and the downstream decoder. A simple readout mechanism, called pooling, sums the activity of input neurons together with equal weights [13] (Figure 1B, left panel). At the other extreme, a decoder may only ‘read’ the activity of the most informative neuron in the population. This scheme, called the ‘lower envelope principle’ [14], [15], gives a weight of 1 to the best input neuron and a weight of zero to all other input neurons. Figure 1C compares the performance of these two decoding schemes applied to the neuronal responses to vibrotactile stimuli of different amplitudes (0 to 33 µm with equal increments of 3 µm), using the discriminability index, A. This index was averaged across all possible stimulus pairs (n = 66) in the non-adapted state. For some populations, pooling outperformed the best neuron, while in other populations pooling performance was not as good as the best neuron.
A third linear decoding scheme takes signal and noise correlations across neurons into account and finds the weights that optimize discriminability (Figure 1B, right panel) by maximizing the average signal-to-noise ratio (SNR). We will refer to this optimal linear decoder as the optimal decoder. Figure 1D quantifies pairwise stimulus discriminability across all possible populations of 8 simultaneously recorded single neurons in our dataset. The optimal decoder achieved a 96.8% improvement in discrimination performance over pooling, as quantified by the average value of A. In this decoding scheme, neurons with a higher SNR are expected to obtain a higher weight and thus make a greater contribution to decoding. To verify this, Figure 1E gives the distribution of weights as a function of SNR. As predicted, the decoder assigns weights of higher absolute value to the neurons with higher SNR.
Figure 2 generalizes the analysis to populations of various sizes. In Figure 2B a distinct set of weights were found for every stimulus pair, thus we refer to this decoder as the ‘pairwise-optimal decoder’. Pairwise-optimal decoding outperformed pooling with the effect becoming more pronounced at larger population sizes. In order to apply the appropriate set of weights, such a decoder requires a priori knowledge about the pair of stimuli to be discriminated. An arguably more biologically plausible decoding scheme is to apply an identical weight vector to discriminate across all stimulus pairs. By analogy with the pairwise-optimal decoder, we refer to this coding scheme as the ‘groupwise-optimal decoder’. Figure 2A provides a comparison of the two schemes. Figure 2C illustrates that the groupwise-optimal decoder outperformed pooling for every population size. Similar to the pairwise-optimal decoder, the improvement over pooling increases with population size. Across all population sizes, the groupwise-optimal decoder was superior to pooling by 54.8%±26.8% (mean ± s.d. across sessions). The rest of the analyses will focus on the groupwise-optimal decoding scheme.
How well does the decoder generalize to new trials? To address this question, we obtained the optimal weight vector from half of the trials (100 random selections of 50 out of 100 trials), and then applied the weights to the other half. In this analysis, we first focus on populations of 8 simultaneously recorded single neurons as a sample population size. On average, the discriminability on untrained trials was 95.7%±1.7% (mean ± s.d. across sessions) that on trained trials. This level of generalization was not specific to the population size of 8 single neurons. Across sessions, the performance of the decoder on untrained trials was 96.8%±2.5% and 96.7%±1.7% of that on trained trials for the whole set of single neurons ranging from 6 to 11 across sessions, and for the whole set of single- and multi-units in each session, respectively.
We further quantified the extent to which the decoder approaches the maximum achievable discriminability in terms of the value of A. In order to do this, we numerically calculated the weight vector which directly maximizes the value of A, using the pattern-search optimization method [16]. We compared the performance of this decoder, A-optimum, with that of the groupwise-optimal decoder. The weight vectors for both approaches were obtained from half of the trials (100 random selections of 50 out of 100 trials), and then were applied to the remaining half. Across sessions, the performance of the groupwise-optimal decoder was 99.1%±3.2% of A-optimum (for the whole set of single neurons) and 99.5%±1.3% (for the combined set of simultaneously recorded single- and multi-units).
To what extent does the decoder tolerate a change in the weight vector? We first examine the relative contribution of individual weights by setting the weight of one unit to zero while maintaining the weight of the other units in the population. This is equivalent to removing one unit from the population. Figure 3A depicts the relative decline in the performance of this decoder (the suboptimal decoder) as a function of the original population size. At the population size of 2, the relative decline in the performance of the suboptimal decoder was 31.0%±1.5% (mean ± s.e.m. across sessions), it reached 5% for the population size of 7 single neurons, and diminished as the population size further increased. The suboptimal decoder still outperformed pooling for the reduced population (Figure 3B). The difference between the performance of suboptimal and pooling increased with population size. We further compared the performance of the suboptimal decoder with the decoder optimized on the reduced set of units. On average, across sessions and population sizes, the performance of the suboptimal decoder was 99.4% that of the optimal decoder, with a minimum of 98.6%±0.3% (mean ± s.e.m. across sessions) observed at the reduced population size of 2.
In order to further quantify the extent to which the decoder tolerates a change in the weight vector, we gradually rotated the weight vector from the optimal direction towards the identity line (Figure 3C). Since decoding along the identity line corresponds to pooling, this analysis provides a characterization of the transition from optimal to pooling. Figure 3D illustrates the effect of rotating the weight vector from optimal direction towards the identity line. As the optimal direction is not perpendicular to the identity line, the 180° trajectory of rotation is not symmetric, but longer on one side (see Figure 3C). The consequence is a minimum in performance at an angle close to 90° (maximum deviation from optimal). For each curve, the two ends of the trajectory correspond to pooling, which in general is neither the best nor the worst decoding strategy.
Setting the identity line as the endpoint of rotation provides an intuitive link between optimal decoding and pooling. However, this represents a specific and rather arbitrary trajectory of rotation. To further characterize the tolerance of the decoder, we systematically rotated the weight vector away from the optimal towards all N-1 other dimensions in the N-dimensional space of population activity (Figure 3E). The optimal weight vector is the eigenvector of corresponding to the highest eigenvalue – where the separation between the population responses to the stimuli is maximal in the SNR space. Likewise, other dimensions correspond to the orthogonalized eigenvectors of (see Methods). The separation of the population responses to the stimuli is correspondingly higher along an eigenvector with a higher eigenvalue. Accordingly, we expect the decoding performance to drop less when the weight vector is rotated toward an eigenvector with a higher eigenvalue. Figure 3F characterizes decoding performance when the weight vector is rotated towards each of the 7 dimensions corresponding to other eigenvectors for a population size of 8 single neurons. Performance dropped by 34.0%±2.9% (mean ± s.d. across sessions) along the second most informative dimension and by 69.7%±6.9% along the least informative dimension. Across all dimensions, for a 30° deviation, we observed an average drop of 10.2%±1.8% (mean ± s.d.).
How does the trial-to-trial correlation in neuronal activity (i.e. the noise correlation) affect the performance of the decoder? To address this question, we first decorrelated the neuronal responses by shuffling the order of trials for every neuron in the population. Shuffling the trial orders eliminates neuronal response covariations while preserving the marginal distribution of population responses and the signal correlation. Thus, any observed effect of trial-shuffling is entirely due to noise correlations. We quantified the effect of noise correlation by ΔAshuffled denoting the percentage difference between the performance of the decoder optimized on the trial-shuffled responses and the performance of the decoder optimized on the true neuronal responses. Previous analysis [5] revealed that neuronal covariability is positive and thus detrimental to the information content of the pooled neuronal responses. Therefore, removing noise correlation is expected to enhance decoding performance. For pooling, removing noise correlation systematically improved decoding, as expected (Figure 4A). This effect increased with pool size, reaching an average improvement of 44.0% in the value of A across sessions in our dataset. However, for optimal decoding, removing noise correlation had no systematic effect, sometimes improving and sometimes impairing performance (the average change in the value of A across sessions was 0.9%, ranging from −9.6% to 12.5% for any population size in our dataset).
The immunity of the optimal decoder to the presence of noise correlation implies that the decoder has incorporated the structure of neuronal covariability. To directly test this idea, we implemented a simpler decoding scheme in which the covariance matrix was forced to be diagonal such that only signal correlation and the variability of individual neurons contributed to the optimization. This is equivalent to optimizing the decoder on the decorrelated population responses and then applying the resulting weights on the true population responses. We denote the value of A for this decoding scheme by Adiag. Figure 4B plots the proportional drop in the average value of A as a result of ignoring noise correlation, as denoted by ΔAdiag. ΔAdiag increased with population size, reaching 30% in our dataset. This finding reveals that the decoder successfully accounts for the noise correlation.
How does sensory adaptation affect the information content of neuronal populations? The original data set contained not only the non-adapted responses analyzed thus far, but also responses collected under two states of adaptation (vibration amplitudes of 6 and 12 µm). This allowed us to investigate how well the optimal linear readout performs under adaptation compared to the non-adapted state. The functional specialty of the whisker-barrel system and the structure of somatosensory cortex as a stand-alone processing stage in rodents [17] suggest that cortical neurons may have access to the network dynamics and the adaptation state. This information can be exploited to optimize the readout under different states of adaptation, leading to an ‘adaptive decoding scheme’.
Figure 5A–D quantify the discrimination performance of an adaptive optimal decoder (in terms of the average value of A) under different adaptation states. The average value of A for optimal decoder is higher under adaptation compared to the non-adapted state. This improvement is most prominent at intermediate A values and diminishes at low and high levels of discrimination performance (Figure 5A and B). These results extend the finding that sensory adaptation enhances coding efficiency from pooling [5] to optimal linear integration.
The enhanced discriminability demonstrated in Figure 5C and D is the average improvement across all pairwise stimulus discriminations (n = 66). To elucidate how sensory adaptation affects the coding efficiency for different stimuli, we quantified the adaptation-induced change in the value of A for individual stimulus pairs. As illustrated in Figure 5E and 5F, there is an elevated discriminability for stimuli higher in amplitude than the adaptor, while there is a decline in discriminability for stimuli lower than the adaptor. This pattern was consistent across sessions (correlations for all pairwise comparisons across sessions were significant, with all p values<0.008, and an average correlation coefficient of 0.68), as well as across both groupwise and pairwise optimal decoding schemes (correlation coefficient between average values across sessions: 0.97). The magnitude of the effect was larger for the groupwise optimal decoder compared with the pairwise decoder (linear regression coefficient of 1.14, significantly higher than 1 with a p value<0.05, regression R2>0.93). Additionally, the peak magnitudes of the decline and the enhancement were close: respectively, −0.23±0.05 (mean ± s.e.m. across sessions) vs. 0.20±0.04 for 6 µm adaptation, and −0.28±0.04 vs. 0.26±0.04 for 12 µm adaptation. These findings represent a shift in discriminability from low amplitudes to amplitudes higher than the adaptor [18], and are consistent with the lateral shift in the amplitude response function of the population [5]. For both adaptation states, the number of stimulus pairs for which discriminability increased was higher than the number of stimulus pairs for which it declined. This led to a net increase in the average value of A.
In order to understand the nature of this improvement in coding efficiency, we quantify the modulation of sensory adaptation on the two components of our optimization objective function (SNR): signal and noise correlations. We then parse out the contribution of each component (signal and noise) to the improvement in coding efficiency through adaptation.
What is the functional effect of noise correlation on the performance of the optimal decoder under different adaptation states? To address this question, Figure 6A and B illustrate ΔAshuffled for the two adapted states and compare it with the non-adapted state. The comparison reveals two main findings. First, the magnitude of the effect of noise correlation was greater in the adapted state. For instance, on average, across populations of 8 single neurons, noise correlation degraded decoding by 3.8%±3.7% (mean ± s.d. across sessions) in the non-adapted state, 8.9%±4.7% in the 6 µm adaptation state and 14.6%±5.1% in the 12 µm adaptation state. This finding is consistent with the results in our previous study that adaptation increased the overall noise correlation [5]. The second difference is in the functional role of noise correlation: contrary to the observation in the non-adapted state whereby noise correlation exhibited positive as well as negative effects on decoding efficiency (abscissa in Figure 6A and B), noise correlation was always detrimental to decoding under adaptation (ordinate in Figure 6A and B). This was in spite of the fact that the decoder was optimized on the adaptation data.
Based on this result, one might expect that ignoring noise correlation to be more detrimental to the performance of a decoder in the adapted state. However, this was not the case. Figure 6C and D illustrate the proportional drop in the value of A when ignoring noise correlation – as captured by ΔAdiag. The detrimental effect of ignoring noise correlation on decoding was less under adaptation. We explore two hypotheses to explain this discrepancy. Sensory adaptation might modulate the population responses in two ways: (i) increase in signal correlation and (ii) decrease in the angle between signal and noise direction. The following section quantifies signal and noise correlations for populations of any size.
What is the effect of sensory adaptation on the redundancy of neurons? As a measure of response redundancy, we quantified the correlation in the average responses to the stimuli, or signal correlation, under each adaptation state. A widely-used measure of signal correlation in the literature is the correlation coefficient between the response functions of two neurons [1], [2], [19]–[26]. However, the cross correlation analysis could not be applied to dimensions beyond two neurons. Therefore, we further scrutinized the correlations in the average response of multiple neurons with principal component analysis (PCA). In mathematical terms, the first eigenvector of the average neuronal spike-count correlation matrix identifies the direction of the greatest correlated variability (signal direction), and the first eigenvalue, denoted by , signifies the magnitude of that variability. The value of normalized to all eigenvalues quantifies the degree of the stretch in the population responses and thus the strength of signal correlation.
We first focus on sample populations of 8 simultaneously recorded single units. Figure 7A shows the 8 eigenvalues of the signal correlation matrix for the stimulus set across the five sessions that contained 8 single units or more. Normalized captures over 57.4% of the covariations in the average population responses to the stimuli in the non-adapted state. The first three eigenvalues represent 91.4% of stimulus-driven cross-neuronal response variability. This is a consequence of the similarity in the intrinsic response pattern of cortical neurons to stimulus intensity; a sigmoidal increase with stimulus intensity [7], [27]. Normalized was higher in the adaptation states compared to the non-adapted state. This finding supports the prediction that sensory adaptation increases signal correlation. This increase in signal correlation is achieved principally by alignment of neuronal response functions through a lateral shift in the amplitude response function of individual neurons [5]. Shuffling the labels of stimuli across neurons reduced signal correlation and essentially eliminated the difference between adaptation states (right panel in Figure 7A). This confirms that the adaptation-induced increase in normalized is not confounded by the sampling structure of neuronal responses, or the response variability of individual neurons in the population, but is a direct consequence of signal correlations across neurons in the population.
As signal correlation analysis captures correlations in the ‘average’ response of neurons across trials, it can also be applied to neurons that were not recorded simultaneously. Thus, we applied this analysis across all single-units (n = 73) in our dataset. Figure 7B represents signal correlation across various population sizes up to 73 single neurons. For this analysis, we calculated the signal correlation index – a rescaled version of normalized adjusted for population size (see Methods). The signal correlation index exhibits a constant relationship with population size signifying that this index is not biased by the number of neurons in the population. The value of signal correlation index was higher in the adapted states, revealing that sensory adaptation increased the homogeneity of cortical neuronal response functions.
We used the same method to quantify noise correlations for the simultaneously recorded units, as explained in detail in [5]. On average, adaptation increased the signal correlation index more than noise correlation index by factors of 4.6 and 3.0 (medians across sessions) for 6 µm and 12 µm adaptation states, respectively. This explains the observed improvement in the performance of the decoder with sensory adaptation. In addition, this result reveals why ignoring noise correlation is less detrimental under adaptation (see Figure 6). We also quantified the angle between the signal and noise direction under each adaptation state and observed no systematic changes across adaptation states. Likewise, across sessions, there was no systematic change in the first eigenvector of signal covariance matrix, with respect to the first eigenvector of the net noise covariance matrix, , over the three states of adaptation.
The performance of the groupwise-optimal decoder approaches its upper bound, pairwise-optimal decoder, when the neuronal responses to sensory stimuli are linearly correlated. This is equivalent to a maximal signal correlation. In this situation, provided that the noise direction is essentially invariant with stimulus, the direction of the optimal weight vector for every stimulus pair is identical, and lies along the groupwise-optimal weight vector. This indicates that the signal correlation can be captured as the difference in the performance of the pairwise and groupwise optimal decoding schemes. Figure 7C verifies this relationship by quantifying the correlation between the signal correlation index and the ratio of the groupwise- to pairwise-optimal decoding performance (Pearson correction coefficient of 0.94; p<0.0001). For over 99% of population sizes and adaptation cases, the correlation coefficient between signal correlation index and the ratio of the groupwise to pairwise optimal decoders' value of A was significant (p values<0.05).
The increased signal correlation through sensory adaptation leads to the following prediction: as a result of the increased homogeneity in neuronal response curves, under adaptation pooling is expected to be closer to the optimal decoding. We tested this prediction in the absence of noise correlation. Figure 8 summarizes different decoding schemes as a function of population size. For each population size, the neurons were selected randomly from all recording sessions. For those neurons in the population that were recorded simultaneously, if any, we shuffled the order of trials in order to eliminate the noise correlation. As predicted, under adaptation, pooling was closer to groupwise-optimal decoding. Adaptation enhanced the performance of all decoding schemes; however this improvement declined with population size (see insets in Figure 8). The improved decoding efficiency was most prominent for pooling.
To what extent does the non-adaptive decoder generalize across states of adaptation? Figure 9 addresses this question by using a fixed set of weights optimized in the non-adapted state. This figure compares the decoding performance of the non-adaptive decoder with the adaptive optimal decoder under each adaptation state. For this comparison, the performances were always quantified using untrained trials. To quantify the performance of the adaptive decoder, non-overlapping sets of training and test trials were obtained from the same adaptation state. For the non-adaptive decoder, the training trials were selected from the non-adapted state while the test trials were from the adapted state. For the 6 µm adaptation state, non-adaptive performance was 93.2%±13.0% (mean ± s.d. across sessions) that of the adaptive decoder for all possible populations of 8 single neurons (Figure 9A). For the 12 µm adaptation state, non-adaptive performance was 83.5%±19.8% that of the adaptive decoder (Figure 9B). In addition to populations of 8 single neurons, we further investigated the level of decoder generalization across adaptation states for the whole set of simultaneously recorded single neurons, as well as the whole set of single- and multi-units in each session. On average, across sessions, for the 6 µm adaptation state, non-adaptive performance was 94.7%±10.9% and 99.3%±2.5% that of the adaptive decoder, for each population set respectively (Figure 9C). For the 12 µm adaptation state, non-adaptive performance was 88.0%±17.4% and 91.5%±7.1% that of the adaptive decoder (Figure 9D).
This level of cross-adaptation generalization could either indicate that the performance of the adaptive decoder is relatively insensitive to changes in weights, or that adaptation does not strongly affect the optimal weights. To investigate this, we first quantified the sensitivity of the adaptive decoder to deviation of the weight vector from its optimal value in the adapted conditions. We systematically rotated the weight vector away from the optimal direction towards all N-1 other dimensions in the N-dimensional space of population activity (see Figure 3E). Figure 10A and B demonstrate the sensitivity of the adaptive decoder for a population size of 8 single neurons. For both adaptation conditions, the discriminability of the decoder consistently degraded with the angle of deviation. Consistent with the non-adapted condition (Figure 3F), the drop in the value of A was greater along the less informative dimensions compared with the more informative dimensions; performance dropped by 36.9%±11.1% and 46.0%±9.8% (mean ± s.d. across sessions) along the second most informative dimension and maximally dropped by 71.8%±11.8% and 64.9%±10.5% towards the least informative dimension for the 6 µm adaptation and 12 µm adaptation, respectively. Across all dimensions, for a 30° deviation, we observed an average drop of 4.7%±1.5% (mean ± s.d.) for the 6 µm adaptation, and a drop of 5.6%±1.3% for the 12 µm adaptation, which is less than the 10.2%±1.8% drop for a 30° deviation in the non-adapted case.
We further quantified the angular difference between the non-adaptive decoding weight vectors and the adaptive one for 6 µm and 12 µm adaptation states. The angular difference directly quantifies the effect of adaptation on the signal and noise directions, and its subsequent effect on the optimal weight vector. Figure 10C and D demonstrate the angular difference between the adaptive decoding weight vector and the non-adaptive one, in terms of the inverse cosine of their dot product. This measure is always positive, leading to a potential positive bias in the estimation of the average angular difference. To estimate this bias, we measured the angular difference between optimal weight vectors of the non-overlapping trial-halves within the adapted state. We then analyzed the correlation between the level of generalization (as in Figure 9) and the bias-subtracted angular differences across states of adaptation. The correlation analysis revealed an anti-correlation between the two measures (Pearson correlation coefficient: −0.6029, p = 0.0007); the higher the generalization across states of adaptation, the lower the changes in the weight vectors.
Here, we characterized the performance of a readout mechanism that linearly combines the responses of neurons in rat barrel cortex. The coefficients of this linear combination represent the synaptic weights between the barrel cortex neurons and the downstream neuron (decoder). We found the weights that maximized the average signal-to-noise ratio taking into consideration correlated variability across neurons. Such a decoder was less sensitive to noise correlations and adaptation state compared to a simple pooling method. In contrast to pooling, where noise correlation was always detrimental to the information content of the pooled population responses, for some populations noise correlation improved the optimal decoding performance. This motif is consistent with similar results of a recent study which quantified texture discrimination accuracy of cortical population responses in awake rats just prior to behavioral responses [28]. Under the optimal coding scheme, the response of less informative neurons could be exploited to provide information about the network state and the structure of noise correlations. We found that adaptation increased noise correlation [5], leading to a greater effect of noise correlation on decoding than in the non-adapted state (Figure 6A and B). Ignoring noise correlation led to a decline in the decoder performance (Figure 4B). This decline increased with population size. Although noise correlation was stronger under adaptation, ignoring it during decoding was less detrimental to the decoding efficiency. This was mainly due to a greater increase in signal correlation through sensory adaptation.
In the present study, we characterized the pairwise discrimination performance using a criterion-free metric, A, in the framework of signal detection theory. Fisher information between neuronal responses and stimuli provides an alternative measure of discriminability [29]–[33]. Fisher information averaged across stimuli is proportional to the value of A when averaged across stimulus pairs with minimum difference (3 µm in our study) – see [33], [34]. Furthermore, we found that the optimal decoding scheme that maximized the signal-to-noise ratio (and not directly the value of A), did identify the maximal value of A (see Figure 3F).
We employed two parallel methods in order to quantify the effect of noise correlation on the information content of the population responses; (1) the effect of trial shuffling on discriminability index, as captured by ΔAshuffled, directly quantifies the effect of noise correlation on coding efficiency, and (2) ΔAdiag quantifies the cost of ignoring noise correlation. These measures are analogous to information theoretic measures such as ΔIshuffled [5], [35]–[37] and Icor-dep [38]–[43], as well as other measures based on signal detection theory such as Δd2shuffled and Δd2diag [44]–[46].
Along the lemniscal pathway, there is a greater than 10 fold increase in the number of neurons representing a whisker from brainstem to cortex; from 160–200 neurons per barrelette [47] and 250–300 neurons per barreloid [48]–[50] to about 2500 cortical neurons per layer IV barrel [51], [52]. One explanation for this increase might be the need to represent multiple features (e.g. a broad range of speeds of whisker motion). For example, the broad range of perceptually discriminable whisker motions [7], [53] can be broken down into narrower ranges. Each of these narrowed ranges of whisker motion intensities could then be represented by a subpopulation of neurons sensitive to that range. The weights of neurons for these combinations could be optimized using the solution applied in the present study. Further experiments are required to investigate the mechanism through which such optimal synaptic weights could potentially be developed across multiple subpopulations.
An important question is whether the readout mechanism adjusts to changes in neuronal response dynamics. This question is not limited to sensory adaptation. In addition to adaptation (temporal context), spatial context can also modulate the response properties of neurons [54], [55] and produce similar perceptual biases and illusions [32]. Likewise, attention also changes the tuning properties of neurons [56]–[59] and induces perceptual illusions [60], [61]. The match between perceptual predictions based on a non-adaptive decoder and psychophysical measures of perceptual biases and thresholds in the visual system is consistent with a fixed non-adaptive readout [32], [33]. However, several attributes of an adaptive readout could potentially produce similar perceptual biases [33]. In addition, cortical neurons may be able to provide information about network dynamics and adaptation state to downstream structures. Further experiments are required to quantify the psychophysical effect of sensory adaptation in the whisker-mediated touch system in rodents.
Here, we observed a remarkable cross-adaptation generalization. In isolation, this could either indicate that the decoding performance is relatively insensitive to changes in weights or that adaptation does not strongly affect the optimal weights. Given the dependence of the decoding performance on the changes in weights as revealed in Figure 10A and B, we conclude that the optimal weights remain relatively unchanged after adaptation. These results can be understood in terms of the changes in the response function of cortical neurons through sensory adaptation. Sensory adaptation shifted the response function and response variability profile of cortical neurons with no systematic modulation on the response saturation level [5]. Thus the set of weights, which maximize discriminability between a pair of stimulus amplitudes in the non-adapted state, are expected to maximize discriminability between a new pair of stimulus amplitudes that are in effect simply shifted by the adaptor.
Our previous study showed that sensory adaptation increases noise correlation across neurons [5]. This increase in noise correlation tends to decrease the overall signal-to-noise ratio. The marked level of cross-adaptation generalization indicates that signal correlation across neurons increases with sensory adaptation as well. This increase in signal correlation can be explained in terms of the adaptation-induced lateral shift in the response of single neurons. In the non-adapted state, neurons exhibit various sensitivity thresholds. However, sensory adaptation tends to equalize the threshold of neurons by aligning their response functions with respect to the adapting stimulus amplitude [5]. This response alignment homogenizes the population of neurons, leading to increased signal correlation.
Here, decoding was performed along the first eigenvector of the . The decoding scheme can however be expanded to other eigenvectors of . As these eigenvectors are not orthonormal (see Methods), the information along them is correlated, leading to redundant population coding. An interesting question is how sensory adaptation changes the direction of these eigenvectors and the amount of information along them in a multi-dimensional feature space of sensory stimuli. If through sensory adaptation the eigenvectors rotate away from each other to form a more orthogonal basis, the information extracted from them is less correlated, leading to an adaptive decorrelated representation of sensory features along these eigenvectors [62], [63].
In the present study, and also in previous relevant studies [28], [36], [46], [64]–[72] the decoder is commonly optimized to maximize the discriminability or minimize the estimation error. However, a behaviorally-relevant question is “which readout mechanism matches the perceptual accuracy of subjects?” To address this question, the optimization objective function should be set to a behavioral measure such as choice probability [73]. Investigating such a perceptually-matched decoder under different temporal (adaptation), spatial or attentional contexts would reveal the extent to which the readout adjusts to context-induced changes in neuronal response dynamics.
|
10.1371/journal.pcbi.1004207 | Thermal Stabilization of Dihydrofolate Reductase Using Monte Carlo Unfolding Simulations and Its Functional Consequences | Design of proteins with desired thermal properties is important for scientific and biotechnological applications. Here we developed a theoretical approach to predict the effect of mutations on protein stability from non-equilibrium unfolding simulations. We establish a relative measure based on apparent simulated melting temperatures that is independent of simulation length and, under certain assumptions, proportional to equilibrium stability, and we justify this theoretical development with extensive simulations and experimental data. Using our new method based on all-atom Monte-Carlo unfolding simulations, we carried out a saturating mutagenesis of Dihydrofolate Reductase (DHFR), a key target of antibiotics and chemotherapeutic drugs. The method predicted more than 500 stabilizing mutations, several of which were selected for detailed computational and experimental analysis. We find a highly significant correlation of r = 0.65–0.68 between predicted and experimentally determined melting temperatures and unfolding denaturant concentrations for WT DHFR and 42 mutants. The correlation between energy of the native state and experimental denaturation temperature was much weaker, indicating the important role of entropy in protein stability. The most stabilizing point mutation was D27F, which is located in the active site of the protein, rendering it inactive. However for the rest of mutations outside of the active site we observed a weak yet statistically significant positive correlation between thermal stability and catalytic activity indicating the lack of a stability-activity tradeoff for DHFR. By combining stabilizing mutations predicted by our method, we created a highly stable catalytically active E. coli DHFR mutant with measured denaturation temperature 7.2°C higher than WT. Prediction results for DHFR and several other proteins indicate that computational approaches based on unfolding simulations are useful as a general technique to discover stabilizing mutations.
| All-atom molecular simulations have provided valuable insight into the workings of molecular machines and the folding and unfolding of proteins. However, commonly employed molecular dynamics simulations suffer from a limitation in accessible time scale, making it difficult to model large-scale unfolding events in a realistic amount of simulation time without employing unrealistically high temperatures. Here, we describe a rapid all-atom Monte Carlo simulation approach to simulate unfolding of the essential bacterial enzyme Dihydrofolate Reductase (DHFR) and all possible single point-mutants. We use these simulations to predict which mutants will be more thermodynamically stable (i.e., reside more often in the native folded state vs. the unfolded state) than the wild-type protein, and we confirm our predictions experimentally, creating several highly stable and catalytically active mutants. Thermally stable active engineered proteins can be used as a starting point in directed evolution experiments to evolve new functions on the background of this additional “reservoir of stability.” The stabilized enzyme may be able to accumulate a greater number of destabilizing yet functionally important mutations before unfolding, protease digestion, and aggregation abolish its activity.
| Protein stability is an important determinant of organismal fitness and is central to the process of enzyme design for industrial applications [1–3]. Most proteins must be folded to carry out their functions in vitro or in vivo. In addition, non-functional aggregation of unfolded or partially-unfolded proteins can have a deleterious effect on the fitness of an organism and can lead to protein aggregation diseases, which include Alzheimer’s and Huntington’s, in humans [4–6]. Aggregation of poorly folded proteins can also hamper protein production for research and technological purposes [7].
While most mutations in a natural protein are destabilizing [8,9], biological proteins are not generally at their highest possible stability; some mutations will stabilize a protein, increasing the equilibrium population of the folded state [10–12]. This stabilization can be achieved by either slowing the rate of unfolding or speeding the rate of folding, depending on the role of the mutated residue in the folding nucleation process [13,14]. The unfolding temperature, Tm, at which the free energy of the folded and unfolded states coincide (ΔG = 0) serves as a common measure of protein stability. Tm is obtainable by experiment and, in theory, from simulation, although current molecular dynamics simulations are limited in their ability to capture full folding or unfolding trajectories of most proteins (except very small fast folding domains [15]) in a tractable amount of simulation time [16].
Several computational methods to predict protein stability or changes in stability upon mutation have been developed and tested [17–19]. However, the performance of these popular methods is still relatively weak [20–22]. Other existing techniques to rationally design proteins with improved stability have involved optimization of charge-charge interactions [23], saturation mutagenesis of residues with high crystallographic B-factors [24], methods based on protein simulation and calculation of free energies [25–27] and comparison to homologous proteins including the ultra-stable proteins of thermophiles [28,29]. We reasoned that better predictions of mutant stability might be obtained by evaluating the unfolding temperature Tm in realistic yet computationally tractable simulations of protein unfolding.
Here, we use a Monte Carlo protein unfolding approach (MCPU) with an all-atom simulation method and knowledge-based potential developed earlier in our lab [16,30,31] to simulate unfolding and predict melting temperatures for all possible single point mutants of E. coli Dihydrofolate Reductase (DHFR). DHFR is an essential enzyme in bacteria and higher organisms, and it is an important target of antibiotics [32] and anti-cancer drugs [33,34]. Its moderate size (18 kDa) makes it amenable to both simulation and experiment. As described in the Materials and Methods section, the Monte Carlo move set consists of rotations about torsional angles. At high temperature, the higher entropy of unfolded states overcomes the increase in energy due to loss of favorable contacts and torsional preferences, leading to unfolding. We experimentally determine melting temperatures and catalytic activities for several predicted stabilizing mutants, and for mutants combining multiple stabilizing mutations. Our approach allows us to identify several stabilized mutants of DHFR, and our prediction method marks an improvement over existing stability predictors such as Eris [19], FoldX [17], and PopMusic [18]. Simulations of non-DHFR proteins likewise indicated that our method is useful as a general approach to simulate protein unfolding and select stabilizing mutations.
Ideally, protein stability for any sequence should be predicted in all-atom equilibrium simulations that cover multiple folding-unfolding events to determine equilibrium populations of various states of the protein. However, despite recent progress in ab initio simulations of protein folding [15] this goal is not attainable for proteins of realistic size and biological relevance. Currently, non-equilibrium unfolding simulations are within reach for sufficiently large proteins and the question arises whether such simulations can be used to assess mutational effects on protein stability, which is an equilibrium property. The following analysis provides an affirmative answer to this question, under certain assumptions. Although the idea of obtaining equilibrium free energy differences from non-equilibrium measurements is not new [35], and protein stabilities have been calculated from molecular dynamics simulations using the Jarzynski equality, e.g., [36–38], such simulations require application of an external steering force; in the present paper we report the use of multi-temperature Monte-Carlo unfolding simulations in obtaining protein stabilities.
Assuming two-state unfolding kinetics [39–42] we can estimate the characteristic time required to cross the unfolding free energy barrier (in fact it is the time spent in the native state waiting for sufficient thermal fluctuation to cross the barrier) as:
τufp=τ0eΔG#kT
(1)
where τufp is first-passage time from the folded to the unfolded state, ΔG# is the free energy barrier between the folded state and the transition state for unfolding (see Fig. 1) and τ0 is the elementary time constant. When simulation time τsim approaches τufp unfolding events are observed in simulation. The apparent “melting temperature”, i.e., the temperature at which unfolding events occur in simulations, therefore depends on the simulation time τsim according to Eq. (1):
kTmapp=ΔG#ln(τsimτ0)
(2)
This analysis suggests that non-equilibrium first passage unfolding simulations are not suitable to predict the temperature at which a protein would unfold at equilibrium. However the effect of mutations on stability can be predicted from unfolding simulations. In order to see this we note that the mutational effect on protein stability ΔΔG is related to the change in the unfolding free energy barrier ΔΔG#, the difference between the WT barrier height and the mutant barrier height, shown in Fig. 1.
ΔΔGi#=(1−φi)ΔΔGieq
(3)
where i denotes the mutated amino acid and φi is the φ-value for residue i which determines the fraction of interactions that this residue forms in the folding/unfolding transition state [40,43,44]. We therefore obtain
kΔTmapp(i)=(1−φi)ΔΔGieqln(τsimτ0)
(4)
where ΔTmapp(i)=Tmapp(i)−Tmapp(WT) is the shift in apparent unfolding temperature upon a specific mutation in the i-th residue. Introducing the relative (to WT) unfolding temperature ΔTmrel(i)=ΔTmapp(i)/Tmapp(WT) we get
ΔTmrel(i)=(1−φi)ΔΔGieqΔG#
(5)
i.e. the mutational shift in observed unfolding temperature, normalized to the observed unfolding temperature of the wild-type at the same simulation condition does not depend on the simulation length, provided that the simulation is sufficiently equilibrated in the native basin so that the rules of transition state theory apply. The analysis of extensive kinetic and equilibrium data for multiple proteins shows that for the majority of mutations (except for a small fraction of residues that participate in the folding nucleus) φi ≈ 0.24 with remarkable accuracy and consistency [45]. We get therefore
ΔTmrel(i)=0.76ΔΔGieqΔG#
(6)
i.e. ΔTmrel(i) is independent of simulation time and proportional to the equilibrium free energy effect of mutations, provided that simulations have equilibrated in the native basin of attraction.
We ran MCPU on DHFR (PDB ID: 4DFR) at a range of temperatures, to generate simulated unfolding curves. Unfolding steps of a sample trajectory are shown in Fig. 2, and a flowchart of the simulation method is shown in S1 Fig. The protein was subject to a brief MD energy minimization, beginning from the WT crystallographic native state, followed by unfolding simulations at each of 32 different temperatures using all-atom Monte-Carlo (see Materials and Methods section). As shown in figures S2 Fig—S4 Fig, the RMSD and total energy increased and the number of contacts decreased as each simulation proceeded, and with increasing temperature. (Here, temperature is given in arbitrary simulation units.) Plots of RMSD and contact number vs. temperature showed sigmoidal behavior, with a clearly identifiable transition temperature, and the melting temperature (Tm) could be determined by fitting to a sigmoidal function (Fig. 3). Plots of energy vs. temperature (S5 Fig) were sigmoid-like, but with an additional rise in energy at low to intermediate temperatures, perhaps indicating pre-melting to a dry-molten globule state with loosened side chains but native-like topology [46,47]. This deviation from sigmoidal behavior becomes clearer as the simulation length is increased (S6 Fig).
All possible single point mutations of DHFR (159 * 19 = 3,021) were simulated with the Monte Carlo protein unfolding simulation protocol. The simulated Tm values were calculated as described above. Of the 3,021 mutations, 523 mutations (17.3%) were predicted to have a stabilizing effect according to all three metrics (energy, contacts, and RMSD), while 42.1% of mutations had a destabilizing effect according to all three metrics. These predictions are in good agreement with statistical analysis of published experimental data and FoldX predictions [8,12]. The simulated Tm values evaluated using RMSD, total energy, and number of contacts are strongly correlated, as shown in Fig. 4A. The distribution of predicted melting temperatures (averaged over the 3 metrics) for all 3021 point mutants is shown in Fig. 4B. Next, we selected a subset of predicted stabilizing mutations for subsequent in depth computational and experimental analysis. To that end we selected the loci where multiple mutations were consistently predicted as stabilizing. Out of this set we selected one mutation at each loci which were predicted as most stabilizing. As a result we arrived at 23 single predicted stabilizing point mutants shown in S1 Table, which we deemed most promising for subsequent in depth computational and experimental analysis. Furthermore, five stabilizing mutations at different sites within DHFR, shown in Fig. 5, were combined to form the multiple mutants listed in Table 1, with the rationale that the combination of individual stabilizing mutants often yields more stable proteins, and these mutants were likewise subjected to computational and experimental analysis.
First we test two predictions that emerge from the theoretical analysis of unfolding simulations. The first prediction is that the apparent unfolding temperature decreases as the length of the unfolding simulation increases (Equation 4). Secondly and most importantly the mutational change in relative (normalized to WT) apparent unfolding temperature is a) robust with respect to simulation time provided that simulations have equilibrated in the native basin and b) directly proportional to the effect of mutations on equilibrium protein stability (Equation 6). We test these predictions using MCPU simulations and experiment.
We carried out two sets of MCPU simulations of different lengths: 2,000,000 and 20,000,000 steps for the 23 predicted stabilizing mutants, 15 mutants studied previously by experiment [48] (the complete set of single mutants is listed in S1 Table), and the 5 stabilizing multiple mutants combining individual mutations listed in Table 1, and compared their predicted absolute and relative simulated unfolding temperatures (Fig. 6). Indeed both predictions of our theoretical analysis are confirmed, i.e., the apparent unfolding temperature decreases with simulation time (Fig. 6A) while the relative unfolding temperature ΔTmrel is remarkably independent of simulation time (Fig. 6B). We note that due to the nature of the energy function used in our simulations, there is no obvious mapping of simulation temperature to real absolute temperature (i.e., in Celsius or Kelvin). Conversion of simulation temperature to physical temperature would require use of experimental data (e.g., WT unfolding temperature and deviation of temperatures over all mutants) and therefore would not provide a completely simulation- or theory-based prediction. Furthermore, as noted above, the apparent absolute value of the transition temperature in the Monte-Carlo unfolding approach depends on simulation time. Therefore, we used relative melting temperature, ΔTmrel(i)=ΔTmapp(i)/Tmapp(WT), when comparing simulation results with experimental results.
As mentioned, the simulated Tm values evaluated using RMSD, total energy, and number of contacts are strongly correlated in our simulations as shown in Fig. 4A and S1 Table. In what follows we define the computational unfolding temperature Tm as averaged over Tm values determined using these three criteria.
We cloned, expressed, and purified the 23 single point mutants of DHFR listed in S1 Table, as well as the multiple mutants listed in Table 1 (see Materials and Methods). The biophysical properties of the mutants were measured and compared with WT DHFR, as shown in S2 Table. As many studies have shown that oligomerization can alter protein stability [23,48,49], we first tested whether mutations induce oligomerization and/or aggregation using the gel filtration method [48,50] and light scattering. The results indicated that all of the 23 mutants were monomeric at studied concentrations except for E154V, which appeared aggregation-prone. We excluded E154V from the subsequent analysis.
As shown in S2 Table, all single mutants are catalytically active except for D27F. D27 is known to be a key catalytic residue of E. coli DHFR [51].
For each mutant we obtained two measures of stability: the apparent melting temperature determined by Differential Scanning Calorimetry (DSC) and the urea midpoint unfolding concentration (Cm) determined by monitoring chemical denaturation by Circular Dichroism (CD) with subsequent fitting to a two-state model (see Materials and Methods). Both measures of stability were highly correlated, despite the fact that thermal unfolding was irreversible (S7 Fig). Of the selected 22 single point mutations, 10 mutations were stabilizing, according to their Tm or Cm values (S2 Table). Given that statistically most random mutations are destabilizing with only a small fraction (less than 18%) stabilizing [8,12], this statistically significant result (p = 0.002 under the null hypothesis that mutations are random) indicates that MCPU is an effective method for selecting stability-enhancing mutations.
As expected, combinations of single stabilizing mutations led to more stable multiple mutant variants, [24,25,52] as predicted by simulation. In particular, the stability of the quintuple mutant (T68N,Q108D,T113V,E120P,S138Y) was found to be substantially higher than that of the wild type protein (Table 1), with Tm 7.2°C higher than WT, and Cm, the urea concentration at the mid-unfolding point, was 0.43M higher than WT. All multiple mutants were catalytically active, and the quintuple mutant and triple mutant (T113V,E120P,S138Y) were found to be more catalytically active than WT. We note that while combination of stabilizing mutations generally increases stability, the effect is less than additive (S8 Fig); for instance, the quintuple mutant is about 4°C less stable than predicted under the assumption of additive ΔTm (a 7.2°C stability increase vs. predicted 9.6°C).
We computationally predicted relative unfolding temperatures of 15 DHFR mutants published earlier [48] and added these mutants to the set for analysis resulting in 42 mutants in total. The correlation coefficient between experimental relative Tm and simulated relative Tm for the 42 mutants was about 0.65, as shown in Fig. 7A. To address the issue that both simulated Tm and DSC measurements are not strictly at equilibrium, we plotted the relation between simulated Tm and equilibrium measurement of stability in chemical denaturation by urea. The denaturation mid-transition urea concentration Cm and computationally determined unfolding temperature exhibit even a slightly higher correlation of r = 0.68 (Fig. 7B), demonstrating that our non-equilibrium simulation method shows good agreement with the equilibrium measurement of urea denaturation, as predicted by Equation 6.
We also used the dataset to evaluate the effect of the number of replications and the number of MC steps on the performance of the method. As shown in Fig. 8A, the prediction accuracy is sensitive to the number of replications. To achieve reliable Tm predictions, at least 20 replications should be used. However, the number of MC steps did not greatly affect prediction accuracy, provided simulations were run for at least ~ 200,000 steps (see Fig. 8B). In the context of the theory developed in the earlier section: “Predicting the effects of mutations on protein stability from non-equilibrium unfolding simulations,” this initial equilibration period may allow time for equilibration within the native basin, after which simulation length does not appreciably affect the consistency of results with equilibrium stability measurements.
It has been proposed that stability imposes a constraint on protein function leading to stability-activity tradeoffs [53,54]. Our data, however, paints a different picture for DHFR—of a weak positive correlation between Tm and kcat or kcat/KM (r = 0.46, p = 0.02 and r = 0.41, p = 0.03 respectively) with one notable outlier D27F, where the stabilizing mutation is made right in the active site (Fig. 9). The D27F mutant has high thermal stability but, as noted above, is not catalytically active, indicating that there is in fact a stability-activity trade-off for this active-site residue.
Using an alignment of 290 bacterial DHFRs, we determined the DHFR consensus sequence (S9 Fig). Mutation of a non-consensus residue to the consensus residue generally resulted in protein stabilization [29]. In 4/16 of the experimentally stabilizing mutations, a residue was changed to the consensus residue, while only 2/29 destabilizing mutations resulted from a change to consensus. Likewise, in 18/29 destabilizing mutations, a residue was changed away from the consensus residue, while this was true for only 5/16 of stabilizing mutations.
We compared the minimum and maximum simulated Tm values obtainable by mutating a single residue to any of the 19 other amino acids (Fig. 10A). There is a weak positive correlation between minimum and maximum melting temperatures (r = 0.30, p = 10−4). Apparently, protein loci where mutations can cause significant stabilization are statistically less susceptible to destabilizing mutations and vice versa, which may be expected: once a residue is already at its most stabilizing amino acid variant, the protein cannot be stabilized further by mutation. Distinct outliers correspond to the loci with the strongest stabilizing or destabilizing effects of mutations. Interestingly, these outliers, which may represent structural weak spots in DHFR, tend to fall on the interface connecting the C-terminal beta hairpin and the rest of the protein (Fig. 10B). This is in fact the interface that is the first to dissociate in the Monte Carlo simulations (see Fig. 2).
We compared our computational DHFR predictions with four popular approaches to predict the effect of a mutation on protein stability: FoldX [17], Eris [26], PopMusic [55], and SDM [56]. (S3 Table). The MCPU performs better than these methods on DHFR mutants. PopMusic shows also strong performance with highly statistically significant r = 0.55 between theory and experiment, however the limitation of this method is that it can consider only single point mutations. To further evaluate MCPU performance we tested it on four additional proteins from four different SCOP structural classes: the Cro repressor protein from bacteriophage lambda (PDB-ID 5CRO), the B. Subtilis major cold shock protein (1CSP), E. coli Thioredoxin (2TRX), and Gln-25 ribonuclease T1 from Aspergillus oryzae (1RN1). Our predictions were compared with Eris and SDM. We did not compare MCPU results with FoldX and PopMusic as these mutations were selected in the training dataset for the two methods. As shown in Table 2, the correlation coefficient between MCPU predictions and the experimental Tm values, averaged over all proteins, is about 0.71, which is higher than that provided by Eris (-0.05), for which predictions were quite poor for both DHFR and other proteins, and SDM (0.63). If we consider only the binary prediction of whether a mutation is stabilizing or destabilizing, MCPU can correctly predict 11 out of 16 mutations, while Eris and SDM correctly classify 9 and 8 mutations respectively.
The theoretical analysis of the unfolding simulations relates the effect of mutations on the equilibrium between folded and unfolded states to the effect of mutations on free energy of the folded and transition states. It is widely believed that in the low-entropy folded state energetic factors dominate. If so that would imply that we can get an equally good correlation between prediction and experiment by estimating the mutational effect on energy of the native state as is the case for most empirical methods. To that end we evaluated the correlation between the energy of the minimized (after long MC equilibration) native state and the experimental Tm and found only a weak correlation with experimental melting temperatures (Table 2, last column), indicating that protein entropy, which is accounted for in the MCPU, in addition to enthalpy, is important in determining protein stability.
Estimates of protein stability using Molecular Dynamics are prohibitive for all but the smallest protein domains. However using MCPU we were able to efficiently explore stabilities of all possible point mutants for an essential enzyme of a typical size (159 amino acids) in a manageable amount of computational time (approx. one hour for every 1,000,000 MC steps). Although the use of rapid Monte Carlo simulations reduces simulation time and allows for a greater number of replicates, our method to predict stability effects of mutations based on non-equilibrium unfolding simulations represents a general approach that could be modified for use with conventional MD simulations, especially given the current rate of improvement in simulation speed and accuracy.
Since our method involves protein unfolding simulations and not equilibrium simulations of both folding and unfolding processes, we expect it to be especially useful for predicting mutations that mostly affect the rate of protein unfolding as highlighted in our theoretical analysis. Low φ-value residues, which acquire contacts with other residues late in the folding process and lose contacts early in the unfolding process [14] constitute the majority of residues in proteins, with φ-value roughly constant around 0.24 as noted in [45]. Combining this observation with assumptions of transition state theory, we found that for the majority of residues (those not part of the folding nucleus [14,57] exhibiting anomalously high φ-values) the observed simulation Tm relative to WT is proportional to the equilibrium stability change ΔΔG, as verified by simulation and experiment. We establish that relative Tm is independent of simulation length, demonstrating that non-equilibrium simulations can in fact be used to quantify relative protein stability.
Many of the experimentally verified stabilizing mutations in DHFR predicted by MCPU are found in the C-terminal beta hairpin region, which is the first to unfold in simulations, prior to the main unfolding event encompassing the entire structure (see Fig. 2). It has been shown that the source of ultra-stability in hyperthermophiles generally arises from slowing the unfolding rate, rather than increasing the folding rate [28], so our method may be particularly suitable for discovering biologically relevant stabilizing mutations. In addition, our results might be particularly applicable to in vivo studies, where protease digestion and/or aggregation proceed from the partially-unfolded state. We note, however, that some stabilizing residues predicted by MCPU lie in the region of the protein that is late to unfold in simulations, including I61V, which raises the experimental melting temperature by 1.7°C. These mutants, along with the destabilized outlier I155A for which relative Tm depends on simulation length (Fig. 6), are appealing candidates for further study, as they may reflect a breakdown in the simplifying assumptions of 2-state kinetic theory for proteins.
It has been hypothesized that there exists a tradeoff between enzyme activity and stability, since certain regions of an enzyme must be sufficiently flexible to promote catalysis [53,54]. This conclusion was reached in [53,58], based on the exploration of stability effects of mutations in the active site of beta-lactamase [53] and rubisco [58]. Fersht and coauthors also found several stabilizing mutations in the active site of Barnase rendering the protein inactive [59]. While we observe a similar effect with the D27F mutation in DHFR, Fig. 9 shows that exploring only mutations in the active site provides a biased view on the tradeoff between activity and stability. Rather a vast majority of mutations throughout the protein show a qualitatively opposite trend. The likely explanation of the distinction between an apparent tradeoff when mutations are made in the active site and the opposite trend for mutations outside of the active site is that “carving” an active site requires special selection of catalytic amino acids, which could indeed have a destabilizing effect, overall. However our observation of a small positive correlation argues against an obligate relation between global protein dynamics and activity for DHFR, at least for the aspects of dynamics that are correlated with stability. Warshel and colleagues reached a similar conclusion in their theoretical analysis of the role of dynamics for DHFR and other proteins in [60]. This point has likewise been made by Bloom et al. [11], who noted that a number of proteins have been stabilized experimentally without loss of activity, and Taverna and Goldstein argued that marginal stability is an inherent property of proteins due to the high dimensionality of sequence space and not due to a requirement of reduced stability in order to generate sufficient flexibility [61].
A straightforward explanation for the weak yet statistically significant positive correlation between activity and stability observed in our case might be that more stable proteins have greater effective concentration of the folded (i.e. active) form. It is also important to note that a weak yet statistically significant positive correlation between activity and stability for DHFR can be revealed only when stabilizing mutations are included in the analysis. Our earlier study [48] analyzed a smaller set of primarily destabilizing mutants and did not reveal any statistically significant trend (positive or negative) in the stability-activity relation for DHFR.
The development of highly-stabilized DHFR mutants through our combined in silico—in vitro approach opens up promising avenues for new in vivo studies. It has been postulated that protein stability places a fundamental constraint on the evolutionary pathways available to a protein [29,62] which has particular significance in the development of antibiotic resistance: higher protein stability can provide the microorganism with an increased capacity to evolve to evade antibiotic drugs [63] or, more generally, with capacity to evolve new functions [62]. We plan to use an approach developed in our lab [48] to endogenously introduce stabilized DHFR mutants into the bacterial chromosome and we will evaluate mutant fitness relative to wild-type using growth rates and competition experiments. These experiments will allow us to assess whether an evolutionary trade-off exists between stability and fitness in vivo, particularly in the presence of antibiotics.
We plan to apply MCPU to predict stability effects of mutations in proteins other than DHFR, in particular to develop highly stabilized mutants. Comprehensive experimental analysis of fitness and/or stability effects of mutations [64] could be useful in assessing the predictive capabilities of this method. In addition to predicting mutant stabilities, MCPU can provide atomic-detail molecular trajectories to rationalize the stability effects of mutations; such analysis is left to future study.
We employed an all-atom Monte Carlo simulation program incorporating a knowledge-based potential, described in previous publications [16,31,65]. Briefly, the energy function is a sum of contact energy, hydrogen-bonding, torsional angle, and sidechain torsional terms, with an additional term describing orientation of nearby aromatic residues. The move set consists of rotations about ϕ, ψ, and χ dihedral angles, with bonds and angles held fixed. Moves are accepted or rejected according to the Metropolis criterion.
Mutations were introduced into the protein using the program Modeller v9.2 [66]. An initial minimization was carried out in NAMD [67] for 5,000 steps, using the default minimization algorithm and par_all27_prot_lipid.inp parameter file (without waters). An additional minimization step was carried out by running the Monte Carlo simulation program at low temperature (0.100 in simulation units) for 2,000,000 steps. A 2,000,000-step simulation was then run at each of 32 temperatures, averaging over all 2,000,000 steps to obtain Energy, RMSD, and number of contacts. These results were averaged over 50 simulations, for each temperature. Data was then plotted and fit to a sigmoid to obtain the computationally-predicted melting temperature, for each of Energy, RMSD, and number of contacts. To assess dependence of melting temperature on simulation length, longer simulations of 20,000,000 steps were carried out with 30 replications, averaging over the final 2,000,000 steps. For DHFR, 1,000,000 steps took approximately one hour of simulation time, on a single CPU.
We evaluated the effect of the number of MC simulation replications on the prediction results. As shown in S9 Fig, the prediction accuracy is sensitive to the number of replications, but converges to a constant value after approximately 20 replications. In addition, we saw that increasing the number of MC steps beyond 2,000,000 steps does not increase prediction accuracy when the protein has been simulated with at least 20 replications, despite the fact that not all simulations have converged by 2,000,000 steps (S2 Fig—S4 Fig).
Sigmoidal fits were accomplished using the module “Sigmoidal, 4PL” using the software program Prism 6. The sigmoid function has the form:
Y = Bottom + (Top-Bottom)/(1+10^((LogIC50-X)*HillSlope))
The tool is accessible from Shakhnovich lab website http://faculty.chemistry.harvard.edu/shakhnovich/software
The wild type dhfr gene was cloned in a pET24 expression vector under the inducible T7 promoter, then transformed into BL21(DE3) cells [69]. Single point mutations of DHFR were constructed based on a two-step PCR-mutagenesis strategy [70], in which the template for the PCR is the plasmid of WT DHFR. The multiple-mutant variants of DHFR were constructed based on the same method with the single point mutation, but the template of PCR was the plasmid of the corresponding dhfr mutant. To verify the mutations of dhfr, DNA sequencing was performed at the GENEWIZ Incorporation (MA, U.S.). The verified plasmids were transformed into competent E. coli BL21(DE3) cells for expression.
WT DHFR and all mutants used in this study were cloned into a pET24 expression vector and overexpressed in the BL21(DE3) pLys E. coli strain.
A single colony of the transformed E. coli carrying the wild type or mutation dhfr was cultured in Luria-Bertani liquid medium containing 50 μg/mL kanamycin (LB-kana) at 30°C overnight, and then inoculated to fresh LB-kana (1:100 dilution) and incubated again at 30°C. When the OD600 of the culture reached 0.6, isopropyl β-D-1-thiogalactopyranoside (final concentration, 0.4 mM) was added. Cultures were incubated for an additional 12–16 h at 25°C. The cells were then collected by centrifugation and disrupted by sonication. The recombinant proteins were purified with Ni-NTA Superflow (QIAGEN, U.S.) according to the manufacturer’s instructions. Then, the collected protein sample was run with Superdex 75pg Column and was desalted with the desalting Column in ÄKTA protein purification system (GE Healthcare, U.S.). The final concentration of the purified protein was determined using the BCA protein assay kit (PIERCE CHEMICAL, USA) or the NanoDrop instrument (GE Healthcare, U.S.).
DHFR kinetic parameters were measured by the progress-curve kinetics, essentially as described [69,71]. A Scientific stopped flow apparatus, RX.2000 Rapid kinetics system (Applied Photophysics, UK) was used with absorbance monitoring at 340 nm, under single-turnover conditions. NADPH was preincubated with DHFR for 5 min in syringe 1 at the temperature 25°C in a thermostated syringe compartment, and then the reaction was initiated by rapidly mixing the contents with dihydropholate (DHF) from syringe 2. The final assay conditions are 25 nM DHFR, 120 μM NADPH(D), and 25 μM DHF in MTEN buffer (50 mM 2-(N-morpholino)ethanesulfonic acid, 25 mM tris(hydroxymethyl)aminomethane, 25 mM ethanolamine, and 100 mM sodium chloride, pH 7.6). The kinetics parameters (kcat and KM) were derived from progress-curves analysis using Global Kinetic explorer [72].
Thermal stability was characterized by differential scanning calorimetry (DSC), essentially as described in references [69,73]. Briefly, DHFR proteins in Buffer A (10 mM potassium-phosphate buffer pH 7.8 supplemented with 0.2 mM EDTA and 1 mM beta-mercaptoethanol) were subjected to a temperature increase of 1°C/min between 20 to 90°C (nano-DSC, TA instruments, U.S.), and the evolution of heat was recorded as a differential power between reference (buffer A) and sample (120 μM protein in buffer A) cells. The resulting thermogram (after buffer subtraction) was used to derive apparent thermal transition midpoints (Tm app). Thermal unfolding appeared irreversible for all DHFR proteins tested [48], and the two state scaled model provided in NanoAnalyze software (TA INstruments, U.S.) was used to fit the Tm app value. The mutants constructed in this study and the ones published earlier [48] were determined with different DSC instruments with slightly different calibration leading to a small offset of about 2°C for the WT DHFR for earlier published data[48].
Urea unfolding was used to measure stability of the DHFR mutants against chemical denaturation. Proteins (25 μM in buffer A) were diluted in urea (0.2 mM increments up to a final urea concentration between 0 and 6 M), preequilibrated overnight at 25°C for 3 hours, and the change in the folded fraction was monitored by a circular dichroism signal at far-uv wavelength (221 nm) at 25°C (J-710 spectropolarimeter, Jasco). Fitting to a two-state model was used to derive the chemical transition midpoint (Cm).
|
10.1371/journal.ppat.1002847 | The Ebola Virus Glycoprotein Contributes to but Is Not Sufficient for Virulence In Vivo | Among the Ebola viruses most species cause severe hemorrhagic fever in humans; however, Reston ebolavirus (REBOV) has not been associated with human disease despite numerous documented infections. While the molecular basis for this difference remains unclear, in vitro evidence has suggested a role for the glycoprotein (GP) as a major filovirus pathogenicity factor, but direct evidence for such a role in the context of virus infection has been notably lacking. In order to assess the role of GP in EBOV virulence, we have developed a novel reverse genetics system for REBOV, which we report here. Together with a previously published full-length clone for Zaire ebolavirus (ZEBOV), this provides a unique possibility to directly investigate the role of an entire filovirus protein in pathogenesis. To this end we have generated recombinant ZEBOV (rZEBOV) and REBOV (rREBOV), as well as chimeric viruses in which the glycoproteins from these two virus species have been exchanged (rZEBOV-RGP and rREBOV-ZGP). All of these viruses could be rescued and the chimeras replicated with kinetics similar to their parent virus in tissue culture, indicating that the exchange of GP in these chimeric viruses is well tolerated. However, in a mouse model of infection rZEBOV-RGP demonstrated markedly decreased lethality and prolonged time to death when compared to rZEBOV, confirming that GP does indeed contribute to the full expression of virulence by ZEBOV. In contrast, rREBOV-ZGP did not show any signs of virulence, and was in fact slightly attenuated compared to rREBOV, demonstrating that GP alone is not sufficient to confer a lethal phenotype or exacerbate disease in this model. Thus, while these findings provide direct evidence that GP contributes to filovirus virulence in vivo, they also clearly indicate that other factors are needed for the acquisition of full virulence.
| Most Ebola virus species cause severe hemorrhagic fevers with high case fatality rates. However, Reston ebolavirus (REBOV) seems to be apathogenic for humans. While the reason for this is unknown, several lines of in vitro research have indicated that the viral glycoprotein (GP) may play a critical role in determining pathogenicity, although until now there was no data to support such a role in the context of an in vivo infection. In order to address this we have generated a novel reverse genetics system to facilitate rescue of REBOV entirely from cDNA, which together with a previously established full-length clone system for the highly pathogenic Zaire ebolavirus (ZEBOV) allowed us to generate chimeras in which the glycoprotein genes from these two viruses have been exchanged (rZEBOV-RGP and rREBOV-ZGP). While the exchange of the viral glycoprotein did not affect virus growth in cell culture, we could show that infection with rZEBOV-RGP resulted in decreased virulence in a mouse model of infection. Further, rREBOV-ZGP did not show any signs of virulence in this model, similar to wild-type recombinant REBOV, showing that while GP contributes significantly to filovirus virulence, it is clearly not the sole determinant of pathogenicity.
| The family Filoviridae, within the order Mononegavirales, contains two genera, Marburgvirus and Ebolavirus (EBOV), with EBOV being currently divided into the species Zaire ebolavirus (ZEBOV), Sudan ebolavirus, TaΪ Forest ebolavirus and Reston ebolavirus (REBOV) [1]. In addition, Bundibugyo ebolavirus is also being proposed as a potential fifth species [2]. Among the Ebola viruses REBOV has long been recognized as being atypical with respect to both its geographical distribution as well as its pathogenic potential.
Unlike other filoviruses, which are endemic to Africa, REBOV first emerged in 1989/90 as the causative agent of an epizootic among a group of cynomolgus macaques (Macaca fascicularis) imported from the Philippines into the United States [3]. Subsequently, two more introductions have been recognized in Italy and again in the United States [4], [5]. As a result of these importations of infected animals epidemiological investigations were conducted in the Philippines and documented active virus transmission in the primate export facility that was the source for all three shipments of infected monkeys [6], [7]. More recently REBOV co-infection was documented in pigs infected with porcine reproductive and respiratory syndrome virus (PPRSV) in the Philippines, however, it remains unclear whether co-infection with REBOV contributed to the particularly high mortality observed in the infected pigs during this epizootic [8]. While experimental work has thus far not shown REBOV alone to cause symptoms in infected pigs [9], infection with ZEBOV has been shown to result in clinical disease in pigs [10] and concerns remain about the possibility of EBOV transmission to humans via the food chain [11], as well as the possibility of adaptation of REBOV to humans as a result of circulation in such intermediate hosts.
The filoviruses are well known as the causative agents of severe, transmittable and untreatable hemorrhagic fever in humans. The case fatality rates associated with Ebola hemorrhagic fever (EHF) range from 30% to 90%, and this is mainly dependant on the virus species involved, with ZEBOV being the most virulent [12], [13]. Unlike infections with the other filovirus species, infection with REBOV has not been linked to human disease despite several documented infections during animal epizootics in the USA and in the Philippines [6], [8], [14]–[16]. The molecular basis and viral determinants responsible for this dramatic difference in pathogenic potential remain unclear; however, GP in particular has been widely speculated to play a key role. To this end numerous in vitro studies have been conducted over the years analysing the possible contributions of putative immunosuppressive motifs [17]–[19], furin cleavage efficiency [20], [21], cytotoxicity [22]–[24] and various other aspects of glycoprotein biology to pathogenesis. However, to date there is no firm evidence that GP is an important factor for virulence and/or pathogenesis in vivo.
With the availability of reverse genetics systems for ZEBOV [20], [24] we have the potential to study mutant viruses in both tissue culture and animal models, but until now this potential has remained largely unrealized. We were interested to develop a similar system for REBOV, both because it provides a much needed tool to study the biology of this important, understudied and recently re-emerged filovirus, but also because it would complement the ZEBOV system and allow for comparative pathogenesis studies, together with ZEBOV-based mutants. In addition, the availability of both REBOV and ZEBOV full-length clones facilitates the exchange of whole genes, an approach that has potential to assess the contributions of entire viral proteins to viral pathogenesis. In this study we describe the development of a novel REBOV full-length clone system as well as the development and characterization of recombinant chimeric filoviruses. These contained either the REBOV GP in the background of ZEBOV (rZEBOV-RGP), or the ZEBOV GP in the background of REBOV (rREBOV-ZGP). Comparison of these viruses was carried out both in cell culture and in vivo using the interferon α/β receptor knock-out (IFNAR−/−) mouse model, a small animal model that accurately reflects the difference in pathogenicity between ZEBOV and REBOV. The results of this study revealed that while the GP exchange is well tolerated in vitro, the rZEBOV-RGP chimera shows decreased lethality and increased time to death in IFNAR−/− mice. However, introduction of ZEBOV GP alone into REBOV resulted in a slightly attenuated phenotype in the mouse model, indicating that while GP is clearly contributing to the virulence of ZEBOV, alone it is not sufficient to confer a more virulent phenotype and suggesting that additional factors must contribute to virulence in this model.
Based on experience with the ZEBOV reverse genetics system it was identified that an important factor in the utility of any novel filovirus full-length clone system would be the ability to easily access manageable portions of the virus genome for downstream mutagenesis. To accomplish this, unique or rare restriction sites within the REBOV genome were identified and it was determined that the combination of restriction sites shown in Fig. 1A would allow easy access to all portions of the viral genome. Sub-genomic cassettes containing each of the fragments shown were generated prior to being used to assemble the full-length plasmid. While rescue of recombinant viruses from the ZEBOV full-length clone system can be readily achieved, recovery of recombinant REBOV from this infectious clone system proved more difficult. Furthermore, while successful rescue could be achieved with helper plasmids encoding the REBOV ribonucleoprotein complex (RNP) components, it was found that the use of ZEBOV helper plasmids to facilitate virus recovery was beneficial. This is consistent with previous work using minigenome systems which also indicated greater activity of the ZEBOV helper plasmids [25]. However, in all cases the proportion of successful rescues remained significantly lower for REBOV than for ZEBOV with only ∼25% of attempted rescues resulting in the production of infectious virus (data not shown).
With the availability of a full-length clone system for REBOV we next sought to construct recombinant chimeric REBOV and ZEBOV in which their respective glycoproteins had been exchanged. Using standard cloning approaches with Type IIS enzymes the glycoprotein ORFs were exchanged while retaining the parental non-coding regions. A schematic illustration of the four recombinant viruses used in this study is shown in Fig. 1B. They include a recombinant REBOV (rREBOV), a recombinant REBOV expressing ZEBOV GP (rREBOV-ZGP), a recombinant ZEBOV (rZEBOV) and a recombinant ZEBOV expressing REBOV GP (rZEBOV-RGP). Despite the generally lower efficiency of the REBOV full-length clone system all four viruses could be recovered. Once rescued the various recombinant viruses were characterized based on both their RNA and protein content. Filovirus species-specific PCRs for GP and NP (Fig. 1C) as well as Western blot analyses for VP40 and GP are shown (Fig. 1D) and clearly demonstrate the chimeric nature of both rREBOV-ZGP and rZEBOV-RGP. In addition, the PCR-based analysis allowed us to exclude the possibility of contamination with parental virus of either recombinant or natural origin (Fig. 1C).
In order to assess the impact of the glycoprotein exchange on the viability and replication efficiency of the chimeric viruses, VeroE6 cells were infected and analyzed for production of progeny virus and for the formation of cytopathic effects (CPE) associated with infection. CPE formation has previously been strongly linked to GP expression in vitro [23], [24], [26]–[28]. Based on the kinetic data obtained no substantial differences in growth could be identified either between the parental wild-type viruses (wt-REBOV or wt-ZEBOV) and the recombinantly-derived viruses (rREBOV and rZEBOV) or between the recombinant viruses and the chimeras with the glycoproteins exchanged (rREBOV-ZGP and rZEBOV-GP) (Fig. 2). In all cases the differences in titre between all related viruses were less than 1 log at any time point with nearly identical end-point titres being reached (Fig. 2). However, it should be noted that while there were no differences between viruses based on the same parental background (i.e. wt-REBOV, rREBOV and rREBOV-ZGP), a 2–3 log difference was observed between REBOV-based viruses and ZEBOV-based viruses, Based on these data none of the recombinant viruses seem to be significantly impaired with respect to in vitro growth. This indicates that both the biomarkers introduced as a part of the full-length clone construction, as well as the GP exchanges, are well-tolerated with respect to fulfilling the basic functions of virus infection and growth. In addition, observations of CPE formation during infection did not demonstrate any obvious differences as a result of the GP exchange (Fig. S1).
Despite being well tolerated with respect to virus growth and replication in vitro, GP has been postulated to have numerous virulence-relevant effects that cannot be well modelled in vitro. Therefore, we next examined the behaviour of our recombinant EBOVs and the GP chimeras in an IFNAR−/− mouse model of infection. Consistent with a previous report [29], we observed a marked difference in outcome in this animal model between ZEBOV and REBOV. rZEBOV infection showed uniform lethality with doses between 0.1 ffu and 104 ffu/animal (Fig. 3 and data not shown) with animals displaying pronounced decreases in activity, marked weight loss, ruffled fur and hunched posture. In contrast, infection with rREBOV did not produce lethal disease at doses of up to 104 ffu (Fig. 3) and with the only signs of infection being transient weight loss between days 4 and 8 post-infection accompanied by slightly decreased activity. Importantly, infection with both rZEBOV and rREBOV produced outcomes very similar, not only in terms of survival, but also in terms of the kinetics of weight loss and mean time to death (Fig. 3 and Fig. S2), to that seen using equivalent doses of wt-ZEBOV and wt-REBOV. This further supports our in vitro data indicating that the parental recombinant EBOVs used in this study are not attenuated compared to the respective wild-type viruses, despite their clonal origins. Using this model to further analyze our chimeric EBOVs, we could see notable differences between rZEBOV and rZEBOV-RGP with respect to outcome using both high (103 and 104 ffu/animal) and low (10 ffu/animal) challenge doses. Following infection with rZEBOV-RGP animals receiving 103 ffu already showed low levels of survival (Fig. 3A), whereas survival was never seen with rZEBOV even at doses as low as 0.1 ffu (data not shown). With the low challenge dose (10 ffu/animal) this difference in outcomes became even more apparent, with 100% of rZEBOV-infected animals still succumbing to infection, while only 47% of the rZEBOV-RGP infected animals succumbed (Fig. 3B). In addition, the mean time to death for these two viruses differed consistently across a range of doses (Fig. S2), with rZEBOV-RGP infected mice dying 1.3–4.2 days later than mice infected with rZEBOV. In contrast all rREBOV and rREBOV-ZGP infected IFNAR−/− mice survived infection without showing prominent signs of disease (Fig. 3). Further, close examination of the weight curves indicates that rREBOV-ZGP may be slightly attenuated compared to rREBOV. Infection was confirmed in all surviving animals by monitoring seroconversion in ELISA (data not shown).
In order to establish a possible basis for alterations in virulence among the chimeric viruses, we next examined known key target organs/tissues (liver, spleen and blood) from animals infected with 10 ffu of each virus at 5 days post-infection with respect to virus load, antigen expression and histopathological changes. This time point represented a phase at which infection was advanced, but prior to the onset of death in the rZEBOV and rZEBOV-RGP infected groups. Despite the observed differences in mortality, no significant changes in virus load were observed between rZEBOV and rZEBOV-RGP in any of the organs/tissues (spleen, liver and blood) examined, either by calculation of the 50% tissue culture infectious dose (TCID50; Fig. 4A) or by quantitative RT-PCR (qRT-PCR; Fig. S3), supporting our in vitro observation that these viruses are comparable in terms of their growth. Further analysis of tissues by immunohistochemical (IHC) staining indicated extensive infection of both liver and spleen with rZEBOV and rZEBOV-RGP. In liver, staining of both Kupffer cells and hepatocytes was observed, while in the spleen samples staining was mainly observed in cells with macrophage-like morphology (Fig. 4B). Interestingly, rZEBOV-RGP infection in the liver seems to be predominantly of Kupffer cells with little spread to surrounding hepatocytes, whereas both cell types are extensively infected during infection with rZEBOV. The ability of both rZEBOV and rZEBOV-RGP to infect macrophage and macrophage-like cells to a similar extent in vivo is further supported by in vitro growth kinetics in RAW 264.7 (mouse macrophage) cells, which show indistinguishable kinetics for these two viruses (Fig. S4). However, the increased ability of rZEBOV to infect hepatocytes could potentially be significant for the development of pathological changes in the liver as well as disease progression and may contribute to the differences in virulence between these two viruses. Consistent with this observation, histopathological analysis by hematoxylin and eosin (H&E) staining revealed that levels of necrosis and inflammation in the liver were markedly lower during infection with rZEBOV-RGP, when compared to infection with rZEBOV, although hepatocellular necrosis was still observed in both groups (Fig. 5). In contrast, in the spleen both rZEBOV and rZEBOV-RGP produced similar levels of necrosis and inflammation with necrosis of both the white and, to a lesser extent, the red pulp being evident.
Among the REBOV-based viruses, analysis of viral titres in target organs by calculation of the TCID50 indicated that titres were 1–2 logs lower for rREBOV-ZGP than for rREBOV in all tissues/organs tested (Fig. 4A). Interestingly, this difference was not as pronounced when samples were analysed by qRT-PCR (Fig. S3). While for both rREBOV and rREBOV-ZGP the IHC staining in liver and spleen was almost exclusively of Kupffer cells or cells with macrophage-like morphology, the extent of staining observed in samples from rREBOV-ZGP infected animals was decreased compared to rREBOV infected samples, again suggesting a markedly lower viral burden in organs from animals infected with rREBOV-ZGP (Fig. 4B). Both rREBOV and rREBOV-ZGP showed only minor pathological changes in liver and spleen samples, with focal inflammation in liver samples being the most prominent observation. Overall, the data suggest that in the spleen both necrosis and inflammation are slightly decreased with rREBOV-ZGP, while the extent of pathological change in the liver is comparable between these two viruses (Fig. 5). These findings are consistent with the survival data in suggesting a slight attenuation of rREBOV-ZGP, in comparison to the parental rREBOV.
Comparison of the pathology data from REBOV-based and ZEBOV-based recombinant viruses revealed uniformly higher levels of both inflammation and necrosis among animals infected with rZEBOV or rZEBOV-RGP, as compared to rREBOV or rREBOV-ZGP (Fig. 5) but also significantly higher levels of staining in IHC (Fig. 4B), findings that most likely reflect an inherent difference between ZEBOV and REBOV in their in vivo growth and spread in this model. This would also be consistent with the large differences observed in the analysis of in vitro growth of these viruses (Fig. 2). Although we cannot exclude that a lower sensitivity of the anti-VP40 antibody towards REBOV also contributes to reduced detection of REBOV infection by IHC, titre analysis by TCID50 and qRT-PCR (Fig. 4A and Fig. S3) also clearly support that REBOV is compromised, compared to ZEBOV, in terms of its in vivo growth.
Most filoviruses cause severe, transmittable and untreatable hemorrhagic fever in humans with high case fatality rates; however, this is not the case for REBOV. Despite its pathogenicity for nonhuman primates, REBOV has never been associated with disease in humans. This is despite investigations that documented at least seven seroconversions among exposed animal handlers, including one who was also positive by RT-PCR, during the early animal importations into the United States [6], [14], [15]. In addition, extensive serosurveys were conducted during the recent Philippine REBOV/PRRSV outbreak, focusing on individuals with a high probability of exposure. Studies identified 6 additional individuals as REBOV seropositive, again in the absence of any notable disease [8], [16]. These individuals were farm workers or butchers, clearly suggesting an occupational exposure. Thus, the currently available data strongly support that REBOV infections lead to either an asymptomatic or subclinical course of disease, at least in healthy adults. While the basis for this lack of human pathogenicity with REBOV remains unknown, the filovirus glycoprotein has long been proposed to play a key role in pathogenesis, with numerous potential mechanisms having been proposed [30]–[32]. However, until now direct evidence for such a role in the context of a filovirus infection has been notably lacking, particularly in vivo.
In this study we have sought to address this issue directly by applying both a previously existing full-length clone system for the highly pathogenic ZEBOV and a novel full-length clone system for REBOV to generate chimeras in which the glycoproteins of these two viruses are exchanged. This study is not only unique in that it examines the effect of an entire open reading frame exchange among filoviruses, but also makes use of a combination of the REBOV and ZEBOV full-length clone systems. This approach is critical to a complete understanding of the role of a given factor in pathogenesis, as we clearly see in this study where GP, while necessary for full virulence in ZEBOV, does not alter the virulence of REBOV. Our in vitro analysis of chimeric EBOV growth indicated that exchange of the GP ORF is surprisingly well tolerated in terms of basic viral functions such as entry, replication and budding, all of which contribute to successful growth in vitro. Interestingly, we also did not see any notable differences in the extent or time of onset of CPE between parental recombinant viruses and the chimeric viruses in which the glycoprotein has been exchanged (i.e. between rREBOV and rREBOV-ZGP or between rZEBOV and rZEBOV-RGP), a finding that may speak against a significant difference in direct GP-mediated cytotoxicity between ZEBOV and REBOV GP when expressed in the context of a filovirus infection. This is in contrast to data using adenovirus vectors where expression of ZEBOV GP was shown to result in significantly more cytotoxicity in vessel explants than REBOV [23], but may rather support the idea that the authentic levels of GP expression associated with the productive stages of virus growth are in fact well tolerated [22]. Despite the absence of any obvious differences in vitro, it remained of interest to examine these chimeric viruses in an in vivo context. This is particularly the case given the various putative immunomodulatory properties of GP (e.g. immunosuppressive motifs, masking of cell surface proteins, glycoprotein shedding) [17]–[19], [33], [34]. In addition, as the receptor-binding protein, GP plays a critical role in target cell selection, and it remains unclear how this might differ between the filovirus species.
A significant limitation in conducting comparative pathogenesis studies with REBOV is the difficulty in selecting an appropriate animal model. While the limited evidence available suggests that REBOV displays a less virulent phenotype than ZEBOV in some species of non-human primate [35], this model is far from ideal for conducting initial animal studies, for both technical and ethical reasons. Further, since it is necessary to adapt filoviruses before they can cause lethal disease in immunocompetent rodent species [36]–[38] this introduces potential problems when attempting to compare viruses that have undergone distinct and only poorly understood adaptation processes [37], [39], [40]. To date the only rodent models that recapitulate the difference in virulence between ZEBOV and REBOV in humans without the need for adaptation are the IFNAR−/−, severe combined immunodeficiency (SCID) and STAT1 knock-out (STAT1−/−) mouse models [29], [41], [42]. Of these systems the IFNAR−/− model has the considerable advantage that it does not have the extensive and broad-ranging defects associated with the SCID and STAT1−/− phenotypes. On this basis we elected to use the previously described IFNAR−/− mouse model [29] in order to examine our chimeric Ebola viruses for alterations in in vivo virulence. In this animal model wild-type ZEBOV (strain Mayinga) has been shown to be uniformly lethal without prior adaptation [29], as was our recombinantly derived rZEBOV. In contrast, rREBOV did not produce disease at doses of up to 104 ffu/animal. This makes the IFNAR−/− mouse a convenient model for recapitulating the differences in pathogenicity between rZEBOV and rREBOV in a small animal model. In contrast to our in vitro findings, in the IFNAR−/− model of infection we observed significant changes in the ability of the chimeric rZEBOV-RGP to cause disease in comparison to rZEBOV. We observed a marked reduction in lethality both at high (103 ffu/animal and 104 ffu/animal) and low (10 ffu/animal) challenge doses, as well as a prolonged time to death. Further analyses aimed at understanding the basis for this in vivo attenuation showed no differences in virus burden in either early (spleen) or later (liver and blood) target organs, again showing that this virus is not compromised in its growth. In spleen samples, infection with both rZEBOV and rZEBOV-RGP was seen mainly in cells with macrophage-like morphology. Similarly, Kupffer cells represented a major target of infection for both rREBOV and rREBOV-ZGP in liver with both viruses showing antigen accumulation in these cells. The ability of both viruses to replicate equally well in macrophage cells is further supported by in vitro data showing comparable growth of these two viruses in a mouse macrophage cell line. However, in the liver of rZEBOV-RGP infected animals the infection appears to have been mainly restricted of Kupffer cells, while liver samples from animals infected with rZEBOV not only showed infection of Kupffer cell but also extensive hepatocyte infection. Since titres in liver samples were similar between these samples, despite the paucity of hepatocyte infection observed in rZEBOV-RGP samples, this indicates that Kupffer cells may actually be the main source of virus production during infection in the liver and that hepatocytes, which are a significant target of virus-induced damage, do not contribute significantly to virus burden in the infected host. Further, decreased infection of hepatocytes with ZEBOV-RGP could potentially explain the markedly decreased necrosis and inflammation observed in the liver. That these findings are observed in liver might be of particular significance given that this organ plays an important role in clotting factor synthesis (reviewed in [43]) and thus tissue damage could have direct implications for coagulation. Unfortunately, this is not an aspect of filovirus pathogenesis that can be reliably modelled in mice [44], [45]. This finding may also suggest that differences in target cell selection exist between the REBOV and ZEBOV GPs, possibly as a result of subtly different receptor usage preferences. Indeed, while EBOV infection has been shown to be enhanced by a number of putative “receptor” molecules [46]–[50] differences in the usage of these molecules by different EBOV species has not yet been examined, but this will be an interesting avenue for future research. Introduction of the ZEBOV GP alone into REBOV led to a slight decrease in the virulence of the resulting chimera, clearly indicating that, despite the many proposed roles of GP for pathogenesis, GP alone is not a decisive determinant of EBOV virulence in this model. While the basis for the slight attenuation seen with the rREBOV-ZGP chimera in vivo remains unclear, it may be related to minor incompatibility on a molecular level between the heterologous GP and other viral proteins. Further, the more significant decrease observed in viral titres in infected organs, compared to viral RNA content, suggests that such a defect may be related to functions in the late steps in the viral lifecycle, such as morphogenesis and budding, processes in which GP plays a prominent role, and which in particular require its interaction with VP40. However, since this virus was not significantly attenuated during in vitro growth it also appears that this is only a factor under conditions present in the in vivo context (e.g. in the presence of an immune response).
In summary, despite the limitations of the mouse model with respect to recapitulating the coagulation defects central of the development of HF in humans and NHPs, our study clearly shows the utility of the IFNAR−/− mouse model for studying the differences in virulence between REBOV and ZEBOV without the need for prior adaptation of the challenge viruses. In addition, we present not only the development of a novel REBOV full-length clone system, but together with an existing ZEBOV full-length clone system, also shed light on the role of GP in pathogenesis. This represents a unique application of filovirus reverse genetics systems to studying the contributions of an entire viral protein to pathogenesis and provides long awaited insight into the contributions of GP to in vivo virulence in an authentic filovirus context. Using this approach we could show that the role of GP in the virulence of ZEBOV is related to inflammatory and necrotic changes in the liver, likely as a result of improved virus spread from infected Kupffer cells into the surrounding hepatocytes, and not to increased virus burden in the various target organs/tissues. However, introduction of REBOV GP into ZEBOV did not completely attenuate the resulting chimera, indicating that other viral proteins also play a significant role in contributing to the virulence of ZEBOV in this model. In particular the enhanced growth of ZEBOV-based viruses, in comparison to those based on REBOV, both in vitro and in vivo speaks for a possible role of efficient replication in pathogenesis, a concept that is also supported by limited studies with filovirus minigenome systems [25]. Consistent with a multifactorial view of filovirus virulence, introduction of ZEBOV GP into REBOV did not affect virulence, supporting the conclusion that while GP is an important determinant of filovirus virulence, alone it is not sufficient for virulence.
ZEBOV (strain Mayinga; Accession #AF272001) and REBOV (strain Pennsylvania; Accession #AF522874) were used as the parental virus strains and provided RNA templates for all experiments. Generation of the mouse-adapted ZEBOV used as a control in the animal experiments has been previously reported [36], [38]. Experiments with both parental and recombinant viruses were performed in the BSL-4 laboratories at the National Microbiology Laboratory of the Public Health Agency of Canada, the Philipps Universität Marburg, Germany and the Rocky Mountain Laboratories (RML), Division of Intramural Research (DIR), National Institute of Allergy and Infectious Diseases (NIAID), National Institutes of Health (NIH), USA.
VeroE6 (African green monkey kidney) and RAW 264.7 (mouse macrophage) cells were maintained in Dulbecco's modified Eagle's medium (DMEM, Life Technologies) supplemented with 10% fetal bovine serum (FBS, PAN Biotech), 2 mM L-glutamine (Life Technologies), 100 U/mL penicillin and 100 µg/mL streptomycin (Life Technologies) and grown at 37°C with 5% CO2.
Two low-copy plasmids containing either a kanamycin resistance gene together with the p15A origin of replication (pKan) or an ampicillin resistance gene together with the p15A origin (pAmp) were generated by using gene-specific primers and standard PCR techniques to amplify the relevant portions of pACYC177 (NEB). A customized multiple cloning site was then generated through hybridization of complimentary overlapping commercial oligonucleotides encoding the following restriction sites: NotI-NcoI-SpeI-XmaI-XhoI-PacI-MluI. In order to facilitate the downstream cloning of some genome fragments the plasmid-encoded XmaI, XhoI and BsmBI sites were deleted from the pKan vector. All sequences as well as details of the cloning strategies can be provided upon request.
Analysis of the REBOV genome revealed several unique and/or rare restriction sites that could be used to generate sub-genomic cassettes. Based on this analysis we selected NcoI, SpeI, XhoI and PacI for the generation of sub-genomic plasmids (Fig. 1A). In addition, a NotI site and MluI site were added flanking the T7 promoter and terminator sequences, respectively, in order to facilitate cloning of the terminal genomic fragments. Further, in order allow differentiation between our recombinant virus and a potential contamination with existing laboratory strains we inserted genetic markers into the REBOV full-length clone to allow genetic identification of this virus as a recombinant. These markers are a silent mutation that abolishes an XhoI site in NP, a silent mutation abolishing a KpnI site in L, and a silent mutation to create an XmaI site in the virion protein (VP) 30 ORF (Fig. 1A). In addition a silent mutation in GP1,2 was retained to allow discrimination between the parental and recombinant GP genes. Initially fragments of the virus genome were cloned into the pKan background using the restriction sites listed above and provided a series of sub-genomic cassettes for use in downstream cloning steps as well as for subsequent assembly of the full-length plasmid. To assemble the full-length genome in pAmp, sub-genomic fragments of the genome were successively introduced into the pAmp vector.
Helper plasmids for full-length genome rescue were produced by cloning the open reading frames (ORFs) encoding NP, VP35, VP30 and L into pCAGGS. Generation of these constructs was previously described with all constructs being validated by sequencing and confirmed to be functional in a REBOV minigenome assay [25], as well as through their ability to mediate rescue of a ZEBOV infectious clone [51]. The plasmids for the wild-type ZEBOV infectious clone system were constructed as previously described [20].
Chimeric ZEBOV/REBOV plasmids, in which the open reading frames for GP were exchanged, were generated from the full-length REBOV and ZEBOV clones using standard cloning techniques and designated pTM1-ZEBOV-RGP and pAmp-REBOV-ZGP.
Recovery of recombinant virus from the ZEBOV full-length genome plasmid was carried out as previously reported [20]. Briefly, VeroE6 cells were split one day prior to transfection into 6-well plates in order to obtain 50% confluent monolayers on the following day. Cells were then transfected with 1 µg of full-length construct, as well as helper plasmids (250 ng pCAGGS-NP, 125 ng pCAGGS-VP35, 75 ng pCAGGS-VP30, 1.0 µg pCAGGS-L) and 250 ng pCAGGS-T7. For rescue of pAmp-REBOV and pAmp-REBOV-ZGP the same approach was followed except that cells were transfected with increased amounts of the helper plasmids (1 µg pCAGGS-NP, 500 ng pCAGGS-VP35, 300 ng pCAGGS-VP30, 4.0 µg pCAGGS-L) and 1.0 µg pCAGGS-T7. For REBOV and rREBOV-ZGP, recovery was attempted with helper plasmids encoding NP, VP35, VP30 and L from both REBOV and ZEBOV. In all cases transfection was carried out using 6 µl FuGENE 6 (Roche) per µg DNA according to the manufacturer's directions with the transfection complexes being removed and the medium replaced 24 h post-transfection. Cells were monitored for the formation of cytopathic effects (CPE) associated with virus infection and a blind passage to fresh 80–90% confluent VeroE6 cells was performed 7 days post-transfection (passage 1, p1). Once these p1 cells showed CPE (7 days for ZEBOV, 14 days for REBOV) fresh VeroE6 cells were again infected (p2), and once CPE formation was observed these supernatants were harvested for use in all further experiments.
VeroE6 cell monolayers with a confluence of 80–90% were infected in 6-well plates with wt-REBOV, wt-ZEBOV, rREBOV, rZEBOV, rREBOV-ZGP or rZEBOV-RGP at an MOI of 0.1 in 1 ml of serum-free DMEM for 1 h at 37°C in a 5% CO2 atmosphere. In addition, RAW 264.7 cells with a confluence of 60–70% were similarly infected with wt-ZEBOV, rZEBOV or rZEBOV-RGP. Following absorption the inoculum was removed and the cells washed with DMEM to remove any unbound virus. Cells were placed in fresh DMEM containing 2% FBS, L-glutamine and penicillin/streptomycin and incubated for 5 (RAW 264.7) or 7 (VeroE6) days. Supernatants were collected on days 1, 2, 3, 4 and 5 post-infection for RAW 264.7 cells and on days 1, 2, 3, 4 and 7 for Vero cells, for analysis of progeny virus release by immunostaining in a focus-formation assay. CPE formation in Vero cells was monitored and photographed on days 1, 2, 3, 4 and 7 post-infection using an Axiovert 200 M microscope (Zeiss).
VeroE6 cell monolayers with a confluence of 80–90% were infected in a 12-well plate format with the various recombinant EBOVs in a 300 µl volume for 1 h at 37°C in a 5% CO2 atmosphere in serum-free DMEM. Following absorption the inoculum was removed and the monolayers were overlaid with 4 ml DMEM containing 1.5% carboxymethyl cellulose (CMC), 2% FBS, L-glutamine and penicillin/streptomycin. After 5 days (ZEBOV) or 10 days (REBOV) cells were fixed in 4% paraformaldehyde (PFA) overnight and then placed in fresh 4% PFA before being removed from the BSL4 facility and incubated for a further 24 h. Fixed cells were then permeabilized in PBS with 0.1% Triton X-100 for 15 min. Staining was performed at room temperature for 1 h, first with a 1∶1,000 dilution of an anti-REBOV VP30 mouse serum [52] or a 1∶200 dilution of an anti-ZEBOV goat serum and then with a 1∶200 dilution of goat anti-mouse Alexa 488 (Molecular Probes) or a 1∶100 dilution of donkey anti-goat FITC, respectively. Foci were counted using an Axiovert 200 M microscope (Zeiss).
To confirm virus rescue whole-cell extracts were prepared by lysing infected cells with sodium docecyl sulfate (SDS) sample buffer [25% glycerol, 2.5% SDS, 125 mM Tris [pH 6.8], 125 mM dithiothreitol, 0.25% bromophenol blue]. Samples were boiled for 10 minutes at 99°C and transferred into a fresh tube before removal from the BSL4 facility at which time the samples were again boiled for 10 minutes at 99°C. Proteins were then separated on 10% SDS polyacrylamide gels and transferred onto polyvinylidene difluoride (PVDF) membranes. Immunostaining was performed with dilutions of primary antibody in phosphate-buffered saline (PBS) containing 1% skim milk and 0.1% Tween-20 as indicated below. The VP40-specific monoclonal antibody 2C4 (1∶50) was used to detect VP40 [53], while the GP-specific monoclonal antibodies 12/1.1 (1∶20,000) and 42/3.7 (1∶5,000) (generously provided by A. Takada, Hokkaido University) were used to detect ZEBOV GP only or both ZEBOV and REBOV GP, respectively. For VP40, detection was performed with an Alexa 680-conjugated anti-mouse IgG secondary antibody (Molecular Probes) using the Odyssey Infrared Imaging System (LI-COR) while for GP detection was performed with a horseradish peroxidase (HRP)-conjugated donkey anti-mouse IgG secondary antibody (Jackson ImmunoResearch) and visualized using the ECL Plus Detection system (GE Healthcare).
Viral RNA was isolated from the infected cells using the QIAamp Viral RNA Mini Kit (Qiagen). The eluted RNA was used for reverse transcription-PCR (RT-PCR) using the Superscript III RT kit (Invitrogen) with subsequent PCR amplification being performed using the iProof PCR kit (Bio-Rad) according to the manufacturer's instructions. This approach was used to generate overlapping fragments that allowed sequencing of the complete viral genomes of all recombinant viruses. Further, in order to visually demonstrate the chimeric nature of rREBOV-ZGP and rZEBOV-RGP and to exclude any contamination with the parental virus, fragments corresponding to the NP and GP genes were amplified using primers specific for REBOV or ZEBOV and analysed by gel electrophoresis.
Groups of C57BL/6 IFNAR−/−mice (n = 5–15) were infected via the intraperitoneal (i.p.) route with 200 µl of DMEM containing the indicated doses (10 ffu, 103 ffu or 104 ffu) of wt-REBOV, rREBOV, rREBOV-ZGP, wt-ZEBOV, rZEBOV, rZEBOV-RGP or MA-ZEBOV. Mice were monitored daily for weight loss and signs of disease. All surviving animals were euthanized at day 28 and final serum samples were collected to determine antibody titers. Additional groups (n = 3) were infected with 10 ffu of rREBOV, rREBOV-ZGP, rZEBOV-RGP or rZEBOV as described above and were sacrificed on day 5 post-infection. Blood, liver and spleen samples were collected and stored at −80°C.
Animals were handled in the RML BSL-4 containment space. Research was conducted in compliance with the guidelines of the NIAID/RML Institutional Animal Care and Use Committee (IACUC). The facility where this research was conducted is fully accredited by the Association for the Assessment and Accreditation of Laboratory Animal Care International (AAALAC) and has an approved Office of Laboratory Animal Welfare (OLAW) Assurance (#A4149-01). Research was conducted under a protocol approved by the IACUC. All procedures were conducted by trained personnel under the supervision of veterinarians and all invasive clinical procedures were performed while animals were anesthetized. Early endpoint criteria, as specified by the IACUC approved scoring parameters, were used to determine when animals should be humanely euthanized.
Tissues were fixed for hematoxylin and eosin staining using 10% neutral buffered formalin. Tissues were then placed in cassettes and processed with a VIP-5 Tissue Tek processor (Sakura Finetek) using a graded series of ethanol, xylene, and ParaPlast Extra. Embedded tissues were sectioned at 5 µm and dried overnight at 42°C prior to staining. Pathological changes were evaluated according to severity: 0 = normal; 1 = minimal change (rare signs of necrosis and/or inflammatory cells); 2 = mild change (isolated small aggregates of necrosis and/or inflammatory cell infiltration); 3 = moderate change (larger aggregates of necrosis and/or inflammatory cells); 4 = marked change (extensive and coalescing foci of necrosis and/or inflammatory cell infiltration); 5 = severe change (diffuse necrosis and/or inflammatory cell infiltration; no remaining normal tissue).
For immunohistochemistry, antigen was detected using a cross-reactive polyclonal rabbit anti-ZEBOV VP40 primary antibody at a 1∶2,000 dilution. The tissues were processed using the Discovery XT automated stainer (Ventana Medical Systems) with a DABMap kit (Ventana Medical Systems) using a Biogenex biotinylated anti-rabbit secondary antibody and were counter-stained with hematoxylin.
VeroE6 cells were seeded into 48-well plates the day before titration. Liver and spleen samples were thawed, weighed and homogenized in a 10-fold volume of DMEM without supplements using a TissueLyser II (Qiagen) prior to the preparation of serial dilutions. Blood samples were thawed and serial dilutions were prepared directly. Media was removed from cells and wells were inoculated in triplicate for each dilution. After one hour DMEM supplemented with 2% FBS, L-glutamine and penicillin/streptomycin was added and cells were incubated at 37°C. Cells were monitored for cytopathic effect (CPE) and the 50% tissue culture infectious dose (TCID50) was calculated for each sample employing the Reed and Muench method [54].
RNA was isolated from mouse blood, liver and spleen samples using the QIAamp Viral RNA Mini Kit (Qiagen). Quantitative RT-PCR was performed as previously described using ZEBOV-or REBOV-specific NP primers and probes [41], [55].
|
10.1371/journal.ppat.1002575 | In Situ Microscopy Analysis Reveals Local Innate Immune Response Developed around Brucella Infected Cells in Resistant and Susceptible Mice | Brucella are facultative intracellular bacteria that chronically infect humans and animals causing brucellosis. Brucella are able to invade and replicate in a broad range of cell lines in vitro, however the cells supporting bacterial growth in vivo are largely unknown. In order to identify these, we used a Brucella melitensis strain stably expressing mCherry fluorescent protein to determine the phenotype of infected cells in spleen and liver, two major sites of B. melitensis growth in mice. In both tissues, the majority of primary infected cells expressed the F4/80 myeloid marker. The peak of infection correlated with granuloma development. These structures were mainly composed of CD11b+ F4/80+ MHC-II+ cells expressing iNOS/NOS2 enzyme. A fraction of these cells also expressed CD11c marker and appeared similar to inflammatory dendritic cells (DCs). Analysis of genetically deficient mice revealed that differentiation of iNOS+ inflammatory DC, granuloma formation and control of bacterial growth were deeply affected by the absence of MyD88, IL-12p35 and IFN-γ molecules. During chronic phase of infection in susceptible mice, we identified a particular subset of DC expressing both CD11c and CD205, serving as a reservoir for the bacteria. Taken together, our results describe the cellular nature of immune effectors involved during Brucella infection and reveal a previously unappreciated role for DC subsets, both as effectors and reservoir cells, in the pathogenesis of brucellosis.
| Brucella are facultative intracellular bacteria chronically infecting humans and animals causing brucellosis, one of the most common zoonotic disease worldwide which can result in infertility and chronic debilitating disease. The cells supporting Brucella growth in vivo remain largely unknown. In order to identify these, we constructed a Brucella melitensis strain expressing a fluorescent protein that allowed us to characterize infected cells by microscopy of the spleen and liver from infected mice. In both tissues, the majority of primary infected cells were cells from the macrophage lineage. The peak of infection correlated with granuloma development. These structures contained the majority of bacteria and were mainly composed of cells expressing CD11b, F4/80, MHC-II, which are specific of activated monocytes/macrophages. A fraction of granuloma cells also expressed CD11c and were similar to inflammatory dendritic cells (DCs). During the chronic phase of infection in susceptible mice, we identified a particular subset of DC expressing CD205 and serving as a reservoir for the bacteria. Overall, our results describe the nature of immune cells infected by Brucella in vivo and reveal an unappreciated role for DC subsets, both as effectors and reservoir cells. These results could help develop new therapeutic strategies to control Brucella infection.
| Brucella (α-proteobacteria) are facultative intracellular Gram-negative coccobacilli that infect humans as well as domestic (goat, sheep, swine, etc.) and wild type mammals. Animal infection leads to abortion and infertility with dramatic economic costs. Brucellosis is mainly transmitted to humans through the ingestion of raw milk or non-pasteurized cheese contaminated with Brucella. It is characterized by undulant fever which, if left untreated, can result in chronic disease with serious clinical manifestations, such as orchitis, osteoarthritis, spondylitis, endocarditis, and several neurological disorders [1]–[5]. Human brucellosis remains a significant public health concern in areas of the world where Brucella infections are endemic in food animals. Indeed, brucellosis has been described as being the most common zoonotic disease worldwide with more than 500,000 new human cases annually [4].
Brucella are highly infectious via oral and aerosol routes and difficult to treat with antibiotics. No safe or effective vaccine is available to prevent human infection. These characteristics justify the classification of Brucella strains as a category B pathogens, which represent a risk for use as bioweapons [6],[7].
B. melitensis is the most frequent cause of human brucellosis [1], [8]. Recently, bioluminescent B. melitensis has been used to visualize the dynamic of bacterial dissemination following intraperitoneal (i.p.) inoculation in mice [9]. Results confirmed that this model parallels human infection and identified major sites of bacterial growth, such as the spleen and the liver, during the early chronic phase of infection. However, cells supporting bacterial growth in vivo in these organs and serving as reservoir are unknown.
Despite recent progress in mouse models of brucellosis, much remains unknown regarding cellular components of the innate and adaptive immune responses induced by B. melitensis infection. We and others have shown that IFN-γ-producing CD4+ T cells [10]–[14] and inducible Nitric Oxide Synthase (iNOS/NOS2)-producing inflammatory dendritic cells (iNOS-DC) [10] are major components of protective immune response against B. melitensis. Activation of these cells involves Toll Like Receptor (TLR) 4 and TLR9 coupled to MyD88 adaptor protein [10]. However, there is little understanding on where these cells are localized in situ so they can initiate cellular interactions to control infection. Several experimental models of infection by intracellular bacteria such as Mycobacteria tuberculosis [15] and Listeria monocytogenes [9], [16] have illustrated the importance of the granulomatous lesion in limiting both tissue damage and bacterial dissemination. The granulomatous lesion is an organized and dynamic structure implicating activated monocytes surrounded by T cells and granulocytes. Its formation involves an orchestrated production of chemokines and cytokines and the upregulation of their cognate receptors along with the expression of addressins, selectins and integrins. Altogether, these elements coordinate the recruitment, migration and retention of cells to and within the granuloma. Chronic granulomatous inflammation has been reported in spleen and liver from natural hosts, human and mice infected by Brucella bacteria [17]. However, the importance of granulomas in the control of Brucella growth, their cellular composition and the signalling pathways implicated in their formation are largely unknown.
To address these issues, we developed a B. melitensis strain stably expressing mCherry fluorescent protein (mCherry-Br). This novel tool has allowed us to determine the phenotype of Brucella-targeted cells and characterize the composition and the dynamic of granuloma in the spleens and livers of infected mice. We observed the formation of iNOS+ granuloma structures surrounding infected cells and identified iNOS-DC (CD11b+ CD11c+ Ly-6G− MHC-II+ iNOS+) and activated monocytes (CD11b+ CD11c− Ly-6G− MHC-II+ iNOS+) as the major cell types constituting granulomatous lesions in both tissues. In addition, we observed that alteration of the MyD88/IL-12/IFN-γ axis deeply affects the cellular composition of granulomas, reducing their ability to control bacterial dissemination and leading to replication and persistence of the pathogen in distinct cellular niches.
In order to detect infected cells on tissue sections by fluorescent microscopy, we constructed a constitutively fluorescent strain of B. melitensis. The mCherry protein, a previously described rapidly maturing variant of the red fluorescent protein DsRed [18], was cloned into the suicide vector pKSoriT-bla-kan downstream of a strong Brucella spp. promoter PsojA [19], [20]. The final construct was transformed into Escherichia coli strain S17-1, and introduced into B. melitensis strain by conjugation (see Material and Method for details).
Flow cytometry (Figure S1.A) and fluorescent microscopy (Figure S1.B) were used to validate mCherry-expressing Brucella (mCherry-Br) strain in vitro. The comparison of wild type and mCherry-Br strains growth curve revealed that mCherry gene expression does not affect ability of B. melitensis to replicate in vitro (data not shown). In addition, following i.p. inoculation of 4×104 CFU, we showed that mCherry-Br was able to infect chronically resistant C57BL/6 mice (Figure S1.C). At 5 days post infection (p.i.), bacterial burden in spleen from mCherry-Br infected mice was reduced by approximately half-a-log when compared to wild type bacteria-infected mice. However, at 25 and 42 days p.i., bacterial counts from both groups were similar, suggesting that bacterial virulence is moderately and transiently affected by the insertion of mCherry tracer.
In order to characterize the spatio-temporal behavior of Brucella in spleen and liver of infected animals, we used immuno-fluorescence microscopy techniques. At low dose of infection (104 to 106 CFU) in C57BL/6 and BALB/c wild type mice, growth of B. melitensis peaked at 5 days p.i. with modest CFU counts (between 105–106 CFU/g of spleen [10]) rendering its visualization on tissue sections very difficult. Our preliminary studies demonstrated that 106 CFU/g of tissue was the limit of sensitivity of our technique, with less than 0.1 infected cells by observation surface (200×230 µm) in tissue sections of 5 µm in thickness and using a 63× objective (data not shown). Therefore, in order to maximize our chances of visualizing the dynamic behavior of mCherry-Br in vivo several days after inoculation, we injected 108 CFU of bacteria in wild type mice. With this dose, no increase in mortality of mCherry-Br infected wild type C57BL/6 and BALB/c mice was observed (data not shown). Importantly, when compared to the routinely used low bacterial inoculum, a high infectious dose (108 CFU) showed similar impact on the composition of recruited spleen cell populations involved in innate inflammatory responses to Brucella. Overall, a high bacterial inoculum only shorten the kinetic of cell recruitment (Figure S2).
First analyses revealed that ten minutes after i.p. inoculation, bacteria were massively present in the spleen and the liver, with an average of 106 CFU/g (Figure S3.A). Both organs maintained a level of 106–108 CFU/g during the 5 first days p.i. (Figure S3.A), allowing the analysis of mCherry-Br infected cells in situ during this time (Figure S3.B). Infected susceptible BALB/c mice displayed significantly higher count of bacteria in the spleen at 3 and 5 days p.i. when compared to C57BL/6 mice. However, at day 12, spleens and livers from wild type mice had less than 105–106 CFU/g and bacteria became undetectable by microscopy analysis (data not shown). Consequently, we limited our analysis to 120 h p.i. in wild type mice.
The functions of the spleen are centered on systemic blood circulation. As such, it lacks afferent lymphatic vessels. It is comprised of two functionally and morphologically distinct compartments, namely the red pulp (r.p.) and the white pulp (w.p.). The r.p. is a blood filter rich in F4/80+ macrophages that removes foreign material as well as damaged and effete erythrocytes. The spleen is also the largest secondary lymphoid organ of the body that initiates immune responses to blood-borne antigens. This function is usually associated to the w.p. compartment. This latter is centered on the central arteriole within the T cell rich area surrounded by B lymphocyte-associated follicles. At the interface between the r.p. and w.p., the marginal zone (m.z.) is a unique region of the spleen. Considered to be a separate compartment rather than part of the w.p., it is designed to screen systemic blood circulation for antigens and pathogens and plays an important role in antigen processing. It also contains several populations of specialized macrophages, such as marginal zone-associated metallophilic macrophages that can be identified with anti-MOMA-1 antibody (for a review see [21]).
Following i.p. inoculation of 108 CFU in resistant wild type C57BL/6 mice, mCherry-Br was detectable in the spleen by 10 minutes p.i. (Figure S3.A). Importantly, at this time, our observation showed that mCherry-Br already localized in the r.p. and the m.z. of infected spleens (data not shown). Then, bacteria increased exponentially during the first 24 h (Figure S3.A) and the number of intracellular bacteria peaked at 120 h p.i. with an average of 8–10 bacteria per cell (Figure 1.A). Importantly, infected cells were mainly located in m.z. and r.p. during the whole kinetic of infection (Figure 1.B–C and Figure 2.A) while infected cells were rarely observed in splenic w.p. area (w.p.).
Further semi-quantitative analyses of infected cell phenotypes in C57BL/6 mice (Figure 2.B and Figure 3.A–B) showed that first infected cells in spleen were mainly r.p. macrophages with a F4/80+ CD11b− phenotype (∼80%) and MOMA-1+ m.z. metallophilic macrophages (∼20%). Between 3 and 6 h after inoculation, the frequency of infected MOMA-1+ cells decreased strongly and CD11c+ dendritic cells progressively became infected (∼30% of total infected cells at 6 h) (Figure 3.A).
Using a home-made rabbit polyclonal antibody specific of B. melitensis (Brucella-Ag), we showed that mCherry-Br fluorescence and Brucella-Ag staining were generally overlapping (Figure S4), although there were some instances where Brucella-Ag staining could be detected in the absence of appreciable mCherry-Br fluorescence, possibly as a consequence of bacterial death or trafficking of antigens away from the bacteria. These results confirm that the majority of Brucella-infected cells can be detected in situ using mCherry-Br.
Confocal analysis in Z-stack demonstrated that bacteria localized inside F4/80+ r.p. macrophages (Figure S5.A–B). Importantly, this was also true for mCherry-expressing VirB-defective Brucella mutant (Figure S5.C). The VirB mutant displays a strongly reduced ability to infect cells in vitro [22] though it is able to persist several days in vivo and colonize mice similarly to wild type bacteria [23]. Therefore, it has been hypothesized that in vivo persistence could be due to extracellular replication [24]. However, our observations are not in favor of this hypothesis since VirB mutant and wild type bacteria both localized within cells of the same phenotype in infected spleen at least in the first 72 h p.i. (data not shown). As expected, VirB mutant is no longer detectable in situ at 120 h p.i.
Our analysis revealed that the phenotype of infected spleen cells presented a drastic evolution between 24 h and 120 h p.i. Indeed, at 24 h p.i., about 80% of infected cells were CD11b− F4/80+ r.p. macrophages (Figure 3.A–B), a fraction of which also expressed MHC-II+ molecules (Figure 3.B and Figure 4.A). Between 24 h and 72 h, the percentage of CD11b+ Ly-6G−-infected monocytes increased progressively (Figure 3.B and Figure S6). At 120 h, these monocytes constituted ∼30–40% of the infected cell population (Figure 3.B) and were mainly found clustered in granulomas (data not Shown).
Granulomas were constituted of CD11b+ F4/80+ Ly-6G− MHC-II+ cells surrounded by CD90+ cells (T cells) and Ly-6G+ cells (granulocytes) (Figure 3.A–B, Figure 4.B, Figure 5.A). A fraction of CD11b+ F4/80+ MHC-II+ cells was also CD11c+ (Figure 5.A), suggesting that these cells were inflammatory DC. In previous reports, these inflammatory DC were detected by flow cytometry and identified as the main iNOS-producing cells during Brucella infection [10]. In agreement with this finding, we observed in situ that CD11b+ and CD11c+ cells located in granulomas were iNOS+ (Figure 5.B) and co-localized with mCherry-Br signal. Importantly, B cells (B220+ cells), T cells (CD90.2+ cells) or granulocytes (Ly-6G+ cells) seemed very rarely infected during the course of infection in the spleen (Figure 3.A–B and data not shown). Of note, a weak expression of CD90.2 was detected on a fraction of highly infected cells within granulomas (Figure 3.A). These cells displayed a cellular morphology (large cells with uncondensed nuclei) distinct from T lymphocytes and co-localized with CD11b and F4/80 (data not shown), suggesting that they were not lymphocytes but rather myeloid cells.
In summary, we concluded that infection seems largely limited to a fraction of the myeloid cell compartment in the spleen. Since the establishment of Brucella infection at day 5 in the spleen correlated with monocytes recruitment and formation of granulomas, we hypothesize that these latter are involved in bacterial containment and/or elimination as these structures contain most of the iNOS-producing cells. Granulomas were also detected in r.p. area of spleen from C57BL/6 mice infected for 5 days with 4×104 CFU (Figure S7.A). As expected, their numbers were reduced when compared to mice infected with 108 CFU. Surprisingly, at 12 days p.i. with both doses, granulomas were mostly located in w.p. area, specifically in T cells zones (Figure S7.A–B). Unfortunately, at this time point, we were not able to determine the presence of Brucella since the bacterial load was under our limit of detection. However, the similarity of the phenotype of these structures (CD11b+ and F4/80+, Figure S7.A and S7.B) with r.p. granulomas suggests that they may surround rare infected cells.
It has been frequently reported that BALB/c mice display higher bacterial loads per organs when compared to C57BL/6 mice during Brucella infection [10], [25]. The reasons for this susceptibility are unclear but it has been associated with lower frequency of IFN-γ-secreting CD4+ cells and iNOS-producing cells in these mice [10]. In accordance with CFU counts measured in spleens (Figure S3.A), the average number of bacteria per cell observed by microscopy analysis in BALB/c and C57BL/6 mice was similar at 48 h, though this number was slightly higher at 120 h in BALB/c mice (Figure 1.A). To determine whether enhanced susceptibility of BALB/c mice could be associated with the persistence of bacteria in distinct cell reservoirs, we further examined the phenotype of infected cells in both mouse strains at that time point. As observed for C57BL/6 mice, the large majority of infected cells were mainly located in splenic r.p. and m.z. in BALB/c mice (Figure 1.B–C) and semi-quantitative analyses of infected cell phenotypes (Figure 3.C) showed a global similarity between C57BL/6 and BALB/c mice. However, the frequency of infected cells located in w.p. was higher in BALB/c mice when compared to C57BL/6 mice (Figure 1.B–C). Further comparative analysis (Figure 1.D–E) of the spatial distribution of infected spleen cells showed that highly infected cells (>20 bacteria/cell) were more frequent in BALB/c infected mice and located mainly in w.p. areas.
In conclusion, infected cell phenotypes are relatively similar in both mouse strains but higher colonization of w.p. cells in infected BALB/c mice seems to explain for their higher CFU counts. These cells frequently expressed MOMA-1 or CD11c markers but were negative for CD11b, F4/80, Ly-6G and CD90.2 (data not shown).
Brucella have frequently been described as silent or stealthy pathogens, able to actively escape the immune response through their ability to grow furtively within cells [26]. This hypothesis is mainly based on the fact that Brucella's lipopolysaccharide (LPS) is a weak activator of macrophages and DC by comparison to conventional LPS from Escherichia coli [27]. In our study, we observed that CD11c+ DC represent 10–30% of spleen cells infected by mCherry-Br during the first 12 h of infection (Figure 2.B and Figure 3.A). Therefore, in our experimental conditions (108 CFU i.p.), we investigated the impact of Brucella infection on DC activation in vivo in C57BL/6 mice. Imaging analysis (Figure S8) of spleen sections revealed that 24 h post inoculation, CD11chigh DCs relocated massively to w.p. T cell area.
Flow cytometry analyses demonstrated that DC migration observed in situ at 24 h p.i. was associated with their maturation as shown by cell surface up-regulation of MHC-II and CD86 co-stimulatory molecule expression on CD11chigh spleen cells (Figure S9.A). Although we observed maturation of all splenic DC subsets (CD8− and CD8+) (data not shown), we also found that direct infection of DC was not required to trigger the phenomenon since DC observed by microscopy in w.p. T cell area at 24 h p.i. were not associated with mCherry-Br signal (data not shown). A lower dose of bacteria (4×104 CFU) was able to induce DC maturation but the peak of maturation was reduced and displaced to 48 h p.i. (data not shown). Importantly, administration of heat-killed (HK) Brucella was also able to induce a dose-dependent DC maturation (Figure S9.B), suggesting that Brucella's pathogen-associated molecular patterns (PAMPs) were sufficient to trigger this process. By comparison, HK Escherichia coli induced DC maturation at a ten-fold lower dose (Figure S9.B). Finally, using genetically deficient C57BL/6 mice, we investigated the role of MyD88 and TRIF adaptor molecules in DC maturation process in vivo and found that MyD88, but not TRIF, was important to Brucella-induced DC maturation (Figure S9.C).
Since the liver constitutes another important site of bacterial growth, we investigated the phenotype of infected cells in this organ in order to identify the common characteristics of infected cells in immunological and non-immunological sites. At 10 min p.i., mCherry-Br signal was already detectable in liver sections from infected C57BL/6 and BALB/c wild type mice and co-localized strictly with F4/80+ MHC-II+ Kupffer cells (data not shown). At 120 h post-inoculation, microscopy analyses revealed that, as observed in the spleen, the liver also displayed numerous granulomas composed of F4/80+ MHC-II+ CD11b+ cells (Figure S10). Granulomas frequently surrounded portal space and harbored the majority of infected cells (data not shown). As it was the case in the spleen, iNOS expression was mainly associated with a fraction of granuloma cells expressing CD11c, suggesting that activated monocytes and inflammatory DC were also major populations of these structures. Importantly, we never observed bacteria inside hepatocytes or Ly-6G+ granulocytes suggesting that, like in the spleen, infection is limited to a specific fraction of the myeloid cell compartment.
As granulomas seem to be the main organized cellular structures containing Brucella in the spleen and the liver, we tried to define the role of T and B lymphocytes in their formation. We infected RAG deficient C57BL/6 mice with 108 CFU of mCherry-Br. RAG−/− mice displayed higher CFU counts in the spleen, but not in the liver, at 120 h p.i. (Figure S11). At the same time, microscopy analysis of tissue sections showed that absence of T and B lymphocytes did not impair F4/80+ CD11b+ CD11c+ iNOS+ granuloma formation in either tissues (Figure 6). Flow cytometry analyses confirmed that recruitment of iNOS-DC was not compromised in the spleen of RAG−/− infected mice (Figure S12). However, in RAG−/− infected spleens, granulomas seemed less organized and less dense. In addition, higher numbers of iNOS+ cells and infected cells were observed outside of granulomas and frequency of F4/80− infected cells was higher when compared to wild type mice (Figure 6). These latter observations could be correlated with the presence of Ly-6G+ infected cells in the spleen of RAG−/− mice (data not shown).
In order to gain insight into immune mechanisms controlling bacterial growth in the spleen, we used a genetic approach to investigate the impact of MyD88, IL-12p35 and IFN-γ deficiencies on the phenotype of infected cells with a particular focus on the ability of the immune response to establish granulomas. These molecules have been described as key elements of Th1-driven immune response controlling Brucella growth in vivo [10],[14], [28],[12]. In agreement with previous published results, all deficient mice displayed higher CFU counts in spleen and liver at 120 h p.i. (Figure S11). As expected, frequencies of IFN-γ+ cells and iNOS-DC in infected spleen were drastically reduced in all Th1 deficient mice when compared to infected wild type mice (data not shown and Figure S12).
Immunofluorescence analysis of liver and spleen sections from 120 h-infected mice showed that all three deficient mouse strains displayed dense aggregates of CD11b+ cells surrounding infected cells and resembling granulomas (Figure 7 for spleen and liver from wild type and MyD88−/− mice and data not shown for IL-12−/− and IFN-γ−/− mice). Whereas iNOS staining was strongly reduced in all three deficient mouse strains (Figure 7 and data not shown), CD90.2+ T cells recruitment around infected cells was normal (data not shown). Moreover, granuloma-like structures in all deficient mouse strains frequently contained higher numbers of infected Ly-6G+ cells but reduced density of F4/80+ and CD11c+ cells (Figure 8 for liver granulomas and data not shown for spleen granulomas). These “unconventional” granulomas, characterized by an inverted ratio of granulocytes versus monocytes/DC, failed to control bacterial growth as demonstrated by an increased mCherry-Br signal in deficient mice, with higher numbers of infected cells outside of granuloma like structures. These “unconventional” granulomas were also detected in infected wild type mice but at a low frequency.
One of our initial objectives was to characterize the phenotype of Brucella reservoir cell types in chronically infected mice. Unfortunately, the detection threshold of mCherry-Br signal in situ required approximately 106 CFU/g of tissue (Figure S3.A and B). Wild type mice initially infected with 108 bacteria displayed CFU counts in spleen and liver inferior to this threshold at 12 days p.i. and MyD88−/−, IL-12−/− or IFN-γ−/− mice showed high mortality rate at the same time point. Thus, in order to increase the lifespan of genetically deficient mice, we tested an inoculum dose of 106 CFU and compared bacterial loads in spleen and liver at 5, 12 and 30 days p.i. (Figure S13). Only IL-12p40−/− BALB/c mice displayed CFU counts superior to the detection threshold in spleens at 12 and 30 days p.i. Interestingly, microscopy analysis of tissue sections from spleens from infected IL-12−/− BALB/c mice (Figure S14.B) showed that mCherry-Br signal, initially located in r.p. at 12 h p.i., progressively relocated in w.p. at 5, 12 and 30 days p.i. This preferential localization in w.p. was also observed in infected IL-12−/− C57BL/6 at 5 days p.i. (Figure S14.C), demonstrating that the BALB/c background is dispensable to observe this phenomenon. Surprisingly, the phenotype of Brucella-containing cells located in w.p. of 12 days-infected IL-12−/− BALB/c mice was strikingly distinct from r.p. infected cells previously described in wild type mice (Figure 9.A and 9.B). Indeed, these cells were highly infected (>20 bacteria/cell) and harbored two distinct cell surface marker combinations. When localized in close proximity of m.z. area, the infected cells expressed MOMA-1, a specific marker of metallophilic marginal zone macrophages, whereas when located deeply within the w.p., they expressed CD11c and DEC205/CD205 (Figure 9.B), both DC-specific molecules. Both types of cells were negative for iNOS, F4/80, CD11b, CD86, CD90.2 and B220 (data not shown) and a fraction of them expressed weakly MHC-II. Although the bacterial load was reduced, the phenotype and location of IL-12−/− infected cells was similar 30 days p.i., (data not shown). Interestingly, highly infected CD11c+ cells located in the w.p. of wild-type BALB/c mice could also be observed 5 days p.i. but at a very low frequency (Figure 9.C and data not shown) suggesting that those cells are not limited to Il-12−/− BALB/c mice.
Given that these heavily infected w.p. cells were not found associated with CD11b+ F4/80+ iNOS+ effector cells, we conclude that they may constitute an important reservoir for B. melitensis under permissive conditions such as absence of Th1 protective response.
Bacteria of the genus Brucella, have been long studied as a relevant experimental model to analyze chronic infections of animals and humans due to the great impact they have on both husbandry practice and human health worldwide [29]. Their virulence and the chronicity of the ensuing disease rely on their ability to modulate both innate and adaptive immune responses and the physiology of the host's cells in which they reside, survive and multiply (reviewed in [30]).
The mouse is considered as a useful animal model to investigate the pathogenesis of brucellosis, to identify specific virulence factors of Brucella spp and to characterize the host immune response [29]. If classical studies [31], [32], and more recently the use of genetically deficient mice, have been useful to uncover the great avenues of protective immune responses against Brucella infection [33], this approach appears insufficient to dissect immune effector mechanisms elaborated by the host to fight B. melitensis. Like Mycobacterium tuberculosis [34], [35],[36], and in contrast to other intracellular bacteria such as Listeria monocytogenes [37] or Legionella pneumophila [38], B. melitensis does not irreversibly affect the health status of the majority of mice genetically deficient for key element of Th1 protective immune response (such as MyD88 [10],[39], IL-12p40 [11], IFN-γ [12] and iNOS [11]) during the first four weeks of infection. These observations suggest that immune effector mechanisms against Brucella are multiple and redundant, and necessitate new approaches to be further characterized. This is the perspective from which we decided to use in situ microscopy techniques for studying the local immune response developed around B. melitensis infected cells. Up to now, only gross morphometric and histopathologic analyses have been conducted [40]–[42] on spleen and liver, that are important sites for colonization and replication of Brucella in the mouse model [29]. Both organs develop the so called “histiocytic infiltrates and microgranulomas” [41], [42]. Of importance, Brucella infection in mice results in lesions mimicking those described in chronic infections in humans which may also develop splenomegaly and hepatomegaly [17], [43]. To our knowledge, our study and the recent work from Archambaud, C et al. [44] are the first in vivo investigations characterizing the in vivo phenotype of Brucella-infected cells. Archambaud, C et al. have focused their study on Brucella abortus-infected pulmonary tissues following intranasal inoculation. In contrast, we used an i.p. systemic model of infection in order to analyse the phenotype of Brucella melitensis-infected cells in spleen and liver. The i.p. route of infection is generally used to establish a persistent infection in the mouse because it leads to a rapid systemic distribution of Brucella sp. and to high bacterial loads in both spleen and liver [33], [45].
Brucella has been initially described as an intracellular bacteria able to replicate in professional phagocytes such as macrophages [46], DC [47] and granulocytes as well as non-professional phagocytes [48], including epithelial, fibroblastic and trophoblastic cells, in the context of cell line cultures. However, the identity of Brucella reservoir cell types during the course of infection in vivo is largely uncharacterized. If macrophages [40] and trophoblastic cells [49] have been clearly associated with Brucella infection in the natural hosts, little has been demonstrated regarding the contribution of other potential cellular niches for Brucella in vivo.
In this study, using a Brucella strain stably expressing mCherry, we demonstrated for the first time that these bacteria present a very restrictive cellular tropism as the majority of infected cells in the spleen and the liver of resistant C57BL/6 wild type mice belongs to a specific fraction of the myeloid lineage. Granulocytes, B cells, T cells, fibroblasts and hepatocytes were never found significantly infected. This result may explain the frequent discrepancies between in vitro and in vivo attenuation of various Brucella mutants. In the physiological cell reservoir and in the complex microenvironment of the host, some of the “virulence factors” identified in vitro may have little or no relevance. A striking example is the mutation affecting the gene virB which encodes the type IV secretion system of Brucella. The virB mutant is unable to grow within host cells in vitro [22] though it can replicate at a similar level as the wild type strain during the first 5 days of infection in vivo [23]. In the present study, virB mutant and wild type bacteria were localized within the same cells in infected spleen at all time-points analyzed during the first three days of infection. This argues in favor of the intracellular localization of virB mutant in vivo and strongly suggests that the type IV secretion system of Brucella is tightly regulated during the infectious process. Since we showed that the nature of the infected cells varied in a time-dependent manner during the first days of infection, it is tempting to speculate that the type IV secretion system depends on the nature of the infected cells and/or its activation status.
Our study also brought to light the complexity and the dynamic of the cellular environment of the pathogen during the course of infection. In the spleen, m.z. (MOMA-1+) and r.p. (F4/80+) macrophages are the first infected cells, followed by DCs (CD11c+) located in m.z. and r.p.. We hypothesized that infection-mediated inflammation is responsible of subsequent recruitment of CD11b+ Ly-6G− monocytes. Our data suggest that these cells are rapidly infected and progressively mature to form complex granulomas in r.p. composed of a mix of activated mature monocytes (CD11b+ CD11c− F4/80+ MHC-II+) and inflammatory DC (CD11b+ CD11c+ F4/80+ MHC-II+) and surrounded by T cells (CD90.2+) and granulocytes (CD11b+ Ly-6G+). This scenario yields major differences when compared with the observations by Archambaud et al. [44] in the context of intra-nasal infection with Brucella abortus. Although macrophages are the first infected cells in spleen, liver, as shown in our study, and lungs [44], DC infection and granulomas formation are not observed in lungs but only in draining lymph node [44]. These discrepancies could be due to differences in dose of infection, route of inoculation and bacterial strain used in both studies.
It has been frequently reported that BALB/c mice display higher bacterial loads per organ when compared to C57BL/6 mice during Brucella infection [10], [25]. The reasons of this susceptibility are unclear. In our experimental model, spleen and liver infected cells displayed a closely similar phenotype and localization. However, higher CFU counts observed in infected BALB/c mice was related to increase colonization of w.p. cells. These latter displayed high number of bacteria and expressed MOMA-1 or CD11c markers but were negative for CD11b, F4/80, Ly-6G and CD90.2,
Transient infection of CD11c+ cells during the early stages of infection in the spleen leads us to characterize the impact of Brucella infection on the maturation of conventional DC. Flow cytometry and in situ staining showed that spleen infection with B. melitensis induced the massive migration of DC in the T cell area of w.p. as well as their maturation during the first 24 h of infection, followed later, by their gradual disappearance. The relationship between Brucella and DCs seems complex and necessitates further experiments to be elucidated. However, our results demonstrate clearly that the course of Brucella infection is associated to DCs activation in spleen. This maturation is dependent of MyD88 adapter molecule and does not require infection as heat killed Brucella induce a similar phenomenon. Importantly, microscopy analysis indicated that none of CD11c+ mature cells migrating to T cell area were infected by B. melitensis (data not shown), suggesting that infection of DCs could impair their migration and maybe their maturation. Again, these results are in agreement with the stealthy strategy [30] attributed to Brucella and with in vitro studies [50],[51] demonstrating the ability of Brucella to regulate DC maturation.
Using genetically deficient mice, we partially clarified some requirements for granuloma formation during Brucella infection. Development of a Th1 response seems a critical step as MyD88−/−, IL-12p35−/− and IFN-γ−/− mice displayed iNOS− altered granulomas, strongly enriched in granulocytes. In contrast, the analysis of RAG−/− infected mice demonstrated that granuloma formation at 5 days p.i. is, for a large part, independent of lymphocytes. We hypothesized, in accordance with the granuloma formation model proposed for Listeria monocytogenes infection [16], that IFN-γ is needed for maturation of monocytes in iNOS-DC and building of fully functional granuloma. As previously described [10], IFN-γ was produced by Natural Killer (NK) cells and T lymphocytes during Brucella infection. In RAG−/− mice, iNOS-producing DC (iNOS-DC) were detectable by flow cytometry and microscopy analysis suggesting that absence of IFN-γ-producing T lymphocyte might be compensated by IFN-γ-producing NK cells. This was confirmed by the absence of iNOS-DC in spleens of RAGγc−/− (deficient for natural killer cells) infected mice (data not shown).
The presence of phenotypically similar granulomas in the infected spleen and liver support the hypothesis that granulomas constitute the “spearhead” of the immune response against Brucella. Correlation between high susceptibility of MyD88−/−, IL-12−/− and IFN-γ−/− mice and strongly altered granuloma structures in these mice suggested a causal link between the presence of granulomas and Brucella growth control. However, the precise role of granuloma during Brucella infection remains to be clarified.
A prominent characteristic of Brucella is its capacity to persist in natural host for life (reviewed in [30]). This property is thought to be related to its capacity to (i) locate intracellularly and avoid the fusion of Brucella-containing vacuoles with host cell lysosomes (ii) limit or modulate the activation of innate and adaptive immune response (mainly due to poorly recognized PAMPs of its cell envelope and poorly characterized secreted effectors), (iii) prolong the lifespan of infected cells. The previously unappreciated restricted specificity of Brucella cell tropism that we observed in vivo allows us to hypothesize that the ability of Brucella to persist could also depend of its capacity to infect cell subsets particularly adapted to sustain its growth and persistence. In absence of Th1 response, granuloma formation was altered and the bacterial burden was significantly higher in spleens of IL-12p40−/− BALB/c mice, making possible microscopy observation in situ and characterization of infected cells during chronic phase of infection (12 and 30 days p.i.). Our analyses revealed that the main infected spleen cells at these time-points are located in the w.p. and express the following cell surface phenotype: CD11b− CD11c+ CD90− CD205+ F4/80− Ly-6G− MHC-IIlow B220− iNOS−. High expression of CD11c and CD205 suggested that these cells are a particular subset of DC. This phenotype is also partially reminiscent of foamy macrophages that express high levels of CD11c, CD205 and low level of MHC-II [52] and constitute nutrient-rich reservoirs in lung granulomas of M. tuberculosis-infected mice [53]. Interestingly, these cells were also detected, albeit with a low frequency, five days post-infection in the w.p. of wild type BALB/c mice infected with 108 CFU. Previously, we demonstrated the reduced ability of BALB/c mice to mount Th1 [10] response in the context of Brucella melitensis infection. Thus, permissive environments, such as the absence of Th1 response, may drive the differentiation of macrophages toward a long-lived and anti-inflammatory phenotype allowing bacterial persistence. Identification of these cells as potential reservoir for Brucella in chronically infected mice could help to ameliorate therapeutic treatment of brucellosis.
In conclusion, this work dissected for the first time the nature of the effectors mechanisms developed in vivo by the immune system after B. melitensis systemic inoculation and described the phenotypic characteristics of infected cells during the initial and chronic steps of the infectious process. These results could help develop new strategies to control Brucella infection.
The animal handling and procedures of this study were in accordance with the current European legislation (directive 86/609/EEC) and in agreement with the corresponding Belgian law “Arrêté royal relatif à la protection des animaux d'expérience du 6 avril 2010 publié le 14 mai 2010”. The complete protocol was reviewed and approved by the Animal Welfare Committee of the Facultés Universitaires Notre-Dame de la Paix (FUNDP, Belgium)(Permit Number: 05-558).
Genetically deficient mice in C57BL/6 background: MyD88−/− [54] were obtained from Dr. S. Akira (Osaka University, Japan). TRIF−/− mice [55] were a kind gift from Dr. B. Beutler (The Scripps Research Institute, CA), IL-12p35−/− mice [56] from Dr. B. Ryffel (University of Orleans, France), IFN-γ−/− mice [57] from Dr. S. Magez (Vrije Universiteit Brussel, Belgium), RAG1−/− mice [58] from Dr. S. Goriely (Université Libre de Bruxelles, Belgium), RAGγc−/− mice from Dr. Michel Y. Braun (Université Libre de Bruxelles, Belgium). IL-12p40−/− BALB/c mice were obtained from Dr. V. Flamand (Université Libre de Bruxelles, Belgium). Wild type C57BL/6 mice and BALB/c mice purchased from Harlan (Bicester, UK) were used as control. All mice used in this study were bred in the animal facility of campus Gosselies from the Free University of Brussels (ULB, Belgium).
B. melitensis strain 16M (Biotype1, ATCC 23456) was isolated from an infected goat and grown in biosafety level III laboratory facility. Overnight culture grown with shaking at 37°C in 2YT media (Luria-Bertani broth with double quantity of yeast extract) to stationary phase was washed twice in PBS (3500×g, 10 min.) before use in mice inoculation.
The mCherry protein, a previously described rapidly maturing variant of the red fluorescent protein DsRed [18], was cloned into the suicide vector pKSoriT-bla-kan downstream a strong Brucella spp. promoter [19], [20], *nouvelle ref* Köhler S, Infect Immun. 1999) previously used to express constitutively the GFP. This promoter was called PsojA and described as controlling the expression of the protein translocase SecE (Köhler S.; personal communication).
The vector was constructed as follows: the mCherry coding sequence was amplified by PCR from pRSET-B-mCherry [18] with the mCherry-up and -down primers and ligated into pGemT-Easy (Promega) to generate pGEM-T-mCherry. The mCherry fragment was then excised from pGEM-T-mCherry by HindIII/XbaI double restriction and subsequently cloned into HindIII/XbaI-cut pKSoriT-bla (pBluescript II KS vector from Stratagene in which RP4 oriT was inserted in order to make this vector mobilizable [59]) to generate pKSoriT-bla-mCherry. Meanwhile, a 500 bp fragment upstream the coding sequence of secE was amplified from the B. melitensis genome by PCR with the PsojA-up and PsojA-down primers and ligated into pGemT-Easy (Promega) to generate pGEM-T-SojA. The PsojA fragment was then excised from pGEM-T-PsojA by NotI/XbaI double restriction and subsequently inserted into NotI/XbaI-cut pKSoriT-bla-mCherry to generate pKSoriT-bla-PsojA-mCherry. Finally, the aphA4 cassette (a promoterless kanamycin resistance gene [60] was excised from pUC4aphA4 with SalI, and subsequently cloned into the XhoI site of pKSoriT-bla-PsojA-mCherry to generate plasmid pKSoriT-bla-kan-PsojA-mCherry. This final construct was transformed into E. coli strain S17-1, and introduced into B. melitensis 16M NalR strain by conjugation. Clones that were kanamycin resistant and fluorescent were further checked by PCR, confirming the insertion of the plasmid at the targeted chromosomal PsojA promoter. Primers used in this study (Sequence (5′-3′)): mCherry-up (XbaI) tctagaatggtgagcaagggcgag, mCherry-down (HindIII) aagcttttacttgtacagctcgtcca, PsojA-up (NotI) gcggccgccttgactatggatgcccgtt, PsojA-down (XbaI) tctagactctgtctgatcaggcacaa.
Similar protocol has been used to construct mCherry-expressing ΔVirB B. melitenis. Construction of ΔVirB mutant has been previously described [61].
Mice were injected intra-peritoneally (i.p.) with indicated dose of B. melitensis in 500 µl of PBS. Control animals were injected with the same volume of PBS. Infectious doses were validated by plating serial dilutions of inoculums. At selected time intervals, mice were sacrificed by cervical dislocation. Immediately after being killed, spleen and liver were collected for bacterial count and flow cytometry and microscopy analyses.
For bacterial count, spleens and livers were recovered in PBS/0.1% X-100 triton (Sigma). We performed successive serial dilutions in PBS to get the most accurate bacterial count and we plated them onto 2YT media plates. CFU were counted after 3 days of culture at 37°C.
Spleen were harvested, cut in very small pieces and incubated with a cocktail of DNAse I fraction IX (Sigma-Aldrich Chimie SARL, Lyon, France) (100 µg/ml) and 1.6 mg/ml of collagenase (400 Mandl U/ml) at 37°C for 30 min. After washing, spleen cells were filtered and incubated in saturating doses of purified 2.4G2 (anti-mouse Fc receptor, ATCC) in 200 µl PBS 0.5% BSA 0.02% NaN3 (FACS buffer) for 10 minutes on ice to prevent antibody binding to Fc receptor. 3–5×106 cells were stained on ice with various fluorescent mAbs combinations in FACS buffer and further collected on a FACScalibur cytofluorometer (Becton Dickinson, BD). We purchased the following mAbs from BD Biosciences: Fluoresceine (FITC)-coupled 145-2C11 (anti-CD3ε), 53-8.7 (anti-CD8ε), M1/70 (anti-CD11b), GL-1 (anti-CD86), Phycoerythrin (PE)-coupled HL3 (anti-CD11c), RM4-5 (anti-CD4). Allophycocyanin (APC)-coupled M5/114.15.2 (anti-IA/IE). The cells were analyzed on a FACScalibur cytofluorometer. Cells were gated according to size and scatter to eliminate dead cells and debris from analysis.
Spleen cells were treated as previously described [62]. Spleen cells were incubated for 4 h in RPMI 1640 5% FCS with 1 µl/ml Golgi Plug (BD Pharmingen) at 37°C, 5% CO2. The cells were washed with FACS buffer and stained for cell surface markers before fixation in PBS/1% PFA for 15–20 min on ice. These cells were then permeabilized for 30 min using a saponin-based buffer (1× Perm/Wash, BD Pharmingen in FACS buffer) and stained with one or a combination of the following intracellular mAbs: allophycocyanin-coupled XMG1.2 (anti-IFN-γ; BD Biosciences), purified M-19 (rabbit polyclonal igG anti-NOS2; Santa Cruz Biotechnology) stained with Alexa Fluor 647 goat anti-rabbit (Molecular Probes). After final fixation in PBS/1% PFA, cells were analyzed on a FACScalibur cytofluorometer. No signal was detectable with control isotypes.
Spleens and livers were fixed for 6 h at 4°C in 2% paraformaldehyde (pH 7.4), washed in PBS, incubated overnight at 4°C in a 20% PBS-sucrose solution under agitation, and washed again in PBS. Tissues were embedded in the Tissue-Tek OCT compound (Sakura), frozen, in liquid nitrogen, and cryostat sections (5 µm) were prepared. Tissues sections were rehydrated in PBS, then incubated successively in a PBS solution containing 1% blocking reagent (Boeringer) (PBS-BR 1%) and in PBS-BR 1% containing any of the following mAbs or reagents: DAPI nucleic acid stain, Alexa Fluor 350 or 488 phalloidin (Molecular Probes), purified 1A8 (anti-Ly-6G), or rabbit polyclonal antibodies anti-NOS2 (Calbiochem) (note that M-19 anti-NOS2; used for cytofluorometry analysis is not use for immunofluorescence microscopy), biotin-coupled HL3 (anti-CD11c, BD Biosciences), NLDC-145 (anti-DEC205/CD205, BMA Biomedical AG), MOMA-1 (anti Marginal Zone Macrophages, BMA Biomedicals), 53-2.1 (anti-CD90.2, BD Biosciences), RA3-6B2 (anti-CD45R/B220, BD Biosciences), Alexa Fluor 647-coupled BM8 (anti-F4/80, Abcam) M1/70 (anti-CD11b, BD Biosciences), M5/114.15.2 (anti-IA/IE, eBiosciences). Uncoupled 1A8 mAb and anti-NOS2 polyclonal antibodies were detected using biotin-coupled R67/1.30 (mouse anti-rat IgG2a, BD Biosciences) and Alexa Fluor 647-coupled goat anti-rabbit IgG (Molecular Probes) in PBS-BR 1%, respectively. Biotin-coupled mAbs were amplified using Alexa Fluor 350 or Alexa Fluor 647 Streptavidin (Molecular Probes) in PBS-BR 1%. Slides were mounted in Fluoro-Gel medium (Electron Microscopy Sciences, Hatfield, PA). Labeled tissues sections were visualized with an Axiovert M200 inverted microscope (Zeiss, Iena, Germany) equipped with high resolution monochrome camera (AxioCam HR, Zeiss). Images, 1384×1036 pixels (0.16 µm/pixel), were acquired sequentially for each fluorochrome with A-Plan 10×/0.25 N.A. and LD-Plan-NeoFluar 63×/0.75 N.A. dry objectives and recorded as eight bit grey levels *.zvi files. Colocalization between two stainings was analyzed using the AxioVision Colocalization module (Zeiss). Double positive pixels were rendered in white, gray or yellow as indicated in the Figures. Images were exported as TIFF files and figures prepared in Canvas 7 program.
For the estimation of the number of bacteria by cells or for the phenotype of infected cells, a minimum of 200 cells by condition were examined. These cells were counted in 6 mice minimum, in two independent experiments. When the number of bacteria by individual cell was too high to be determined, the number of bacteria was assumed to be of 20 or more (see Figure 1.E).
Confocal analysis were performed with LSM510 NLO multiphoton confocal microscope fitted on an Axiovert M200 inverted microscope equipped with C-Apochromat 40×/1.2 N.A. water immersion objectives (Zeiss). Optical sections of 1 µm thick, 568×568 pixels (0.1 µm/pixel), were collected sequentially for each fluorochrome and recorded as eight bit grey levels *.lsm files.
We have used a (Wilcoxon-) Mann-Whitney test provided by GraphPad Prism program to statistically analyze our results. Each group of deficient mice was compared to wild type mice. We also compared each group to each other and displayed the result when it is required. Values of p<0.05 were considered to represent a significant difference. *, **, *** denote p<0.05, p<0.01, p<0.001, respectively.
|
10.1371/journal.pgen.1002460 | Cohesin Protects Genes against γH2AX Induced by DNA Double-Strand Breaks | Chromatin undergoes major remodeling around DNA double-strand breaks (DSB) to promote repair and DNA damage response (DDR) activation. We recently reported a high-resolution map of γH2AX around multiple breaks on the human genome, using a new cell-based DSB inducible system. In an attempt to further characterize the chromatin landscape induced around DSBs, we now report the profile of SMC3, a subunit of the cohesin complex, previously characterized as required for repair by homologous recombination. We found that recruitment of cohesin is moderate and restricted to the immediate vicinity of DSBs in human cells. In addition, we show that cohesin controls γH2AX distribution within domains. Indeed, as we reported previously for transcription, cohesin binding antagonizes γH2AX spreading. Remarkably, depletion of cohesin leads to an increase of γH2AX at cohesin-bound genes, associated with a decrease in their expression level after DSB induction. We propose that, in agreement with their function in chromosome architecture, cohesin could also help to isolate active genes from some chromatin remodelling and modifications such as the ones that occur when a DSB is detected on the genome.
| Genomic stability requires that deleterious events such as DNA double-strand breaks (DSBs) are precisely repaired. The natural compaction of DNA into chromatin hinders DNA accessibility and break detection. Therefore, cells respond to DSBs by triggering multiple chromatin modifications that promote accessibility and facilitate repair. We have recently developed a novel system whereby a restriction enzyme can be induced to inflict multiple DSBs across the human genome. This system permits high-resolution characterization of changes in the chromatin landscape that are induced around DSBs. While we previously reported the profile of H2AX phosphorylation (a primary event in chromatin remodelling that takes place in response to DSBs), we now provide the high resolution mapping of cohesin, a complex implicated in the 3-D organisation of chromosomes within the nucleus. Unexpectedly, we have discovered that cohesins play a role in the maintenance of gene transcription in regions where chromatin has been remodelled during the DSB response.
| DNA packaging into chromatin hinders detection and repair of DNA Double Strand Breaks (DSBs), and therefore DSB repair occurs simultaneously with multiple chromatin modifications, including histone acetylation, ubiquitylation and phosphorylation, as well as ATP dependant nucleosome remodelling and chromatin protein deposition or exclusion (for review [1],[2]). These chromatin changes not only generate a chromatin state permissive to DNA repair, but also contribute to DSB signalling and checkpoint activation. Phosphorylation of H2A in yeast or H2AX in mammals (referred to γH2AX) occurs rapidly, within a few minutes, and is considered to be one of the first DSB-induced chromatin modifications. While γH2AX is not required for the initial recruitment of repair proteins onto DSBs, it is necessary for the proper assembly of repair foci (also called IRIF, for IRradiation Induced Foci) and full activation of the DNA Damage Response (DDR) [3], [4]. H2AX deficient mice are radio-sensitive and subject to increased genomic instability [5], highlighting the critical function of γH2AX in vivo. Remarkably, γH2AX spreads across large chromatin domains surrounding DSBs, around 50 kb in yeast [6] and up to 2 Mb in vertebrate cells [7]–[11]. Until recently, the mechanism(s) underlying such wide spreading, as well as its consequences on chromatin activity and gene transcription were unclear. Indeed, several lines of evidence indicated that DSB generation triggers RNA Pol II and Pol I exclusion/pausing at break sites and inhibits transcription of proximal genes in an ATM dependent manner [12], [13]. However, whether and how transcription was affected further distally from the break in γH2AX domains remained elusive [6], [14]. Recently, we developed a stable human cell line, designed for controlled, sequence-specific DSB induction, based on the expression of an 8 bp restriction enzyme (AsiSI) fused to the oestrogen receptor ligand binding domain. Using this system, we monitored γH2AX distribution and changes in transcription, around more than 20 DSBs located on chromosomes 1 and 6 using ChIP-chip [10].We uncovered that γH2AX spreads unevenly over megabases of surrounding chromatin, avoiding transcribed genes. Within γH2AX domains, we found that gene transcription remained unchanged upon DSB induction [10]. We suggested that the γH2AX profile reflects the spatial organisation of chromatin and proposed a 3-dimensional model, which accounts for the accurate maintenance of gene transcription proximal to DSBs via their exclusion outside of γH2AX foci.
In addition to γH2AX, evidence suggests that cohesin plays a critical role in DSB repair (for review [15], [16]). Cohesin is a multi-subunit complex, thought to embrace DNA as a ring-shaped structure, that mediates sister chromatin cohesion and ensures accurate chromosome segregation. It consists of the proteins SCC1 (also termed Rad21/Mcd1p), SCC3 (SA1 and SA2 in human somatic cells) and the heterodimer SMC3/SMC1. In yeast, cohesin is recruited over a 50 kb chromatin domain surrounding an HO-induced break [17]–[19]. In vertebrate cells, cohesins are targeted to chromatin upon ionizing radiation [20] and to DSBs induced by X ray stripes and laser tracks during G2 [21]–[23], although this may only occurs at very high power settings [22]. However, ChIP studies clearly showed that SMC1 and SCC1 are recruited to an I-SceI-induced DSB [24], suggesting that loading of cohesin at DSBs also occurs in mammalian cells. Cohesin promotes equal homologous recombination between sister chromatids and prevents homologous recombination between repeats or homologous chromosomes [24]–[28]. In addition, its function in DSB repair depends upon cohesion establishment, a phenomena known as DIC (Damage Induced Cohesion) ([29]–[32] for review [33]). This led to the proposal that cohesin may participate in post-replicative DNA repair by ensuring proper cohesion between sister chromatids thus facilitating homologous recombination with the sister locus.
Importantly, beyond its role in DSB repair and sister chromatid cohesion, another function for cohesin has recently emerged. In vertebrates, the cohesin complex accumulates at specific loci, mainly enhancer/promoters and sites bound by the CTCF insulator protein [34]–[36]. There, it participates in the transcriptional control of neighbouring genes, most likely through its ability to mediate long-range interactions between chromatin fibers, thereby allowing enhancer/promoter interaction and/or insulation from the surrounding chromatin [34]–[37]. More generally, cohesins are now believed to play a critical role in genome organization, participating in loop formation and thus affecting various DNA-based processes such as transcription and replication [38].
Given the multiple roles of cohesin in DSB repair, higher-order chromatin structure and transcriptional control, we decided to characterize the cohesin profile around AsiSI-induced DSBs in order to both further refine its function in DSB repair and its potential impact on γH2AX spreading. Here we show that, in contrast to yeast, cohesin is only moderately recruited to AsiSI-induced DSBs in human cells and does not spread over more than 5 kb. Remarkably, cohesin binding antagonizes γH2AX accumulation within γH2AX domains. Depletion of the SCC1 cohesin subunit leads to both an increase in γH2AX and a DSB-dependent transcriptional downregulation of genes within γH2AX domains, suggesting that cohesins are, at least in part, responsible for the accurate transcriptional control observed in γH2AX domains. Finally, we also analyzed the consequences of cohesin depletion on the positions of γH2AX domain boundaries, and found that while most of these boundaries remained unaffected, at some genomic locations cohesin helped to confine γH2AX spreading.
We recently developed a human cell line that stably expresses an AsiSI-ER fusion restriction enzyme (the AsiSI-ER-U20S cell line). Treatment with hydroxytamoxifen (4OHT) triggers nuclear localisation of the enzyme and induces DSBs at defined genomic loci, enabling ChIP analyses of protein recruitment at DSBs [10]. In order to better understand the function of cohesin in DSB repair, we thus performed ChIPs against various human cohesin subunits before and after break induction. The specificity of homemade antibodies was first confirmed using western blot, immunoprecipitation and ChIP assays on a known cohesin-binding site [35] (Figure S1 and Figure S2.). We found that 4OHT treatment induced the targeting of SMC3, SCC1 and SCC3 (SA1 and SA2) at AsiSI-induced DSBs (Figure 1A, 1B, 1C respectively) indicating that the full complex is likely to be recruited at DSBs.
Since it was previously reported that cohesins may target DSBs preferentially in the G2 phase of the cell cycle [21], we monitored SCC1 recruitment in G2 arrested AsiSI-ER-U20S cells following a RO-3306 treatment. We did not find a major difference in loading of SCC1 onto DSBs when compared with asynchronous cells (Figure S3A). In addition, we also used the AsiSI-ER-T98G cell line [10], [11] that can easily be synchronized by serum starvation, to monitor cohesin recruitment in G1 and G2 synchronized cells. Again, SCC1 DSB-targeting was similar in G1 and G2 (Figure S3B). A ChIP performed at 14 hour after 4OHT treatment ensured that SCC1 recruitment did not change drastically at a later time point (Figure S4).
We therefore decided to perform SMC3 ChIP-chip experiments in asynchronous cells, before and after 4 hours of 4OHT treatment, using human Affymetrix tiling arrays covering chromosomes 1 and 6, in order to simultaneously investigate the distribution of cohesins around multiple DSBs with high resolution. On these two chromosomes, the SMC3 distribution in untreated AsiSI-ER-U20S cells was similar to the distribution of SCC1 reported for HeLa cells [35] (see examples Figure S5A). 37.6% of SMC3 binding sites identified in AsiSI-ER-U20S were also identified using the SCC1 dataset from HeLa cells. Both SCC1 and SMC3 signals showed a clear enrichment at transcription start sites (TSS) (Figure S5B), consistent with the fact that a significant proportion of cohesin binding sites are located in close proximity to promoters [34]–[36], results which confirm the validity of our ChIP-chip data.
Strikingly, we found that recruitment of SMC3 at DSBs induced by 4OHT treatment was moderate and did not spread widely around the DSB to form a γH2AX-like domain, but rather localized within close proximity to the break (Figure 1D). When averaged around the 24 AsiSI-induced DSBs on chromosomes 1 and 6 ([10]; see Table S1 for a list of AsiSI sites), the SMC3 profile showed a weak increase upon 4OHT addition over a ∼5 kb region surrounding the DSB (Figure 1E). Although weak, we found that this increase of SMC3 after 4OHT treatment at the vicinity of AsiSI sites (on a 2 kb window) was significant (p<0.05) (Figure 1F and Figure S6).
In order to confirm that cohesin did not spread around DSBs in our cell line, we performed ChIP followed by Q-PCR analyses using primer pairs located at various positions from a DSB. Both SMC3 and SCC1 showed a clear increase upon 4OHT treatment at the immediate vicinity of the break, recruitment that was undetectable further away from the DSB (Figure 1G). Importantly, several labs previously reported an extended recruitment of cohesin over 50 kb domains around a single HO-induced DSB in yeast [17]–[19]. Since in our cell line, 4OHT treatment induced over a hundred DSBs [11], we wondered whether the lack of cohesin spreading observed here could be due to a limiting amount of free cohesin or/and available cohesin loaders, for targeting at DSBs. In order to address this point, we first controlled the amount of soluble cohesin (unbound to chromatin) in the nucleus before and after 4OHT treatment. Both SCC1 and SMC3 were still present in the soluble fraction after DSB induction (Figure S7A), indicating that free cohesins are not a limiting factor in these conditions. In addition, we also performed a SCC1 ChIP in an I-SceI-ER U20S cell line, in which one single DSB is induced upon 4OHT treatment. As observed on AsiSI-induced DSBs, we could detect a 4OHT-dependant increase of SCC1 at the I-SceI-induced DSB (300 bp), but not at 2.4 kb from the DSB (Figure S7B). Thus the high amount of DSBs induced by AsiSI over the human genome is not responsible for the lack of spreading observed in human cells.
Altogether, our data indicate that in human cells, cohesin is moderately targeted to DSBs and that it does not spread over wide chromosomal domains in contrast to yeast.
During the course of previous studies, we noticed that γH2AX within domains tended to decrease on cohesin peaks identified in HeLa cells [35] (Figure S8). Thus we next compared the cohesin distribution obtained in our AsiSI-ER U20S cells in absence of DSB induction, with our previously reported γH2AX profile. Within γH2AX domains, areas showing low levels of γH2AX (“holes”) often coincided with peaks of SMC3 monitored before 4OHT treatment (Figure 2A). We retrieved the γH2AX peak/hole positions within domains (see Material and Methods) and averaged the profile of SMC3 across their borders. γH2AX peak/hole transition coincided with a change in the SMC3 profile (Figure 2B). In addition, we also found that the genes showing high levels of SMC3 rather harbour low level of γH2AX (Figure S9). In order to confirm these data we also profiled SCC1 in our cell line under normal conditions. Again, the SCC1 distribution in AsiSI-ER U20S cells was similar to the profile characterized in HeLa cells (Figure S10A–S10B), and 43% of the binding sites in AsiSI-ER U20S cells, overlapped with binding sites in HeLa cells. We found that, as observed with SMC3, SCC1 peaks coincided with γH2AX holes, and that SCC1 rich genes showed low levels of γH2AX (Figure S11).
Altogether these results suggest that the cohesin present onto chromatin before any DSB induction antagonizes γH2AX establishment/maintenance.
To test this hypothesis we analysed by ChIP-chip the γH2AX profile upon SCC1 depletion by siRNA. Depletion of this subunit has been shown to also trigger an almost complete disappearance of SMC3 from chromatin [35]. SCC1 siRNA [35] was highly efficient since both RNA and protein levels were strongly reduced (Figure S12A–S12B). In addition, chromatin-bound SCC1 was also efficiently depleted by siRNA as shown by ChIP (Figure S12C). We observed that within domains, γH2AX signals increased in SCC1 depleted cells when compared to cells transfected with control siRNA (Figure 3A left panel, Figure S13 upper and middle panels, and Figure S14A). This was also confirmed by Q-PCR analyses of γH2AX ChIP in control and SCC1 depleted cells (Figure S15), using primers pairs located at various positions from the DSB in five different γH2AX domains.
This increase was not detected elsewhere on the genome indicating that it was not due to an effect of SCC1 depletion on basal levels of γH2AX (Figure 3A right panel, Figure S13 lower panels and S14B). Our cleavage assay indicated that SCC1 depletion did not change the efficiency of AsiSI site cutting (Figure S16). Therefore, the enhanced phosphorylation of H2AX observed in SCC1 depleted cells was not due to an increase in AsiSI-ER activity, but rather to some modification(s) of the establishment or maintenance of γH2AX on chromatin. Importantly, we could also detect this increase by immunofluorescence (Figure S17), and changes in γH2AX levels upon SCC1 depletion have also been observed by western blot using irradiated cells [39], which further support our findings.
Furthermore, we found that the γH2AX increase observed upon SCC1 siRNA transfection occurred preferentially on cohesin-bound chromatin (Figure 3B). The ratio of γH2AX in SCC1-depleted versus control cells, averaged over an 80 kb window around each of the 24 AsiSI sites, correlates with the level of both SMC3 and SCC1 averaged over the same window (Figure 3C and Figure S18). This strongly suggests that the effect of cohesin on γH2AX is direct and mediated in cis in chromatin, rather than due to a global increase of signalling and kinase activity within the cell.
We next examined in more detail the behaviour of γH2AX in SCC1 depleted cells, more specifically on the genes contained within γH2AX domains. We reported previously a decrease in γH2AX signal at Transcriptional Start Sites (TSS) within γH2AX domains [10]. This decrease was practically undetectable in SCC1-deficient cells, when compared to siRNA control cells (Figure 4A). Accordingly, in cells transfected with SCC1 siRNA we could observe a significant increase of γH2AX at promoters compared to control cells, whereas this increase was much less pronounced upstream or downstream TSS (Figure S19). This indicates that SCC1 depletion triggers an abnormal accumulation of γH2AX at TSS. We observed that this behaviour preferentially affects genes normally bound by cohesin (Figure 4B). For each of the 359 genes embedded in γH2AX domains, we calculated the SMC3 signal and the ratio of γH2AX in siRNA SCC1/siRNA CTRL transfected cells. When plotted against each other we could see a significant correlation (Figure 4C). The same was true when SCC1 signal was plotted (Figure S20). Along the same line, genes on which γH2AX increased the most after SCC1 depletion, significantly showed more SMC3 (upper panel) and SCC1 (lower panel) (Figure S21). This strongly suggests that the presence of cohesin prevents γH2AX spreading on genes.
We confirmed these data by Q-PCR on selected SMC3-bound (ARV1, CTNNBIP1, GNAI3, ATXN7L2, and AMIGO1) and two SMC3-unbound (GBP5 and GBP6) genes (Figure S22). Transfection with SCC1 siRNA increased γH2AX levels up to twofold on the SMC3-bound genes but did not affect the SMC3-unbound regions (Figure 4D). Altogether our data indicate that cohesin directly controls the accumulation of γH2AX on chromatin and at promoters.
Gene transcription remains unaffected within AsiSI-induced γH2AX domains and active genes harbour low levels of γH2AX [10]. Since cohesin depletion led to an increase in γH2AX at cohesin-bound genes, we wondered whether transcription was still maintained after DSB induction in this cohesin-deficient context. We performed RT-QPCR for eight genes located within γH2AX domains, before and after break induction in control and SCC1 depleted cells. As expected, since cohesin plays a role in transcriptional regulation, SCC1 depletion affected the transcription of some of the tested genes, without DSB induction (Figure S23). As previously reported [10], 4OHT treatment did not alter gene expression in SCC1-proficient cells (CTRL siRNA). In contrast, gene expression decreased after 4OHT treatment in an SCC1 depleted background (Figure 5), indicating that cohesin helps to ensure normal gene expression in γH2AX domains after DSB induction.
Finally, we examined the behaviour of γH2AX upon SCC1 depletion at γH2AX domain boundaries. Cohesin has been proposed to mediate long range interactions and to play a role in chromosome looping and 3-dimensional organisation. Thus, it appears as an intriguing candidate for restraining γH2AX spreading within defined chromosomal domains. Using γH2AX domain boundaries identified in control transfected cells (Table S2), we observed a wider spreading of γH2AX in SCC1 depleted cells than in control cells (Figure 6A). However, when we looked individually at each AsiSI-induced γH2AX domain, we found that some domains appeared to be cohesin-independent while others showed extended spreading upon SCC1 depletion. This difference was not a consequence of elevated γH2AX within domains, since among domains that incurred a similar increase in γH2AX upon SCC1 depletion, some domains showed extended spreading (Figure 6B top panel) while others did not (Figure 6B bottom panel). The extended spreading observed on this domain was further confirmed by QPCR analysis using primers pairs at various locations (Figure 6C). One possibility is that cohesins are directly involved in a subclass of domain boundaries where they act to constrain spreading. However, we could not find a correlation between cohesin distribution and boundary positions either globally (Figure S24A) or individually (Figure S24B). Thus, it is unlikely that cohesins are physically involved in defining the limits of γH2AX domains. Alternatively, the global increase of γH2AX that occurs upon SCC1 depletion could account for the extended spreading on chromatin (such as Figure 6B top panel) unless some other specific features constrain this spreading (such as on the domain Figure 6B bottom panel).
Taking advantage of our recently described inducible system to generate sequence specific DSBs at multiple positions, we have investigated the recruitment of cohesin at DSBs in human cells. Consistent with previous reports [21]–[24], we observed an increase of several cohesin subunits at break sites. As in yeast this recruitment likely depends on H2AX phosphorylation, since significant decrease in SMC3 and SCC1 targeting was observed when using an ATM inhibitor ([20] and our unpublished data). However, we found that cohesin recruitment was very moderate and restricted to the immediate vicinity of the DSB which is in stark contrast to the 50-kb wide cohesin loading that occurs in yeast around HO-induced DSBs [17]–[19]. Importantly, since in our system doing ChIP after 4H of 4OHT treatment allows studying all recruitment events that occur at a DSB between 0H and 4H of repair (as once in the nucleus the enzyme cuts and re-cuts the site), this difference is unlikely to be due to a difference in the kinetics of cohesin recruitment at DSBs. We also performed cohesin ChIP at 14H post-break induction, in order to make sure that in human cells cohesin targeting does not occurs at very late time point (Figure S4). In addition, we also controlled that such a restricted cohesin recruitment was not due to the high amount of DSBs induced in our cell line. We showed that soluble cohesins were not limiting after DSB induction and that a similar cohesin recruitment pattern was also observed in an I-SceI cell line (single cut) (Figure S7). Thus, altogether, our data show that cohesins are only recruited to the vicinity of a DSB in human cells contrarily to the extended cohesin spreading observed in yeast.
Accumulation of cohesin around DSBs has been proposed to enhance cohesion between sister chromatids in order to promote efficient repair by homologous recombination (HR) (for review [33]). While HR accounts for the majority of repair events in yeast, DSBs are mainly repaired by Non Homologous End Joining events in mammalian cells, even during G2 phase [40]. This could thus account for the difference of cohesin spreading observed between yeast and mammalian cells. Several additional differences exist in the behaviour of cohesin complexes between yeast and metazoan. For example, yeast cohesins have been proposed to translocate along chromatin fibers, eventually accumulating at sites of convergent transcription [41]. In contrast, Drosophila and mammalian cohesins do not show any preference for convergent genes and accumulate at promoters and CTCF binding sites [34]–[36], [42]. These differences in cohesin distribution may reflect basic differences in the organization of yeast and metazoan genomes, the former being smaller and more compact, with a higher density of transcribing genes. They might also be indicative of different cohesin targeting mechanisms, which could also partake in the different localizations observed at DSBs. Finally, the absence of cohesin spreading in human cells may be compensated for by post-translational modifications that increase cohesion. Acetylation and phosphorylation of cohesin subunits at various residues are suspected to play critical roles in regulating the ATPase and translocase activity, as well as the cohesion properties of the cohesin complex (for review [15]). Thus, follow up investigations into the distribution of cohesin modifications upon DSB induction may reveal the molecular basis for the observed differences in localization between yeast and human cells. More specifically, residues 966 and 957 of SMC1, which are phosphorylated by ATM in response to damage [43]–[46], are not conserved in yeast and it is thus tempting to speculate that they could act to promote cohesion using preloaded cohesins in mammalian cells.
We found that depletion of cohesin leads to a global increase of γH2AX after DSB induction (both using ChIP and immunofluorescence), in agreement with reports of γH2AX increase in irradiated, SCC1- and SMC3-depleted cells [39]. While this increase was moderate, it was reproducible and observed at several γH2AX domains (Figure S15). Our data indicate that the removal of cohesin from chromatin triggers an accumulation of γH2AX in cis, since this increase is found preferentially on regions normally enriched in cohesin. One hypothesis is that cohesin inhibits the establishment of H2AX phosphorylation, for example by counteracting ATM activation or/and recruitment. Alternatively, the increase in γH2AX upon SCC1 depletion could reflect impairment in the recruitment of phosphatases at breaks, such as PP2A [47]. It is interesting to note that Sugoshin, a protein that interacts with the cohesin complex and regulates cohesion in mitosis and meiosis, also interacts with PP2A [48]–[50]. One could thus envisage that cohesin recruits PP2A to chromatin and thereby regulates γH2AX levels.
Both Pol II and Pol I transcription are down regulated in the vicinity of a DSB in an ATM-dependent manner [10], [12], [13]. Whether this extinction is induced by γH2AX is not clear, since inhibition of Pol I is independant of H2AX [12] and inhibition of Pol II at least in yeast, appears to be dependent on resection rather than on γH2AX spreading [6]. We recently reported that the transcription of genes within γH2AX domains, but further away from a break, remains unchanged and that these active genes harbour reduced levels of γH2AX [10]. Here we found that this maintenance of transcription in γH2AX domains is impaired upon cohesin depletion.
First we observed a moderate but general increase of γH2AX levels on the cohesin-target genes encompassed in γH2AX domains after cohesin depletion, indicating that cohesins contribute to maintain reduced γH2AX level on genes. Secondly, for eight genes located in various domains, this was associated with a significant DSB-dependant transcriptional decrease. Since the effect of cohesin depletion on γH2AX levels occurred on most genes of the domains, it is likely that the trend observed on these eight genes is a general feature illustrating the role of cohesin in transcriptional maintenance, although further genome wide studies would be required to generalize our findings. It is also important to underline that both γH2AX increase on genes and DSB-dependant transcriptional decrease in cohesin depleted cells were quite moderate, and thus, while this could be to due siRNA efficiency, or to an asynchronous cleavage of AsiSI sites in the cell population, we also cannot exclude that other unrelated factor participate in the protection of active genes in γH2AX domains.
This cohesin-dependant gene protection is unlikely to be a damaged–induced process since cohesin recruitment at DSB only occurs on the surrounding 2kb. Instead we favour the hypothesis that the cohesins already present on a normal, undamaged genome could protect active genes from the chromatin changes induced by DSBs, such as γH2AX which has been proposed to enhance chromatin compaction [51] and could therefore be deleterious for transcription. Interestingly, many recent studies have established a clear link between the ability of cohesin to regulate transcription and its ability to mediate chromosome looping. It is thus tempting to speculate that cohesin could protect transcription in γH2AX domains, by maintaining transcribed loci outside of γH2AX foci. This would allow to both keep low levels of γH2AX on active genes and to ensure their correct transcription post DSB induction (Figure 7).
Since cohesins are known to mediate chromatin looping, they could also be involved in anchoring the chromosomal domain, within which γH2AX would spread. While we found that cohesin depletion triggers boundary expansion at some domains, we could not find a corresponding enrichment in cohesin at those positions (Figure S24). Thus, it is unlikely that cohesin plays a direct role in anchoring γH2AX domains. We believe that the global increase in γH2AX levels that occurs in the absence of cohesin, leads to extended spreading farther away from the break unless some specific constraints counteract γH2AX propagation. In order to get insights into the nature of these potential constraints, we have compared our data with the recently published Hi-C mapping of long range chromosomal interactions [52], which identified the positions of chromosomal domains, amongst other features. Remarkably, a significant proportion of γH2AX domain boundaries correlated with chromosomal domain transitions (Figures S25 and S26). In conclusion, we believe that γH2AX spreads around DSBs until it naturally fades away or it encounters a chromosomal domain transition. Fading is likely dependent on factors such as the distance from the break and the intensity of γH2AX induction, and thus cohesin depletion would trigger extended γH2AX spreading due to higher levels of γH2AX. In contrast, chromosomal domain transition stops propagation regardless of γH2AX levels, and it is unlikely that cohesins are involved in these domain transitions, since these boundaries were intact upon SCC1 depletion (not shown).
In summary, our results suggest that phosphorylation of H2AX after DSB is established on a pre-existing chromatin/chromosomal organization (Figure 7). While further investigations are required to validate such a hypothesis, it is interesting to point out that if true, γH2AX spreading might thus be used as read-out of 3-dimensional chromosome structure.
Rabbit polyclonal antibodies against SMC3 were raised using recombinant SMC3 (αSMC3-A) or an SMC3 peptide (αSMC3-B) and have been described in [53]. They were further tested and validated in human cells in [38], and in the present manuscript (Figure S1). The rabbit polyclonal antibody against SA1 was raised using a C-terminal peptide as immunogen (CEDDSGFGMPMF) and has been validated by ChIP in mouse cells (Remeseiro et al., submitted), and in human cells (this manuscript, Figure S2). The rabbit polyclonal antibody against SA2 was made against a peptide within the C-terminal region of hSA2 “EPKRLRPEDSFMSV”, and affinity purified against the antigen. This antibody was validated against human proteins in Figure S2.
AsiSI-ER-U20S and I-SceI-ER-U20S cells were cultured in Dulbecco's modified Eagle's medium (DMEM) supplemented with antibiotics, 10% FCS (Invitrogen) and 1 µg/mL puromycin at 37°C under a humidified atmosphere with 5% CO2. AsiSI-ER-T98G cells were cultured in Minimum Essential Media (MEM) GlutaMAX, supplemented with MEM Non Essential Amino Acid (NEAA), antibiotics, and 10% FCS (Invitrogen).
Synchronization of AsiSI-ER-T98G cells was achieved by 72 hours of serum starvation (0% FBS). Cells were collected in G1 and G2 phase after 10H and 28H, respectively, of 20% FBS re-induction. Synchronisation of AsiSI-ER-U20S in G2 was achieved by an 18H R0-3306 (Calbiochem) 9 µM treatment. For siRNA transfection, 5.0×106 cells were electroporated with 10 µL of 100 µM siRNA using the Cell Line Nucleofector kit V (Amaxa), according to the manufacturer instructions, and collected 48H after transfection. Sequences for siRNA are displayed Table S3. When indicated, cells were treated with 300 nM 4OHT for 4H or 14H.
RNA was extracted using the RNAeasy kit (Qiagen) following manufacturer instructions. 1 µg of RNA was reverse transcribed using Im-PromII RT (Promega) with random hexamers. cDNAs were analyzed by Q-PCR using primers described in Table S3 and normalized against P0 cDNA levels.
Cell pellets (5.106 cells) were fractionated as reported [54]. Briefly, cells were first resuspended for 15 min on ice in 200 µl of 50 mM HEPES pH 7.5, 150 mM NaCl, 1 mM EDTA supplemented with the mini protease (Roche) and phosphatase inhibitor cocktail (Sigma). Following centrifugation at 14000 rpm for 5 min, the supernatant was collected (fraction I), and pellets were incubated in 200 µl of the same buffer supplemented with 0.1% triton for 15 min at 4°C. The supernatant was collected as before (fraction II). The pellets were further extracted in 200 µL of the same buffer supplemented with 0.1 mg/mL RNAse (Abcam) for 30 min on ice. The extracts were clarified by centrifugation at 14 000 rpm for 5 min (fraction III). Pellets were next resuspended in 200 µl extraction buffer supplemented with 10 nM MnCl2 and 0.07 mg/mL DNAse1 for 30 min at room temperature Western blot were performed using Invitrogen precast gels and buffer following manufacturer instructions, and using an anti-SMC3 (αSMC3 A), an anti-SCC1 (Ab992 rabbit), or an anti-H3 (Ab1791-100 rabbit).
ChIP assays were carried out according to the protocol described in [55] with the following modifications. 200 µg of chromatin was immunoprecipitated using 2 µg of anti-γH2AX (Epitomics), anti-rad21 (SCC1) (Abcam ab992), anti-SMC3 (a mixture of the two rabbit homemade antibodies), anti SA1 (rabbit homemade antibody), anti SA2 (rabbit homemade antibody) or without antibody (mock). For ChIP-Q-PCR, immunoprecipitated and input DNA were analysed in triplicate by real time Q-PCR (primer sequences are provided Table S3). IP efficiency was calculated as percent of input DNA immunoprecipitated, on positive loci (such as close to a DSB) and on a negative locus (devoid of DSB). Data were expressed relative to the signal obtained on the negative locus. For ChIP-chip, DNA was amplified, labelled, and hybridized to high density oligonucleotide tiling arrays covering human chromosome 1 and 6 (Affymetrix Human Tiling 2.0R-A), using the standard Affymetrix procedure, by the GeneCore facility at EMBL Heidelberg.
Scanned array data were normalized using Tiling Affymetrix Software (TAS) (quantile normalization, scale set to 500) and analyzed as described in [10]. Peaks and boundaries of γH2AX domains were determined using our home made algorithm (described in [10]). Briefly, this algorithm was inspired from [56] and allows determining enriched domains of any size. Domains are determined through a two-step process. The first step defines zones of interest as contiguous sections of N probes in which x% of the probes are above a certain threshold. Second step allows bidirectional zones extension from theses seeds, to refine their limits (also based on % of probes above a certain threshold). These zones can next be merged and filtered based on their size and values. For cohesin peaks identification, we used the following settings: Contiguous sections of 20 probes with at least 17 probes above the threshold were identified (threshold was based on the percentage of graph values greater than 90% on individual chromosomes).
In order to plot data with respect to transcription start sites (TSS), gene transcript positions and orientations were obtained from the refFlat table from UCSC (hg18). All genomic coordinates were from the genome assembly NCBI Build 36.1, and annotations were retrieved from the UCSC genome browser http://genome.ucsc.edu.
Microarray probe coordinates and data have been submitted to Array Express under accession number E-TABM-1164.
The full procedure for the cleavage assay has been previously described [10]. Briefly biotynilated double stranded oligonucleotide were ligated overnight to genomic DNA extracted from 4OHT treated or untreated AsiSI-ER-U20S cells. T4 ligase was heat inactivated at 65°C for 10 min, and DNA was fragmented by EcoRI digestion at 37°C for 2 h followed by heat inactivation at 70°C for 20 min. After a preclearing step, DNA was pulled down with streptavidin beads (Sigma) at 4°C overnight, and then washed 5 times in RIPA buffer and twice in TE. Beads were resuspended in 100 µL of water and digested with HindIII at 37°C for 4 h. After phenol/chloroform purification and precipitation, DNA was resuspended in 100 µL of water, and submitted to Q-PCR, using primers described in Table S3.
After transfection with siRNA, and 4OHT treatment for 4H, cells were fixed in PBS containing 3.7% formaldehyde for 15 min at RT, permeabilized in PBS-0.5% Triton X100 for 10 min, and blocked with 3% bovine serum albumine (BSA) for 30 min. After 2 h incubation with γH2AX antibody (Cell Signalling), cells were washed with PBS and probed for 1H with an Alexa Fluor 594 anti mouse antibody (Molecular Probes). Slides were mounted with Vectashield (Vector Laboratories), and images were acquired using a Leica microscope equipped with a charge-coupled device camera (CoolSNAP ES; Roper Industries), and the MetaMorph software (MDS Analytical Technologies).
Quantification of fluorescence levels was done on a least 100 nuclei using home-developed macros in ImageJ software (National Institutes of Health, Bethesda, MA) to normalize background, thresholds and measures.
|
10.1371/journal.pgen.1004625 | BMPs Regulate msx Gene Expression in the Dorsal Neuroectoderm of Drosophila and Vertebrates by Distinct Mechanisms | In a broad variety of bilaterian species the trunk central nervous system (CNS) derives from three primary rows of neuroblasts. The fates of these neural progenitor cells are determined in part by three conserved transcription factors: vnd/nkx2.2, ind/gsh and msh/msx in Drosophila melanogaster/vertebrates, which are expressed in corresponding non-overlapping patterns along the dorsal-ventral axis. While this conserved suite of “neural identity” gene expression strongly suggests a common ancestral origin for the patterning systems, it is unclear whether the original regulatory mechanisms establishing these patterns have been similarly conserved during evolution. In Drosophila, genetic evidence suggests that Bone Morphogenetic Proteins (BMPs) act in a dosage-dependent fashion to repress expression of neural identity genes. BMPs also play a dose-dependent role in patterning the dorsal and lateral regions of the vertebrate CNS, however, the mechanism by which they achieve such patterning has not yet been clearly established. In this report, we examine the mechanisms by which BMPs act on cis-regulatory modules (CRMs) that control localized expression of the Drosophila msh and zebrafish (Danio rerio) msxB in the dorsal central nervous system (CNS). Our analysis suggests that BMPs act differently in these organisms to regulate similar patterns of gene expression in the neuroectoderm: repressing msh expression in Drosophila, while activating msxB expression in the zebrafish. These findings suggest that the mechanisms by which the BMP gradient patterns the dorsal neuroectoderm have reversed since the divergence of these two ancient lineages.
| The trunk nervous system of both vertebrates and invertebrates develops from three primary rows of neural stem cells whose fate is determined by neural identity genes expressed in an evolutionarily conserved dorso-ventral pattern. Establishment of this pattern requires a shared signaling pathway in both groups of animals. Previous studies suggested that a shared signaling pathway functions in opposite ways in vertebrates and invertebrates, despite the final patterning outcomes having remained the same. Here, we employ bioinformatics, biochemistry, and transgenic animal technology to elucidate the genetic mechanism by which this pathway can engage the same components to generate opposite instructions and yet arrive at similar outcomes in patterning of the nervous system. Our findings highlight how natural selection can act to conserve a particular output pattern despite changes during evolution in the genetic mechanisms underlying the formation of this pattern.
| In both Drosophila melanogaster and vertebrates, Bone Morphogenetic Proteins (BMPs) are expressed in the epidermal ectoderm abutting the dorsal border of the neuroectoderm [1]. The genetic network that underlies formation of a centralized nervous system consisting of segregated motor and sensory centers appears to have been conserved across bilaterians (animals with right-left symmetry) [2]. BMPs are thought to exert a common function in the early epidermal ectoderm during neural induction (i.e., suppressing expression of neural genes in epidermal regions that experience peak BMP levels). BMP signaling also acts subsequently in a dose dependent fashion to pattern dorsal versus medial regions of the neuroectoderm. For example, the trunk Central Nervous System (CNS) of both invertebrates and vertebrates consists of three primary rows of neuroblasts that are determined by the expression of three conserved transcription factors. In metazoan species spanning all three primary branches (e.g., Ecdysozoa -Drosophila, lophotrochozoa – annelids, and deuterostomes - vertebrates) “neural identity” genes (vnd/nkx2.2, ind/gsh and msh/msx) are expressed in the same relative order and orientation with respect to the dorsal-ventral axis and an epidermal BMP source. Moreover, in a broad range of organisms, BMPs and opposing antagonists have been found to play a key role in patterning the ectoderm and establishing neuronal fates. These commonalities suggest an ancestral origin for the CNS among bilateria [1]–[4] and raise the possibility that BMPs play a conserved role in patterning the CNS axis.
Despite their consistent role in promoting epidermal over neuronal cells fates in diverse species, BMPs and other extracellular factors are deployed in diverse patterns and may act by distinct mechanisms to achieve D/V patterning [5]. For instance, in Drosophila, BMPs originating in the presumptive epidermis act to repress expression of neural genes during both neural induction [6] and subsequent neuroectodermal patterning [3], [7]. In vertebrates, however, the prevailing view is that BMPs act as they do in flies to repress expression of neural genes within epidermal regions early during neural induction [1], [8] but switch function later to activate expression of orthologous neural identity genes in dorsal regions of the neural tube (e.g., the msh orthologs Msx1/2) [9]. Thus, in mice, ectopic BMP signaling leads to ventral expansion of msx expression in the neural tube [10]. In contrast, in Drosophila, the absence of BMPs leads to msh expanding dorsally into non-neural domains [11]. In zebrafish, there is evidence that BMPs act in a bimodal fashion where intermediate BMP levels are necessary for activating Msx genes, while both low and high levels of BMPs repress or fail to activate these target genes [12]. Similarly, in amphioxus, a basal chordate, msx is expressed more broadly but at reduced levels in response to ectopic BMP signaling [13]. In Echinoderms, where BMPs and chordin are co-expressed in the ventral ectoderm that gives rise to neural tissue [14], msx is expressed dorsally and is activated by peak levels of BMPs that diffuse dorsally from their ventral source into non-neural regions while Chordin remains restricted to ventral regions where it blocks the BMP response in neural cells [15]. While these conserved suites of gene expression strongly suggest a common ancestral origin for BMPs in axial patterning, it is unclear whether the regulatory mechanisms establishing these patterns have been similarly conserved during evolution.
BMPs signal via hetero-tetrameric receptor complexes consisting of two type-I and two type-II subunits, which in turn phosphorylate the cytoplasmic transducing-SMAD proteins (Mothers Against Dpp (Mad) in Drosophila, SMAD1/5/8 in vertebrates). Once phosphorylated, pMad/pSMAD1/5/8 translocates into the nucleus in a complex with Medea/Smad4 whereupon they act as transcription factors to regulate expression of BMP target genes (reviewed in [16]). Mad and Medea (Med) bind DNA as a heteromeric complex consisting of two Mad subunits and one Med subunit to regulate genes through interactions with binding sites composed by a Mad (GC-rich) site separated, by a variable length spacer, from a Med (Smad Binding Element or SBE) site. One of the best characterized such sites in Drosophila is the brinker (brk) Silencer Element (SE) which has a spacer length of 5 nucleotides [17]–[20]. Brk encodes a transcriptional repressor protein and the brk gene itself is repressed by Dpp (the Drosophila BMP4 homologue) signaling. Repression of brk through its SEs requires the presence of the zinc-finger protein Schnurri (Shn) [21]–[23], which is provided maternally and is also expressed zygotically in dorsal epidermal regions of the early embryo. Hence, in Drosophila, genes that are repressed by BMPs have been found to have binding sites for pMad/Med/Shn (henceforth, pMMS) complexes in their cis-Regulatory Modules (CRMs) while genes that are directly activated by BMPs, such as the inhibitory SMAD daughters-against-dpp (dad), contain activating elements (AE) in their CRMs [24]. These AE elements also share a bipartite configuration (GC-rich/spacer/SBE), but have configurations (spacing and sequence constraints) that do not allow for Shn binding and lead instead to the recruitment of activating transcriptional co-factors.
Here, we compare BMP-mediated regulation of CRMs controlling the expression of the Drosophila msh and zebrafish and mouse msx genes in the early dorsal nerve chord. We identify zebrafish and mouse msx neuroectodermal CRMs that drive expression in the dorsal neuroectoderm. We find that both Drosophila msh and zebrafish msxB CRM-reporter transgenes respond to BMPs and characterize BMP responsive sites within these elements. Consistent with prior genetic studies [7], the Drosophila msh CRM contains Shn-dependent SE sites that are required for BMP repression. Surprisingly, it also harbors sites that resemble known BMP-responsive activation sites, which, however, do not bind to pMad/Medea (pMM) complexes in vitro, but are nonetheless required for msh expression. In addition, we characterize a single SMAD binding site with a novel spacing of SMAD1/5/8 and SMAD4 binding motifs in a minimal zebrafish msxB CRM that is required for dorsal neuroectodermal expression. This comparison suggests that while overall gene expression patterns have been conserved between flies and zebrafish and are both regulated by BMP signaling, distinct mechanisms have evolved to generate the shared output patterns in these two widely separated metazoan lineages.
A 700 bp msh CRM (henceforth referred to as ME for Msh Element) has been identified that is directly repressed by Ind [25]. The response of the ME to BMP-mediated regulation has not yet been investigated, however. As is the case for the endogenous msh gene (Fig. 1A), the expression of a ME-lacZ construct expands throughout the dorsal region of the embryo in dpp- mutants (Fig. 1B). In order to determine whether Dpp regulates msh directly or indirectly, we analyzed BMP regulation of the ME element. Consistent with a direct role of BMP signaling on this CRM, genome wide chromatin immune precipitation (ChIP) data [26], [27] revealed DNA binding sites for the BMP effectors Mad, Medea and Shn within the ME region in blastoderm stage embryos (available on the UCSC genome browser - http://genome.ucsc.edu/ or the Berkeley Drosophila Transcription network Project - http://bdtnp.lbl.gov/Fly-Net/chipchip.jsp) (Fig. 1D). We confirmed the involvement of Shn in regulating msh within the neuroectoderm by examining homozygous zygotic shn- mutant embryos, which exhibit a partial dorsal expansion of msh expression (Fig. 1C).
To identify BMP responsive sites within the ME, we first scanned this element for known consensus binding sites for Mad, Med, Shn, and Brk. The two best characterized BMP responsive elements are the Silencer Element (SE), which binds a trimeric complex comprised of pMMS (GNCGNC(N)5GNCTG), and the activator element (AE), which binds pMM heteromers (GGCGCCA(N)4GNCV). Brk binding sites ((T)GGCGYY) overlap with a subset of AE elements [24]. Although there are no perfect consensus SE, AE, or Brk sites within the ME, we identified several candidate sites with either single base-pair mismatches to the SE or AE elements or variable spacer length (N)5–6. We defined three such candidate SE sites (SE1, SE2 and SE3) with a single nucleotide mismatch and two conserved candidate AE sites with a spacer of 6 nucleotides (conforming to the expanded consensus: GNCGNC(N)6GNCV) and tested each of these sites for direct DNA binding of pMM or pMMS complexes in vitro using Electrophoretic Mobility Shift Assays (EMSAs).
The SE1 and SE2 candidate silencer sites (Fig. 2A) both conform to the relaxed consensus of GNYGNC(N)5GNCTG (where Y can be either C or T). EMSA experiments using DNA oligonucleotide probes reveal that pMM and pMMS complexes assembled on the SE1 and SE2 sites in a BMP dependent fashion (Fig. 2B) but not on the SE3 site (Fig. S1C). As expected, mutation of the Med (SBE) motif within the SE1 (SE1SBE) or SE2 (SE2SBE) sites abolished binding of all BMP responsive complexes in vitro. In contrast, none of the candidate AE or Brk sites bound pMM, pMMS, or Brk complexes (Fig. S1C) (see below however, regarding effects of mutating or deleting the candidate AE sites).
In order to test the in vivo roles of the SE sites, we mutated each site (i.e., using the same SBE mutations that abolished all BMP responsive DNA binding in vitro described above) and generated a series of small deletions spanning virtually the entire ME (i.e., all but 36 bp). These mutant constructs were inserted into the same chromosomal integration site as the reference ME construct using the PhiC31 transgenesis system [28]. Deletion of the 5′ most 100 bp of ME, which contains both SE sites, led to dorsal expansion of reporter gene expression (Fig. 2C). Transgene expression, however, was also weaker within its normal neuroectodermal domain, suggesting that contributing activation sites are also present within this region. Targeted mutation of the individual SE1 and SE2 sites also led to discernable dorsal expansion of reporter gene expression, which was more pronounced for the SE2 mutant. Mutating both SE sites in combination (SE1, SE2 double mutant) resulted in more prominent dorsal expansion than observed for either mutant alone, but still less than that observed for the wild-type ME (or the endogenous msh gene) crossed into a dpp- mutant background. We conclude that SE elements mediate direct BMP-dependent repression of the ME and that additional direct or indirect BMP-dependent inputs also contribute to negatively regulating this CRM.
Our prior genetic studies revealed that BMP signaling is more effective in repressing expression of ind than msh [7]. One possible explanation for this differential response is that the ind CRM might contain higher affinity SE sites than those in the msh CRM. Indeed, a single perfect consensus matching SE site in the ind CRM (Fig. 2B) has been shown to be required for repression of this element dorsally [20], [29]. In line with the possibility that SE sites in the ind and msh CRMs have differing affinities for binding pMMS complexes, modifying the SE2 site by one base-pair to adhere to the optimal SE consensus resulted in greater pMMS binding (Fig. 2B - SE2*), which was most evident in competition experiments (Fig. S1D). We tested whether the optimized ind-like SE2* site would result in repression of msh CRM activity in vivo. In support of this site being more effective at recruiting repressive pMMS complexes, reporter gene expression driven by the SE2* ME was greatly reduced relative to that of the wild-type ME. This reduced expression was BMP-dependent since SE2*ME-driven reporter gene expression was restored and expanded throughout the dorsal region in a dpp- mutant background to a degree comparable to that observed for the intact ME (Fig. 2C). Taken together, these results suggest that differential affinities of pMMS complexes for SE sites in the ind and msh CRMs contribute to the mechanism by which silencer elements mediate graded BMP responses of these two genes in the Drosophila neuroectoderm.
As mentioned above, in our initial search for BMP-responsive sites in the ME we identified two sites that were similar to activation elements (AE) but that did not bind pMM complexes in EMSA assays (Fig. S1C). We nonetheless tested for potential roles of these sites by deleting them or creating a point mutation in one of them (AE2). Deletions encompassing either AE1 or AE2 or the AE2 point mutation greatly reduced ME-lacZ expression (Fig. 3B,C,E), while deletion of the 3′ most region containing a previously reported Ind site [25] resulted in ventral expansion of reporter gene expression as expected. We tested the possibility that activation of the ME via AE2 might be balanced against repression mediated by the SE1 and SE2 sites by constructing a triple mutant in which all three sites were eliminated. We reasoned that if the AE2 site, which acts as a bonafide activation site, functions in a BMP independent manner, combining it with the double SE site mutant might result in loss of expression (the AE mutant phenotype). On the other hand, if the AE2 site were providing an important activation function in the neuroectoderm via BMP signaling, the triple mutant should at least show ectopic expression dorsally (e.g., if this was a BMP-dependent activation site, relieving repression would give rise to normalized expression since we would be removing both activating and repressing components). We found loss of expression in this triple mutant comparable to that of the AE2 single mutant (Fig. 3F), suggesting that activation via the AE2 site is BMP-independent.
Although the above analysis suggests that the AE2 site acts in a BMP-independent fashion, we further examined the possibility that BMPs might play an activating as well as repressive role in regulating msh expression. Embryos that are dorsal- (maternal); dpp- (zygotic) double mutants express msh ubiquitously [11]. To test whether there might be a threshold at which Dpp enhances rather than suppress msh expression, we attempted to augment msh expression locally by generating embryos that lack Dorsal and whose only source of Dpp is one copy of dpp driven in a narrow stripe by the eve 2 CRM (Fig. 3G,H) or by, adding progressive amounts of Dpp (by varying copy number of the dpp locus – Fig. S2A,B). In both cases, we observed only a diminution in msh expression, further arguing against any activating role for Dpp. Finally, we considered the possibility that BMPs might act indirectly to regulate msh expression via non-canonical mechanisms (e.g. via ETS or the HMG-box Cic transcription factors) by altering EGF-R signaling. We found no evidence, however, for a role of EGFR signaling in influencing the position of the dorsal border of msh expression (Fig. S2C-E). In aggregate, our experiments suggest that BMP-dependent regulation of the ME is mediated by SE sites and by additional inhibitory inputs, which may act either directly or indirectly.
The above analysis of the msh CRM in Drosophila is consistent with genetic data indicating that BMPs act by dosage sensitive repression of neural identity gene expression [7]. To determine the mechanism by which BMPs regulate expression of orthologous vertebrate Msx genes we sought to identify the zebrafish (Danio rerio) msxB CRM using the powerful tol2 transgenesis system [30]. We choose to focus on regulation of the msxB gene among the zebrafish Msx paralogs as this gene has the earliest onset and most specific pattern of expression in the dorsal neuroectoderm [31]. We identified a 2.4 Kb region of DNA immediately upstream of the zebrafish msxB coding region that drives faithful reporter gene expression in the dorsal neuroectoderm in both neural plate and early neural tube stages (i.e., 3–6 somite stage embryos) of a stable transformant line (Fig. 4A,B). This fragment has two peaks of strong sequence conservation among vertebrates, which overlap regions of predicted open chromatin [32] (Fig. 4A). Later during neural tube stages, the early neural plate expression pattern fuses into a single dorsal zone (e.g., top panels in Fig. 4D).
We also tested a 5 Kb genomic fragment upstream of the mouse (Mus musculus) Msx1 gene, which like the zebrafish msxB CRM carries sequences lying immediately upstream of the transcriptional start site (Fig. 4A). When the mouse CRM-GFP construct was introduced into zebrafish embryos, it drove expression in a pattern (Fig. 4B) very similar to that of the fish msxB gene as well as that observed endogenously in mice. These results suggest that both the zebrafish and mouse CRMs contain sufficient information to correctly direct expression to the dorsal ectoderm despite the fact that they show only limited sequence conservation. These observations provide another clear example of the highly conserved function of vertebrate CRMs from lineages that diverged over 400 MYA in the absence of obvious sequence conservation in these non-coding regions [33], [34].
We pared down the zebrafish msxB CRM in transient transformant embryos and identified a minimal 671 bp fragment containing the most conserved island that also faithfully recapitulates msxB expression in dorsal neuroectodermal/neural crest progenitor cells (Fig. 4D). Paralleling our approach in Drosophila, we searched for BMP responsive sites within the minimal msxB CRM by first scanning bioinformatically for candidate SE or AE sites using the SMAD1/5/8 consensus GNCKNC and SMAD4 consensus GNC(T/V) with relaxed spacing constraints, and then testing by EMSA whether oligonucleotides containing these sites could indeed assemble Drosophila pMM and/or pMMS complexes in response to BMP signaling in vitro (Fig. S3). This analysis identified a single highly conserved site (zAE) to which BMP signal-dependent pMM (but no pMMS) DNA binding was observed. The zAE contains candidate SMAD1/5/8 and SMAD4 binding sites separated by an unusually long 16 bp spacer (Fig. S3A,B). These sites are also present in mouse albeit with different spacing (12 bp). Further analysis of this binding motif revealed that the SMAD1/5/8 and SMAD4 sites are each required, suggesting that the functional zAE includes both sites (Fig. S3C). Changing the sequence or length of the spacer DNA linking the two sites did not affect the ability to form pMM complexes in vitro indicating that the exceptional length of the zAE spacer is not required for SMAD complex formation in vitro. Interestingly, however, changing the linker length to 5 bp allowed the formation of trimeric pMMS complexes (Fig. S3D).
We generated a 36 base pair deletion spanning the zAE (and both the SMAD1/5/8 and SMAD4 candidate binding sites – DEL mutant) in the context of the 671 bp msxB CRM and observed that GFP reporter gene expression was lost in transient transformant embryos (Fig. 4D). Similarly, mutation of two core base pairs in the GC-rich region of the zAE (GCR1 mutant), which abolished pMM binding in vitro (Fig. 4C; Fig. S3C), also reduced reporter expression in vivo in transient transformant embryos (Fig. 4D). These results indicate that a single BMP responsive site within the 671 bp zebrafish msxB CRM is required for mediating reporter gene activation by this element in vivo.
The above dissection of BMP-responsive sequences within the Drosophila and zebrafish msh/msx CRMs suggests that they are under opposing forms of BMP regulation: repression in Drosophila versus activation in zebrafish. To test this hypothesis further, we compared the response of these CRMs to alterations in BMP signaling in vivo (Fig. 5). In Drosophila, we examined msh and ME reporter gene expression in both a dpp- mutant background and in embryos ectopically expressing dpp in the dorsal epidermis. As mentioned above, msh and ME-reporter gene expression both expand dorsally in dpp- mutant (Fig. 1B; Fig. 5A). Conversely, ectopic dpp expressed from a Heat Shock-dpp construct (HS-dpp) resulted in loss of msh expression within its normal domain (Fig. 5C). In zebrafish, a stable transgenic line carrying the 2.4 kb msxB-GFP reporter construct was crossed to lines carrying either a Heat Shock-chordin (HS-CHD) or a Heat Shock-BMP (HS-BMP) construct. When the BMP antagonist Chordin was induced by heat treatment (Fig. 5G), msxB-GFP reporter expression was strongly suppressed, as was endogenous msxB expression (Fig. 5D). The opposite effect was observed in HS-BMP embryos, however, where expression of endogenous (Fig. 5F) and reporter (Fig. 5I) genes was broadened compared to control embryos (Fig. 5E and 5H, respectively) that were subjected to the same conditions. Thus, consistent with the inverse effects of mutagenizing BMP-responsive sites in the Drosophila msh and zebrafish msxB CRMs, these two elements respond in an opposing fashion to equivalent manipulations of BMP signaling in vivo. Our analysis strongly suggests that BMPs pattern the neuroectoderm primarily via repression in Drosophila, while in zebrafish, BMPs function, at least in part, to activate the orthologous msxB gene.
BMPs play a highly conserved role in neural induction and also contribute to establishment of dorsal-ventral polarity within the CNS. In the latter case, however, it has not been established whether they act through common or distinct mechanisms to effect dose-dependent patterning of neural identity genes. Since BMPs regulate expression of highly conserved members of the ancient msh/msx family in the dorsal neuroectoderm of both Drosophila and vertebrate embryos, comparing cis-regulation of these genes by BMPs provides an excellent paradigm for addressing whether cis-regulatory processes are maintained across distant taxa. Our analysis of the Drosophila msh embryonic CRM suggests that BMPs act in part through two SE-type binding sites that mediate repression, while activation sites do not appear to mediate responses to BMP signaling. In contrast, we identified a single SMAD binding site within an embryonic zebrafish msxB CRM that is required for BMP-dependent activation. These findings suggest that BMPs act on msh/msx CRMs by opposite mechanisms in these two lineages, while nonetheless driving similar output expression patterns in the dorsal neuroectoderm.
Mutational analysis of the Drosophila msh CRM in this study supports a direct role for BMP repression acting via the SE1 and SE2 sites to suppress activity of this element in the dorsal ectoderm where there are likely to be moderate levels of BMP signaling. Mutation of either of these sites results in modest dorsal expansion of reporter gene expression while elimination of both sites by point mutations or a deletion spanning both sites causes more pronounced ectopic dorsal expression. The dorsal expansion in SE1, SE2 double mutants is less complete, however, than that observed when the intact ME is crossed into a dpp- background, indicating that additional inputs are also involved in repressing the activity of this element dorsally. These additional BMP-dependent inputs might act either directly or indirectly. Since each of the three deletions spanning the remaining portions of the CRM (i.e. sequences outside of the deletion covering the SE1 and SE2 sites) result in reduced CRM activity it is possible that the effects of such hypothetical additional BMP responsive sites are canceled out by the deletion of necessary adjacent activation sites (e.g., deletion of the A2 site in the D3 region, Fig. 3C). If these hypothetical repressor sites act directly on the msh CRM they would presumably bind Mad, Medea and Schnurri, MAPK pathway transcriptional effectors, or possibly yet unknown BMP mediators alone or in conjunction with other transacting factors. Our detailed bioinformatic analysis and systematic experimental EMSA surveys have failed to identify any such sites, however. It is also possible that part of the BMP response of the msh CRM is mediated indirectly. For example, we have previously reported that localized overexpression of Brk can de-repress msh expression dorsally [7], yet there are no consensus Brk sites in the ME and we were unable to detect Brk protein binding to any closely related candidate Brk sites by EMSA (Fig. S1). Thus, Brk may act via regulating expression of other components required for BMP signaling such as the BMP type-1 receptor Thick veins [35]. Alternatively, activators of the ME may be under negative BMP/Brk regulation.
The SE1 and SE2 sites that play a role in repressing ME activity dorsally are imperfect matches to the consensus SE sites determined by Pyrowolakis and colleagues [36]. The ind CRM, however, which according to genetic data is more sensitive to BMP repression than msh [7], contains a perfect SE site required for repressing activity of this element dorsally [20]. When the SE2 site in the ME was mutated to similarly match the ideal SE consensus sequence (SE2*) it repressed ME expression in its normal dorsal ectodermal domain in a dpp-dependent fashion (Fig. 2C). In addition, competition experiments indicate that Mad/Schnurri/Medea bind to the ind-like SE element with higher affinity than the msh SE2 element (Fig. S1D). These combined findings suggest that differences in affinity of SE sites for forming Mad/Med/Shn complexes contribute to the distinct responses of the two CRMs to BMP-mediated threshold-dependent repression.
Using a combination of bioinformatics and efficient transgenesis in zebrafish we identified genomic fragments upstream of the zebrafish msxB and mouse Msx1 genes that drive neuroectodermal GFP-reporter gene expression at the open neural plate stage in zebrafish embryos. Further analysis of a minimal 671 bp zebrafish CRM identified a single conserved SMAD binding site that is required for activity of this element. An novel feature of this BMP-activation site is that the SMAD1/5/8 and SMAD4 binding site motifs are separated by a 16 bp spacer, which interposes approximately one and half turns of the DNA helix between these two sites, thus differing from other characterized vertebrate BMP activation sites in which these SMAD binding sites are closer [37]. Interestingly, deletion of 11 bp (about one turn of the helix) endows this modified site with the ability to bind the pMMS repressor complex in vitro. Whether this unique architecture of the msxB BMP activation site is relevant to activity within the neuroectoderm remains to be explored.
We also examined the in vivo response of the endogenous zebrafish msxB gene and the msxB-CRM to inhibition of BMP signaling or ectopic expression of BMPs and compared these responses to equivalent manipulations of BMP signaling in Drosophila. In Drosophila, msh or ME-lacZ expression expands dorsally in a dpp- mutant while msh expression is repressed within its normal dorsal neuroectodermal domain by ectopic dpp expression. In contrast, expression of the zebrafish msxB gene, which is mirrored by activity of the msxB-CRM, is lost upon inhibition of BMP signaling and expanded or elevated in response to ectopic BMPs. Thus, both mutational analysis and in vivo testing suggest opposing mechanisms for BMP-dependent regulation of the msh and msxB genes in the early neuroectoderm.
Given the opposing mechanisms by which the msh and msxB CRMs respond to BMPs, it is intriguing that a site within the msh CRM closely resembling an activation site (AE2) is required for activation of this CRM. Also, another AE-like site (AE1) lies within a region which when deleted greatly reduces ME driven reporter gene expression, although the role of that AE1 site remains to be examined. These AE-like sites, while having only single mismatches to consensus Mad-Medea binding sites, did not bind Mad-Medea complexes in vitro, indicating that they are most likely not involved in mediating a BMP response. Additionally, experiments designed to identify potential positive roles of BMP signaling in regulating ME activity provided no evidence for such an effect. Given the known role of AE sites in other genes to BMP-dependent activation and the evidence that BMPs can act positively to promote msx gene expression in vertebrates, it is tempting to speculate that these sites could once have been BMP responsive activation sites and were subsequently co-opted by different transcription factors (possibly a TAGteam motif [38] binding protein) in the course of evolution to maintain msh expression in a BMP-independent fashion. Identifying such transcriptional activators is an interesting goal for future experiments.
In Drosophila, Evo/Devo studies of the even-skipped stripe 2 CRM have suggested that regulatory mechanisms that lead to a particular gene expression pattern are extremely flexible, i.e., the same pattern can be achieved in multiple ways [39]. Accordingly, in the current case of BMP-dependent regulation of msh/msxB expression, natural selection may have operated similarly to maintain relevant gene expression patterns that fulfill a particular function (i.e. dorsal neuroectodermal expression) while allowing the upstream mechanisms generating that pattern to change over time.
As summarized above, our analysis strongly suggests that BMPs pattern the neuroectoderm primarily via repression in Drosophila, while in zebrafish, BMPs function, at least in part, to activate the orthologous msxB gene. Genetic studies and exogenous BMP treatment in zebrafish suggest that msx gene expression may also be repressed by high levels of BMP signaling. Whether the BMP-responsive site in the 671 bp msxB CRM together with other potential BMP-responsive elements mediate such a biphasic response will be interesting to address in future experiments. In the future, it will also be important to determine whether expression of other msx paralogs in the dorsal CNS of zebrafish (e.g., msxC,E [31]) or msx genes in other vertebrates (e.g., the murine Msx1 neuroectodermal CRM identified here) are similarly regulated by BMPs. Analysis of these additional vertebrate msx CRMs should reveal whether distinct evolutionary trajectories have shaped the BMP responsiveness of these elements. Such comparative studies may also shed light on whether there is a single or multiple independent origin(s) of BMP regulation of vertebrate msx genes. Furthermore, analysis of the CRM driving BMP-dependent expression of an echinoderm Msx homolog in regions of peak BMP activity [14] will be informative since this gene is expressed in the non-neural ectoderm. In this case, one might predict finding only positively acting AE-like BMP-responsive sites.
There are two possible explanations for distinct mechanisms of BMP-regulation of msh/msxB expression in flies versus fish. One is that these genes independently evolved BMP responsiveness. Alternatively, BMP-dependent regulation may be an ancestral trait dating back to the first bilaterians with a condensed CNS. We favor the latter alternative for the following reasons. First, the co-linearity of msh-msx, ind-gsh/pax, and vnd-Nkx2.2 genes relative to the source of BMPs and the BMP responsiveness of these genes in species from all three primary branches of bilateria - flies (ecdysozoa), vertebrates (deuterostome chordates), and annelid worms [40] (lophotrochozoa) - provides a compelling argument for this arrangement reflecting the ancestral state. Second, a polarized source of BMPs was present in diploblasts (e.g., corals [41], [42], jellyfish [43], and the sea anemone [44], [45]) and therefore preceded evolution of bilaterian triploblasts and a condensed CNS. Thus, it is plausible that a single species evolved a condensed CNS which deployed neural identity genes along the DV axis in much the same way that Hox genes are expressed in sequential order along the AP axis. Finally, if one looks more broadly among the 30 bilaterian phyla, a striking trend is that at least some clades within most of these phyla have a condensed CNS with three primary axon bundles [46], suggestive of an ancestral tripartite subdivision of the CNS. It is true that there are also many examples of species scattered among these phyla that either secondarily lost a condensed polarized CNS or retained a prior ancestral state in which there was only a distributed nervous system. Echinoderms in which Msx genes are expressed in the non-neural ectoderm (see above) or the hemichordate Saccoglossus kowalevskii which has lost bilateral symmetry to become radially organized [47] may be examples of such derived simplifications of the nervous system. Thus, in our view, the most likely scenario is that the ancestral bilaterian CNS was a condensed nervous system partitioned into at least three DV domains and that loss of centralization has occurred numerous times in different lineages undergoing morphological simplification.
If one assumes a common ancestral origin for BMP-regulation of msx genes, one can imagine various scenarios under which BMP-mediated regulation of msh/msx genes could have switched its effect during evolution. In vertebrates, BMP targets frequently contain Drosophila SEs that activate rather than repress transcription. This might be due to Shn proteins losing their repressive activity through changes in the Shn amino acid sequence and/or the lack of components required for repression downstream of Shn. The molecular relatedness of SEs and AEs raises the possibility that ancestral SE-mediated repressive effects on msh/Msx expression may have been relatively easy to convert into activating effects in the vertebrate lineage by the loss of the Shn repressor function. Consequently, the increased linker length of zAE could be accounted for by the lack of evolutionary pressure on the SE to meet the sequence requirements for Shn recruitment.
Since the Drosophila msh gene is weakly repressed by BMPs (e.g., relative to ind and other neural genes such as AS-C, scrt or sna [6]), while vertebrate msx genes are weakly activated by BMPs (i.e., high neuroectodermal levels of BMPs are required to activate msx genes) an intermediate CRM state may have existed in which BMPs both weakly activated msx gene expression within the neuroectoderm at moderate levels while repressing gene expression at the peak BMP levels present in the adjacent epidermis. Indeed the zebrafish msxB gene may represent such a bifunctional intermediate condition since in vivo studies indicate that high levels of BMPs can inhibit msxB expression [12]. It remains to be determined whether such proposed positive and negative inputs are mediated by a single or multiple independent CRM(s). Within different evolutionary lineages such biphasic responses could have then been rendered monophasic in opposing directions to account for the observed differences in the Drosophila versus vertebrate or echinoderm Msx CRMs. In vertebrates, one potential driving force for reducing the effect of BMP-mediated inhibition may have been the incorporation of BMP expression within the dorsal neural tube itself since this would be expected to generate much higher BMP levels than would result from BMPs diffusing in from the adjacent epidermal ectoderm (e.g., as is the case in Drosophila).
In future analyses it will also be important to examine BMP-mediated regulation of additional neural identity genes expressed along the dorsal-ventral axis including the Gsh ≈ ind and Nxk2.2 ≈ vnd genes as CRMs controlling expression of each of these genes will have undergone independent evolutionary trajectories. Since there is evidence that laterally and ventrally expressed genes in vertebrates are inhibited by BMPs [48]–[52], and because the more ventrally expressed ind gene in Drosophila is more sensitive to BMP-mediated repression than msh [7], one might expect to find similar, and perhaps conserved ancestral modes, of BMP-mediated repression of these genes across bilateria.
It will also be interesting to understand how flexible the ancestral metazoan state was by investigating the relationship between BMPs and msx genes in basal metazoans such as jellyfish. In these diploblastic animals, although the BMP-msx relationship has not been tested, BMP2/4 [53] and msx [54] homologues are expressed in adjacent regions during development, as is the case in the majority of triploblastic animals.
We identified candidate SE and AE sites in the msh, msxB and msx1 CRMs using binding site consensus sequences curated from the literature referenced and used Gene Palette [55]. For this analysis, we used the consensus sequence GNCGNC(N)5GNCTG to identify candidate Silencer Elements (SE) and the consensus GGCGCCA(N)4GNCV for Activator elements (AE) allowing for single base-pair mismatches to these consensus sequences. We identified candidate zebrafish msxB and mouse Msx1 CRMs by using genome wide alignments for multiple vertebrate species, which indicates regions of high sequence conservation as provided by the UCSC genome browser (http://genome.ucsc.edu).
The 700 bp msh CRM is described in Von Ohlen et al., 2009 [25]. All primers used in this study and the corresponding constructs generated can be found in Table S1. The various Drosophila msh-CRM constructs were subcloned in pCR-TOPO vectors (Invitrogen) and subsequently cloned into the [P]acman vector [28] as NotI and KpnI restriction fragments. Site-directed mutagenesis PCR methods were adapted from [56]. The primers used to isolate the zebrafish msxB CRMs and the mouse msx1 CRM can be found in Table S1. Zebrafish constructs were cloned into pENTR-TOPO (Invitrogen), transferred to pTol2 by Gateway Recombination and injected in zebrafish embryos as previously described [30].
The Drosophila dpph46 null allele used in this study is Flybase stock number 2061. The 8x HS-dpp stock and its use are described in Biehs et al 1996 [6]. The schnurri mutant allele is shn04738. To generate the dl dpp st2-dpp+ embryos, females that are Dpdpp/+; dl1 cn1 sca1/dpph46 wgsp dl1 were crossed to yw/Y; dpph46 wgsp st2-dpp+,w+/CyO males. The fly strain used to inject all constructs has genotype PBac{yellow[+]-attP-3BVK00002 and injections were outsourced to BestGeneInc (http://www.thebestgene.com/).
The zebrafish strains containing the hsp70:bmp2b [57] and hsp70:chd [58] transgenes were crossed to stable transgenic lines containing the msxB-CRM construct. Embryos at the sphere stage (4hpf) were subjected to heat shock at 37°C for 1 hour and then returned to normal temperature of 28.5°C until they were fixed at the bud – 6 somite (10–11hpf) stage as necessary.
For electrophoretic mobility shift assays (EMSA), Drosophila S2 cells were co-transfected with 50 ng TkvQD and 175 ng Mad- and Med-expression plasmids or with 400 ng of a ShnCT-expression plasmid. Cells were harvested 72 hr after transfection and lysed for 10 min at 4°C in 100 µl of 100 mM Tris (pH 7.5), 1 mM DTT, 0.5% TritonX100 and 1%NP40. Radioactively labeled probes were generated by annealing and filling in partially overlapping oligonucleotides in the presence of [P]-32 ATP. Binding reactions were carried out in a total volume of 25 µl containing 12.5 µl 2x binding buffer (200 mM KCl, 40 mM HEPES (pH 7.9), 40% glycerol, 2 mM DTT, 0.6% BSA and 0.02% NP40), 10000 cpm of radioactively-labeled probe, 1 µl poly dIdC (1 mM) and 7 µl of cleared S2 cell extracts. After incubation for 30 min at 4°C, the reactions were analyzed by non-denaturing 4%polyacrylamide gel electrophoresis followed by autoradiography.
Fluorescent in situ hybridization methods used were performed according to [59] in Drosophila embryos and adapted to zebrafish embryos by increasing the hybridization temperature: 55°C in Drosophila to 65°C in zebrafish embryos. Antibodies used: dpERK (Cell Signaling Technology #5683), anti-digoxigenin (Roche), anti-biotin (Roche), Alexa fluor 488, 594, 647 (Invitrogen). We also used colorimetric staining methods performed according to O'Neill and Bier [60]. The DNA template used to generate the msxB probe was a generous gift from the Riley lab. Histochemical stain images were acquired using a Nikon optical microscope and fluorescent stain images were collected using a LEICA SP2 confocal microscope. Images were adjusted for color, brightness and contrast using Adobe Photoshop software.
|
10.1371/journal.pgen.1005676 | Ectodysplasin/NF-κB Promotes Mammary Cell Fate via Wnt/β-catenin Pathway | Mammary gland development commences during embryogenesis with the establishment of a species typical number of mammary primordia on each flank of the embryo. It is thought that mammary cell fate can only be induced along the mammary line, a narrow region of the ventro-lateral skin running from the axilla to the groin. Ectodysplasin (Eda) is a tumor necrosis factor family ligand that regulates morphogenesis of several ectodermal appendages. We have previously shown that transgenic overexpression of Eda (K14-Eda mice) induces formation of supernumerary mammary placodes along the mammary line. Here, we investigate in more detail the role of Eda and its downstream mediator transcription factor NF-κB in mammary cell fate specification. We report that K14-Eda mice harbor accessory mammary glands also in the neck region indicating wider epidermal cell plasticity that previously appreciated. We show that even though NF-κB is not required for formation of endogenous mammary placodes, it is indispensable for the ability of Eda to induce supernumerary placodes. A genome-wide profiling of Eda-induced genes in mammary buds identified several Wnt pathway components as potential transcriptional targets of Eda. Using an ex vivo culture system, we show that suppression of canonical Wnt signalling leads to a dose-dependent inhibition of supernumerary placodes in K14-Eda tissue explants.
| Mammary glands are the most characteristic feature of all mammals. The successful growth and function of the mammary glands is vital for the survival of offspring since the secreted milk is the main nutritional source of a new-born. Ectodysplasin (Eda) is a signaling molecule that regulates the formation of skin appendages such as hair, teeth, feathers, scales, and several glands in all vertebrates studied so far. In humans, mutations in the EDA gene cause a congenital disorder characterized by sparse hair, missing teeth, and defects in exocrine glands including the breast. We have previously shown that excess Eda induces formation of supernumerary mammary glands in mice. Here, we show that Eda leads to extra mammary gland formation also in the neck, a region previously not thought to harbor capacity to support mammary development. Using Eda loss- and gain-of-function mouse models and transcriptional profiling we identify the downstream mediators of Eda. The presence of extra nipples is a fairly common developmental abnormality in humans. We suggest that misregulation of Eda or its effectors might account for some of these malformations. Further, the number and location of the mammary glands vary widely between different species. Tinkering with the Eda pathway activity could provide an evolutionary means to modulate the number of mammary glands.
| The murine mammary gland development initiates at around embryonic day 10.5 (E10.5) with the establishment of bilateral milk or mammary lines [1]. Between E11-E12, five pairs of mammary placodes, local thickenings of the epithelium, emerge at conserved positions. By E13.5, the placodes have transformed via hillock stage to buds that have submerged downward and are surrounded by several layers of a specialized dermis, the primary mammary mesenchyme [1]. As the tip of the primordium begins to elongate, at E15.5, it forms a primary sprout that invaginates into the more distal secondary mammary mesenchyme. Branching morphogenesis begins a day later, and by birth a small ductal tree with several branches has formed.
The murine mammary line is not externally visible but only detectable from histological sections or molecularly identifiable by expression of Wnt pathway genes such as Wnt10b or TOP-gal, a transgenic reporter of the canonical Wnt pathway [2, 3]. Initially, the milk line is not a continuous structure but instead three independent Wnt10b-positive stripes arise: in axillary and inguinal regions, and the third one in the flank between the fore and hind limb buds. The axillary milk line gives rise to placode 1, inguinal to placode 5, and placodes 2, 3 and 4 form from the milk line of the flank [2]. Establishment of placodes is asynchronous and expression analysis of the Wnt pathway mediator Lef1 revealed a designated order: 3, 4, 1/5 and 2 [4]. As the placodes form, low level of Wnt10b expression transiently combines all three milk lines but by E12.5 Wnt10b expression becomes confined to mammary buds [2, 3]. Placode morphogenesis is thought to rely mainly on migration of the progenitor cells along and from the immediate vicinity of the milk line and not on proliferation [5, 6].
Similar to other ectodermal appendages such as hair follicles and teeth, reciprocal interactions within and between the epithelium and the underlying mesenchyme are a necessity for proper development and pattering of mammary glands [1, 7, 8]. These interactions are mediated by conserved signaling pathways, of which at least the fibroblast growth factor (Fgf), Wnt/β-catenin, and Neuregulin (Nrg)/ErbB pathways regulate mammary placode formation. Mammary gland initiation relies on a complex interplay between these pathways and transcription factors Gli3 and Tbx3 and their absence disrupts formation of one or more placode pairs (reviewed in [9]).
Fgf10, emanating from the tip of the thoracic somites and the limb buds, has been proposed to function as one of the earliest signals for milk line specification. In the absence of Fgf10 or its receptor FgfR2b, only placode pair four develops [2, 4]. Wnt/β-catenin signaling is required for all mammary placodes to form. Ectopic ectodermal expression of the secreted Wnt inhibitor Dkk1 abolishes all signs of mammary placodes [3]. Disruption of the Hedgehog pathway mediator Gli3 leads to loss of placodes 3 and 5 [10, 11]. In Tbx3 null embryos, all mammary placodes are absent with the exception of occasional presence of placode 2 [12]. Tbx3 has been proposed to act both up- and downstream of Fgf and Wnt pathways but the details of these interactions are not well understood [12–14]. Finally, hypomorphic Nrg3 mutant mice display frequently missing or hypoplastic placode 3 but also supernumerary placodes [15], whereas ectodermal overexpression of Nrg3 induces multiple supernumerary mammary glands along and adjacent to the milk line [16].
Another important player in embryonic mammary gland development is the tumor necrosis factor (Tnf) superfamily ligand Ectodysplasin-A1 (hereafter Eda) and its receptor Edar. The Eda pathway has a well characterized role in the development of diverse set of ectodermal organs [17, 18]. The ectodermal appendage phenotype of Eda null mice (Tabby mice) and mice with compromised activation of transcription factor NF-κB is highly similar [19], and biochemical and genetic studies have confirmed the importance of NF-κB downstream of Eda [18, 20]. In humans, mutations in the genes encoding EDA, EDAR, or the cytosolic signal mediator EDARADD cause a condition known as hypohidrotic ectodermal dysplasia (HED). In addition to tooth, hair, and sweat and salivary gland defects, breast anomalies such as hypoplastic/absent/supernumerary nipples and even absence of breast tissue have been reported in HED patients [21–23]. Studies using Eda loss- and gain-of-function mouse models have shown that Eda regulates embryonic and prepubertal mammary gland branching morphogenesis via NF-κB [24]. However, all five mammary glands form in Eda null mice suggesting that Eda is dispensable for mammary placode formation [24, 25]. Strikingly, ectodermal overexpression of Eda (K14-Eda mice) leads to formation of supernumerary mammary placodes along the milk line, in particular in the region between mammary buds 3 and 4, and give rise to supernumerary mammary glands in the adult [26, 27]. Beyond this, little is known about the importance of Eda in the initial stages of mammary gland development.
We report here that NF-κB is dispensable for mammary placode induction, yet it is necessary for the ability of Eda to induce supernumerary mammary primordia. Using an unbiased genome-wide approach, we identify several transcriptional targets of Eda. We provide evidence indicating that Eda promotes mammary cell fate by enhancing canonical Wnt signaling activity. Furthermore, we find that Eda induces supernumerary mammary glands not only between the endogenous mammary glands, but also in the neck region. Based on analysis of wild-type and K14-Eda embryos we propose that the murine mammary line extends more anteriorly than previously recognized.
Embryonic mammary primordia exhibit high Eda-dependent NF-κB activity from E12 onwards [24, 25]. To gain further insights on the role of the Eda/NF-κB pathway in early mammogenesis, we assessed NF-κB signaling activity with reporter mice expressing β-galactosidase under an NF-κB–responsive element in control and K14-Eda embryos. We detected NF-κB activity in the mammary placode forming region from E11 onwards (Figs 1 and S1). At E11.0, a low level reporter expression was detected in the region of future mammary placode 3 and the interface of the forelimb bud and the thorax where placode 1 will later appear (Figs 1A and S1). At E11.25 faint expression was detected also at the border of the hind limb bud and ventrum (prospective placode 5), as well as at the site of future primordium 4 (Figs 1B and S1B). By E11.5 reporter expression had intensified in placode 3 and become more condensed at placodes 1, 4, and 5 (Fig 1C). Dispersed X-gal-positive cells were detected at the location of prospective placode 2 (Figs 1C and S1D). In addition, modest amount of reporter positive cells were observed along the entire milk line, from placode 1 to 5. At E12.0 high localized reporter expression was confined to the mammary buds and low level NF-κB activity was found throughout the dorsal side of the embryo whereas the ventrum appeared devoid of reporter expression (Fig 1D). At these early stages, the reporter expression was constantly stronger in K14-Eda background (Fig 1A–1D).
Similar to previous reports on expression of the Wnt pathway genes and TOP-gal reporter, NF-κB reporter positive cells disappeared from the milk line between E12.5 and E13.5 in control embryos (Fig 1E and 1F). In K14-Eda embryos, elevated NF-κB signaling was observed along the milk line, as well as in the dorsum, yet they exhibited no obvious focal clustering of X-gal-positive cells between buds 3 and 4, i.e. at the site of prospective supernumerary primordia, until at ~E12.5 (Fig 1C–1E). These foci were more pronounced at E13.5, although reporter expression was markedly less intense than in endogenous buds (Fig 1F).
The milk line, the area possessing mammary inductive capacity, is considered to extend from the axilla to the groin [1]. To our surprise, we observed faint NF-κB reporter activity from E11.0 onwards also in the neck area, anterior to mammary bud 1, which was substantially more pronounced in K14-Eda embryos (arrowheads in Fig 1). In K14-Eda embryos, reporter expression was confined to one or up to four small foci suggesting that supernumerary placodes were induced in the neck region. In situ hybridization analysis revealed high focal expression of Wnt10b, as well as Dkk4, another placode marker [28] in endogenous placodes of E11.25 wild type and K14-Eda embryos, as well as in the neck region (Fig 1G and 1H). The latter coincided with the site of ectopic placodes marked by NF-κB reporter expression in K14-Eda embryos (compare Fig 1G and 1H to Fig 1A–1C).
To analyze more in detail NF-κB activity, we sectioned whole mount stained reporter embryos (Fig 2). NF-κB activity was present throughout the developing mammary epithelium in control and K14-Eda embryos at E12.5, similar to expression of Edar (Fig 2A and 2B). At E13.5, NF-κB reporter activity was mainly confined to the basal cells in control embryos (Fig 2B), but remained high throughout the bud in K14-Eda embryos (Fig 2C). Further, sectioning confirmed that supernumerary neck placodes were truly thickened at E12.5 (Fig 2D, left column) and showed that supernumerary mammary buds, in particular those between buds 3 and 4, consisted of both reporter positive and negative cells (Fig 2D, right column).
Supernumerary mammary placodes forming between gland 3 and 4 give rise to nipples with an associated ductal system in K14-Eda adults, and are responsive to pregnancy hormones [26]. As suggested by embryonic analyses (Fig 1), a nipple was observed also in the neck region and was often accompanied by accessory, smaller nipple-like structures (Fig 3A). However, the nipple-like structures in the neck region did not express keratin 2e, a specific marker of nipple epithelium [29] indicating defective differentiation of the nipple epithelium (Fig 3B). Surprisingly, the neck region was also capable of supporting ductal morphogenesis (Fig 3C). Similar to the supernumerary glands located between glands 3 and 4 [26], the ductal trees in the neck were considerable smaller than those of the endogenous glands and displayed typical pregnancy-associated morphological changes (Fig 3D and 3E).
As discussed above, engagement of Edar leads to activation of NF-κB. In unstimulated cells, inhibitory IκB proteins, most commonly IκBα, retain NF-κB is in the cytosol [30]. Ligand binding leads to phosphorylation and degradation of IκBα thereby releasing NF-κB. To elucidate the importance of NF-κB signaling in the embryonic mammary placode development, we utilized the IκBαΔN mouse strain which displays suppressed NF-κB activity as a result of ubiquitous expression of a non-degradable IκBα [19]. Analysis of NF-κB reporter expression in IκBαΔN embryos at E11.25 revealed absence of reporter expression (Fig 4A). This indicates that NF-κB signaling is fully suppressed in this mouse model at the time of mammary placode induction, similar to later developmental stages (E12-E16) [24]. In situ hybridization analysis of Lef1, which is expressed both in the mammary epithelium and the mesenchyme at E12.5 [31, 32], confirmed the presence of normal number of mammary primordia in IκBαΔN embryos (Fig 4B), yet Wnt10b expression suggested that mammary buds may be somewhat smaller (Fig 4C). Taken together, these data show that NF-κB activity is dispensable for mammary placode induction.
To address the necessity of NF-κB for the ability of Eda to induce supernumerary placodes, we crossed K14-Eda strain with the IκBαΔN mice. Mammary placode markers Tbx3, Wnt10b, and PTHrP [2, 24, 33] were expressed in mammary buds of wild type, IκBαΔN, K14-Eda and compound K14-Eda;IκBαΔN embryos at E13.5 (Fig 5A–5C). Expectedly expression of PTHrP and Wnt10b appeared slightly downregulated in IκBαΔN background as they have been identified to be transcriptional targets of Eda/NF-κB [24, 34]. All three were also detectable in the supernumerary primordia of K14-Eda embryos, Tbx3 showing a circular expression pattern around the placodes though. Expression of all marker genes was completely abolished in the supernumerary placode forming region in the compound mutants (Fig 5A–5C). Analysis with scanning electron microscope (SEM) showed no morphological signs of supernumerary placodes in K14-Eda;IκBαΔN mutants (Fig 5D). Further, supernumerary nipples or ductal trees were never observed in the adult compound mutants. Our findings show that even though NF-κB is not needed for the formation of endogenous mammary placodes, it is indispensable for formation of Eda-induced supernumerary mammary placodes.
In order to identify the immediate downstream targets of Eda/NF-κB, we performed microarray profiling of genes expressed in Eda-/- E13.5 mammary buds exposed to control medium or to recombinant Fc-Eda protein. Using the same setup, but quantitative real-time reverse-transcriptase–PCR (qRT-PCR) and candidate gene approach, we have previously shown that Eda upregulates expression of Wnt10a, Wnt10b, Dkk4, and PTHrP in mammary buds [24]. Altogether 245 probes were upregulated (including Wnt10a, Wnt10b, Dkk4, and PTHrP) and 78 probes downregulated by Eda treatment (Tables 1 and S1).
Genes in several different signaling pathways including Wnt, Fgf, Tnf, Tgfβ, chemokine, and hedgehog pathways were differently expressed. In addition, adhesion molecules Madcam1 and Icam1, extracellular matrix degrading metalloproteinases Adamts15 and Mmp9, chloride channel proteins clca1 and clca2 (recently reannotated as a1 and a2 variants of clca3, respectively), and transcription factor Foxi3 were among the upregulated genes (Table 1). To validate the microarray results, we performed qRT–PCR analysis and in situ hybridization (ISH) or immunostaining of selected genes, both strongly and modestly induced ones (Table 1, Figs 6A, 6B and S2). Of the 7 genes tested all showed the same tendency as in the microarray, the difference between control and Eda-treated specimen being statistically significant for 5 genes.
Both Madcam1 and Icam1 are known to be expressed in hair placodes and their transcripts are upregulated by Eda in E14 back skin [35]. We did not detect Madcam1protein or Icam1, Adamts15, or Mmp9 transcripts in developing mammary primordia of control embryos by whole mount analysis, yet Madcam1 and Mmp9 (but not Icam1 or Adamts15) were readily observed in K14-Eda embryos (S2A and S2B Fig) suggesting that they lie downstream of Eda in mammary buds. Mmp9-deficient mice have no overt mammary gland phenotype [36] possibly owing to redundancy with other Mmps. There are no reports on the function of the other genes in mammary gland development.
Transcription factor Foxi3 was one of the most highly induced genes by Eda. Foxi3 is mutated in several dog breeds, a condition described as canine ectodermal dysplasia [37]. We have previously identified Foxi3 as an Eda-induced gene in developing hair follicles and teeth and shown augmented expression in K14-Eda mammary buds in vivo [38]. The finding prompted us to analyze whether Foxi3 could play a role in mammary gland induction. However, the mammary glands of Foxi3 null embryos were indistinguishable from control littermates and formation of Eda-induced supernumerary mammary primordia was unaffected by loss of Foxi3 (S3A–S3D Fig).
Our microarray and previous qRT-PCR analyses revealed that several Wnt pathway genes are induced by Eda (Fig 5, [24]). Further, we have earlier reported that Lef1 is expressed very early on in the emerging supernumerary placodes of K14-Eda embryos [27]. Given the importance of the Wnt pathway in mammary placode formation, we wanted to study more closely whether expression of the Wnt pathway genes is altered in response to diverse levels of Eda by comparing Eda-/-, wild type, and K14-Eda embryos at E12.5, when ectopic placodes are becoming apparent between buds 3 and 4.
Wnt10b, one of the earliest markers of the milk line, becomes gradually restricted to the placodes as they emerge [2, 3]. Kremen2 (Krm2) is a transmembrane protein that inhibits Wnt signaling in the presence of Dkk proteins [39] whereas Lgr4 is a receptor for R-spondins, which are potent Wnt pathway stimulators [40, 41]. Both Krm2 and Lgr4 have been localized to E12.5 mammary buds [3, 42]. Wnt10a, Wnt10b, Krm2 and Lgr4 were all present in the endogenous mammary buds of all three genotypes (Fig 7A–7D). Expression of all four genes revealed a correlation with Eda levels: reduction in Eda-/- and up-regulation in K14-Eda mammary buds. Notably, Wnt10b and occasionally Lgr4 and Kremen2 were clearly upregulated as a continuous streak in K14-Eda embryos at the site where supernumerary placodes form. Further, we analyzed expression of β-catenin, which also exhibited a streak-like expression pattern between buds 3 and 4 in K14-Eda embryos (Fig 7E).
Next, we studied the expression of other genes critical for mammary placode formation. Tbx3 and Nrg3 are first detected in the mesenchyme but at the onset of placode formation they become upregulated (Tbx3) or completely restricted (Nrg3) to the mammary epithelium [12, 15, 33]. Expression of both genes was expectedly found in the endogenous mammary buds in all three genotypes (Fig 7F and 7G). However, neither of them could be detected in the ectopic mammary forming region in K14-Eda embryos at E12.5 (Fig 7F and 7G), yet Tbx3 was observed in the ectopic primordia at E13.5 ([24]; Fig 5).
Our results show that all Wnt pathway genes studied exhibited early upregulation in the region where ectopic mammary placodes arise, whereas expression of other genes implicated in mammary placode formation (Tbx3, Nrg3) was detectable in this region only at a later developmental stage. Although different probes cannot be directly compared with each other, these data might suggest that especially Wnt pathway activation is critical for induction of ectopic placodes downstream of Eda. Further, our microarray associated Eda with several Wnt pathway genes, but revealed no link between Eda and Tbx3 or Nrg3. Two Fgf ligands (Fgf17 and Fgf20) were upregulated by Eda, but these Fgfs are thought to signal mainly via the mesenchymally expressed c isoforms of Fgfrs, not Fgfr2b [43], and are thus unlikely to function in a manner similar to Fgf10.
In order to be able to manipulate and follow mammary placode formation more precisely, we developed an ex vivo tissue culture setup. In brief, ventrolateral skin explants containing the milk line region were dissected from E12.5 embryos and grown in a Trowell-type culture system as described previously [44]. Explants isolated from K14-Eda and control littermate embryos were cultured for a period of two days. After one day (E12.5+1d), wild type and K14-Eda samples appeared almost identical (Fig 8A and 8B). By E12.5+2d, K14-Eda explants were clearly distinguishable from controls due to the presence of supernumerary bud-like structures that had formed between buds 3 and 4, and occasionally also between buds 2 and 3 (Fig 8A and 8B). Typically, 2–3 supernumerary primordia formed between buds 3 and 4. The explants thus recapitulated the in vivo phenotype very closely [27].
Next, we tested whether recombinant Fc-Eda protein had the capacity to induce formation of ectopic buds ex vivo. After one day (E12.5+1d), control and Eda-treated samples appeared fairly similar (Fig 8C and 8D) although incipient supernumerary placodes were observed in Eda-treated specimen. A day later, similar to K14-Eda explants, several ectopic bud-like structures had developed within the milk line in Eda-treated specimen, whereas the controls showed no morphological changes in this region (Fig 8C and 8D). Increased NF-κB reporter activity was evident in response to Eda treatment at the sites of presumptive supernumerary placodes (Fig 8E and 8F).
The endogenous mammary buds of both control and Eda-treated explants expressed Wnt10b, Krm2 and PTHrP. As in vivo, expression of these genes was observed in Eda treated samples between buds 3 and 4 (S4A–S4C Fig). Sonic hedgehog (Shh) is a hair lacode-specific marker whose expression is barely detectably in mammary buds [45]. No Shh expression was observed in endogenous buds or in the region where supernumerary mammary primordia formed in control or Eda-treated samples (S4D Fig).
Upregulation of several Wnt pathway genes by Eda suggests involvement of canonical Wnt signaling in the induction of supernumerary placodes. However, the effects of Wnt pathway would be difficult to assess genetically due to several putative target genes of Eda that could act redundantly. Instead, we cultured E12.5 wild type and K14-Eda explants in the presence of XAV939, an inhibitor of the canonical Wnt-pathway [46]. Application of XAV939 on tissues of TOP-gal Wnt reporter embryos confirmed significant downregulation of Wnt signaling in all treated explants (17/17 explants) (Fig 9A). Supernumerary placodes were always observed in non-treated K14-Eda samples at E12.5 + 2d whereas their formation was greatly reduced by low (10 μM) and almost completely inhibited by high (40 μM) concentration of XAV939, respectively (Fig 9B–9F). At these concentrations, XAV939 had no apparent effect on endogenous buds in wild-type or K14-Eda explants. We also tested the effect of XAV939 on endogenous placodes at the time when they emerge (E11.0) and visualized forming mammary primordia with the aid of K17-GFP transgene [47]. 40 μM of XAV939 did not prevent formation of endogenous placodes, although placode size was clearly reduced (S5 Fig) indicating that supernumerary placodes are more sensitive to Wnt inhibition than endogenous ones. In conclusion, these data suggest that Eda signaling upregulates Wnt activity within the milk line which leads to formation of ectopic mammary placodes.
We report here that mice overexpressing the Tnf-like ligand Eda develop supernumerary mammary glands not only along the milk line [26, 27], but additionally in the neck region, anterior to mammary gland 1. Further, based on pregnancy-associated morphological changes, these glands are functional. Traditionally the region possessing mammary potential has been thought to be limited to the area between the axilla and the genital tubercle [1]. The murine mammary line has been identified by a streak of Wnt10b expressing cells [2]. Initially three separate streaks form: a central streak between the limbs appears first followed by independent stripes at the ventral border of each of the limbs where placodes 1 and 5 form [2]. We also observed similar stripes of NF-κB reporter expressing cells in the axilla and groin, and as reported for Wnt10b [2], they only later became connected with the central streak of the milk line.
In addition to the mammary line, a separate streak of Wnt10b-positive cells, named the dorsal line, has been noted but the importance of these cells has remained elusive [2]. This streak is located dorsally to the milk line, encircles the fore limb bud from the dorsal side and ends at the anterior edge of the fore limb bud. Intriguingly, this is exactly where ectopic placodes form in K14-Eda embryos. We observed a discernible cluster of Wnt10b-positive and Dkk4-positive cells, and a less pronounced aggregate of NF-κB reporter expressing cells, in this location also in control embryos. Inspection of published pictures reveals that this domain is also positive for several other placode markers including Wnt6, Tbx3, and s-SHIP-GFP [2, 33, 48]. Further, it is characterized by high TOP-Gal activity [3, 49]. Collectively, these data and our new findings indicate that the murine milk line extends past the axillary area.
Our data suggest that Eda induced mammary cell fate supporting signals result in the maintenance of a normally transient group of mammary cells at the far end of the dorsal line. Similarly, Eda has been proposed to sustain a transient signaling center in the dental lamina resulting in the formation of an ectopic tooth in K14-Eda mice [27]. On the other hand, in wild-type embryos the area between buds 3 and 4, another region where supernumerary mammary glands develop in K14-Eda mice, is characterized by scattered rather than clustered Wnt10b-positive cells. This may explain why Eda can readily overcome the developmental threshold for placode induction in the neck region leading to the appearance of supernumerary placodes in this position substantially earlier than elsewhere in the milk line. At the time when supernumerary placodes arise between buds 3 and 4, the streak of Wnt10b-positive cells in no longer detectable in control embryos, but it is maintained/reappears in K14-Eda embryos. However, similar to the endogenous buds, Wnt10b expression becomes later confined to the newly formed buds. This suggests that similar mechanisms account for the formation of both endogenous and supernumerary mammary primordia.
Our analysis on Eda-/- embryos showed that lack of Eda does not interfere with the patterning of endogenous mammary placodes. NF-κB is thought to be activated by all Tnf receptors, but also JNK and p38 pathways can be employed by many Tnfrs [50]. JNK pathway has been suggested to mediate Edar signaling, at least in some cultured cell lines [51]. Though NF-κB can be activated by multiple stimuli we found no evidence that other NF-κB activating cues besides Eda operate during mammary placode formation. As NF-κB was shown to be dispensable for mammary gland induction, it was surprising that formation of Eda-induced supernumerary placodes was NF-κB -dependent. We find it plausible that Eda/NF-κB has a role in the formation of endogenous mammary placodes as well, but loss of function may be compensated for by other pathways, in particular those that enhance Wnt signaling activity (see also below). Nrg3/ErbB4 is one potential pathway that could exert this function as Nrg3-coated beads induce Lef1 expression concomitant with ectopic placodes-like structures in vitro [15].
Our microarray profiling of genes differentially expressed in mammary buds upon short exposure to recombinant Eda protein revealed members of several signaling pathways. This implies that Eda may act by tinkering the activity of multiple mammary-associated pathways during morphogenesis. We found significant changes in expression of Wnt, Fgf, Tnf, and chemokine pathway genes, similar to our previous findings in Eda-regulated genes in hair placodes [35, 52]. Although these two studies cannot be directly compared due to different microarray platforms used, it seems that the gene regulatory network governed by Eda is largely shared between hair follicles and mammary glands.
We found changes in several Wnt pathway genes upon Eda treatment; both agonists (Wnt10a, Wnt10b, Lef1, and Lgr4) and antagonists (Lrp4, Kremen2 and Dkk4) were upregulated, as also observed in hair placodes [24, 34, 52]. Proper spacing of many ectodermal appendages is believed to be achieved by combinatory regulation of positive and negative signals [53, 54]. The reaction-diffusion model suggests that soluble factors that either promote or inhibit placode fate are co-expressed in placodes. However, unequal diffusion/stability of these substances may result in higher activator activity in the placodes whereas in the surrounding tissue, the opposite, higher inhibitor to activator ratio, prevents acquisition of placode fate. Further, it seems plausible that these cues are fine-tuned by several different pathways. The ability of Eda to modulate the expression of both placode activators and inhibitors in combination with the input from other signaling pathways may also explain the puzzling finding that HED patients may have both missing and supernumerary nipples [21, 23].
We propose that maintenance and/or enhancement of Wnt pathway activity is the critical molecular mechanism whereby Eda induces the formation of mammary placodes. Our conclusion is based on the following findings: 1) Canonical Wnt signaling is absolutely necessary for mammary placode induction and genetic deletion of Wnt pathway antagonists (Lrp4, Sostdc1) causes more epidermal cells to adopt mammary cell fate along the mammary line [3, 31, 49]; 2) Several Wnt pathway genes show an early upregulation at the prospective site of ectopic mammary placodes in K14-Eda embryos; and 3) Pharmacological inhibition of Wnt signaling suppresses formation of supernumerary placodes in K14-Eda mammary explants in a dose-dependent manner at doses that however, do not yet prevent formation of endogenous placodes. Eda and Wnt signaling pathways are intertwined during development of several ectodermal organs [24, 34, 52, 55, 56]. In primary hair placodes, Wnt/β-cat signaling enhances Edar expression which in turn is required for upregulation of Wnt10a/b to levels high enough for placode morphogenesis to proceed. In the absence of Eda, placode formation is halted at a rudimentary ‘pre-placode’ stage characterized by severely reduced levels of Wnt activation [19, 34, 52, 57]. However, mammary placodes are largely insensitive to loss of Eda. Interestingly, Lgr4 deficient embryos display a similar defect in primary hair placodes as Eda null embryos [58] but all mammary glands form in Lgr4 deficient mice [59]. Collectively, these data indicate that in the absence of Eda, cues other than Eda/NF-κB are responsible for maintenance of Wnt10a/b/Lgr4 expression and thereby sufficient Wnt signaling activity to support early mammary morphogenesis. Alternatively, other Wnt ligands/Lgrs that are insensitive to Eda levels may have a more critical role in mammary primordia than in hair placodes.
The number of mammary glands is usually considered to be a species-typical invariant trait [60–62] indicating that the balance between mammary fate promoting signals must be tightly balanced with inhibitory cues to ensure the development of the correct number of mammary glands. Yet, in some species such as the pig, dairy cattle, and multimammate mice (a.k.a. African soft-furred rat), Mastomys natalensis and its closely related species, a notable intraspecific variability has been observed [60, 63, 64]. Even in humans, accessory nipples and breast tissue are found at relatively high prevalence (estimates ranging from 0.2% to 5.6%) [65, 66]. These findings show that the mammary line has the capacity to produce more than the species-typical number of organs. Misregulation of the Eda/Wnt pathway could offer an explanation for some sporadic cases of polythelia or absence of breast.
The number and location of mammary glands vary widely between mammals [60, 62, 67]. In humans and other primates mammary glands are located at the thoracic region, in most ungulates at the inguinal region, in mice, cats and dogs at both regions, whereas pigs have them along the entire length of the milk line. Usually the number of pairs corresponds to the average number of offspring born at a time [61, 67]. Highest mammary gland numbers are seen in some marsupials and domesticated pigs, whereas mice and rats have maximally 6 pairs [60–62]. Multimammate mice are a striking exception with 8 to 12 pairs, or even more, scattered throughout the mammary line [64, 68] thereby greatly resembling K14-Eda mice. However, it is not known whether the milk line is expanded anteriorly as in K14-Eda mice. Changes in the Eda pathway activity have been linked to intraspecies evolutionary adaptations in the numbers of skin appendages in two species: the amount of armor plates in marine vs. freshwater threespine sticklebacks and the sweat gland density in modern human populations [69, 70]. It is tempting to speculate that differential expression levels of the Eda pathway components account for some of the interspecific differences observed in the number and position of mammary glands.
The generation and genotyping of the following mouse strains have been described: K14-Eda [26], IκBαΔN [19], Eda null (Tabby) (Jackson Laboratories; stock no. 000314), TOP-gal (Jackson laboratories; stock no. 004623), K17-GFP, Foxi3-deficient, and NF-κB reporter mice [24, 47, 71, 72]. K14-Eda, Foxi3-deficient, K17-GFP, and NF-κB rep mice were maintained on the C57Bl/6 background. IκBαΔN mice were bred in the C57BL/6 or a mixed C57BL/6 and FVB background. Eda null and TOP-Gal mice were on B6CBA and NMRI backgrounds, respectively. The appearance of a vaginal plug was considered the embryonic day (E) 0.5. The age of the embryos were further staged according to the limb morphogenesis [73] and other external criteria.
All mouse experiments were approved by the local ethics committee and National Animal Experiment Board of Finland under licenses KEK13-020 and ESAVI/2984-04.10.07–2014. The mice were sacrificed with CO2 followed by cervical dislocation.
Embryos or dissected tissues were fixed overnight in 4% PFA at 4°C, processed through rising ethanol series and xylene into paraffin and sectioned at 5 μm. Whole mount X-gal staining was done according to a published protocol [74]. The samples were postfixed with 4% PFA. When sectioned, the counterstain was performed with Nuclear fast red. Processing of the mammary glands and the Carmine alum staining was performed as previously described [24]. Whole embryos and tissues were photographed using the Olympus SZX9 stereomicroscope and slides with the Zeiss Imager.M2.
For the keratin2e immunostaining, sections were deparaffinised and citrate-treated in 6mM sodium-citrate buffer (pH 6). The blocking was done with 5% goat serum in 3% BSA in PBS. The samples were incubated overnight with a primary mouse antibody against keratin2e (10R-C166a, 1:200, Fitzgerald) followed by a goat anti-mouse-HRP secondary antibody (1:500; Jackson Immuno Research). Detection was done with the Vectastain Elite ABC Kit (Vector Laboratories) and counterstain with haematoxylin. The whole mount immunostaining for Madcam1 was performed with a primary rat antibody against Madcam1 (550556, 1:25; BD Pharmingen) and a secondary anti-rat-HRP antibody (1:200, Santa Cruz Biotechnology). The DAB substrate kit for peroxidase (Vector Laboratories) was used for detection. Unspecific staining was blocked with 1% dry milk in 1xPBS/0.1% Tween-20.
The ventrolateral skins that contained the mammary forming region and at least the endogenous buds 2, 3 and 4 were dissected from E12.5 embryos and half embryo explants were prepared from E11.0 embryos as indicated in the text. The explants were cultured for 1 to 2 days in a Trowell-type culture setting [24, 52]. The medium consisted of 1:1 mixture of DMEM and F12 (Ham’s Nutrient Mix: Life Technologies) and was supplemented with 10% (vol/vol) FCS (PAA Laboratories), 2 mM l-glutamine, penicillin-streptomycin and ascorbic acid (75 mg/L). When indicated, recombinant Eda protein (Fc-Eda-A1) [75] was added to the growth medium to achieve a final concentration of 250ng/mL. Wnt inhibitor XAV939 in DMSO (Stemgent) was used as 10 μM or 40μM concentrations. Two separate stock solutions were generated for the inhibitor in order to avoid DMSO concentrations higher than 0.25% in the culture medium. Each time, one side of the embryo was used as the control and the other was treated with XAV939.
The embryos or tissue culture samples were fixed overnight in 4% PFA at 4°C and processed for whole mount in situ hybridization or for paraffin-embedding. The whole mount in situ hybridization was performed with inSituPro robot (Intavis AG). The following digoxigenin-labelled RNA probes were used: PTHrP [76], Wnt10b [77], Wnt10a [78], Lef1, β-catenin, Shh [79], Tbx3, Nrg3 [15], Kremen2 (nucleotides 1306–1703 of NM_028416.2), Mmp9 (nucleotides 527–1131 of NM_013599.3) and Lgr4 (nucleotides 3408–3823 of NM_172671.2). The detection was achieved by using BM Purple AP substrate Precipitating Solution (Boehringer Mannheim). Radioactive in situ hybridization on paraffin sections was carried out according to previously described protocols using 35S-UTP labelled (Amersham) probe specific to Edar [80].
The hanging drop culture has been described in detail elsewhere [24, 52]. In short, two pools of 15–20 E13.5 Eda-/- mammary buds from 4 or 5 embryos were collected for each sample pair: one pool was treated with 250 ng/mL of Fc-Eda for 4h, whereas the other one was maintained in a control medium for 4h. RNA extraction and cDNA synthesis was performed as described previously [24, 52]. qRT-PCR was done in a LightCycler 480 (Roche, Indianapolis, IA) and the following analysis was done with software provided by the manufacturer. The expression data were normalized against Ranbp1 gene. For primer sequences see (S2 Table).
Unpaired Student’s t- test was used for statistical analysis of all data. P-values of ≤0.05 were considered to be statistically significant.
E13.5 mammary buds were dissected from Eda-/- embryos and used either as a control or exposed to 250 ng/mL of Fc-Eda as described above. 15–20 mammary buds were pooled in one sample, and three biological replicates were collected. RNA was extracted as previously described [24, 28] and RNA quality was monitored using a 2100 Bioanalyzer (Agilent Technologies). RNAs were processed and hybridized on Affymetrix Mouse Exon 1.0 ST arrays (Santa Clara, CA) in the Biomedicum Functional Genomics unit (University of Helsinki, Finland). Significance analysis between treated and control samples was done using three statistical tests. In each test, a paired t-test (pairing was done over treatment-control pairs) was applied to the data. Differentially expressed genes were detected using Limma, IBMT (intensity based moderated t-test), and Cyber-T. All methods were applied with default parameters. Obtained p-values were adjusted for multiple testing using Storey’s q-value method. The data discussed in this publication have been deposited in NCBI's Gene Expression Omnibus [81] and are accessible through GEO Series accession number GSE69781 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE69781)
|
10.1371/journal.pgen.1008119 | Temporal genomic contrasts reveal rapid evolutionary responses in an alpine mammal during recent climate change | Many species have experienced dramatic changes in their abundance and distribution during recent climate change, but it is often unclear whether such ecological responses are accompanied by evolutionary change. We used targeted exon sequencing of 294 museum specimens (160 historic, 134 modern) to generate independent temporal genomic contrasts spanning a century of climate change (1911–2012) for two co-distributed chipmunk species: an endemic alpine specialist (Tamias alpinus) undergoing severe range contraction and a stable mid-elevation species (T. speciosus). Using a novel analytical approach, we reconstructed the demographic histories of these populations and tested for evidence of recent positive directional selection. Only the retracting species showed substantial population genetic fragmentation through time and this was coupled with positive selection and substantial shifts in allele frequencies at a gene, Alox15, involved in regulation of inflammation and response to hypoxia. However, these rapid population and gene-level responses were not detected in an analogous temporal contrast from another area where T. alpinus has also undergone severe range contraction. Collectively, these results highlight that evolutionary responses may be variable and context dependent across populations, even when they show seemingly synchronous ecological shifts. Our results demonstrate that temporal genomic contrasts can be used to detect very recent evolutionary responses within and among contemporary populations, even in the face of complex demographic changes. Given the wealth of specimens archived in natural history museums, comparative analyses of temporal population genomic data have the potential to improve our understanding of recent and ongoing evolutionary responses to rapidly changing environments.
| Museum specimens represent an irreplaceable archive that can be used to understand how species respond to rapid environmental change. We recovered genomic data from archived samples spanning a century of climate change in co-distributed declining versus stable species of montane chipmunks. Applying novel statistical methods, we find evidence for strong positive selection on a physiologically relevant gene despite increased population fragmentation in the declining species. Our results reveal rapid evolutionary responses, but also highlight that genetic adaptation has been insufficient to prevent range collapse in this endemic alpine species. These findings illustrate how biological archives can be used to pinpoint genetic responses through time to better understand how species are responding to rapidly changing environments.
| Rapid environmental change threatens global biodiversity and has led to population declines in many species [1–5]. Although phenotypic plasticity may enable populations to track rapidly changing climates, evolutionary adaptation will often be essential for long-term persistence [6]. Disentangling plasticity from evolutionary responses ultimately requires resolving the genetic basis of adaptation. However, it remains challenging to differentiate recent or ongoing positive selection from stochastic genetic drift in contemporary populations that are also undergoing extreme demographic changes [7, 8]. Natural history museum collections may hold the key to overcoming many of these difficulties by providing crucial temporal information on species distributions, phenotypes, and population genetic variation spanning periods of recent environmental change [9–11]. Temporal genomic contrasts have begun to yield powerful insights into recent evolutionary responses in humans [12, 13] and other species [14–17], indicating that genetic analyses of biological archives will be an effective tool for understanding evolutionary responses to rapid anthropogenic climate change [18].
Using contrasts between early 20th century and modern museum surveys, Moritz and colleagues [1] showed that the ranges of several high elevation small mammal species in the Yosemite National Park (YNP) region of the Sierra Nevada mountains (California, USA) have retracted upward over the past century. This and associated studies [19, 20] demonstrated the potential of using museum archives to understand species and community-level ecological responses during periods of recent climate change. Contemporary range shifts towards higher latitudes, elevations, or both have now been documented in many terrestrial species [5, 19, 21, 22], and are generally thought to reflect direct and indirect population responses to warming temperatures [22, 23]. However, these works have also highlighted that closely related species can differ markedly in the magnitude and direction of their ecological responses, for reasons that are often not clear [1, 20]. Currently, we know very little about how recent range shifts have affected evolutionary processes within species, or the extent to which evolutionary genetic responses have been synchronous within and between co-distributed species.
Here we focus on two chipmunk species within the YNP montane mammal community that show different ecological and phenotypic responses over the last century of climate change (Fig 1). Western chipmunks have long been considered models for niche partitioning by elevation and habitat type [24–26]. The alpine chipmunk (Tamias alpinus) is an ecological specialist endemic to the high elevation alpine habitats of the Sierra Nevada Mountains. The lodgepole chipmunk (T. speciosus) occurs more broadly across mid- to high-elevation subalpine coniferous forests of California. Tamias alpinus has undergone severe range contraction driven by extirpation of lower elevation populations [1, 20] combined with pronounced shifts in diet and cranial morphology [27] across the alpine zone of YNP and elsewhere in the Sierra Nevada mountains. Spatial modeling of current versus historical ranges across YNP suggests that the strong upward contraction of T. alpinus is best explained by increases in minimum winter temperatures; competing models including co-occurrence with other species of chipmunks or changes in the distribution of preferred vegetation types did not improve prediction of the observed range contraction in this species [28, 29]. That increasing minimum temperature alone had the strongest impact makes sense in that there has been little change in vegetation across the high elevation talus slopes preferred by this species. By contrast, range contraction in a mid-elevation chipmunk species, T. senex, is best explained by changes in its preferred vegetation types [28, 29]. For YNP T. alpinus, there is also evidence of strong directional selection on skull morphology over the past century [30], whereas the range, diet, and morphology of the partially overlapping lodgepole chipmunk (T. speciosus) has remained relatively stable within YNP [20, 27, 31]. As is generally the case [32], it remains unclear why the montane specialist, T. alpinus, has contracted during the past century whereas its more widely-distributed congener, T. speciosus, has not.
These two species set up a natural contrast, allowing us to understand how differing ecological responses during periods of rapid environmental change correspond to differing evolutionary responses. A previous temporal survey of eight microsatellite markers revealed increased subdivision and declining allelic diversity in YNP T. alpinus, but no significant changes in overall population genetic variation of T. speciosus over the same interval [33]. Synthesizing descriptions of phenotypic [30], behavioral [31], and genetic variation [33] into a detailed understanding of demographic and evolutionary responses in these species requires genomic data. Towards this end, Bi et al. [9] used a custom exon capture platform to enrich and sequence ~11,000 exons (~4 Megabases or Mb across 6,249 protein-coding genes) from 20 early 20th century and 20 modern YNP T. alpinus. These genome-wide data confirmed signatures of increasing population subdivision in this retracting species and illustrated the potential for targeted genomic experiments to generate high quality data from archived specimens. However, neither of these preceding analyses had sufficient sampling to determine whether evident increases in population structure were due to reductions in local population size, migration rates, or both.
Here, we build on these previous works [9, 33] by generating ~9.4 Mb of targeted exome sequence data from 303 chipmunks (194 T. alpinus, 100 T. speciosus, and 9 samples from 4 other species). We used these comparative population genomic data to quantify evolutionary responses over the past century at two scales (Fig 1). First, we sequenced 96 modern (48 T. alpinus and 48 T. speciosus collected between 2003–2008) and 108 historic (56 T. alpinus, 52 T. speciosus collected in 1915–1916) samples collected across geographic transects in YNP for the focal species. Second, we generated an independent geographic transect of 38 modern (2003–2012) and 52 historic (1911–1916) T. alpinus in the Southern Sierras (SS), where this species also shows range contraction [20], to test to what extent evolutionary responses across the range of this declining alpine specialist have been consistent. To analyze these temporal data, we developed a novel analytical framework based on Approximate Bayesian Computation (ABC) that allowed us to quantify changes in population sizes and migration rates in the context of demographic history, and then to localize positive selection on standing genetic variation at specific genes. Our results provide new insights into recent evolutionary responses in this system and demonstrate how a genomic time-series approach can be broadly applied to other archived specimens to improve understanding of evolutionary responses over the past few centuries.
High sequencing coverage is necessary to reliably genotype ancient DNA samples [34] due to extensive DNA degradation [35]. This persistent technical challenge makes whole genome resequencing of historic mammalian populations impractical, especially in species without high-quality reference genomes (e.g., Tamias). Therefore, we designed a custom targeted capture to enrich and sequence exons from over 10,107 protein-coding genes (9.4 Megabases or Mb total) in 294 T. alpinus and T. speciosus samples (S1 Table) and nine samples from four other chipmunk species (total n = 303). We sampled geographic transects of modern and historic (~100 year-old) populations in YNP for both species as well as an independent SS transect of T. alpinus, where this species has also contracted [20]. An average of 49 individuals were sequenced per population (Fig 1). This design allowed us to (i) compare stable and retracting species within the same montane mammal community (YNP), and to (ii) determine to what extent the same evolutionary responses have occurred across two transects (YNP and SS) spanning the latitudinal range of the range-retracted species T. alpinus.
Exome enrichment was highly specific (90–93% of cleaned reads on target) and sensitive (>92% of the target regions sequenced), resulting in high coverage of targeted regions (26–35× average individual coverage per population, S2 Table). Although historic DNA samples are notorious for poor technical performance, all targeted T. speciosus and T. alpinus individuals yielded moderate to high coverage data with similar capture success between modern and historic samples. Analysis of mitochondrial DNA indicated that empirical error rates were ~fourfold higher in historic (0.16%) versus modern samples (0.04%), due primarily to DNA damage typical of century-old museum samples [9, 36]. Although nucleotide misincorporations associated with deamination of methylated cytosine bases (C-to-T and G-to-A) were most common near the ends of DNA fragments, such changes remained elevated throughout the sequence (S1 Fig). Therefore, we applied several quality filters to remove all putative misincorporations and to mitigate other common sources of genotyping error [9] (S3 Table). All filters were uniformly applied to historic, modern, and simulated data (see below) to facilitate comparisons across time periods and species. After filtering, we identified 20,395, 10,395, and 10,954 high-quality single nucleotide polymorphisms (SNPs) in YNP T. speciosus, YNP T. alpinus, and SS T. alpinus, respectively.
Upwards range contraction in montane environments can lead to increased population structure and reduced genetic diversity due to decreased population size. We observed a consistent trend towards relatively minor reductions in nucleotide diversity in modern versus historic samples of T. alpinus and T. speciosus estimated at two spatial scales: metapopulations (e.g., YNP or SS) and demes of spatially clustered sampling localities (θπ and θW; S2 Fig, S4 Table), consistent with prior results [33]. We then quantified the degree of population genetic structure within each species by estimating the global fixation index (FST) for historic and modern populations. Within YNP, population structure was relatively low overall but increased nearly two-fold in modern T. alpinus (FST,historic = 0.032, FST,modern = 0.058), consistent with increased population fragmentation as this species has contracted upwards [33]. By contrast, population structure in T. speciosus increased only slightly over time (FST,historic = 0.027, FST,modern = 0.030). We detected very little uncertainty in our estimates of the site frequency spectrum (SFS), and estimates of global FST showed non-overlapping 0.95 bootstrapped confidences intervals for all pairwise temporal contrasts (S4 Table). Collectively, these patterns suggest that the stronger increase in FST observed for YNP T. alpinus is not simply explained by reductions in per deme nucleotide diversity, which are of similar magnitude across YNP and SS T. alpinus and YNP T. speciosus (S2 Fig).
To further evaluate changes in the spatial patterning of population genetic structure, we used NGSadmix to conduct a maximum likelihood (ML) analysis of historic and modern population structures [37]. We detected a substantial increase from two to six genetic clusters for YNP T. alpinus, compared to stable overall population structure (K = 2 across both time points) within T. speciosus (Fig 2A; S5 Table). Principal component analyses of these data also indicate less genetic similarity among modern YNP T. alpinus samples (Fig 2B). These genome-wide estimates incorporate genotype uncertainty to provide an accurate overview of changes in genetic diversity and structure over recent timescales. Although consistent in the direction of change, the increases in YNP T. alpinus population genetic structure appear even more striking than previously detected using lower resolution data [33, 38]. By contrast, there was minimal structure in T. speciosus within the same ecosystem whether considering a few microsatellite loci [33] or thousands of SNPs (Fig 2A).
Consistent with patterns of increased population genetic structure, spatial models of occurrence indicate that range contraction has reduced local connectivity between suitable habitat patches for YNP T. alpinus [33]. Given that range contraction has also been detected at the southern limit of the range of T. alpinus [20], we next tested to what extent similar temporal signatures of increased genetic structure were apparent in SS populations (Fig 1). Population structure in the SS transect has also increased (FST,historic = 0.034, FST,modern = 0.044), but to a lesser extent than observed in YNP, and with no overall increase in the number of distinct genetic groups (modern and historic K = 2; Fig 2A). This suggests some variation in the local genetic consequences of seemingly synchronous, range-wide contractions. However, we note that our power to detect changes in overall population structure may have been limited by the fact that historic and modern sampling localities were more spatially clustered in SS when compared to YNP (Fig 1).
Rapid range shifts could also increase the likelihood of hybridization if there are changes in the degree of sympatry or in relative densities among closely related species [39, 40]. Occasional hybridization appears to be common in western chipmunks [41], including evidence for ancient mitochondrial introgression from the broadly distributed least chipmunk, T. minimus, into T. speciosus [42]. The alpine chipmunk is very closely related to T. minimus with evidence of historical gene flow [38], raising the possibility that recent range fluctuations have induced hybridization between T. alpinus and either T. minimus or T. speciosus. However, we found no evidence for recent appreciable nuclear gene flow between T. alpinus and adjacent populations of T. speciosus or with neighboring (lower elevation) populations of T. minimus (S3 Fig). Thus, recent range collapse does not appear to have led to the breakdown of reproductive barriers between this high elevation endemic and other co-distributed species [38]. We did, however, detect nuclear introgression from T. speciosus into at least one T. minimus sample (S3 Fig), suggesting that reproductive isolation remains incomplete between these species despite strong ecological partitioning [26].
These basic descriptions of genetic variation expand the scope and resolution of previous analyses of a few microsatellite loci genotyped in T. alpinus, T. minimus, and T. speciosus [33, 38], and from more limited exon capture data from T. alpinus [9]. However, disentangling changes in migration versus local effective population size, and identifying genes under positive selection in the context of recent demographic change, requires more comprehensive analyses. Therefore, we developed a novel analytical framework to fully exploit our temporal dataset. Traditional population genetic analyses often assume even sampling across space and time, yet studies using museum specimens are typically imbalanced because of limited availability of samples. ABC is well suited for demographic inference under such circumstances because biased temporal sampling and sample processing can be simulated for populations that have experienced complex demographic histories. Accordingly, simulations can be filtered in the same way as observed data (e.g., removal of SNPs associated with errors in historic DNA), permitting meaningful comparisons between expected and observed results. A major difficulty of ABC is the choice of statistics that sufficiently describe demographic parameters of interest. Multiple, jointly informative, summary statistics are often used to sufficiently estimate parameters while reducing the risk of any particular statistic biasing the results [43]. The site frequency spectrum (SFS) is often an optimal choice for fitting demographic histories since many commonly used summary statistics can be derived from it. In practice, high dimensionality and low count categories of joint site frequency spectra make them difficult to fit. Consequently, we developed an effective means of fitting binned two-dimensional SFS (2D-SFS) using an ABC framework designed to infer population histories from serially sampled metapopulations (S4 Fig).
We constructed 2D-SFS for each of the three pairwise temporal contrasts by pooling individuals across sampling localities within YNP or SS into a single metapopulation per time period (S5 Fig). While allele frequencies should be highly correlated over such short time scales, the overall shape of the joint spectrum should change in predictable ways in response to various population-level processes. We fitted multiple demographic models to the 2D-SFS (S6 Fig; S7 Fig) describing population size (constant size, bottlenecks, expansions) and connectivity (migration) among subpopulations or demes representing spatially clustered localities (see S6–S9 Tables and S1 Text for details on model selection, evaluation, and inference). The best fitting population history for YNP T. alpinus was characterized by relatively small but constant deme effective sizes through time (~1,350 individuals) and an approximately threefold decrease in migration within the past ~90 years (Fig 3). This fitted demographic model fits well with our observation of increased population structure in YNP T. alpinus (Fig 2) despite only minor reductions in nucleotide diversity per deme through time (S2 Fig, S4 Table). In contrast, both SS T. alpinus and YNP T. speciosus were found to have much larger effective deme sizes (~4,600 and ~4,560 individuals respectively) and higher migration rates overall. Consistent with the observed increase in FST, SS T. alpinus showed some evidence for a recent, very small decline in effective size (Fig 3). YNP T. speciosus was the only population that showed a clear signature of size change, albeit related to a historic (pre 20th C) population expansion (Fig 3).
We found that modern samples tended overall to be less genetically similar to each other than did historic samples in all three comparisons (Fig 2B; S8 Fig). Consistent with this observation, the fitted demographic histories for each species and transect support a recent (< 90 years ago) decrease in migration among demes (Fig 3; see S1 Text, S6 Table). Decreased migration is expected if climate change is broadly affecting the amount of connectivity between suitable habitat patches available to species within this montane community [33]. However, as noted above, substantially increased genetic structuring was most evident in YNP T. alpinus (Fig 2A). Long-term changes in gene flow should ultimately affect the genetic composition of metapopulations across these landscapes. It is possible that higher overall migration rates and larger effective population sizes have so far buffered the population genetic effects in T. speciosus and SS T. alpinus. However, genetic structure could accumulate over time according to the inferred histories.
These results emphasize the utility of high-resolution demographic inference from genomic data not only for reconstructing population histories, but also as a potentially powerful conservation management tool [18, 44, 45]. An ABC framework is generalizable to other temporally sampled genetic datasets, allowing high-resolution inference into demographic histories over shallow evolutionary timescales that are relevant to recent anthropogenic climate change. In species of conservation concern, signatures of reduced migration could be used to motivate introductions between populations prior to significant genetic erosion, buffering against the future loss of genetic diversity and the accumulation of deleterious variation [46]. The benefits of such proactive efforts would have to be weighed carefully relative to the potential risks of introducing locally maladaptive variation [47].
Connections between broad ecological patterns and the genetic structure of populations are often intuitive and predictable. Upwards range contraction in T. alpinus is associated with reduced connectivity between suitable montane habitats [33], which we infer has reduced migration between patches and increased genetic drift. However, a priori expectations for patterns of recent adaptive evolution are far less predictable in these species. Recent range shifts [20] and temporal changes in diet and skull morphologies [27, 30] are both consistent with stronger directional selection gradients in YNP and SS T. alpinus relative to T. speciosus. On the other hand, lower effective population sizes and reduced migration (at least in YNP) should make selection relatively less effective in T. alpinus. Likewise, more effective adaptive responses could explain why larger and more connected populations of T. speciosus have remained stable in the face of common environmental stressors.
To begin to tease these issues apart, we tested for specific genetic changes that might underlie recent adaptive responses in these species by directly comparing genetic differences between historic and modern populations. All three temporal population pairs are very closely related (Fig 2), however, they are also separated by changes in population structure and sizes (Fig 3) that may confound standard signatures of positive selection [48]. Therefore, we tested for individual SNPs that had undergone large frequency shifts between historic and modern populations using an approach that is robust to the confounding influence of complex population histories on the genomic distribution of FST [49–51]. We found no significant allele frequency shifts over time in YNP T. speciosus or SS T. alpinus. In contrast, we identified five outlier SNPs in YNP T. alpinus populations (false discovery rate [FDR] q-value < 0.01) relative to the inferred null distribution of per-site, genome-wide FST between the temporally sampled populations (genome-wide temporal FST = 0.012; Fig 4A). To verify the inference of positive selection on these SNPs, we compared the observed FST values to null distributions simulated under the best ABC-fitted demographic history for YNP T. alpinus (Fig 3). Our simulated FST distributions were in close agreement with the overall observed FST values. Thus, it is very unlikely that demography alone could produce the extreme changes in allele frequencies that we observed at the outlier loci (p-value < 3e-7; S9 Fig).
Derived allele frequencies at all five differentiated SNP positions increased ~threefold in the modern populations (average frequencies of 0.22 historic versus 0.65 modern; Fig 4B) and all were located in the protein-coding gene, Arachidonate 15-Lipoxygenase (Alox15) (Fig 4D). Alox15 is a broadly expressed lipoxygenase involved in the resolution of acute inflammation through the generation of lipid-derived signaling molecules known as resolvins [52–54]. Alox15 expression has been associated with cardiovascular disease, oxidative stress, and response to hypoxia [55–57] as part of the Hypoxia-inducible factor-1α (HIF-1α) regulation pathway [58]. Two of the outliers represent synonymous changes in non-adjacent exons (positions a, b; Fig 4D) while the three other SNPs (positions c-e) were at non-coding positions within the same intron. All five positions were in strong linkage disequilibrium (historic r2 = 0.86; modern r2 = 0.93) in YNP T. alpinus but invariant in all other populations except for one site (b) that was at similar frequency across the SS T. alpinus temporal contrast (historic = 0.13, modern = 0.2).
Given an ~500m contraction of the low elevation range limit in YNP T. alpinus over the last century [20], the temporal allele frequency shifts at Alox15 could simply reflect non-sampling of extinct low elevation populations. To test this, we first estimated Alox15 allele frequencies as a function of elevation by pooling individuals into discrete 100-meter elevation bands. We observed the largest increases in derived allele frequencies at low to mid-elevation localities (Fig 4C), and mean derived allele frequencies were not significantly correlated with elevation in either of the temporal samples (historical R2 = 0.46, p-value = 0.14; modern R2 = 0.34, p-value = 0.42). Furthermore, all five positions remained strong outliers in our temporal FST contrasts when we excluded low elevation sampling localities that were present only in the historic YNP T. alpinus transect (OutFLANK FDR q < 0.05, ABC-fitted FST distribution p-value = 4e-7). Thus, evolutionary responses at Alox15 are consistent with in situ evolutionary change primarily among remnant demes below the upper bound of the modern YNP T. alpinus range (<3200 meters elevation).
In principle, the large shift in Alox15 allele frequencies observed between historic and modern samples could be driven by changes in habitats, food availability, or some other non-climate related environmental factor. However, based on previous modeling of changes in the elevational range [28] and the function of Alox15, we suggest that physiological response to warming is the strongest current hypothesis. Winter temperature appears to be a primary limiting factor in the distribution of T. alpinus, with range contractions strongly tracking upslope shifts in minimum winter temperatures [28]. Increases in minimum winter temperatures at mid-elevations are resulting in reduced YNP snowpacks [59, 60], which Rubidge and colleagues suggested might reduce over-winter survival of T. alpinus through loss of critical thermal insulation of hibernacula [28]. Interestingly, arousal from hibernation has been shown to induce oxidative stress and hypoxia [61] and Alox15 shows strong seasonal induction in other species of hibernating squirrels [62].
A potential link between the intensity of selection on variation at Alox15 and changes in winter snowpack could also explain why we did not detect selection at the same gene in the SS transect. Tamias alpinus populations in the Southern Sierra are fixed for ancestral alleles at all but one of the outlier YNP SNPs, suggesting that these populations may lack genetic variation at Alox15 that is putatively adaptive in YNP. Moreover, SS T. alpinus populations are currently found above ~3200 meters—above the elevation range showing the largest allele frequency shifts in YNP—and overall snowpack has been more stable in the southern Sierra during the last century [59]. Though speculative, these scenarios help illustrate how evolutionary responses among populations may depend on both adaptive potential (i.e., standing genetic variation) and local environmental conditions.
Temporal sampling of genomic data has the potential to provide powerful insights into the evolutionary effects of rapid environmental change [18]. Here we built on previous works [1, 9, 33] by generating targeted genome-wide sequence data from 294 chipmunks spanning a century of climate change. By integrating high throughput sequencing, cost and time-effective targeted enrichment technologies, and sophisticated inference methods, we provide powerful comparative insights into demographic and evolutionary responses of two montane species experiencing rapid environmental change. Our genomic time-series approach demonstrates one way that historical archives can be used to study biological responses during recent environmental change [9, 11, 18]. Temporal genomic data can provide a means to understand the current state of populations and their potential evolutionary trajectories, providing powerful tools to inform the conservation of populations experiencing changing environments.
The identification of targets of positive selection during the recent upslope range contraction in T. alpinus points to a candidate gene and potential phenotypes associated with physiological stress that warrant further study. We caution that further evidence, such as differences in over-winter survival across genotypes or other functional studies, are necessary to demonstrate a causal relationship between Alox15 and response to climatic-induced stress. Further, our capture experiment only covered a subset of protein coding genes (~50%) and did not include extensive coverage of regulatory regions that may often modulate rapid evolutionary responses [63]. That said, alpine chipmunks also show greater stress response to changes in external conditions [64], a narrower range of activity patterns [31], and more pronounced shifts in diet and functional aspects of cranial morphology when compared to T. speciosus over the past century [27, 30]. Thus, the combination of phenotypic, behavioral, and now genetic evidence points to some component of physiological stress as a key factor in the greater sensitivity of T. alpinus to environmental change.
Even in the absence of links to specific phenotypes or fitness, the identification of evolutionary responses at specific genes should help inform future on-ground studies focused on identifying the proximate causes of warming-related population declines across the range of this or other affected species [32]. Indeed, the potential for adaptive evolution to rescue populations in decline has emerged as an important concept in conservation biology [65], with increasing efforts to directly incorporate evolutionary principles into conservation planning [66]. As a cautionary note, our results suggest that putatively adaptive responses in T. alpinus at Alox15 (Fig 4), as well as rapid shifts in functional morphology and diet [27, 30], have nonetheless been insufficient to prevent extensive extirpation of lower elevation populations of this alpine specialist.
Comparative analyses of species range shifts over the past century have provided powerful insights into the ecological impacts of and biological responses to rapid environmental changes [1, 19, 21–23]. Here we have begun to extend these ideas to a comparative population genomic framework. Moving forward, we suggest that the true power of analyzing genomes and phenotypes of historical museum archives lies in the potential to extend across species [9, 11]. Though the occurrence of museum records tend to be highly punctuated through space and time for a given species, historic collection efforts, such as those led by Joseph Grinnell and other early naturalists, usually surveyed many co-distributed species. With comparable contemporary sampling efforts, these invaluable archives will enable comparative community level insights into the impacts of and evolutionary responses to rapidly changing environments.
All animals sampled in the modern era were collected following procedures approved by the University of California, Berkeley Animal Care and Use Committee (Permit number R278–0315). Permits were provided by Yosemite National Park and Sequoia-Kings Canyon National Park.
Tamias speciosus and T. alpinus surveyed in this study were collected from montane transects in Yosemite National Park (YNP) and the Southern Sierras (SS). Historic samples were collected by Joseph Grinnell and his colleagues from 1911 to 1916, and are preserved as dried skins in the Museum of Vertebrate Zoology (MVZ), at the University of California, Berkeley. Modern samples were collected from the same sites by the ‘Grinnell Resurvey’ team led by MVZ researchers and collaborators from 2003 to 2012 (Fig 1; S1 Table). We examined 100 YNP T. speciosus (52 historic, 48 modern), 104 YNP T. alpinus (56 historic, 48 modern), and 90 SS T. alpinus (52 historic, 38 modern) from each transect. We also sampled six T. minimus (the Least chipmunk) collected east of YNP, which were used to test for potential hybridization between T. alpinus and T. minimus [38]. Furthermore, we included one sample each of three other species (T. striatus, T. ruficaudus, and T. amoenus) in order to polarize SNPs identified in our focal populations. Historic DNA was extracted from toe pad tissue (~3 x 3 mm) in a separate dedicated laboratory using a previously described protocol [9]. DNA was extracted from modern samples using Qiagen DNeasy Blood and Tissue kits following the manufacturer’s protocol. Genomic libraries for all samples were constructed following Meyer and Kircher [67] with slight modifications [9].
We used RNA-seq [68] to sequence and assemble [69] transcriptomes for multiple tissues sampled from a single modern SS T. alpinus to serve as a reference for exome capture probe design. We targeted exonic regions (6.9 Mb, including flanking introns and intergenic regions) corresponding to 8,053 T. alpinus genes targeted by our previous array-based capture experiments in chipmunks [9, 70, 71]. In addition, we extracted a broad set of candidate genes from the AmiGO and NCBI protein databases with functional annotations that were potentially relevant to environmental stress responses (e.g., HSP/HSF, hemoglobin, cytokines, apoptosis, immunity, oxidative stress, oxidative phosphorylation). We then used a BLASTx search against these genes to locate 2,054 orthologous transcripts (2.4 Mb) from the Tamias transcriptome and included these transcripts in our capture. We also targeted the complete mitochondrial genome (~16.4 Kb) to assess empirical error rates and five previously sequenced nuclear genes [42, 72] to use as positive controls in post-capture qPCR assays of global enrichment efficiency. Probes were designed and manufactured by NimbleGen (SeqCap EZ Developer kits).
Barcoded genomic libraries were pooled together and hybridized in seven independent reactions with Tamias Cot-1 DNA and barcode-specific blocking oligonucleotides. Six hybridization experiments were used for the focal species (one per time point for each of the three temporal contrasts) and one additional capture was performed on pooled libraries from six T. minimus and three outgroup samples (T. striatus, T. ruficaudus, and T. amoenus). After hybridization, each of the enriched genomic libraries were amplified using PCR and sequenced using one lane of Illumina HiSeq2000 per capture (100-bp paired-end).
Bioinformatic processing of exon capture data followed our previous protocols [9, 70]. All raw sequencing reads were treated to remove adapters, exact duplicates, low complexity (i.e., runs of ambiguous or mononucleotide sequence), and reads sourced from bacteria and human contamination. Overlapping paired reads were merged to avoid inflated estimates of coverage and biased genotype likelihoods. We used filtered sequencing reads (28.9 Gb) from 48 modern YNP T. alpinus samples to generate de novo assemblies with ABySS [73] that were then merged using Blat [74], CD-HIT [75], and CAP3 [76] to remove redundancies. This total assembly was then compared to the original targets to construct a non-redundant target reference of 21,128 assembled contiguous sequences (contigs) totaling 20.8 Mb, and error-corrected following Bi and colleagues [9]. We then aligned cleaned reads from T. alpinus, T. speciosus, and T. minimus samples to the T. alpinus reference using Novoalign (http://www.novocraft.com). Nucleotide positions were filtered at individual, contiguous sequence, and position levels of quality control following our previously described methods [9] (S3 Table) using the script snpCleaner (https://github.com/tplinderoth/ngsQC/tree/master/snpCleaner). For each of the three temporal transects, we retained the intersection of filtered contigs between all historic and modern populations. As a result, 2,569, 2,451, and 2,738 contigs (11.6–13% of the total) were eliminated from YNP T. speciosus, YNP T. alpinus, and SS T. alpinus datasets, respectively. At the site level, we removed sites showing unusually high or low coverage, excessive strand bias, end distance bias, base quality bias, and map quality bias. We also filtered out sites with extensive missing data among samples within each population.
We were particularly attentive to errors associated with long-term DNA degradation. Postmortem nucleotide damage from hydrolytic deamination causes conversion from cytosine (C) to uracil (U) residues resulting in misincorporation of thymine (T) during PCR amplification [34, 35, 77]. We conservatively removed all C-to-T and G-to-A (i.e., the reverse complement of the C-to-T change with respect to the original PCR template molecules) SNP positions from the datasets to avoid inaccurate population genetic inferences stemming from base misincorporation. In total, 9.0, 9.3, and 8.5 Mb of data from YNP T. speciosus, YNP T. alpinus, and SS T. alpinus passed all quality controls and were used in subsequent analyses.
We used probabilistic methods for variant discovery and allele frequency estimation as implemented within ANGSD [78]. Using a population-specific SFS estimated from allele frequency likelihoods as a prior, we obtained allele frequency posterior probabilities and called SNPs using a 0.95 probability cutoff of being variable. The realSFS function was used to generate 1,000 bootstrap replicates of the folded site frequency spectrum (SFS) for each metapopulation by resampling per site allele frequency likelihoods. We then used ANGSD to estimate the number of segregating sites (S), Watterson's theta (θW), pairwise nucleotide diversity (θπ), and Tajima's D in the historic and modern T. alpinus and T. speciosus metapopulations. For each metapopulation, we generated 100 estimates of θW and θπ using randomly chosen SFS bootstrap replicates as priors to evaluate sensitivity of these point estimates on the SFS prior. Additionally we estimated diversity statistics for demes within metapopulations, the former representing spatially clustered sampling localities. For each transect, population differentiation within and between the modern and historic populations was determined using probabilistic methods for estimating FST [79] and individual covariance matrices for principal component analysis (PCA) as implemented in ngsTools [80]. Confidence intervals (0.95) for global FST were generated from 1,000 bootstrap replicates of per-site FST values. To compare allele frequencies over time, we estimated the 2D-SFS between the pooled modern and pooled historic demes of each transect (i.e., three 2D-SFS comparisons). SNPs identified in T. speciosus and T. alpinus were polarized relative to T. striatus, T. ruficaudus, and T. amoenus. We further examined population genetic structure using NGSadmix [37], which estimates admixture proportions from genotype likelihoods. We ran 10 replicates for K (number of clusters) ranging from 1–10 and summarized results (S5 Table) across runs to determine the best K [81]. To test for hybridization between T. alpinus and T. minimus samples, we used the program ADMIXTURE [82] to estimate individual ancestries using one randomly sampled SNP per contig.
Next we developed an ABC framework for fitting binned 2D-SFS from serially sampled populations or metapopulations (S4 Fig) and used this approach to test hypotheses about the demographic histories of the sampled chipmunk populations. Additional details on demographic model construction, simulations, model selection, and inference are provided in S1 Text. Briefly, we fitted 5–9 explicit demographic models (S6 Fig) characterized by possible changes in migration and population size to each of the temporal contrasts. We performed 25,000 simulations per model, drawing parameter values from uniform or log-uniform prior distributions and then simulating ~20.2 Mb of sequence data for each individual under the specified history using the coalescent simulator fastsimcoal [83]. Lineages from the different demes were sampled at the present (modern sample) and 90 generations in the past (historic sample) according to the actual number of sampled individuals. Then all samples within a respective time period were pooled and the historic versus modern 2D-SFS was calculated. A custom script was used to calculate diagonal and off diagonal bins of the joint SFS (S4 Fig), which served as our ABC summary statistic. We used the R package 'abc' [84] to calculate model posterior probabilities and to evaluate the reliability of our model selection procedure. We considered the best fitting models for each species/transect to be those with the highest posterior probabilities (S7 Table) and we used a cross validation procedure to determine error rates associated with model choice (S8 Table). To aid model choice, we also considered the fit of the maximum likelihood (ML) estimate for each model to our observed data. We evaluated goodness-of-fit for the selected models by comparing the Euclidean distance between our observed and simulated 2D-SFS bins (S9 Table).
We considered SNPs with large allele frequency shifts between the modern and historic time periods that could not be attributed to demography as evidence for positive selection. We used the program OutFLANK [50] to detect FST outlier SNPs (FDR q-value < 0.01), empirically adjusting the degrees of freedom of χ2-distributed FST values to account for the influence of demography. We then compared the observed SNP FST values to null exome-wide and per-site FST distributions generated by performing 1,500 neutral simulations under the best fitting population history for YNP T. alpinus.
|
10.1371/journal.pntd.0001726 | Efficacy of Praziquantel against Schistosoma mekongi and Opisthorchis viverrini: A Randomized, Single-Blinded Dose-Comparison Trial | Schistosomiasis and opisthorchiasis are of public health importance in Southeast Asia. Praziquantel (PZQ) is the drug of choice for morbidity control but few dose comparisons have been made.
Ninety-three schoolchildren were enrolled in an area of Lao PDR where Schistosoma mekongi and Opisthorchis viverrini coexist for a PZQ dose-comparison trial. Prevalence and intensity of infections were determined by a rigorous diagnostic effort (3 stool specimens, each examined with triplicate Kato-Katz) before and 28–30 days after treatment. Ninety children with full baseline data were randomized to receive PZQ: the 40 mg/kg standard single dose (n = 45) or a 75 mg/kg total dose (50 mg/kg+25 mg/kg, 4 hours apart; n = 45). Adverse events were assessed at 3 and 24 hours posttreatment.
Baseline infection prevalence of S. mekongi and O. viverrini were 87.8% and 98.9%, respectively. S. mekongi cure rates were 75.0% (95% confidence interval (CI): 56.6–88.5%) and 80.8% (95% CI: 60.6–93.4%) for 40 mg/kg and 75 mg/kg PZQ, respectively (P = 0.60). O. viverrini cure rates were significantly different at 71.4% (95% CI: 53.4–84.4%) and 96.6% (95% CI: not defined), respectively (P = 0.009). Egg reduction rates (ERRs) against O. viverrini were very high for both doses (>99%), but slightly lower for S. mekongi at 40 mg/kg (96.4% vs. 98.1%) and not influenced by increasing diagnostic effort. O. viverrini cure rates would have been overestimated and no statistical difference between doses found if efficacy was based on a minimum sampling effort (single Kato-Katz before and after treatment). Adverse events were common (96%), mainly mild with no significant differences between the two treatment groups.
Cure rate from the 75 mg/kg PZQ dose was more efficacious than 40 mg/kg against O. viverrini but not against S. mekongi infections, while ERRs were similar for both doses.
Controlled-Trials.com ISRCTN57714676
| Parasitic worm infections are of public health importance in Southeast Asia. Particularly, the blood-dwelling Schistosoma mekongi worm, which is acquired by skin contact with the infectious cercariae in freshwater, can lead to liver enlargement. An infection with Opisthorchis viverrini is obtained by consumption of undercooked freshwater fish, and this infection increases the risk of developing cholangiocarcinoma. A single oral dose of 40 mg/kg praziquantel is recommended for mass treatment of schistosomiasis and opisthorchiasis, while at the individual level, a total dose of 75 mg/kg divided into three doses, is currently common practice to treat O. viverrini infection. Diagnosis is based on stool examination under a microscope for detection of worm eggs, but is limited by the low sensitivity of the widely used Kato-Katz technique. In this study, we showed that a 75 mg/kg total dose of praziquantel (50 mg/kg+25 mg/kg given 4 hours apart) cleared significantly more O. viverrini infections than a single 40 mg/kg dose, but no difference was observed for S. mekongi. Solicited adverse event profiles were mainly mild and similar in both groups. Repeated stool examination before and after treatment was essential for an accurate assessment of drug efficacy in terms of cure rate, but showed no effect on assessing egg reduction rates.
| Schistosomiasis, food-borne trematodiasis, and soil-transmitted helminthiasis are neglected tropical diseases that are of considerable public health relevance in Southeast Asia [1]. In Lao People's Democratic Republic (Lao PDR), approximately 80,000 individuals are at risk for schistosomiasis mekongi, 2 million individuals are at risk for food-borne trematodiasis (particularly opisthorchiasis), and 1 million school-aged children are at risk for soil-transmitted helminthiasis [1]. Praziquantel (PZQ) is the current drug of choice in the treatment of schistosomiasis and most of the food-borne trematode infections [1]. Deworming programs against schistosomiasis aim at morbidity control [2]. The World Health Organization (WHO) recommends a standard single dose of oral PZQ between 40 and 60 mg/kg for both schistosomiasis and food-borne trematodiasis [1], [2]. In Lao PDR, a single dose of 40 mg/kg PZQ is recommended for mass treatment of schistosomiasis and opisthorchiasis [3]. For individual treatment, the PZQ dose to treat Opisthorchis viverrini infection is a total dose of 75 mg/kg divided into three doses [4].
PZQ is known to be effective against all six Schistosoma species causing disease in humans. However there have been just two small published clinical trials on PZQ cure rates against Schistosoma mekongi [5], [6]. Both were non-randomized studies involving individuals relocated to non-endemic areas and given 60 mg/kg PZQ divided into two or three doses. To our knowledge, a controlled trial to treat S. mekongi using 40 mg/kg, the recommended dose for mass treatment in Lao PDR, and any comparison between different PZQ doses for superiority has so far not been undertaken.
Several clinical trials have assessed PZQ efficacy against O. viverrini at the following dosages: single dose of 25, 40, or 50 mg/kg, or repeated 25 mg/kg doses for a total dose of 50, 75, or 150 mg/kg [7]–[13]. However, none has been conducted in Lao PDR, which also has S. mekongi co-endemic areas, and 40 mg/kg has not been compared with 75 mg/kg.
Diagnosis of schistosomiasis, opisthorchiasis, and other intestinal or hepatobiliar helminth infections in epidemiological studies is commonly based on the detection of parasite eggs in stool specimens under a microscope. The Kato-Katz technique [14], [15] is the recommended field method [16] and permits estimation of infection intensity expressed in eggs per gram of feces (EPG). It is a relatively simple and rapid diagnostic method, but unfortunately, a single Kato-Katz thick smear has low sensitivity, particularly for light infections, and hence repeated stool examinations are necessary to improve the sensitivity of this technique [17]–[20]. This is especially important after treatment to avoid overestimation of cure rates. The low sensitivity of a single Kato-Katz thick smear results from the small amount of stool examined (usually 41.7 mg), variation in helminth egg excretion over time in the same individual, and from variation in egg density within a stool specimen depending on sampling location, as recognized for Schistosoma mansoni [19], [21], [22]. The relative contribution of day-to-day and intra-specimen variation in fecal egg counts has been investigated for S. mansoni [19], [21] where examination of repeated stool specimens, rather than examination of multiple Kato-Katz thick smears derived from a single stool specimen, was shown to be more appropriate to improve the sensitivity of detecting an infection [19], [22]. While it is documented for S. mansoni that diagnostic sensitivity depends on the sampling effort, other helminth species are less well investigated. Repeated or multiple stool specimen collection is difficult in practice, particularly in rural community field surveys [20], due to logistical requirements and cost implications.
The current study pursued two objectives. First, we assessed the efficacy of two oral PZQ regimens (i.e., 40 mg/kg single dose, and 75 mg/kg divided dose, given as 50 mg/kg then 25 mg/kg 4 hours apart) against S. mekongi and O. viverrini infections. Second, we determined the effect of multiple stool sampling on the diagnostic accuracy of the Kato-Katz technique before and after treatment, and assessed its impact on drug efficacy evaluation, considering both cure and egg reduction rates.
Ethical clearance was obtained from the National Ethics Committee, Ministry of Health (MoH) in Vientiane, Lao PDR (reference no. 027/NECHR) and by the Ethics Committee of Basel, Switzerland (EKBB; reference no. 255/06). The study protocol is registered with Current Controlled Trials on controlled-trials.com (identifier ISRCTN57714676). Written informed consent was obtained by the parents or guardians of all pupils before participation in the study. The children had the opportunity to withdraw from the study at any time.
Both doses of PZQ (i.e., single 40 mg/kg dose or total of 75 mg/kg dose) are accepted within Lao MoH published guidelines. The 40 mg/kg single dose is mainly used in mass drug administration programs, while 75 mg/kg (divided into three dosages) is used for the treatment of individuals. In our study the 75 mg/kg dose was divided into two doses (50 mg/kg plus 25 mg/kg given 4 hours apart) to simplify the regimen for a school setting where classes ended by the early afternoon. At the end of the follow-up period, all children were treated against soil-transmitted helminth infections with a single oral dose of 400 mg albendazole [3].
The primary objective of this study was to compare the efficacy of two different dose regimens of oral PZQ in school-aged children from southern Lao PDR in a S. mekongi and O. viverrini co-endemic area. The two regimens compared were (i) 40 mg/kg single dose and (ii) 75 mg/kg divided dose, given as 50 mg/kg then 25 mg/kg 4 hours apart. The secondary objectives were to determine the effect of multiple stool sampling to assess cure and egg reduction rates and to estimate the increased diagnostic sensitivity by multiple Kato-Katz thick smears from a single stool specimen compared with additional stool specimens obtained over several days before and after treatment. S. mekongi and O. viverrini were the species of primary interest, but hookworm was also included for the baseline analyses. Finally, the prevalence of the other intestinal helminth infections among our cohort of schoolchildren was also assessed.
The dose comparison study was a randomized trial with 1∶1 allocation. It was conducted in February and March 2007 in the primary and secondary schools on Don Long Island, Khong district, Champasack province, Lao PDR. The 308 children registered at the Don Long school were invited for the dose comparison trial. Most of the pupils (60%) lived in one of the four villages of Don Long Island, whereas the remaining children traveled from four villages on surrounding islands. In-depth stool examination was limited to 93 children aged 10–15 years (two classes). Based on the asymptotic normal method (formula 7) of Sahai and Khurshid [23], this sample size has a 70% power to demonstrate a superiority of 20% of the highest PZQ dosage (type I error: alpha = 5%; 1-tailed test) when considering a 20% dropout rate. Analyses of the present paper are restricted to this in-depth cohort. Acutely ill or febrile children were excluded from the study.
Don Long is a rural island in the Mekong River with about 1,500 inhabitants who practice subsistence farming and fishing. Previous studies on this island found the area to be co-endemic for S. mekongi and O. viverrini infections [24], [25]. Laboratory facilities were established in Khong district hospital in Muang Khong, a village on the east side of Don Khong, the main island of Khong district.
The children were assigned into two treatment arms following a 1∶1 allocation regardless of the baseline examination. Randomization was generated using a random number table in blocks of 10. Randomization and supervision of the trial were conducted by the study leaders (LL, TKM). Based on the child's weight, the dose was rounded to the nearest 150 mg by splitting the 600 mg PZQ tablets (Distocide®, Korea) in quarters using a pill cutter. Doses were prepared in advance by team members not involved in administrating the intervention. Each preparation was double verified for name, dose, and recorded weight for each child. Twelve hours before treatment, all doses were prepared and sealed in opaque envelopes that were labeled with the dose number, study unique identification number, the child's name, and weight. After the dose envelopes were prepared, the randomization and allocation list was sealed in an opaque envelope. Box 1 contained the envelopes with the first (and only) dose for children allocated in the 40 mg/kg arm and the first dose for those assigned to the 75 mg/kg arm, organized by school class and name. Box 2 contained the prepared envelopes for the second dose (25 mg/kg) only for those children allocated for the total dose of 75 mg/kg PZQ.
The drugs were administered by one of two paired teams of health care workers. The team confirmed that the child matched the identification on the drug envelope and then directly observed treatment. The drug administering teams were not involved prior to or after the study and not in any outcome assessments. As the different regimen was apparent (single vs. a divided dose 4 hours apart) neither the two health care teams nor the children were masked during treatment administration. The Lao physicians who assessed the children for adverse events following treatment were unaware of the dose allocation and were not involved with administering the intervention (KP, PAS). Laboratory technicians assessing infection status were blinded to the dose allocation.
The purpose and procedures of the study were explained to the school director, teachers, and to the village chief, who all agreed to participate. The study was explained during class to the children and written informed consent was received from their parents or guardians.
Clinical baseline measurements and baseline laboratory determination of infection status were performed prior to treatment for each participating child. Clinical measurements included a morbidity questionnaire and physical examination. For laboratory procedures, plastic bags with pre-labeled 30 ml plastic containers were distributed to the children at enrolment and pupils were asked to return the containers the following day with a thumb-sized portion of their morning stools. Containers were collected each morning at the school from 07:30 to 08:30 hours, recorded on a line listing, and children were given new empty plastic containers for the following day. This procedure was repeated until 3 morning stool specimens per child were received. Fresh stool specimens were transferred daily to the laboratory on Khong Island for examination. From each stool specimens, triplicate Kato-Katz thick smears using standard 41.7 mg templates were prepared on microscope slides in accordance with the kit instructions (Vestergaard Frandsen; Lausanne, Switzerland). The slides were quantitatively examined under a microscope within 1 hour following slide preparation. The number of eggs of O. viverrini, S. mekongi, hookworm, Trichuris trichiura, Ascaris lumbricoides, Taenia spp., Enterobius vermicularis, and other helminths were counted and recorded separately. For quality control, 10% of the slides were randomly selected and re-examined by a senior technician without prior knowledge of the results. When discrepancies were observed (e.g., egg counts differing by more than 10%), the technicians received closer supervision by a more experienced colleague. Since O. viverrini cannot be easily distinguished from minute intestinal flukes (MIF) microscopically by the Kato-Katz technique [26], infections reported here as O. viverrini infections are assumed to include some MIF co-infections.
Following baseline data collection, children were treated with 40 mg/kg or 75 mg/kg oral PZQ as described. Immediately following the dose, the children were given two soupspoons of sticky rice (∼40 g) to increase PZQ bioavailability and minimize potential adverse events [27]. Adverse events spontaneously reported within 3 hours after administration of the first dose were recorded. Additionally, a solicited questionnaire on adverse events was administered 24 hours following PZQ administration and graded for severity. All clinical and laboratory assessments were repeated 28–30 days after PZQ administration.
Data were entered in EpiData software version 3.1 (EpiData Association; Odense, Denmark) and double-checked against the original data sheets. Data analysis was performed using Intercooled STATA release 9.0 (StataCorp; College Station, TX, USA).
For each helminth species, an infection was defined as the presence of one or more eggs in at least one of the Kato-Katz thick smears examined. Cumulative prevalence of each helminth infection detected after examination of 9 Kato-Katz thick smears (3 stool specimens with triplicate Kato-Katz per specimen) was calculated. Tests for significant associations with gender were analyzed by negative binomial regression. Intensity of infection (expressed in EPG) was calculated by multiplying the observed number of eggs by a factor of 24. Geometric mean intensity of infection was calculated on EPG. Infections with O. viverrini were classified into three groups [28]: light (1–999 EPG), moderate (1,000–9,999 EPG), and heavy infections (≥10,000 EPG). S. mekongi infections were grouped into the following three categories [29]: light (1–99 EPG), moderate (100–399 EPG), and heavy infections (≥400 EPG). Negative binomial regression was applied to compare infection intensities of S. mekongi and O. viverrini at baseline among the two treatment groups.
Cure rates of S. mekongi and O. viverrini were calculated as the proportion of children with no egg excretion after treatment among those with eggs in their stool at baseline. Children found egg-negative prior to treatment but egg-positive after treatment were considered to be false negative and counted as infected at baseline. These infections were assumed to have been missed at baseline because the 28–30 days follow-up would not have provided adequate time for re-infection and patency between the two surveys. Cure rates obtained with the two tested doses were compared with Fisher's exact test. Egg reduction rates were determined by comparing the geometric mean egg output before and 28–30 days after treatment among children infected at baseline (1 - geometric mean egg output posttreatment/geometric mean egg output at baseline, multiplied by 100).
The effect of multiple sampling on the sensitivity of the Kato-Katz technique to detect S. mekongi and O. viverrini infections was assessed before and after drug administration. Hookworm infections were also included at baseline. Prevalences with 95% confidence interval (CI) were calculated for each sampling effort, the minimum effort being defined as the first Kato-Katz thick smear derived from the first stool specimen. The sampling effort increased with additional Kato-Katz thick smear examinations from the same stool specimen and with additional stool specimens. The McNemar test was used to compare prevalences assessed by different sampling efforts. The maximum sampling effort, 9 Kato-Katz thick smears, was taken as the diagnostic ‘gold’ standard to assess the sensitivity of increasing sampling efforts.
Adverse event frequencies depending on treatment doses were compared with the exact χ2 test. Additionally, infection intensities were expressed in EPG and for each child the arithmetic means were computed for each sampling effort. At the cohort level, geometric mean fecal egg counts were calculated for each sampling effort considering only the children with complete datasets at each time point separately. The analysis was restricted to the egg-positive children, based on the examination of 9 Kato-Katz thick smears (maximum sampling effort).
The 93 children (54 boys, 39 girls) included in the in-depth cohort all agreed to participate and written parental or guardian consent was received. Participants had a median age of 12 years (range: 10–15 years). Eighty-five children provided at least one stool specimen during the baseline survey and during the 28–30 day posttreatment follow-up. Among them, 64 children provided three stool specimens at both time points and had therefore complete datasets, with a compliance of 69% (64/93) (see Figure 1).
All schoolchildren were given treatment, according to their randomized treatment allocation. In the in-depth cohort, 46 children received 40 mg/kg PZQ and 47 received 75 mg/kg divided dose. The effect of multiple sampling on the sensitivity of the Kato-Katz technique was analyzed before and after treatment and was restricted to children with complete datasets at each time point separately, with a compliance of 97% (90/93) at baseline and 71% (66/93) at the 28–30 day posttreatment follow-up. There were no significant differences in the gender ratio, average age, or infection prevalence between the baseline and the posttreatment follow-up groups (all P>0.05).
Table 1 summarizes baseline infection prevalences and intensities of all helminth species diagnosed in the present study before PZQ administration. Results pertained to those children who had complete data records (9 Kato-Katz thick smears) prior to treatment (n = 90) and before and after treatment combined (n = 64). S. mekongi, O. viverrini, and hookworm were the most common parasitic infections at baseline, with prevalences above 85% for each helminth species, as assessed with the maximum sampling effort. Other intestinal parasitic infections, in descending order of prevalence, were T. trichiura, A. lumbricoides, E. vermicularis, and Taenia spp. One infection with Hymenolepis diminuta was detected. Infection prevalences for any of the aforementioned helminths did not differ between boys and girls.
Cure and egg reduction rates were compared between two cohorts (Figures 1a and 1b). First, children who complied with the maximum diagnostic effort (9 Kato-Katz thick smears before and after treatment, n = 64) and, second, children with a minimum diagnostic effort (1 Kato-Katz thick smear at each time point, n = 85). Results are summarized in Tables 2 and 3. For both cohorts, there was no significant differences in the infection intensities of S. mekongi and O. viverrini at baseline between the two treatment groups (all P>0.05).
S. mekongi cure rates among children who had provided three stool specimens at baseline and follow-up were 80.8% (21/26; 95% CI: 60.6–93.4%) after 75 mg/kg PZQ and 75.0% (24/32; 95% CI: 56.6–88.5%) after 40 mg/kg PZQ, which was not significantly different (P = 0.754). With the minimum diagnostic effort, observed cure rates were considerably higher, 94.7% (18/19; 95% CI: not defined) and 85.7% (18/21; 95% CI: not defined), respectively. S. mekongi egg reduction rates in both cohorts were >93%. Slightly higher egg reduction rates were observed at the minimum sampling effort (97.9% and 99.6% in the 40 mg/kg and 75 mg/kg treatment group, respectively), compared to the highest sampling effort (96.4% and 98.1%, respectively).
Based on the maximum sampling effort, O. viverrini cure rates were 96.6% (28/29; 95% CI: not defined) after 75 mg/kg PZQ and 71.4% (25/35; 95% CI: 53.4–84.4%) after 40 mg/kg PZQ, showing a statistically significant difference (P = 0.009). Considering the minimum diagnostic effort, observed cure rates were 100% (35/35; 95% CI: not defined) and 94.3% (33/35; 95% CI: not defined), respectively, with no statistically significant difference (P = 0.493). Egg reduction rates, regardless of treatment group and diagnostic efforts, were above 99%.
Solicited 24-hour adverse event profiles in the two treatment groups are summarized in Table 4. Fourteen children were not available to be interviewed (n = 6, 40 mg/kg dose; n = 8, 75 mg/kg dose), corresponding to 15.1% lost to follow-up, but no serious adverse events were reported by the community when we returned days 28–30 for post-treatment follow-up. Most children reported one or more adverse events (76/79, 96%). More cases were reported for most types of adverse events in the 75 mg/kg treatment arm, but did not reach statistical significance in this small sample when comparing the total number of events or those graded as severe. There were a total of 7 cases recorded as hypotension (below 100 mm Hg systolic blood pressure) in the 75 mg/kg treatment group compared with a single case in the 40 mg/kg group, which was statistically higher (P<0.02) but no case was graded severe (e.g., no syncope). Children with hypotension associated with dizziness and vomiting were given rest and monitored; all cases were self-limiting. No serious adverse events required hospitalization.
Figure 2 shows the cumulative prevalence of infected children over repeated stool specimens according to the number of Kato-Katz thick smears examined per stool specimen for S. mekongi and O. viverrini infections both at baseline and at the 28–30 day posttreatment follow-up survey. Baseline results for hookworm infections were also recorded although not the primary outcome of the study (nor were hypotheses made on the efficacy of PZQ against this helminth species). The sensitivity of three different sampling efforts (considering the maximum diagnostic effort of 9 Kato-Katz thick smears as the diagnostic ‘gold’ standard) is presented in Table 5.
Figure 3 illustrates the results of increased sampling effort on the geometric mean fecal egg counts before and after PZQ treatment of all children infected with S. mekongi and O. viverrini. This was also assessed for hookworm at baseline.
At baseline, the mean fecal egg counts gradually increased with increasing sampling efforts. Thus, egg count estimates for S. mekongi, O. viverrini, and hookworm increased 4, 2.4 and 1.7-fold, reaching values of 25 EPG, 342 EPG and 321 EPG, respectively, when assessed with the maximum sampling effort. S. mekongi and O. viverrini mean fecal egg count estimates were considered low-intensity infections.
After PZQ treatment, the benefit of the maximum sampling effort for EPG was 9-fold and 8-fold for S. mekongi and O. viverrini, respectively. When comparing the pretreatment baseline with the 28–30 day posttreatment follow-up, the mean fecal egg count for O. viverrini sharply decreased from 342 to 9.1 EPG. The decrease was less marked for S. mekongi, from 25 to 8 EPG.
A 3×1 sampling effort yielded substantially higher estimates than a 1×3 sampling effort for S. mekongi and hookworm egg counts. By contrast, the same efforts showed only a minimal increase for O. viverrini egg counts.
PZQ is the drug of choice against most trematode infections, including schistosomiasis and opisthorchiasis. To our knowledge, PZQ dose comparison studies have not been described for S. mekongi. Dose comparison studies for O. viverrini have been conducted, but most studies relied on an insensitive diagnostic approach, i.e., single stool specimen examination before and after drug administration. The accuracy of diagnosis, which is particularly important for estimating cure rates, can be improved by examining multiple Kato-Katz thick smears derived from a single or multiple stool specimens [30].
In this study, S. mekongi cure rate after administration of 75 mg/kg PZQ (80.8%) was not significantly higher than the cure rate obtained after a single dose of 40 mg/kg (75.0%) when assessed with the maximum sampling effort of 9 Kato-Katz thick smears. The cure rate from either regimen was largely overestimated if diagnosis was based on a single Kato-Katz thick smear. Studies based on fewer Kato-Katz thick smears are more likely to overestimate cure rate and be less diagnostically sensitive to detect any differences in dose comparisons. Two small studies carried out in the 1980s on S. mekongi infection reported high cure rates with 60 mg/kg PZQ (90.9% and 97.5%, respectively) [5], [6] when analyzing 2–3 stool specimens but using different stool diagnostic techniques (Kato-Katz+modified Ritchie's and Stoll's, respectively). Similarly in a recent multi-country randomized trial comparing single 40 mg/kg and 60 mg/kg PZQ in children aged 10–19 years, with infections diagnosed by two stool specimens (duplicate Kato-Katz thick smears per specimen), the 21-day posttreatment follow-up was reported as 92.8% with 60 mg/kg, which was not a significant improvement against S. mansoni, S. haematobium, or S. japonicum infections compared to the standard 40 mg/kg [31]. Consistent with results obtained from this recent trial, our study did not document a significantly improved cure rate (days 28–30 posttreatment) with an even higher total dose (75 mg/kg dose) for S. mekongi, even with higher diagnostic sensitivity from greater stool sampling efforts. However our additional sampling effort did observe a cure rate for 40 mg/kg about 15% lower than rates reported in the multicenter trial.
O. viverrini cure rate after administration of 75 mg/kg PZQ (96.6%) was significantly higher than the cure rate obtained after a single dose of 40 mg/kg (71.4%) when assessed with the maximum sampling effort. However, if the cure rate had been based on results of single Kato-Katz thick smear before and after drug administration, as often the case in community-based surveys, no significant difference would have been found. Cure rate was particularly overestimated when based on a single Kato-Katz thick smear in this study for a 40 mg/kg dose (94.3%), similar to high, and most likely overestimated cure rates (91–100%) reported from previous studies using the same dosage and only a single stool examination [9], [10], [12]. Cure rates which were reported as 100% after administration of 75 mg/kg PZQ (divided into three doses) were also likely overestimated in previous studies [7], [13].
Our study therefore provides supportive evidence that a 75 mg/kg total dose of PZQ is highly efficacious against O. viverrini and S. mekongi infections in school-aged children from Lao PDR. The total dose was divided into two doses instead of three and had a 24-hour profile of common adverse events similar to a single 40 mg/kg dose. Two doses, instead of three, are operationally and logistically more feasible, but clearly single-dose regimens are the preferred option for large-scale preventive chemotherapy programs. The small size of our study, however, limits detecting a difference in the nature or frequency of adverse events between the two regimens.
The non-significant difference between the two doses to cure S. mekongi infections should be interpreted with caution. Again, this may result from the study's small sample size and it would therefore be valuable to investigate a larger sample. In addition, most of the children included in our study only had low intensity infections while cure rate achieved by PZQ has been shown to be influenced by the infection burden [32]. Some authors have argued that egg reduction rate is a more appropriate indicator than cure rate for drug efficacy evaluation [33], [34]. We assessed both cure and egg reduction rates. Importantly, we found very high egg reduction rates (>99%) against O. viverrini for both treatment regimens regardless of the sampling effort. For S. mekongi, considering 9 Kato-Katz thick smears as the diagnostic ‘gold’ standard, a somewhat lower egg reduction rate was observed with a single 40 mg/kg dose of PZQ compared to the higher split dose (96.4% vs. 98.1%). At the lower sampling effort, higher egg reduction rates were observed (97.9% and 99.6%, respectively). These data suggest that the worm burden sharply declined from either dose regimen, which was found using either minimal or maximal diagnostic effort. This may be explained by the low posttreatment infection intensity of the non-cured children given either dose. The geometric mean egg counts in the two PZQ regimens were very similar. The public health goal of preventive chemotherapy is to reduce morbidity, which is indirectly assessed using egg reduction rates. Our results suggest that PZQ, given at a single oral dose of 40 mg/kg, is suitable to achieve this goal, particularly against O. viverrini.
At baseline, the relative increase of sensitivity by multiple sampling was relatively low, especially for O. viverrini and hookworm infections. By contrast, multiple sampling was important after treatment, when infection prevalence and intensity were much lower. As a result, the sensitivity of the first Kato-Katz thick smear was much lower after treatment than at baseline, with a 4-fold lower and 3-fold lower sensitivity to detect O. viverrini and S. mekongi infections, respectively.
A single Kato-Katz thick smear is known to have a low sensitivity for the diagnosis of O. viverrini, especially for low intensity infections [20]. For S. mekongi, the low sensitivity of a single Kato-Katz thick smear to detect this fluke observed in the present study agrees with previous findings obtained from investigations focusing on S. mansoni and S. japonicum [18], [19], [22], [35], [36]. Studies on the sensitivity of the Kato-Katz technique for diagnosis of S. mekongi are generally lacking.
For O. viverrini and hookworm diagnosis, the sensitivity of a single Kato-Katz thick smear to detect infection at baseline was fairly high. For hookworm, this was in contrast to previous studies from Côte d'Ivoire [37], [38], Ethiopia [18], and Tanzania [39], where the sensitivity of a single Kato-Katz thick smear varied from 18% to 53%. However, after drug administration, when the overall O. viverrini infection intensity of our cohort of children became low (<10 EPG), this study indicates the need for multiple Kato-Katz thick smear examinations, ideally performed on stool specimens collected over consecutive days for a more accurate estimation of the cure rate.
Helminth eggs are non-randomly distributed within a stool specimen because the intestinal content is not uniformly mixed [40] and may affect the sensitivity of detecting an infection and fecal egg count estimates from a single Kato-Katz thick smear. Important day-to-day variation in egg output has been thoroughly documented for S. mansoni and S. japonicum [19], [21], [22], [35]. By contrast, O. viverrini egg output was found to be relatively consistent over a period of several days in hospitalized patients [41]. Of note, Schistosoma egg shedding dynamics are additionally affected by retention of eggs in intestinal and liver tissues and the lower fecundity of female worms.
We have compared the relative importance of intra-specimen and day-to-day variation of fecal egg counts before and after PZQ administration and determined its effect on evaluating anthelmintic drug efficacy. Previous research has shown that the examination of fewer specimens from different days proved to be superior than examining multiple Kato-Katz thick smears from a single stool specimen for more accurate estimates of the ‘true’ infection status for S. mansoni [19], [22]. In the present study for S. mekongi and hookworm infections, examination of one Kato-Katz thick smear per stool specimen, with specimens collected over a 3-day period (3×1 sampling scheme), resulted in higher prevalence and mean infection intensity than three Kato-Katz thick smears taken from the first stool specimen (1×3). For O. viverrini, however, the 3×1 and 1×3 sampling scheme revealed the same prevalence estimates. Since repeating the collection of a stool specimen over consecutive days is more costly, logistically more cumbersome, and negatively impacts on study compliance, examination of multiple Kato-Katz thick smears from a single stool specimen should be considered as a suitable approach for community surveys of helminth infections. Similar observations have been made before for the diagnosis of Clonorchis sinensis [42].
S. mekongi is known to be endemic in certain areas of the Mekong River basin [25], [43]–[45], while O. viverrini and hookworm species are widely distributed across Lao PDR [46]–[48]. Point prevalences as high as those observed in the present study for S. mekongi (87.8%), O. viverrini (98.9%), and hookworm (96.7%), based on a rigorous diagnostic effort, have rarely been described in the literature. Yet, our findings corroborate with a recent risk profiling study in more than 50 villages of Champasack province, where O. viverrini prevalences were above 80% in most villages, with particularly high prevalences observed in villages in close proximity to the Mekong River [24]. WHO surveyed selected villages on Khong Island (an island also situated along the Mekong River, only 10 km from our study site) prior to starting schistosomiasis control campaigns in the late 1980s, and found a similarly high S. mekongi prevalence (87.8%) as reported here [49].
Studies carried out in rural provinces of southern Lao PDR (Champasack and Saravane) reported prevalences of O. viverrini and hookworm ranging from 18.8% to 70.8% and from 12.5% to 46.1%, respectively [46], [47], [50], [51]. Infection prevalence is known to vary locally [46], which may partially explain the difference between prior estimates and those found in this study. However, previous prevalence estimates were based on a single Kato-Katz thick smear, while 9 Kato-Katz thick smears were examined in the present study. O. viverrini infection prevalence probably includes MIF infections since co-infections are common, and polymerase chain reaction (PCR) techniques on stool specimens taken from the same study area in southern Lao PDR [52] have demonstrated that MIF eggs cannot easily be distinguished microscopically from O. viverrini by the Kato-Katz technique [26].
In conclusion, the present study found that the added benefit of multiple Kato-Katz thick smear examination and repeated stool sampling depends on the helminth species and baseline infection intensity. Thus, in the present setting in Lao PDR, where O. viverrini, S. mekongi, and hookworm are all highly endemic, estimating the baseline prevalence and intensity of infection for these species with a single Kato-Katz examination may be acceptable. By contrast, estimating the prevalence of infection after treatment by the Kato-Katz technique requires multiple thick smears, ideally taken from multiple stool specimens because the positive predictive value is lower (both lower prevalence and lower geometric mean fecal egg count after treatment). A single Kato-Katz thick smear after treatment will considerably overestimate cure rate, but only minimally influences egg reduction rates. A rigorous diagnosis approach is necessary for estimating ‘true’ cure rates, as it has been previously demonstrated in studies on S. mansoni [30], [53]. For anthelmintic drug evaluations with emphasis on egg reduction rates, a single Kato-Katz thick smear before and after treatment might suffice. In our view, multiple stool examination should nonetheless be considered in a subsample of the population surveyed in order to improve the monitoring of large-scale control programs, provide reasonable estimates on infection prevalence and intensity, and detect subtle changes in drug efficacies that might indicate the emergence of drug resistance development.
|
10.1371/journal.pgen.1002617 | The Caenorhabditis elegans HEN1 Ortholog, HENN-1, Methylates and Stabilizes Select Subclasses of Germline Small RNAs | Small RNAs regulate diverse biological processes by directing effector proteins called Argonautes to silence complementary mRNAs. Maturation of some classes of small RNAs involves terminal 2′-O-methylation to prevent degradation. This modification is catalyzed by members of the conserved HEN1 RNA methyltransferase family. In animals, Piwi-interacting RNAs (piRNAs) and some endogenous and exogenous small interfering RNAs (siRNAs) are methylated, whereas microRNAs are not. However, the mechanisms that determine animal HEN1 substrate specificity have yet to be fully resolved. In Caenorhabditis elegans, a HEN1 ortholog has not been studied, but there is evidence for methylation of piRNAs and some endogenous siRNAs. Here, we report that the worm HEN1 ortholog, HENN-1 (HEN of Nematode), is required for methylation of C. elegans small RNAs. Our results indicate that piRNAs are universally methylated by HENN-1. In contrast, 26G RNAs, a class of primary endogenous siRNAs, are methylated in female germline and embryo, but not in male germline. Intriguingly, the methylation pattern of 26G RNAs correlates with the expression of distinct male and female germline Argonautes. Moreover, loss of the female germline Argonaute results in loss of 26G RNA methylation altogether. These findings support a model wherein methylation status of a metazoan small RNA is dictated by the Argonaute to which it binds. Loss of henn-1 results in phenotypes that reflect destabilization of substrate small RNAs: dysregulation of target mRNAs, impaired fertility, and enhanced somatic RNAi. Additionally, the henn-1 mutant shows a weakened response to RNAi knockdown of germline genes, suggesting that HENN-1 may also function in canonical RNAi. Together, our results indicate a broad role for HENN-1 in both endogenous and exogenous gene silencing pathways and provide further insight into the mechanisms of HEN1 substrate discrimination and the diversity within the Argonaute family.
| Small RNAs serve as sentinels of the genome, policing activity of selfish genetic elements, modulating chromatin dynamics, and fine-tuning gene expression. Nowhere is this more important than in the germline, where endogenous small interfering RNAs (endo-siRNAs) and Piwi-interacting RNAs (piRNAs) promote formation of functional gametes and ensure viable, fertile progeny. Small RNAs act primarily by associating with effector proteins called Argonautes to direct repression of complementary mRNAs. HEN1 methyltransferases, which methylate small RNAs, play a critical role in accumulation of these silencing signals. In this study, we report that the 26G RNAs, a class of C. elegans endo-siRNAs, are differentially methylated in male and female germlines. 26G RNAs derived from the two germlines are virtually indistinguishable, except that they associate with evolutionarily divergent Argonautes. Our data support a model wherein the methylation status and, consequently, stability of a small RNA are determined by the associated Argonaute. Therefore, selective expression of Argonautes that permit or prohibit methylation may represent a new mechanism for regulating small RNA turnover. As we observe this phenomenon in the germline, it may be particularly pertinent for directing inheritance of small RNAs, which can carry information not encoded in progeny DNA that is essential for continued transgenerational genome surveillance.
| Argonautes are an evolutionarily conserved family of proteins implicated in diverse cellular processes. They function as effector proteins in the RNA-induced silencing complex (RISC), a gene regulatory complex that binds small, non-coding RNAs to target its silencing effects. Small RNAs are broadly segregated into groups that differ in their mechanisms of biogenesis and silencing, as well as in the subsets of Argonaute effectors that bind them. The microRNAs (miRNAs) are highly conserved small RNAs processed from endogenous hairpin precursors that regulate networks of mRNAs primarily through post-transcriptional repression [1], [2]. The piRNAs, so named for the Piwi Argonautes that bind them, function predominantly in maintenance of germline integrity, often through repression of repetitive transposable elements. The small interfering RNAs comprise a more heterogeneous group that includes small RNAs derived from cleavage of exogenous double-stranded RNA (exo-siRNAs) or generated endogenously (endo-siRNAs).
Chemical modification has emerged as an important theme in regulation of small RNA function (for a review, see Kim et al., 2010 [3]). Internal editing has been found to occur in select miRNA precursors through the action of ADAR (adenosine deaminase acting on RNA) enzymes, with consequences for miRNA processing efficiency, stability, and targeting [4]–[8]. Some siRNAs generated in fly and mouse also show evidence of editing by ADARs [9], [10], but the significance of such internal editing among siRNAs is not yet known. In contrast, terminal editing through 2′-O-methylation, addition of untemplated nucleotides, or exonucleolytic trimming plays a more general role in small RNA metabolism. These terminal modifications are not unrelated. Evidence in plants and animals suggests that methylation of the 3′ terminal nucleotide protects small RNAs from polyuridylation and polyadenylation, signals that direct exonucleolytic degradation [11]–[16]. Thus, terminal methylation plays an important role in regulating small RNA turnover. Formation of the 2′-O-methyl group is catalyzed by HEN1, a methyltransferase discovered in Arabidopsis thaliana that is conserved across metazoa, fungi, viridiplantae, and bacteria [17]. Although plant and animal HEN1 orthologs exhibit 40–50% amino acid similarity in the conserved methyltransferase domain [18], the proteins differ in their substrate specificity. Plant HEN1 acts on small RNAs in duplex and methylates both siRNAs and miRNAs [19]–[21]. In contrast, animal HEN1 orthologs modify only single-stranded small RNAs [22]–[24], enabling methylation of small RNAs such as piRNAs, which are not derived from double-stranded RNA intermediates [25]–[29]. While animal piRNAs appear to be universally methylated [24], [26], [27], [30]–[32], animal miRNAs are generally not methylated [19], [26], [31], and the mechanisms by which animal HEN1 orthologs discriminate between substrates are not entirely clear. HEN1 orthologs that catalyze terminal methylation of small RNAs have been characterized in mouse, fish, and fly, among other organisms [15], [22]–[24], [33], yet the orthologous methyltransferase in worm [18] has yet to be investigated. With its expanded Argonaute family and diverse small RNA classes, Caenorhabditis elegans provides an advantage for studying HEN1 substrate specificity.
Since the discovery of the founding members of the microRNA family in C. elegans [1], [2], [34], many additional classes of small RNAs have been characterized. A large-scale small RNA sequencing effort revealed a class of terminally methylated 21-nucleotide RNAs with 5′ uridines [27]. These 21U RNAs were subsequently determined to represent the piRNAs of C. elegans based on their germline-specific expression, association with worm Piwi Argonautes PRG-1 and PRG-2, and function in transposon silencing and maintenance of temperature-dependent fertility [35]–[38]. Also found through small RNA cloning and deep sequencing were populations of 26- and 22-nucleotide RNAs with a 5′ preference for guanosine (the 26G RNAs and 22G RNAs, respectively) that constitute the endo-siRNAs of C. elegans [27], [39]. The 26G RNAs are primary endo-siRNAs generated in the germline to regulate spermatogenic and zygotic gene expression. They are divided into two non-overlapping subclasses named for the Argonautes that bind them: the ERGO-1 class 26G RNAs, which are generated in the maternal germline and distributed into the embryo, and the ALG-3/ALG-4 class 26G RNAs, which are specific to the male germline and required for sperm function [40]–[42]. The 22G RNAs are composed of many small RNA classes, all of which are bound by worm-specific Argonautes (Wagos). A large population of 22G RNAs are secondary endo-siRNAs whose production by RNA-dependent RNA polymerases is triggered by the activity of 21U RNAs and 26G RNAs [36], [41]–[43]; however, many other 22G RNAs are independent of these primary small RNAs [44], [45]. Secondary siRNAs serve to amplify the signal of primary small RNAs to effect robust silencing. Production of 22G secondary siRNAs is also triggered by exogenously introduced dsRNAs [43], [45]–[47], suggesting convergence of endogenous and exogenous RNAi pathways at the level of the secondary siRNA response.
Among C. elegans small RNAs, only 21U RNAs and 26G RNAs are known to be methylated [27], [42]; 22G RNAs triggered by either primary endo- or exo-siRNAs appear to be unmethylated [45], [46]. Although the significance of worm small RNA methylation is unknown, loss of terminal methylation has been shown to decrease stability of piRNAs in many animal models [15], [22], [24] and both endo- and exo-siRNAs in fly [22], [48]. Methylation may therefore represent an essential step in stabilization of some classes of worm small RNAs.
In this study, we characterize the C. elegans hen1 ortholog, which has been named henn-1 (hen of nematode), as the name hen-1 has already been assigned to an unrelated C. elegans gene. We demonstrate that HENN-1 methylates small RNAs bound by Piwi clade Argonautes: the 21U RNAs and the ERGO-1 class 26G RNAs. However, we show that 26G RNAs bound by Ago clade Argonautes ALG-3 and ALG-4 are not methylated and are therefore henn-1-independent. Differential methylation of 26G RNAs provides evidence for an existing model [13], [22], [23], [49], [50] wherein evolutionarily divergent Argonautes either direct or prohibit HEN1-mediated methylation of associated small RNAs. In further support of this Argonaute-dictated methylation model, we find that small RNAs are likely methylated after associating with an Argonaute: the Argonaute ERGO-1 is required for 26G RNA methylation, but methylation is not required for ERGO-1 to bind a 26G RNA.
In the henn-1 mutant, levels of both 21U RNAs and ERGO-1 class 26G RNAs drop precipitously after their deposition into embryo, suggesting that HENN-1-mediated methylation is essential for perdurance of the maternal small RNA load during filial development. Accordingly, the henn-1 mutant shows enhanced somatic sensitivity to exogenous RNAi, a phenotype associated with loss of ERGO-1 class 26G RNAs. Surprisingly, however, the henn-1 mutant germline exhibits an attenuated response to RNAi, suggesting that HENN-1 may also function in the exogenous RNAi pathway. Altogether, our study supports a role for HENN-1 in diverse small RNA pathways in C. elegans and offers further insight into the mechanisms governing substrate discrimination for animal HEN1 orthologs.
To examine small RNA methylation in C. elegans, we began by characterizing C02F5.6, the gene previously predicted to encode the HEN1 ortholog in worm [18]. This gene, subsequently named henn-1, encodes a protein that exhibits significant amino acid similarity across the conserved HEN1 methyltransferase domain relative to established members of the HEN1 family (Figure S1). Although two henn-1 gene models with differing 3′ ends have been proposed, 3′RACE and protein studies using a rabbit polyclonal antibody generated against a common N-terminal HENN-1 epitope detected only the longer isoform (Figure S2A, S2B).
To facilitate our studies of the function of HENN-1, we isolated and characterized the henn-1(tm4477) allele. This allele carries a deletion that encompasses henn-1 exon four, which encodes 65% of the conserved methyltransferase domain as annotated by Kamminga et al. [15]. Sequencing of the henn-1(tm4477) mRNA indicates that loss of exon four activates a cryptic splice donor site in the third intron, resulting in an extended third exon that encodes a premature termination codon (Figure S2B). The henn-1(tm4477) mRNA is readily detected by RT-PCR but does not produce a detectable protein product (Figure S2A) or exhibit methyltransferase activity (see below), suggesting that henn-1(tm4477) (hereafter, henn-1) represents a functional null allele.
Like piRNAs in fly [22], [23], [32], mouse [30], [31], and zebrafish [26], the C. elegans 21U RNAs are terminally methylated [27], but the factor catalyzing this modification has not yet been identified. To determine if 21U RNA methylation depends on henn-1, we assessed methylation status using the β-elimination assay [51]. A small RNA molecule whose terminal nucleotide has been 2′-O-methylated is resistant to this treatment, whereas the cis-diols of an unmodified 3′ terminal nucleotide are oxidized by sodium periodate, rendering the nucleotide susceptible to β-elimination under basic conditions. The resulting size difference can be resolved on a polyacrylamide gel to determine methylation status. All 21U RNAs examined were found to be terminally methylated in a henn-1-dependent manner (Figure 1A, Figure S3A), whereas a control miRNA was not methylated in either wild-type or henn-1 mutant animals (Figure 1B). Although 21U RNAs are still detectable in the henn-1 mutant, the abundance of the full-length species is visibly decreased for some 21U RNAs; this correlates with the appearance of putative degradation products of unmethylated, unprotected 21U RNAs. To demonstrate that loss of 21U RNA methylation in the henn-1 mutant is specifically due to the absence of henn-1, we used the Mos1-mediated single copy insertion technique [52] to introduce a henn-1::gfp transgene driven by the promoter of the polycistronic mRNA that encodes henn-1 (xkSi1) or by the germline-specific pie-1 promoter (xkSi2) into the henn-1 mutant (Figure S2C). Both endogenous and germline-specific expression of henn-1::gfp restore 21U RNA methylation in the henn-1 mutant (Figure 1A).
To investigate the relationship between terminal methylation and piRNA accumulation, we used Taqman RT-qPCR to assess 21U RNA levels in wild-type and henn-1 mutant animals across development at 25°C. Importantly, the Taqman stem-loop RT primer is capable of distinguishing between full-length and terminally degraded small RNAs [53]. For example, the let-7e miRNA differs from let-7a only in the absence of the final nucleotide and U>G substitution at the ninth nucleotide, a position likely not represented in the stem-loop Taqman primer. Absence of this final nucleotide decreases detection of let-7e by the let-7a Taqman assay by more than a thousandfold [53]. henn-1 mutant embryo and early larva show dramatically reduced detection of female germline-enriched piRNA 21UR-1848 (Figure 2A), consistent with decreased embryonic detection for some 21U RNAs observed by northern blot (Figure 1A, Figure S3A). 21U RNA levels recover to wild-type in late larval stages, coincident with the onset of germline proliferation and de novo 21U RNA biosynthesis; however, in gravid animals at 56 hours, 21UR-1848 levels in the henn-1 mutant have declined to less than 50% of those observed in wild-type (P = 0.0005; two-tailed t-test). Eight additional 21U RNAs examined show a similar pattern (Figure S4). These data suggest that henn-1 is dispensable for piRNA biogenesis but essential for robust inheritance of piRNAs. Parallel analysis of miR-1 and several additional miRNAs across development shows that effects of loss of henn-1 are specific to its substrates and not due to generalized small RNA dysregulation in the henn-1 mutant (Figure 2B, Figure S5).
We next sought to determine the extent to which decreased abundance of piRNAs in the henn-1 mutant compromises activity of the piRNA pathway. Unlike in fly, where many selfish genetic elements are desilenced in the absence of piRNAs [32], C. elegans at present has only a single established molecular readout for piRNA pathway function: increased expression of transposase mRNA from Tc3, a Tc1/mariner family transposon [35], [36]. Two 21U RNAs have been found to map to Tc3, but both map in the sense direction and thus are unlikely to act directly in Tc3 repression via canonical RNAi mechanisms [35], [36]. Rather, 21U RNAs likely mediate their repressive effects through triggering production of secondary siRNAs, 22G RNAs, that engage worm-specific Argonautes (Wagos) to effect Tc3 gene silencing [36], [45]. We therefore identified a 22G RNA that shows complete antisense complementarity to Tc3 and can be classified as a Wago-dependent, 21U RNA-dependent secondary siRNA based on its total depletion both in the MAGO12 mutant, which lacks all Wagos, and in the prg-1(n4357); prg-2(n4358) double mutant, which lacks piRNAs [36], [45]. Levels of this 22G RNA in the henn-1 mutant are reduced by 44% in embryo but not significantly altered in hatched L1 larva (Figure S6A). This suggests that the low embryonic and early larval levels of 21U RNAs in the henn-1 mutant are still sufficient to trigger production of secondary siRNAs, although to a lesser degree than in wild-type.
Consistent with the modest effect of loss of henn-1 on accumulation of piRNA-triggered secondary siRNAs, henn-1 mutant animals exhibit only a small increase (35% in starved L1 larva, 25% in L1 larva fed for 4 hours at 25°C) in Tc3 transposase mRNA levels relative to wild-type (Figure 2C). This is not unexpected due to the poor coincidence of the time intervals corresponding to piRNA dysregulation in the henn-1 mutant and Tc3 sensitivity to 21U RNAs; the henn-1 mutant shows the greatest disparity in piRNA levels in early larval development, whereas Tc3 levels are most sensitive to piRNAs in germline and embryo (Figure 2A, 2C). These findings suggest that HENN-1 is not strictly required for piRNA target repression, but contributes to robust silencing of Tc3.
In addition to Tc3 dysregulation, loss of prg-1 also results in a temperature-sensitive sterility phenotype [38], [43]. To determine if the henn-1 mutant also exhibits a fertility defect, we assessed fertility at 20°C and 25°C. At 20°C, brood size of the henn-1 mutant does not differ significantly from that of wild-type. In contrast, henn-1 mutant animals maintained at 25°C exhibit a 25% decrease in brood size relative to wild-type (P = 0.0059; two-tailed t-test) that can be rescued by germline expression of henn-1::gfp from the xkSi2 transgene (Figure S7). The impaired fertility of the henn-1 mutant is consistent with abnormal fertility phenotypes associated with loss of HEN1 methyltransferase activity in other animals. Loss of HEN1 in Tetrahymena thermophila depletes Piwi-interacting RNAs called scan RNAs, impairing DNA elimination and, consequently, the viability of progeny [24]. The zebrafish hen1 mutant fails to maintain a female germline, resulting in an exclusively male population [15]. Nevertheless, we cannot conclude that the temperature-sensitive fertility defect of the henn-1 mutant is due exclusively to compromise of the 21U RNA pathway.
26G RNAs were reported to be methylated in the first C. elegans small RNA deep sequencing study [27]. Subsequent studies concluded that the species assessed was an ERGO-1 class 26G RNA [40]. Consistent with these data, we found that ERGO-1 class 26G RNAs, found in female germline and embryo, are methylated. As was the case for piRNAs, this methylation occurs in a henn-1-dependent manner (Figure 3A, Figure S3B). Surprisingly, however, ALG-3/ALG-4 class 26G RNAs, specific to the male germline, showed no evidence of methylation even in wild-type animals (Figure 3B, Figure S3C). One potential explanation for this observation would be that female germline small RNAs are universally methylated, whereas male germline small RNAs are not. To explore this possibility, we assessed 21U RNAs in male and female germlines. Both were methylated (Figure 3C), indicating that differential 26G RNA methylation cannot be explained simply by a lack of methyltransferase functionality in the male germline.
Because the two classes of 26G RNAs bind unique Argonautes in male and female germlines, we hypothesized that the Argonaute ERGO-1 might direct methylation of 26G RNAs. To address this question, we sought to assess methylation of an ERGO-1 class 26G RNA in the absence of ERGO-1. As 26G RNAs are dramatically depleted in the absence of their respective Argonautes [40], we queried published wild-type and ergo-1(tm1860) gravid adult deep sequencing libraries [42] to identify an ERGO-1 class 26G RNA that still accumulates to levels sufficient for visualization by northern blotting in the ergo-1(tm1860) mutant. 26G-O1, an extremely abundant ERGO-1 class 26G RNA, is present at roughly 0.5% wild-type levels in the ergo-1(tm1860) mutant, but still abundant enough to detect by northern blotting. Consistent with our hypothesis that ERGO-1 is required for 26G RNA methylation, we found that 26G-O1 is unmethylated in the ergo-1(tm1860) mutant embryo (Figure 4A). We next asked the converse question: Is 26G RNA methylation required for association with ERGO-1? We immunopurified ERGO-1 complexes from wild-type and henn-1 mutant embryo lysates (Figure 4B) and extracted RNA. In both wild-type and henn-1 mutant samples, ERGO-1 class 26G RNAs are readily detected (Figure 4C), indicating that ERGO-1 effectively binds both methylated and unmethylated 26G RNAs. Taken together, these data suggest that 26G RNAs bind ERGO-1 and are subsequently methylated by HENN-1.
To test whether HENN-1-mediated methylation is required to maintain levels of all substrate small RNAs, we assessed ERGO-1 class 26G RNAs for defects in accumulation in the henn-1 mutant. Loss of henn-1 has more severe consequences for this class of small RNAs than are observed for 21U RNAs: ERGO-1 class 26G RNA 26G-O3 fails to accumulate to wild-type levels at any stage of development, although the disparity is less pronounced in adulthood, during peak 26G RNA biogenesis (Figure 5A). For comparison, we assayed levels of ALG-3/ALG-4 class 26G RNA 26G-S5 across the developmental window during which it is readily detected by Taqman RT-qPCR. Levels of 26G-S5 are similar in the henn-1 mutant relative to wild-type (Figure 5B), consistent with the idea that HENN-1 is required for accumulation of ERGO-1 class 26G RNAs but dispensable for that of ALG-3/ALG-4 class 26G RNAs. Analysis of seven additional ERGO-1 class 26G RNAs and two additional ALG-3/ALG-4 class 26G RNAs corroborated these observations (Figures S8, S9).
To determine the effect of loss of henn-1 on the silencing of ERGO-1 class 26G RNA targets, we assayed levels of a panel of mRNAs targeted by ERGO-1 class 26G RNAs for desilencing in henn-1 mutant animals. During time points at which ERGO-1 class 26G RNAs are abundant, only modest upregulation of some, but not all, targets was detected; furthermore, no single target shows consistent desilencing in the henn-1 mutant (Figure 5C, Figure S10A). This is not unexpected, however, as the targets themselves vary in both expression and sensitivity to small RNA-mediated silencing across development [40]. To determine the specificity of this effect, two non-targets were examined in parallel. The maximal upregulation for either non-target does not exceed the maximal upregulation observed for any target, suggesting that the upregulation of ERGO-1 class 26G RNA targets in the henn-1 mutant may be a consequence of 26G RNA depletion (Figure 5C, Figure S10B). This connection is supported by our observation that a Wago-dependent and ERGO-1 class 26G RNA-dependent secondary siRNA that presumably enhances target silencing also shows defects in accumulation in embryo (Figure S6B). The effect is modest, indicating that, as observed for the piRNA pathway, the depleted pool of ERGO-1 class 26G RNAs in the henn-1 mutant is still sufficient for triggering fairly robust production of secondary siRNAs. Nevertheless, in an accompanying manuscript, Montgomery et al. observe that HENN-1 is required for silencing activity of a similar secondary siRNA upon a sensor transgene [54], suggesting that this pathway may indeed be compromised by loss of henn-1.
ALG-3/ALG-4 class 26G RNAs are restricted to the male germline, and their mRNA targets are enriched for genes involved in spermatogenesis [40]. Accordingly, loss of ALG-3/ALG-4 class 26G RNAs results in male-associated sterility at non-permissive temperatures due to defects in sperm activation that are thought to arise from target dysregulation [41]. ERGO-1 class 26G RNAs, in contrast, are dispensable for fertility and target mostly poorly conserved and incompletely annotated genes, many of which reside in duplicated regions of the genome [42]. It is therefore not unexpected that the ergo-1(tm1860) mutant, which lacks ERGO-1 class 26G RNAs, exhibits no overt phenotypes that can be traced to target dysregulation. Rather, the ergo-1(tm1860) mutant exhibits an enhanced RNAi sensitivity (Eri) phenotype that is attributed to effects of loss of the ERGO-1-dependent small RNAs themselves; presumably, depletion of ERGO-1 class 26G RNAs and dependent secondary siRNAs liberates limiting RNAi factors shared between the endogenous and exogenous RNAi pathways [43], [55], [56].
To determine whether loss of henn-1 depletes ERGO-1 class 26G RNAs sufficiently to produce an Eri phenotype, as observed in the ergo-1 mutant, we subjected L1 larvae from a panel of strains to feeding RNAi targeting various genes in the soma or germline. In order to expose subtle differences in RNAi sensitivity, we modulated the degree of knockdown, attenuating the dose of dsRNA trigger by diluting the bacterial RNAi clone with a bacterial clone harboring empty vector. RNAi of the somatic gene lir-1 causes larval arrest and lethality in wild-type animals at full strength, but dilution 1∶1 with empty vector largely eliminates the effect. In contrast, the eri-1(mg366) mutant, which lacks 26G RNAs, is affected severely by even dilute lir-1 RNAi. The henn-1 mutant also shows dramatically increased sensitivity to lir-1 feeding RNAi relative to wild-type (Figure 6A, 6B). A henn-1; eri-1 double mutant, however, shows RNAi sensitivity that is virtually identical to that of the single eri-1 mutant, suggesting that the Eri phenotype of each allele likely stems from the same defect, namely, loss of ERGO-1 class 26G RNAs. While the somatic Eri phenotype of the henn-1 mutant shows partial rescue by the germline-specific henn-1::gfp transgene xkSi2, henn-1::gfp expression under the native promoter from transgene xkSi1 rescues wild type RNAi sensitivity completely in the henn-1 mutant (Figure 6B). These findings suggest that loss of henn-1 in both germline and soma contributes to the Eri phenotype of the henn-1 mutant. The henn-1 mutant exhibits a similar somatic Eri response to RNAi of dpy-13 and lin-29 (Figure S11).
While the somatic Eri phenotype of the henn-1 mutant was expected, knockdown of genes required for germline development or embryogenesis revealed that, incongruously, the henn-1 mutant maternal germline exhibits an RNAi defective (Rde) phenotype. Animals subjected to pos-1 RNAi lay dead embryos because maternally loaded pos-1 mRNA is required for specifying cell fate of many tissues during embryonic development [57]. On pos-1 RNAi diluted 1∶2 with empty vector (1/3 strength), knockdown in wild-type animals is still sufficiently robust to reduce average brood size to fewer than five offspring per animal. henn-1 mutant animals at this dilution, however, produce an average brood greater than tenfold that of wild-type, suggesting that loss of henn-1 confers resistance to RNAi-mediated knockdown of this maternally deposited mRNA (Figure 6C). A lesser but statistically significant effect was observed for RNAi of the germline-expressed transcripts par-1, par-2, pie-1, and glp-1 (Figure S12). Sensitivity to pos-1 RNAi is effectively rescued by either endogenous or germline-specific expression of henn-1::gfp, likely due to the fact that both transgenes are expressed in germline.
HEN1 orthologs appear to be restricted to the germline in vertebrates [15], [33]; however, we observe phenotypes in both the germline and soma of the henn-1 mutant that suggest broader activity. To investigate expression of HENN-1 in C. elegans, we assessed henn-1 mRNA and protein levels throughout development. henn-1 mRNA levels are lowest in young larva and increase as the germline proliferates, peaking in gravid adult (Figure 7A, Figure S13A). Germline-deficient glp-4(bn2) adult hermaphrodites show approximately a 50% reduction in henn-1 mRNA levels relative to wild-type (Figure S13B), indicating that henn-1 mRNA is expressed in both germline and soma. Embryonic levels of henn-1 are high but decrease rapidly; this pattern suggests that, unlike in zebrafish [15], henn-1 mRNA may be maternally deposited into the embryo. HENN-1 protein is detectable throughout development and in both hermaphrodite and male adults (Figure 7B).
We next assessed the distribution of HENN-1::GFP fusion protein expressed from xkSi1, the rescuing henn-1::gfp transgene driven by the endogenous promoter, in the henn-1 mutant background. Although single copy transgene expression levels are too low for direct visualization by fluorescence microscopy, HENN-1::GFP is readily detected using a mouse monoclonal anti-GFP antibody. Whole-mount immunostaining of transgenic L4 larvae reveals that HENN-1::GFP is expressed broadly in diverse somatic tissues and germline (Figure S13C). Non-transgenic larvae show no signal, indicating that detection of HENN-1::GFP is specific. In extruded gonads of xkSi1; henn-1 hermaphrodites, HENN-1::GFP is detected throughout the germline. Notably, the proximal oocytes show cytoplasmic and intense nucleoplasmic HENN-1::GFP expression (Figure 7C). Although nucleoplasmic enrichment is lost following fertilization, HENN-1::GFP is also abundant in embryo, with ubiquitous expression prior to gastrulation (Figure S13D). HENN-1::GFP is also expressed throughout the germline of xkSi1; henn-1 males (Figure 7D). During sperm maturation, we detect enrichment of HENN-1::GFP in residual bodies, but we cannot definitively conclude that it is excluded from sperm (Figure 7D, inset). In wild-type animals, studies of endogenous HENN-1 using the rabbit polyclonal antibody generated against an N-terminal HENN-1 epitope corroborate the above findings, although the signal is more difficult to detect (Figure 7E). Staining in the henn-1 mutant yields no signal for anti-GFP and anti-HENN-1 antibodies (Figure 7F); this demonstrates that detection of transgenic and endogenous HENN-1 proteins is specific. Together, these data define an expression pattern consistent with a role for HENN-1 in modifying small RNAs in both male and female germlines as well as in soma.
The 21U RNAs and 26G RNAs appear to be significantly stable only in the presence of their respective Argonaute proteins [35], [36], [40]; accordingly, the localization patterns of the Argonaute proteins reflect the distribution of the different classes of small RNAs. We therefore wanted to compare the expression patterns of HENN-1 and the 26G RNA-binding Argonautes to determine whether the small RNA substrate specificity of HENN-1 could be explained by differential access to Argonaute-bound small RNAs. ERGO-1, which binds methylated 26G RNAs, is abundant in embryo [42], and its transcript is enriched during oogenesis [58], but its localization has not yet been reported. We assessed the staining pattern of ERGO-1 in hermaphrodite gonad and embryo using a polyclonal antibody generated against a C-terminal ERGO-1 epitope. ERGO-1 expression in the hermaphrodite germline begins at pachytene exit and persists in embryo (Figure S13D, S13E). ERGO-1 shows cytoplasmic enrichment both in germline and embryo, suggesting that the cytoplasmic pool of HENN-1 may act in methylating 26G RNAs bound by ERGO-1. This interaction may, however, be transient, as we were unable to identify HENN-1 by mass spectrometry of immunopurified ERGO-1 complexes, nor could we detect ERGO-1 in immunopurified HENN-1::GFP complexes by western blot (data not shown). Notably, both HENN-1 and ERGO-1 remain abundant in early embryo (Figure S13D). This is consistent with the proposed existence of a somatic endo-siRNA pathway that promotes continued biosynthesis of ERGO-1 class 26G RNAs after fertilization [59].
We next assessed co-localization of HENN-1 and ALG-3. ALG-3 and its close paralog, ALG-4, bind unmethylated 26G RNAs, and their transcripts are enriched during spermatogenesis [58]. In the male gonad, a rescuing gfp::alg-3 transgene was reported to express in the proximal male germline, with localization to P granules beginning at late pachytene [41]. During sperm maturation, GFP::ALG-3 is relegated to residual bodies. Dual immunostaining of GFP::ALG-3 and endogenous HENN-1 demonstrates a large region of overlap (Figure S13F), but HENN-1 does not appear to localize to P granules. This does not explain why ALG-3/ALG-4 class 26G RNAs are not methylated, because it is likely that HENN-1 can access P granules transiently: PRG-1 localizes predominantly to P granules [35], [37], and the PRG-1-bound piRNAs are methylated. This is in contrast to zebrafish Hen1, which carries a poorly conserved C-terminal domain (Figure S1) that directs localization of Hen1 to nuage, perinuclear granules similar to C. elegans P granules, to methylate piRNAs [15].
We have shown that HENN-1 is essential for methylating select classes of C. elegans small RNAs, namely, 21U RNAs and ERGO-1 class 26G RNAs. As is the case in other animals, small RNAs in C. elegans that associate with Piwi clade Argonautes require HENN-1 for maintenance of wild-type levels. Ago clade Argonaute-associated microRNAs and ALG-3/ALG-4 class 26G RNAs, in contrast, are HENN-1-independent (Figure S14A). It has been proposed that spatial and temporal regulation of HEN1 ortholog expression may contribute to small RNA substrate specificity in metazoans [24]. However, our immunostaining studies indicate that HENN-1 is coexpressed in the same tissues and subcellular compartments as Argonautes ERGO-1, PRG-1, and ALG-3 and their respective small RNAs (Figure 7, Figure S13). Therefore, differences in gross sub-cellular localization cannot explain the failure of ALG-3/ALG-4 class 26G RNAs to be methylated. Furthermore, although the two subclasses of 26G RNAs are generated in different germlines from non-overlapping targets, their sequences exhibit no obvious distinguishing characteristics that might account for their non-uniform methylation status.
One model of small RNA methylation posits that animal HEN1 orthologs only methylate small RNAs bound by Argonautes [15], [22]–[24], [49]. In support of this, work in fly shows that siRNA methylation requires assembly of DmAgo2 RISC [22], [50], and in vitro studies using lysate from a silkworm ovary-derived cell line show that methylation of synthetic RNA only occurs after the longer substrate is bound by a Piwi protein and trimmed to piRNA size [60]. This model predicts that all 26G RNAs are bound as unmethylated species by either ERGO-1 in the female germline or ALG-3/ALG-4 in the male germline and subsequently methylated or not, respectively. This is consistent with our findings in vivo that ERGO-1 is required for methylation of 26G RNAs (Figure 4A) and associates with 26G RNAs of either methylation status (Figure 4C). It has been further proposed that the identity of the Argonaute determines whether bound small RNAs are methylated [22], [23], [49], [50]. An elegant illustration of this is provided by fly miR-277, which associates with both Ago1, the canonical fly miRNA Argonaute, and Ago2, which binds methylated siRNAs [61]. The miR-277 pool contains both methylated and unmethylated species. Depletion of Ago2 in cell culture results in loss of methylated miR-277, whereas Ago1 depletion results in a completely methylated miR-277 population [22]. Similarly, fly hairpin derived hp-esiRNAs sort into Ago1 and Ago2, but accumulate mainly in Ago2 because only hp-esiRNAs bound by Ago2 are methylated and therefore protected against degradation triggered by their extensive target complementarity [50]. In C. elegans, the model of Argonaute-dictated methylation can be invoked to explain the disparate methylation of the 26G RNAs: in the male germline, only ALG-3/ALG-4 are expressed, resulting in an unmethylated male 26G RNA population, whereas exclusive expression of ERGO-1 in the female germline and embryo directs methylation of female and zygotic 26G RNAs. This raises the intriguing possibility that selective expression of Argonautes that permit or prevent methylation could represent a new mechanism for differentially regulating small RNA turnover.
It is important to note that our results do not definitively exclude an alternative model wherein 26G RNAs are methylated prior to association with Argonautes and subsequently bound by ALG-3/ALG-4 only if unmethylated or by ERGO-1 only if methylated. In this model, HEN1 would methylate 26G RNAs in both germlines, but degradation of labile unbound siRNAs would result in a purely unmethylated or methylated population of 26G RNAs in male and female germlines, respectively. Because 26G RNAs assessed in embryo are fully methylated (Figure 3A, Figure S3B), such a mechanism would require that ERGO-1 exhibit very unfavorable kinetics for association with unmethylated small RNAs. We do not find this to be the case, as ERGO-1 binds some 26G RNAs with similar efficiency when methylated and unmethylated (Figure 4C). Our data therefore provide stronger evidence for a model of Argonaute-dictated methylation of small RNAs.
Differential germline expression of Argonautes could have evolved in C. elegans because of advantages conferred by selective stabilization of female germline 26G RNAs. Unlike ALG-3/ALG-4 class 26G RNAs, which appear to function exclusively during sperm development [40], [41], ERGO-1 class 26G RNAs exert much of their influence during embryonic and larval development, well beyond initiation of their biogenesis in the hermaphrodite germline [40]. Accordingly, their targets are depleted of germline-enriched genes [40], [59]. The oocyte contributes the vast majority of the initial zygotic cellular contents; therefore, methylation of 26G RNAs originating in the female germline may ensure robust inheritance and perdurance of primary small RNAs. Methylation of 26G RNAs in the male germline would likely not significantly increase their representation in sperm or zygote, as ALG-3/ALG-4 are relegated to residual bodies during spermatogenesis and exert effects in mature sperm only indirectly through dependent secondary 22G RNAs [41]. Nonetheless, it would be interesting to express ERGO-1 ectopically in sperm and determine whether ALG-3/ALG-4 class small RNAs are methylated. Such a strategy may reveal unexpected consequences related to inappropriate methylation and stabilization of ALG-3/ALG-4 class 26G RNAs.
In the absence of henn-1, we show that response to RNAi-mediated knockdown is enhanced for somatic genes (Figure 6A and 6B, Figure S11). This is likely due to destabilization of ERGO-1 class 26G RNAs in the henn-1 mutant, which reduces competition with primary exo-siRNAs for stimulating secondary siRNA activity mediated by somatic Argonautes such as SAGO-1 and SAGO-2 [43], [55]. While germline-specific expression of henn-1::gfp only partially rescues this somatic Eri phenotype, henn-1 mutant animals rescued with an endogenous henn-1::gfp transgene, which drives both somatic and germline expression, show wild-type RNAi sensitivity. Under the model of competing endo- and exo-RNAi pathways, this suggests that HENN-1-mediated methylation of ERGO-1 class 26G RNAs in the germline alone cannot maintain small RNA levels sufficient to sequester an appropriate proportion of the limiting RNAi factors. It is possible that ERGO-1 class 26G RNA biogenesis continues in embryo and larva, as previously suggested [59], and that high concentrations of HENN-1 are necessary for continued stabilization of these small RNAs. Such a model would be consistent with our characterization of the distributions of HENN-1 and ERGO-1, both of which are still detected in abundance in developing embryo (Figure S13D, S13E).
While the majority of the phenotypes observed in the henn-1 mutant can be attributed to destabilization of endogenous small RNA substrates, the germline Rde phenotype suggests a role for HENN-1 in exogenous RNAi. It is unclear why HENN-1 is dispensable for robust exogenous RNAi in the soma but required in the germline. While this may be an indirect effect, as suggested in concurrent work by Kamminga et al. [62], one possible explanation is that HENN-1 stabilizes primary exo-siRNAs or dependent 22G secondary siRNAs. There is support in fly for methylation of exo-siRNAs and transgenic hairpin-derived siRNAs [22], [63], but this has not yet been demonstrated in C. elegans. 22G RNAs triggered by primary exo-siRNAs appear not to be methylated [47], consistent with our and others' observations that Wago-dependent 22G RNAs from diverse endogenous sources are unmethylated (Figure 3A, Figure S3B, and [45]). The methylation status of worm primary exo-siRNAs has not been definitively established, although a 22-nucleotide siRNA generated from a transgene encoding a perfect hairpin was not found to be methylated [46].
All Argonautes contain two signature domains, PAZ and Piwi [64]. The Piwi domain, unique to Argonautes, adopts an RNase H-like configuration and serves as the catalytic core of RISC [65], [66]. The PAZ domain recognizes and anchors the 3′ end of the small RNA [67], [68]. Comparison of Piwi and Ago clade Argonautes reveals that Piwi proteins contain a small insertion in their PAZ domains in a loop connecting two β strands [69]. Crystal structures of a human Piwi Argonaute PAZ domain suggest that this insertion results in the formation of a more spacious binding pocket capable of accommodating the 2′-O-methyl group of a piRNA. Interactions between the methyl group and hydrophobic residues lining the pocket confer a threefold to sixfold higher binding affinity for 2′-O-methyl than 2′-OH [69]. In C. elegans, only PRG-1/PRG-2 and ERGO-1 show evidence of a PAZ domain insertion (Figure S14B), consistent with their designation as Piwi clade Argonautes and association with methylated small RNAs.
In spite of their shared classification, ERGO-1 exhibits far less homology than PRG-1/PRG-2 to mammalian and insect Piwi proteins (Figure S14A) [43]. Similarly, among worm, fly, and human Argonautes, DmAgo2 and C. elegans Argonaute RDE-1 are among the most divergent members of their clades [43]. In fact, so divergent is RDE-1 that its cladistics are ambiguous, with our and other published alignments variably assigning it to each of the three clades (Figure S14A and [43], [70]). Both DmAgo2 and RDE-1 bind exo-siRNAs, although only the former has been shown to permit methylation [22]. Interestingly, both lack the insertion found in Piwi Argonaute PAZ domains (Figure S14B). The absence of this insertion in DmAgo2 suggests that it is not required for association with methylated small RNAs, raising the possibility that RDE-1 too may permit methylation of associated small RNAs. If HENN-1 does not methylate RDE-1-bound small RNAs, it is unclear what specific role HENN-1 plays in exo-RNAi in the germline. Nevertheless, its dual functions in endogenous and exogenous RNAi place HENN-1 in the company of DCR-1 and the Wago proteins at the intersection between these two RNAi pathways.
C. elegans were maintained according to standard procedures. The Bristol strain N2 was used as the standard wild-type strain. The alleles used in this study, listed by chromosome, are: unmapped: neIs23[unc-119(+) GFP::ALG-3]; LGI: glp-4(bn2), prg-1(tm872); LGII: xkSi1[PC30A5.3::henn-1::gfp::henn-1 3′UTR cb-unc-119(+)] II, xkSi2[Ppie-1::henn-1::gfp::tbb-2 3′UTR cb-unc-119(+)] II; LGIII: rde-4(ne301), henn-1(tm4477); LGIV: eri-1(mg366), fem-1(hc17), him-8(e1489); LGV: ergo-1(tm1860). The neIs23[unc-119(+) GFP::ALG-3] strain was generously provided by Craig Mello (University of Massachusetts, Worcester, MA).
For embryo samples, L1 larvae were grown at 20°C until gravid. Embryos were isolated using sodium hypochlorite solution; an aliquot of embryos was allowed to hatch overnight at room temperature to determine viability. For male samples, synchronized him-8(e1489) L1 larvae were grown at 20°C for 72–75 hours. Males were isolated by filtering through 35 µm mesh [71]. For female samples, synchronized fem-1(hc17) L1 larvae were plated and grown at 25°C for 52 hours. For time course samples, synchronized wild-type (N2) and henn-1(tm4477) L1 larvae were grown at 25°C until gravid; embryos were extracted and harvested for RNA or hatched overnight at room temperature and then grown at 25°C for the specified number of hours before harvest. The prg-1(tm872) time course samples were prepared in the same way, except that animals were grown for the first generation at 20°C to evade temperature-sensitive sterility. Samples were processed by either three rounds of freeze/thaw lysis or two rounds of homogenization for 15 sec using the Tissue Master-125 Watt Lab Homogenizer (Omni International) and the RNA was extracted in TriReagent (Ambion) following the vendor's protocol, with the following alterations: RNA was precipitated in isopropanol for one hour at −80°C; RNA was pelleted by centrifugation at 4°C for 30 min at 20,000× g; the pellet was washed three times in 75% ethanol; the pellet was resuspended in water.
For detection of small RNAs, 10 or 40 µg of total RNA were β-eliminated as described [51]; control samples were processed in parallel without sodium periodate. Northern blot analysis was performed as described [72]. In brief, 5 or 10 µg of β-eliminated total RNA were resolved on 17.5% or 20% denaturing Urea-PAGE gels (SequaGel, National Diagnostics) and transferred to Hybond-NX membrane (Amersham). 21 and 26 nt synthetic RNAs were run as size markers and visualized in tandem with rRNA by ethidium bromide staining. Pre-hybridization/hybridization and washes were performed at 48°C or 50°C. Oligonucleotides corresponding to the antisense sequences of the small RNAs (Table S1) were synthesized and end-labeled with [α-32P]-dATP using the miRNA StarFire kit (Integrated DNA Technologies).
To test the response to exogenous RNAi, bacterial clones from the Ahringer RNAi library [73] were diluted with bacteria harboring the empty vector L4440 to achieve a level of RNAi sensitivity that allowed us to differentiate the RNAi responses in the strains examined. To determine lir-1 RNAi sensitivity, the lir-1 RNAi bacterial clone diluted with L4440 bacterial clone at a 1∶1 or 1∶2 ratio (1/2 or 1/3 strength) was used; >50 L1 larvae were plated per plate and the number of total animals assayed per plate was determined at day two after plating; the percent of animals exhibiting the larval arrest phenotype was determined at 70 hours at 20°C. Sensitivity to RNAi of dpy-13 and lin-29 was also assessed using this method, where animals subjected to dpy-13 RNAi were imaged at 70 hours and those subjected to lin-29 RNAi were evaluated for the absence of protruding vulva or bursting phenotype. For pos-1 RNAi, synchronized L1 larvae were singled onto plates with pos-1 RNAi diluted with empty vector at a 1∶2 ratio (1/3 strength) that had been induced overnight at 25°C. Animals were grown at 20°C for six days and progeny were counted. Sensitivity to RNAi of pie-1, par-1, and par-2 was assessed similarly at the indicated dilutions with 4 plates of 4 P0 animals per strain. Sensitivity to glp-1 RNAi was determined at the indicated dilutions by plating 4 plates of >50 L1 larvae per strain per gene and scoring for the absence of oocytes and embryos in both arms of the germline at 70 hours at 20°C. For all RNAi sensitivity assays, data are representative of at least two independent experiments.
To determine brood size, synchronized L1 larvae from gravid adults grown at 20°C or shifted to 25°C for two generations were singled onto plates with OP50 and grown to adulthood at their respective temperatures. Once egg-laying began, animals (N≥13 per strain) were transferred to fresh plates daily until the supply of fertilized eggs was exhausted. Progeny of the singled parents were counted as late larvae/adults. Results are representative of two independent experiments.
Taqman small RNA probes were synthesized by Applied Biosystems (Table S2) [74]. For each reaction, 50 ng of total RNA were converted into cDNA using Multiscribe Reverse Transcriptase (Applied Biosystems). The resulting cDNAs were analyzed by a Realplex thermocycler (Eppendorf) with TaqMan Universal PCR Master Mix, No AmpErase UNG (Applied Biosystems). We could not identify a small RNA whose levels were consistent across development for use in normalization. Therefore, to preserve the developmental profile of each of the small RNA assessed, back transformation was used to calculate relative small RNA levels from qRT-PCR cycle numbers. As a control for RNA quality, miR-1 Taqman assays were run in parallel for all samples excluding the ERGO-1 RNA immunoprecipitation samples, in which miRNAs are absent. For quantification of mRNA levels, 100 ng of total RNA were converted into cDNA with Multiscribe Reverse Transcriptase (Applied Biosystems) following the vendor's protocol with the following changes: 25 units of RT and 7.6 units of RNAse OUT (Invitrogen) were used per reaction. cDNAs were analyzed using Power Sybr Green PCR Master Mix (Applied Biosystems) (primers, Table S3). Relative mRNA levels were calculated based on the ΔΔ2Ct method [75] using eft-2 for normalization. For all qPCR, 40 cycles of amplification were performed; reactions whose signals were not detected were therefore assigned a cycle number of 40. All results presented are the average values of independent calculations from biological triplicates unless indicated. To determine average upregulation of ERGO-1 26G RNA targets in henn-1 relative to wild-type (Figure 5C), the mean was calculated for all of the ratios generated by dividing each henn-1 biological replicate by each wild-type biological replicate.
3′ RACE was performed using the 3′ RACE System for Rapid Amplification of cDNA ends (Invitrogen) according to the manufacturer's protocol. henn-1 gene-specific primer (5′ GCAGTATGTCGCCTCCAAGTAGAT 3′) was used to amplify henn-1 3′ ends from cDNA generated from embryo. Product corresponding to only the seven-exon gene model of henn-1 was observed, consistent with detection of a single protein isoform corresponding to this model on western blot analysis.
The endogenous henn-1::gfp reporter construct (xkSi1) was generated by introducing the following fragments into pCFJ151: endogenous promoter of the henn-1-containing operon CEOP3488 [76] (3.9 kb PCR fragment immediately upstream of the C30A5.3 start codon), henn-1 genomic coding region (1.8 kb PCR fragment with mutated termination codon), gfp coding region (0.9 kb fragment with multiple synthetic introns and termination codon), and henn-1 endogenous 3′UTR (1.1 kb PCR fragment immediately downstream of henn-1 termination codon). The germline-only henn-1::gfp reporter construct (xkSi2) was generated as above with the following substitutions: CEOP3488 operon promoter was replaced with the pie-1 promoter (2.4 kb PCR fragment immediately upstream of pie-1 start codon) and henn-1 endogenous 3′UTR was replaced with the C36E8.4 3′UTR (0.3 kb PCR fragment downstream of C36E8.4). Constructs were cloned into the pCFJ151 vector, confirmed by sequencing, and used to generate single-copy integrated transgenes via the MosSCI technique [52]. Gene fusion products of the expected size were specifically detected by western blot with both anti-HENN-1 and anti-GFP antibodies.
Synthetic antigenic peptides were conjugated to KLH and each was used to immunize two rabbits (Proteintech). Antisera were subsequently affinity purified using Affi-Gel 15 gel (Bio-Rad). Antigenic peptide sequences are as follows: N-terminal HENN-1 peptide with N-terminal added cysteine (CTYVEAYEQLEIALLEPLDR), C-terminal ERGO-1 peptide (CEVNKDMNVNEKLEGMTFV).
Proteins immobilized on Immobilon-FL transfer membrane (Millipore) were probed with anti-HENN-1 rabbit polyclonal antibody (1∶2000), anti-γ-tubulin rabbit polyclonal antibody (LL-17) (Sigma) (1∶2000), or anti-ERGO-1 rabbit polyclonal antibody (1∶1000). Peroxidase-AffiniPure goat anti-rabbit IgG secondary antibody was used at 1∶10000 (Jackson ImmunoResearch Laboratories) for detection using Pierce ECL Western Blotting Substrate (Thermo Scientific).
Wild-type, henn-1, or eri-1(mg366) embryos isolated from gravid adults grown at 20°C were frozen in liquid nitrogen and homogenized with a Mixer Mill MM 400 ball mill homogenizer (Retsch) Homogenates were suspended in lysis buffer (50 mM HEPES (pH 7.4), 1 mM EGTA, 1 mM MgCl2, 100 mM KCl, 10% glycerol, 0.05% NP-40 treated with a Complete, Mini, EDTA-free Protease Inhibitor Cocktail tablet (Roche Applied Sciences)) and clarified by centrifugation at 12,000× g for 12 minutes at 4°C. Aliquots of homogenate were reserved as crude lysate for western blot to confirm that immunoprecipitations were performed in lysates of equivalent protein concentration (2 mg/mL). For immunoprecipitations, embryo homogenates were incubated at 4°C for one hour with 75 µg anti-ERGO-1 rabbit polyclonal antibody conjugated to Dynabeads Protein A (Invitrogen), after which the beads were washed (500 mM Tris-HCl (pH 7.5), 200 mM KCl, 0.05% NP-40) and associated proteins were eluted with 200 µL glycine. Three quarters of each eluate were precipitated overnight at 4°C in trichloroacetic acid, pelleted, washed with acetone, and resuspended for western blot analysis. The remaining eluate was treated with 2 mg/ml Proteinase K (Roche) and incubated at 37°C for 30 minutes. RNA was isolated from the eluate by incubation with TriReagent and processed as described above. RNA pellets were resuspended in 10 µL water and 5 µL were used for each Taqman RT reaction.
Primary antibodies were applied according to the following specifications: anti-GFP mouse monoclonal antibody 3E6 (Invitrogen) was diluted 1∶1500 to detect HENN-1::GFP and 1∶200 to detect ALG-3::GFP; anti-ERGO-1 rabbit polyclonal was diluted 1∶200; and anti-HENN-1 rabbit polyclonal antibody was preabsorbed as described [77] with henn-1(tm4477) mutant extract and diluted 1∶200. Alexa Fluor 555 goat anti-rabbit IgG and Alexa Fluor 488 goat anti-mouse IgG (Molecular Probes) secondary antibodies were diluted 1∶500. All antibodies were diluted in 0.5% bovine serum albumin (Sigma). For immunostaining of gonads and embryos, synchronized gravid hermaphrodites or adult males grown at 20°C were dissected on Superfrost Plus positively charged slides (Fisherbrand) with 27 G×1/2 inch BD PrecisionGlide needles (Becton, Dickinson and Company) as described by Chan and Meyer in WormBook [78] Protocol 21 with 1.5% paraformaldehyde (Sigma). Slides were incubated with primary antibodies overnight at 4°C and with secondary antibodies for three hours at room temperature. Slides were mounted with VECTASHIELD Mounting Medium with DAPI (Vector Laboratories). For whole-worm immunostaining, synchronized late L4 larvae grown at 20°C were transferred to subbed slides [77] in M9, fixed for six minutes in 1.5% paraformaldehyde, freeze-cracked, and incubated for 15 minutes in ice cold methanol. After fixation, slides were processed as above. Images were captured on an Olympus BX61 epifluorescence compound microscope with a Hamamatsu ORCA ER camera using Slidebook 4.0.1 digital microscopy software (Intelligent Imaging Innovations) and processed using ImageJ.
|
10.1371/journal.pntd.0003725 | Overexpression of Cytoplasmic TcSIR2RP1 and Mitochondrial TcSIR2RP3 Impacts on Trypanosoma cruzi Growth and Cell Invasion | Trypanosoma cruzi is a protozoan pathogen responsible for Chagas disease. Current therapies are inadequate because of their severe host toxicity and numerous side effects. The identification of new biotargets is essential for the development of more efficient therapeutic alternatives. Inhibition of sirtuins from Trypanosoma brucei and Leishmania ssp. showed promising results, indicating that these enzymes may be considered as targets for drug discovery in parasite infection. Here, we report the first characterization of the two sirtuins present in T. cruzi.
Dm28c epimastigotes that inducibly overexpress TcSIR2RP1 and TcSIR2RP3 were constructed and used to determine their localizations and functions. These transfected lines were tested regarding their acetylation levels, proliferation and metacyclogenesis rate, viability when treated with sirtuin inhibitors and in vitro infectivity.
TcSIR2RP1 and TcSIR2RP3 are cytosolic and mitochondrial proteins respectively. Our data suggest that sirtuin activity is important for the proliferation of T. cruzi replicative forms, for the host cell-parasite interplay, and for differentiation among life-cycle stages; but each one performs different roles in most of these processes. Our results increase the knowledge on the localization and function of these enzymes, and the overexpressing T. cruzi strains we obtained can be useful tools for experimental screening of trypanosomatid sirtuin inhibitors.
| Sirtuins are a family of deacetylases, evolutionary conserved from bacteria to mammals. They participate in the regulation of a wide range of nuclear, cytoplasmic and mitochondrial pathways, and are considered pro-life enzymes. In the last years the search for sirtuin inhibitors was a very active field of research, with potential applications in a large number of pathologies, including parasitic diseases. We are interested in the study of the two sirtuins present in the protozoan parasite Trypanosoma cruzi, being our objective to understand their function. First, we determined the localization of these enzymes in the parasite: TcSIR2RP1 is a cytoplasmic enzyme and TcSIR2RP3 localizes in the mitochondrion. When we overexpress cytoplasmic TcSIR2RP1, the transgenic parasites differentiate to metacyclic trypomastigotes and infect mammalian cells more efficiently. In contrast, the overexpression of mitochondrial TcSIR2RP3 does not affect metacyclogenesis but modifies epimastigotes growth and slightly increases the proliferation of the parasite in the intracellular stage. We also used these transgenic lines to test their sensibility to previously described sirtuin inhibitors.
| Acetylation is a ubiquitous protein modification present in prokaryotic and eukaryotic cells that participates in the regulation of many cellular processes. A limited set of acetyltransferases and deacetylases, and of the acetyl-lysine “reading” domain (bromodomain) are the principal components of the acetylation/deacetylation machinery. Among them, protein deacetylases are enzymes that catalyze the removal of acetyl groups from the ε-amino group of lysine residues and are classified into four classes. Sirtuins, the class III (NAD+-dependent) protein deacetylases, are homologous to the yeast transcriptional repressor, Sir2 [1]. Sir2, as well as all sirtuins, deacetylates lysine residues in a unique chemical reaction that consumes nicotinamide adenine dinucleotide (NAD+) and generates nicotinamide, O-acetyl-ADP-ribose (OAADRr), and the deacetylated substrate [2].
Saccharomyces cerevisiae Sir2, the founding member of the group, is a histone deacetylase (reviewed in [3]) involved in a range of chromatin-mediated processes; namely, gene silencing at telomeres and mating-type loci, DNA repair [4–5], suppression of recombination within ribosomal DNA (rDNA)[6], DNA replication [7], chromosome stability [8] and plasmid segregation [9]. However, the identification and characterization of new members of this protein family in other organisms led to the discovery of more diverse functions and localizations. It is now recognized that sirtuins remove acetyl groups from lysines in nuclear, cytosolic and mitochondrial protein substrates [10]. Sirtuins are evolutionarily conserved enzymes present in all kingdoms of life, ranging from bacteria to higher eukaryotes including humans. Members of this family share a core domain of ~250 amino acids that exhibits 25–60% sequence identity between different organisms. Genes coding for seven sirtuins (SIRT 1–7) have been found in the human genome, with subcellular distribution, substrate specificity, and cellular functions quite diverse [11].
Trypanosoma cruzi is a hemoflagellate protozoan parasite, branched early from the eukaryal lineage. It is an intracellular pathogen responsible for Chagas’ disease, or American Trypanosomiasis, a chronic infectious disease affecting 8 million people [12]. While Chagas disease is endemic in Latin America, a significant increase in confirmed cases of Chagas has recently been reported in the USA, Canada, Japan, Australia and Europe, indicating that it is an emerging disease [13]. Current therapies rely on a very small number of drugs, most of which are inadequate because of their severe host toxicity and numerous side effects. The identification of new biotargets is essential for the development of more efficient therapeutic alternatives. The structural basis for inhibition of sirtuins has been established through previous structural and functional studies [14–17]. Involvement of sirtuins in the cell cycle strongly suggests a role for these enzymes in cancer and the potential use of their inhibitors as anticancer drugs [18]. In addition, inhibition of sirtuins from Trypanosoma brucei and Leishmania ssp. showed promising results, indicating that these enzymes may be considered as targets for drug discovery in parasite infection [19–22].
T. cruzi belongs to the Kinetoplastida order, Trypanosomatidae family, as well as Trypanosoma brucei and Leishmania ssp., and together they are termed TriTryps. Genes encoding three Sir2 related proteins (SIR2RPs) were found in the TriTryps. The trypanosomatid genes were designated SIR2-related proteins, SIR2RP1–3. A previous phylogenetic analysis places SIR2RP1 in a group with ScSir2, HsSIRT1 and HsSIRT2, while SIR2RP2 and SIR2RP3 are more closely related to bacterial proteins and to HsSIRT4 and HsSIRT5 respectively [23]. However, a more recent extensive analysis places SIR2RP1 in the sirtuins subgroup Ib, together with cytoplasmic HsSIRT2, and is now clearly differentiated from the nuclear HsSIRT1 [24].
SIR2RP1 from several Leishmania species and all three SIR2RPs from T. brucei have been characterized [16, 23]. SIR2RP1 is found in cytoplasmic granules in different stages of L. major, L. infantum and L. amazonensis; and, under certain conditions, it is observed in the excreted/secreted fraction [25–27]. LiSIR2RP1 is also found associated with the cytoskeleton network and deacetylates α-tubulin, a function that resembles that of human SIRT2 and HDAC6. In contrast, TbSIR2RP1 is a nuclear chromosome-associated protein. It is expressed throughout T. brucei life cycle, catalyses NAD+-dependent ADP ribosylation and deacetylation of histones and in the mammalian-infective bloodstream-stage controls DNA repair and repression of RNA polymerase I-mediated expression immediately adjacent to telomeres [16, 23]. TbSIR2RP2 and 3 localize in the single mitochondrion of the parasite and it was reported that their interference do not produce growth or differentiation defects. In Trypanosoma cruzi, the gene coding for SIR2RP2 is lacking. This general landscape suggests that in spite of their sequence similarity, sirtuin variants from TriTryps have evolved to different functions.
Here, we report the first characterization of the two sirtuins present in T. cruzi. In epimastigotes, TcSIR2RP1 localizes in the cytoplasm while TcSIR2RP3 localizes to the parasite’s single mitochondrion. Overexpression of TcSIR2RP1 causes no alteration to epimastigote growth, but it increases the number of trypomastigotes obtained in in vitro metacyclogenesis and the infectivity rate of Vero cells. In contrast, overexpression of TcSIR2RP3 slightly decreases epimastigote growth and the infectivity rate of Vero cells, it does not affect the in vitro differentiation to metacyclic trypomastigotes, and it increases the proliferation rate of intracellular amastigotes. Finally, overexpression of either of these sirtuins protects the parasite from the effect of sirtuin inhibitors.
All experiments were approved by the Institutional Animal Care and Use Committee of the School of Biochemical and Pharmaceutical Sciences, National University of Rosario (Argentina) (File 6060/227) and conducted according to specifications of the US National Institutes of Health guidelines for the care and use of laboratory animals. Rabbits were only used for the production of polyclonal antibodies. The rabbits were immunized three times with the protein and an equal volume of Freund´s adjuvant, and bled two weeks after the final injection [28].
TcSIR2RP1 and TcSIR2RP3 genes were amplified using the following oligonucleotides: SIR2RP1ss (5’ AAAGGATCCATGAATCAAGATAACGCCAAC), SIR2RP1HAas (5’ AACTCGAGAGCATAATCCGGCACATCATACGGATATTTTCGGT CTGTCAG), SIR2RP3ss (5’ AAGGATCCATGAAGCCGCGGCGTCAGAT) and SIR2RP3HAas (5′ AACTCGAGAGCATAATCCGGCACATCATACGGATACACCGCGT CTTGAAG). DNA purified from T. cruzi epimastigotes was used as template. The PCR products obtained with a proofreading DNA polymerase were inserted into pCR2.1-TOPO vector (Invitrogen) and sequenced. TcSIR2RP1 and TcSIR2RP3 coding regions were inserted into a pENTR3C vector (Gateway system Invitrogen) using the BamHI/XhoI restriction sites included in the oligonucleotides (underlined) and then transferred to pDEST17 (Gateway system Invitrogen) and pTcINDEX-GW vectors by recombination using LR clonase II enzyme mix (Invitrogen). The pDEST17 constructs were transformed into Escherichia coli BL21 pLysS and recombinant proteins, fused to a six histidine tag, were obtained by expression-induction with 0.5 mM IPTG for 5 h at 37°C. The proteins were purified by affinity chromatography using a Ni-NTA agarose resin (Qiagen) following the manufacturer’s instructions.
Rabbit polyclonal antisera against TcSIR2RP1 and TcSIR2RP3 were obtained by inoculating subcutaneously the recombinant proteins to these animals as described above.
T. cruzi epimastigote forms (Dm28c strain) were cultured at 28°C in liver infusion tryptose (LIT) medium (5 g/L liver infusion, 5 g/L bacto-tryptose, 68 mM NaCl, 5.3 mM KCl, 22 mM Na2HPO4, 0.2% (w/v) glucose and 0.002% (w/v) hemin) supplemented with 10% (v/v) heat-inactivated FCS, 100 U/ml penicillin and 100 mg/l streptomycin. Cell viability was assessed by direct microscopic examination.
For inducible expression of Sir2rp1-3 genes in the parasite, we first generated a cell line expressing T7 RNA polymerase and tetracycline repressor genes by transfecting epimastigotes with the plasmid pLew13 using a standard electroporation method. Briefly, epimastigote forms of T. cruzi Dm28c were grown at 28°C in LIT medium, supplemented with 10% FCS, to a density of approximately 3 × 107 cells/ml. Parasites were then harvested by centrifugation at 2,000 × g for 5 min at room temperature, washed once in PBS and resuspended in 0.35 ml of transfection buffer pH 7.5 (0.5 mM MgCl2, 0.1 mM CaCl2 in PBS) to a density of 1 × 108 cells/ml. Cells were then transferred to a 0.2 cm gap cuvette (Bio-Rad) and ~50 μg of DNA was added in a final volume of 40 μl. The mixture was placed on ice for 15 min and then subjected to 2 pulses of 450 V and 500 μF using GenePulser II (Bio-Rad, Hercules, USA). After electroporation, cells were transferred into 3 ml of LIT medium containing 10% FCS, maintained at room temperature for 15 minutes and then incubated at 28°C. Geneticin (G418; Life Technologies) was added at a concentration of 200 μg/ml, and parasites were incubated at 28°C. After selection, transfected epimastigotes were grown in the presence of 200 μg/ml of G418. This parental cell line was then transfected with pTcINDEX-GW constructs and transgenic parasites were obtained after 3 weeks of selection with 100 μg/ml G418 and 200 μg/ml Hygromycin B (Sigma).
To obtain metacyclic trypomastigotes, epimastigotes were differentiated in vitro following the procedure described by Contreras and coworkers [29] using chemically defined conditions (TAU3AAG medium). Briefly, cells were washed with PBS and incubated in TAU medium (190 mM NaCl, 17 mM KCl, 2 mM MgCl2, 2mM CaCl2, 8 mM phosphate buffer pH 6.0) in the absence or presence of 0.25 μg/ml Tetracycline, reaching a density of 5 x 108 parasites/ml at 28°C for 2 hours. Then they were diluted 1:100 in TAU3AAG Medium (TAU medium plus 10 mM Glucose, 2 mM L-Aspartic Acid, 50 mM L-Glutamic Acid and 10 mM L-Proline) and incubated at 28°C for 72 hours, again in the absence or presence of Tetracycline. Finally, the parasites were fixed, stained with Giemsa, visualized with a Nikon Eclipse Ni-U microscope and counted using ImageJ software [30]. Only parasites with a fully elongated nucleus and a round kinetoplast at the posterior portion end of the parasite were considered metacyclic forms [31]. Five hundred parasites from each triplicate were counted and the experiment was repeated tree times.
Vero cells were cultured in DMEM medium (Life Technologies), supplemented with 2 mM L-glutamine, 10% FCS, 100 U/ml penicillin and 100 μg/ml streptomycin.
Metacyclic trypomastigotes were obtained by spontaneous differentiation of epimastigotes at 28°C. Cell-derived trypomastigotes were obtained by infection with metacyclic trypomastigotes of Vero cell monolayers. After two rounds of infections, the cell-derived trypomastigotes were used for the infection and intracellular amastigotes proliferation experiments. Trypomastigotes were collected by centrifugation of the supernatant of previously infected cultures at 2,000 x g at room temperature for 10 minutes and incubated for 3 hours at 37°C in order to allow the trypomastigotes to move from the pellet into the supernatant. After this period, the supernatant was collected and trypomastigotes were counted in a Neubauer chamber. The purified trypomastigotes were pre-incubated in the presence or absence of 0.25 μg/ml Tetracycline for 3 hours and then used to infect new monolayers of Vero cells at a ratio of 10 parasites per cell. After 6 h of infection at 37°C, the free trypomastigotes were removed by successive washes using saline solution. Cultures were incubated in complete medium with or without Tetracycline (0.25 μg/ml) for 2 days post-infection. Infections were performed in DMEM supplemented with 2% FCS. Cells were then fixed in methanol and the percentage of infected cells and the mean number of amastigotes per infected cell, were determined by counting the slides after Giemsa staining using a Nikon Eclipse Ni-U microscope, by counting ~1000 cells per slide. The significances of the results were analyzed by a two-way ANOVA using GraphPad Prism version 6.0 for Mac. Results are expressed as means ± SEM of triplicates, and represent one of three independent experiments performed.
Exponentially growing epimastigotes were washed twice with cold PBS, pellets were resuspended in urea lysis buffer (8 M Urea, 20 mM Hepes pH 8, 1 mM phenylmethylsulphonyl fluoride (PMSF), and Protease Inhibitor Cocktail set I, Calbiochem), incubated at room temperature for 20 minutes and boiled for 5 minutes with protein loading buffer. Insoluble debris was eliminated by centrifugation. The same procedure was applied to amastigote and trypomastigote cellular pellets.
Transfected Dm28c epimastigotes in exponential growth phase were centrifuged for 10 min at 2,000 x g and washed twice in homogenization buffer (25 mM Tris-HCl pH 8, 1 mM EDTA, 0.25 M sucrose, 1 mM PMSF). Subcellular fractions were obtained following the procedure described by Opperdoes and coworkers [32]. The parasites were grinded in a pre-chilled mortar with 1 x wet weight silicon carbide until no intact cells were observed under the light microscope. The lysate was diluted and centrifuged at 100 x g for 10 min to remove the silicon carbide. Unbroken cells, nuclei and debris were sedimented at 1,000 x g for 10 min (Fraction N). From the resulting soluble extract a large-granule fraction (LG) was separated at 5,000 x g for 15 min, a small-granule fraction (SG) at 20,000 x g for 20 min and microsomal fraction (M) at 139,000 x g for 1 h. All the sediments were resuspended in urea lysis buffer.
Protein extracts were fractioned in SDS-PAGE and transferred to nitrocellulose membranes. Transferred proteins were visualized with Ponceau S. Membranes were treated with 10% non-fat milk in PBS for 2 hours and then incubated with specific antibodies diluted in PBS for 3 hours. Antibodies used were: rat monoclonal anti-HA (ROCHE), rabbit polyclonal anti-TcSIR2RP1 and anti-TcSIR2RP3, rabbit polyclonal anti-Acetyl-lysine (Millipore), mouse monoclonal anti-acetylated α-tubulin clone 6-11B-1 (Sigma Aldrich), mouse monoclonal anti-trypanosome α-tubulin clone TAT-1, rabbit polyclonal anti-T. cruzi mitochondrial Malate Dehydrogenase (TcMDHm), rabbit and mouse polyclonal anti-T. cruzi Tyrosine Amine Transferase (TcTAT), mouse polyclonal anti-T. cruzi Aspartate Transaminase (TcASAT) and rabbit polyclonal anti-T. cruzi Bromodomain Factor 2 (TcBDF2). Bound antibodies were detected using peroxidase labeled anti-mouse, anti-rabbit IgGs (GE Healthcare) or anti-rat IgG (Thermo Scientific) and developed using ECL Prime kit (GE Healthcare) according to manufactures protocol.
Trypomastigotes and exponentially growing epimastigotes were centrifuged, washed twice in PBS, settled on polylisine-coated coverslips and fixed with 4% para-formaldehyde in PBS at room temperature for 20 minutes. For the mitochondrial staining, parasites were resuspended in PBS and incubated with 1 μM MitoTracker (Invitrogen) for 30 minutes at 28°C, washed twice in PBS and fixed with 4% para-formaldehyde. Fixed parasites were washed with PBS and permeabilized with 0.1% Triton X-100 in PBS for 10 minutes. After washing with PBS, parasites were incubated with the appropriate primary antibody diluted in 5% BSA in PBS for 2 hours at room temperature. In colocalization experiments both antibodies were incubated together. Non-bound antibodies were washed with 0.01% Tween 20 in PBS and then the slides were incubated with fluorescent-conjugated anti-mouse (FITC, Jackson Immuno Research) or anti-rat (FITC, Life Technologies) and anti-rabbit (Cy3, Life Technologies) IgG antibodies and 2 μg/ml of DAPI for 1 hour. The slides were washed with 0.01% Tween 20 in PBS and finally mounted with VectaShield (Vector Laboratories). To analyze intracellular amastigotes, Vero cells monolayers were grown on coverslips and infected with T. cruzi trypomastigotes as described above. Two days post-infection cultures were washed with PBS and fixed with methanol at room temperature for 3 minutes. The same procedure described above was followed for immunodetection. Images were acquired with a confocal Nikon Eclipse TE-2000-E2 microscope using Nikon EZ-C1 Software. Adobe Photoshop CS and ImageJ software were used to process all images.
To determine the IC50 values of the sirtuin inhibitors, epimastigotes of T. cruzi Dm28c strain were cultured at 28°C in liver infusion tryptose medium (LIT) supplemented with 10% FCS in the absence or presence of Nicotinamide, Cambinol and Ex-527 (Sigma) at various concentrations, in triplicates. Cell growth was determined after culture for 72 hours by counting viable forms in an automatized hemocytometer adapted to count epimastigotes (WL 19 Counter AA, Weiner Lab). Then, Dm28c wt, Dm28c pTcINDEXGW-SIR2RP1HA and Dm28c pTcINDEXGW-SIR2RP3HA strains (uninduced and induced with 0.5 μg/ml Tetracycline), were cultured at 28°C in LIT with FCS in the absence or presence of the sirtuin inhibitors at concentrations above their IC50 values.
Experiments were performed in triplicate, and at least three independent experiments were performed. Data are presented as the mean ± SEM. Statistical analysis of the data was carried out using two-way ANOVA and unpaired Mann-Whitney, two-tailed Student t test. Differences between the experimental groups were considered significant as follows: p<0.05 (*), p<0.005 (**), p<0,001 (***) and p<0.0001 (****). To determine the IC50 values, we used nonlinear regression on Prism 6.0 GraphPad software. Student’s t test was applied to ascertain the statistical significance of the observed differences in the IC50 values.
Two protein coding sequences (TcCLB.507519.60 and TcCLB.506559.80) corresponding to Sir2 related proteins were identified in the T. cruzi genome, termed TcSIR2RP1 and TcSIR2RP3 respectively (http://www.tritrypdb.org/tritrypdb/). TcSir2rp1 and TcSir2rp3 encode proteins of 359 and 241 amino acids, with predicted molecular weights of ~ 39.6 and 26.8 kDa and pIs of 6.39 and 6.51, respectively. The alignment of T. cruzi sirtuins with human SIRTs and ScSir2 (S1 Fig) shows that although TcSIR2RPs lack the N-terminal portion, which is required for nucleolar localization in ScSir2, they contain a complete catalytic domain (Pfam: PF02146). SIR2RP1 contains a Serine-rich motif towards the C-terminus and one of the Cys residues from the zinc-binding motif (CX2CX20CX2C type) is absent in SIR2RP3. The GAD and NID motifs as well as other residues important for catalysis are conserved (HG, arrowheads in S1 Fig). The catalytic domain of TcSIR2RP1 and TcSIR2RP3 share a sequence identity/similarity of 17.5%/26.3% with each other and 23.1%/33.9% and 20.8%/ 33.7% with ScSIR2 respectively. Similarly to what Greiss and Gartner observed [24], TcSIR2RP1 grouped with the cytoplasmatic human SIRT2 whereas TcSIR2RP3 is more related to mitochondrial HsSIRT5 and bacterial sirtuins (S2 Fig). Despite the discrepancies observed for each sirtuin from different Tritryp species regarding their localization and function, they seem to be conserved at the sequence level.
In order to evaluate TcSIR2RP1 and TcSIR2RP3 expression in T. cruzi, antibodies were raised against the recombinant proteins and purified by affinity chromatography. After confirming the specificity of the antibodies (S3 Fig showed a single band of the expected molecular weights), they were used in Western blots to test total lysates of epimastigotes, amastigotes and trypomastigotes. As can be observed in Fig 1, the expression of sirtuins is developmentally regulated throughout T. cruzi life cycle. TcSIR2RP1 shows similar expression levels in epimastigotes and amastigotes, but lower in trypomastigotes. TcSIR2RP3 expression levels are higher in epimastigotes than in amastigotes and it is not detected in trypomastigotes under the conditions assayed.
Overexpression of TcSIR2RP1 and TcSIR2RP3 enzymes was performed using the T. cruzi inducible vector pTcINDEXGW [33]. Epimastigote cell lines expressing each sirtuin with a C-terminal HA tag under the control of a Tetracycline-regulated promoter were generated (Materials and Methods). The induction of the expression by Tetracycline was tested by western blot (Fig 2A and 2B) and immunofluorescense (Fig 2C). Western blot analysis of whole-cell extracts with rat monoclonal anti-HA antibodies revealed the expression of both constructs after the addition of Tetracycline, at their expected molecular weights. No leaky expression was observed in the uninduced parasite lines (Fig 2A). The western blots with the specific antibodies against TcSIR2RP1 and TcSIR2RP3 show a high degree of overexpression (20-fold) in the induced lines (Fig 2B). We also tested the inducible expression of the sirtuins in intracellular amastigotes and trypomastigotes by western blot of total lysates (Fig 3A) and immunofluorescense (Fig 3B) with anti-HA (quantification of results from Fig 3A are shown in S4 Fig). The tagged sirtuins are expressed only in the presence of Tetracycline, throughout T. cruzi life cycle.
Different technical approaches were performed to determine the localization of TcSIR2RPs, using the specific antibodies raised in rabbit against the two recombinant proteins and commercial anti-HA monoclonal antibodies. Confocal immunocolocalization microscopies performed with cytosolic (anti-TAT) [34] and mitochondrial (MitoTracker) markers, together with anti-TcSIR2RP1 and anti-TcSIR2RP3, showed that TcSIR2RP1 co-localizes with TAT (cytosol) and TcSIR2RP3 with MitoTracker (Fig 4A). In parallel, the subcellular distribution of tagged sirtuins was analyzed by immunofluorescense of induced epimastigotes of each cell line with cytosolic (anti-TAT) and mitochondrial (anti-MDHm) markers together with anti-HA (Fig 4B). As can be observed in Fig 4B, SIR2RP1-HA colocalized with TAT and SIR2RP3-HA with MDHm, supporting the results obtained with specific antibodies.
To study in further detail their localizations, we performed subcellular fractionation by differential centrifugation of the transfected lines. The fractions obtained were analyzed by western blot with different subcellular markers (cytosolic TAT, mitochondrial MDH and nuclear BDF2 [35]), and with anti-HA (Fig 4C). The nuclear marker was enriched in fraction N, the mitochondrial marker in fractions N and LG, and the cytosolic marker in fraction S, as reported by Opperdoes et al [32]. In agreement with our previous results, SIR2RP1-HA exhibits a cytosolic pattern while SIR2RP3-HA exhibits a mitochondrial one.
To test the deacetylation activity of TcSIR2RPs, we performed western blot of uninduced and induced total lysates of each cell line with anti-Acetyl-lysine antibodies. Fig 5A shows the amount of protein loaded for each condition. Overexpression of both sirtuins reduced the acetylation levels of specific proteins (Fig 5B). The differentially acetylated proteins are depicted with arrowheads. It is worth noticing that the deacetylated proteins are different for each overexpressed sirtuin.
Alpha-tubulin is one of the most abundant acetylated proteins in trypanosomes. In fact, the rate of acetylated/non-acetylated α-tubulin in trypanosomatids is higher than in other eukaryotic cells like yeast or mammalian cells [36–37]. To test if any of the TcSIR2RPs can deacetylate α-tubulin, we measured acetylated α–tubulin in uninduced and induced parasites by western blot analysis. When normalized to total α-tubulin, the results reflect significant diminutions of the acetylated form of the protein of 27% and 35% in TcSIR2RP1HA and TcSIR2RP3HA overexpressing lines respectively (Fig 5C). As already mentioned, deacetylation of α-tubulin mediated by SIR2RP1 was reported in Leishmania. However, even though our results are very confident, we consider that we cannot be conclusive enough to assign to TcSIR2RP1 nor to TcSIR2RP3 the function of being the T. cruzi tubulin deacetylase. The observed hypoacetylation could be an unspecific result due to the overexpression of a sirtuin. Furthermore, it has been demonstrated that in mammalian cells, tubulin is deacetylated not only by SIRT2, but also by the non-sirtuin deacetylase HDAC6 [38]. Since there are more than one deacetylases with similarity to HDAC6 in trypanosomatids, a deeper study of the whole set of deacetylase enzymes is needed to determine the most relevant tubulin deacetylase in this organism.
We monitored the effect of sirtuins overexpression on epimastigote growth by counting cell numbers daily after protein induction. Fig 6 shows that Dm28c TcSIR2RP1HA cell line grew at similar rates in the absence and presence of Tetracycline (which was re-added every 5 days), but those harboring TcSIR2RP3HA showed a delay in the growth rate when induced. Even more surprising is the fact that the TcSIR2RP3HA expressing line reached stationary phase at a smaller number of parasites per ml. This culture continues in a stationary phase for the same period of time as the uninduced line. This phenomenon needs to be further studied in order to be completely understood, and the existence of a quorum sensing mechanisms recently described in T. brucei opens a novel possibility of interpretation [39].
Sirtuins are considered as fundamental life sustaining biocatalysts and under various conditions were found to be pro-survival [40]. Therefore, there is an increasing interest in sirtuins as therapeutic targets. Several structures of complexes involving sirtuins and their inhibitors have been reported [17, 41]. In this work, we tested three sirtuin inhibitors: Nicotinamide (NAM) [42], Cambinol and Ex-527 [43]. All of them inhibited T. cruzi Dm28c epimastigotes growth in a concentration-dependent manner in axenic cultures. The IC50 values obtained for each inhibitor are shown in Table 1. Ex-527 IC50 (206.2 μM) is significantly higher compared to the other tested sirtuin inhibitors. Since, Ex-527 is a potent and selective HsSIRT1 inhibitor, with a reported IC50 of 0.1–1 μM [43], the low toxicity we observed might indicate that neither of the sirtuins present in T. cruzi is related to HsSIRT1, consistent with our phylogenetic analysis. On the contrary, NAM and Cambinol exhibited IC50 values similar to those obtained for purified sirtuins of other organisms such as: NAM IC50 for PfSIR2: 51.2 μM [44]; NAM IC50 for hSIRT5: 46.6 μM, hSIRT1: 50–100 μM, hSIRT3: 30 μM; Cambinol IC50 for hSIRT1: 56 μM, hSIRT2: 59 μM [45].
We considered the possibility that sirtuin-overexpressing lines might be “protected” or less sensitive to the sirtuin inhibitors. To test this hypothesis, both parasite lines were treated with the inhibitors, in the absence and presence of Tetracycline. As seen in Fig 7, overexpression of sirtuins protects epimastigotes from the growth inhibition of NAM and Cambinol. Moreover, the treatment with these inhibitors reversed the growth defect of the induced TcSIR2RP3HA cell line. In contrast, Ex-527-driven growth inhibition may be due to a pleiotropic effect and not to a reduction of the parasite’s sirtuin activities, since the inhibitory effect on trypanosomes is only observed at high concentrations. We also calculated the IC50 values of NAM and Cambinol for the overexpressing strains. The values obtained are shown in Table 1, as expected they are higher than for the wild type parasites.
In vitro metacyclic trypomastigotes were produced from epimastigotes using TAU medium, in the absence (-Tet) or presence (+Tet) of Tetracycline. TcSIR2RP1 overexpression resulted in a meaningful increase of metacyclogenesis (59%), whereas TcSIR2RP3 strain showed similar levels of metacyclic trypomastigotes formation in both the induced and uninduced conditions (Fig 8).
To study the importance of sirtuins expression in trypomastigotes´ infectivity and in the replicative form present inside the mammalian host, we investigated how the transgenic lines induced with Tetracycline performed in vitro for invasion and replication in host cells. First, we performed the experiment with Dm28c wild type parasites to rule out any undesired effect of the Tetracycline treatment. Indeed, there was no significant difference in the infectivity rate nor in the number of amastigotes/cell (S5 Fig). Trypomastigotes were pre-incubated in the presence or absence of 0.25 μg/ml Tetracycline and NAM (100 μM) and then used to infect Vero cells at a ratio of 10 parasites per cell. After 6 h of infection at 37°C, the free trypomastigotes were washed out and replaced by complete medium alone or with Tetracycline (0.25 μg/ml) for 2 days post-infection. Microscopic observation of Vero cells stained with Giemsa showed that treatment of uninduced T. cruzi trypomastigotes with NAM [(-/-, +NAM) vs (-/-)] caused a significant reduction in the percentage of infected cells (as previously reported by Soares and coworkers [42]) (Fig 9A). The overexpression of both sirtuins protected the trypomastigotes from the negative effect of NAM [(+/+, +NAM) vs (-/-, +NAM)]. These results suggest that sirtuin activity is necessary for an effective infection of mammalian cells. To analyze the effect of sirtuin overexpression on the infectivity rate of trypomastigotes, we focused on the condition in which the expression was induced only in the trypomastigote stage during infection (+/-). As can be seen in Fig 9A, overexpression of TcSIR2RP1HA increased the infectivity rate of trypomastigotes [(+/-) vs (-/-)], while overexpression of TcSIR2RP3HA slightly diminished it [(+/-) vs (-/-)]. To test the effect of sirtuins overexpression on the proliferation of intracellular amatigotes (Fig 9B), we only added Tetracycline after the infection, when the trypomastigotes were washed out, for 48 hours post-infection (-/+). The number of amastigotes per infected cell is slightly decreased by the overexpression of TcSIR2RP1HA, but increased by TcSIR2RP3HA [(-/+) vs (-/-)]. The reduction in the number of amastigotes per cell observed when expression of TcSIR2RP3 was induced during the infection [(+/-) vs (-/-)], might be a consequence of the diminished infectivity or it might indicate that the overexpression of this enzyme at the trypomastigote stage results in an inefficient differentiation to amastigotes, thus delaying the amastigotes´ replication. In contrast, when inducing at all times (+/+), the overexpression of amastigotes increases the proliferation rate hiding the trypomastigote to amastigote differentiation delay.
We present herein the first experimental characterization of Trypanosoma cruzi sirtuins TcSIR2RP1 and TcSIR2RP3. The expression of these enzymes is developmentally regulated throughout T. cruzi life cycle. TcSIR2RP1 is highly expressed in epimastigotes and amastigotes, but at lower levels in trypomastigotes. On the other hand, TcSIR2RP3 expression levels are higher in epimastigotes than in amastigotes, and it seems not to be expressed in trypomastigotes. The fact that the two sirtuins are differentially expressed along the life cycle of the parasite suggests that acetylation levels could play a role in the regulation of the biology of the parasite forms. It is remarkable that the life cycle pattern expression of Tetracycline induced/T7 transcribed sirtuins is similar to those of the wild type enzymes. This observation suggests that the protein levels could be regulated by a post-transcriptional mechanism independent of the 3´ and 5’ non-coding regions, which are absent in the pTc-INDEX-GW constructions or by a post-translational mechanism.
Our results clearly demonstrate that TcSIR2RP1 and TcSIR2RP3 are, respectively, cytoplasmic and mitochondrial enzymes. These observations were expected, since both proteins lack the N-terminal portion responsible for the nuclear localization of ScSir2 and other related sirtuins. However, the possibility that under certain conditions, the trypanosomal sirtuins could be imported temporarily to the nucleus by an alternative targeting pathway cannot be completely ruled out. Taken together, our data suggest that sirtuin activity is important for the proliferation of T. cruzi replicative forms, for the host cell-parasite interplay, and for differentiation among life-cycle stages; but each one performs different roles in most of these processes.
Considering its cellular localization, TcSIR2RP1 seems to be functionally more related to Leishmania than to T. brucei ortholog, even though TbSIR2RP1 is more similar at the sequence level (68% identity) than LmSIR2RP1 (55% identity). The fact that T. brucei has a nuclear sirtuin that is absent in the other TriTryps can be explained by some well-known differences existing among these species. TbSIR2RP1 participates at the epigenetic-mediated silencing of RNA polymerase I-transcribed telomeric regions, but nothing similar occurs in T. cruzi or in Leishmania. In Plasmodium falciparum, a nuclear sirtuin (PfSIR2A), phylogenetically unrelated to TbSIR2RP1, is also implicated in telomeric gene silencing. Taken together, these results suggest that the participation of a sirtuin in histone deacetylation represents the exception, associated to telomeric gene silencing, rather than the rule of the function of sirtuins in these organisms, and it could be an example of convergent evolution.
In spite of their cellular localization and way of action, sirtuins are considered pro-survival regulators of metabolism and lifespan. These general functions are related with the use of NAD+ as substrate, which together with acetyl-CoA, the acetyltransferases substrate, are considered sensors of the energetic state of the cell. Nuclear sirtuins, like HsSIRT1, regulate transcriptional response to starvation or redox stress and under certain conditions, cytoplasmic and mitochondrial sirtuins are imported to the nucleus with the same purpose (Reviewed in [46]). Since transcriptional regulation is absent in T. cruzi, it is reasonable to think that only non-nuclear functions will be found for these enzymes. One of the functions of cytoplasmic HsSIRT2 is to deacetylate the enzyme phosphoenolpyruvate carboxykinase (PEPCK), increasing its stability, upon glucose deprivation [47]. Under caloric restriction, human mitochondrial sirtuins may also regulate gluconeogenesis from amino acids through glutamate dehydrogenase (GDH), an enzyme that converts glutamate to α-ketoglutarate, thereby controlling glucose production via the TCA cycle [48–50]. HsSIRT3 and HsSIRT4 modulate the activity of GDH through deacetylation and ADP-ribosylation, respectively (even though the existence of gluconeogenesis was only proved in amastigotes from Leishmania, it was already proposed that this pathway should be present in all trypanosmatids [51]). Human SIRT3 also decreases reactive oxygen species (ROS) production by stimulating superoxide dismutase 2 (SOD2), and enhances cellular respiration by increasing the activities of complex I, complex II (via succinate dehydrogenase (SDH)), complex III and isocitrate dehydrogenase 2 (IDH2), affecting both glucose and lipid metabolism. Finally, mitochondrial HsSIRT3 stimulates β-oxidation and ketone body formation by targeting and activating long-chain acyl CoA dehydrogenase (LCAD) and 3-hydroxy-3-methylglutaryl-CoA synthase 2 (HMGCS2), respectively [52–53]. Although more specific research is needed to determine whether trypanosome sirtuins share these functions, some of our results can be interpreted under this general rationale: TcSIR2RP1 improves metacyclogenesis, a differentiation process induced by starvation of the parasites [54–55]. The increased infectivity of this strain resembles that observed for LmSIR2RP1 by Sereno and coworkers [56]. As this sirtuin was detected in the excreted/secreted material of parasites, they constructed LmSIR2RP1-overexpressing fibroblasts, which were more permissive towards Leishmania invasion than control ones. These results suggest a function for parasite cytoplasmic sirtuins at the host-parasite interplay. However, the secretion of TcSIR2RP1 is yet to be confirmed. In contrast, TcSIR2RP3 seems not to be implicated in metacyclogenesis or cell infection but improves the replication of intracellular amastigotes, which occurs in an oxidant environment, suggesting the participation of this enzyme in redox stress response.
The activity of sirtuins in protecting cells from stress and starvation and so improving life span, led to the establishment of their role in cell proliferation. Currently, many sirtuin inhibitors are being assayed against different types of cancer. Conversely, sirtuin activators are considered as potential anti-aging drugs. It has also been proposed that sirtuins are promising targets for the development of anti-trypanosomal, anti-plasmodial and anti-leishmanial drugs and many well-characterized sirtuin inhibitors have already shown anti-parasitic activity [21–22, 40, 57]. In addition, an in silico structural and surface analysis of trypanosomal and human sirtuins determined potentially important structural differences in the corresponding inhibitor binding domains, indicating a possible selectivity of an inhibitor for a specific protein [58]. In a very recent in silico study, Sacconnay and coworkers [59] assayed a list of 50 phytochemicals previously described as anti-trypanosomals by docking into TcSIR2RP1 and TcSIRRP3, revealing that the activity of four of these compounds could be explained by the inhibition of the sirtuins activity. The results presented herein, contribute to the knowledge of these enzymes localization and function, and our TcSIR2RP1HA and TcSIR2RP3HA overexpressing T. cruzi strains can be useful tools for experimental screening of trypanosomatid sirtuin inhibitors.
|
10.1371/journal.pcbi.1000133 | Stroke Rehabilitation Reaches a Threshold | Motor training with the upper limb affected by stroke partially reverses the loss of cortical representation after lesion and has been proposed to increase spontaneous arm use. Moreover, repeated attempts to use the affected hand in daily activities create a form of practice that can potentially lead to further improvement in motor performance. We thus hypothesized that if motor retraining after stroke increases spontaneous arm use sufficiently, then the patient will enter a virtuous circle in which spontaneous arm use and motor performance reinforce each other. In contrast, if the dose of therapy is not sufficient to bring spontaneous use above threshold, then performance will not increase and the patient will further develop compensatory strategies with the less affected hand. To refine this hypothesis, we developed a computational model of bilateral hand use in arm reaching to study the interactions between adaptive decision making and motor relearning after motor cortex lesion. The model contains a left and a right motor cortex, each controlling the opposite arm, and a single action choice module. The action choice module learns, via reinforcement learning, the value of using each arm for reaching in specific directions. Each motor cortex uses a neural population code to specify the initial direction along which the contralateral hand moves towards a target. The motor cortex learns to minimize directional errors and to maximize neuronal activity for each movement. The derived learning rule accounts for the reversal of the loss of cortical representation after rehabilitation and the increase of this loss after stroke with insufficient rehabilitation. Further, our model exhibits nonlinear and bistable behavior: if natural recovery, motor training, or both, brings performance above a certain threshold, then training can be stopped, as the repeated spontaneous arm use provides a form of motor learning that further bootstraps performance and spontaneous use. Below this threshold, motor training is “in vain”: there is little spontaneous arm use after training, the model exhibits learned nonuse, and compensatory movements with the less affected hand are reinforced. By exploring the nonlinear dynamics of stroke recovery using a biologically plausible neural model that accounts for reversal of the loss of motor cortex representation following rehabilitation or the lack thereof, respectively, we can explain previously hard to reconcile data on spontaneous arm use in stroke recovery. Further, our threshold prediction could be tested with an adaptive train–wait–train paradigm: if spontaneous arm use has increased in the “wait” period, then the threshold has been reached, and rehabilitation can be stopped. If spontaneous arm use is still low or has decreased, then another bout of rehabilitation is to be provided.
| Stroke often leaves patients with predominantly unilateral functional limitations of the arm and hand. Although recovery of function after stroke is often achieved by compensatory use of the less affected limb, improving use of the more affected limb has been associated with increased quality of life. Here, we developed a biologically plausible model of bilateral reaching movements to investigate the mechanisms and conditions leading to effective rehabilitation. Our motor cortex model accounts for the experimental observation that motor training can reverse the loss of cortical representation due to lesion. Further, our model predicts that if spontaneous arm use is above a certain threshold, then training can be stopped, as the repeated spontaneous use provides a form of motor learning that further improves performance and spontaneous use. Below this threshold, training is “in vain,” and compensatory movements with the less affected hand are reinforced. Our model is a first step in the development of adaptive and cost-effective rehabilitation methods tailored to individuals poststroke.
| Stroke is the leading cause of disability in the US, and about 65% of stroke survivors experience long-term upper extremity functional limitations [1]. Although patients may regain some motor functions in the months following stroke due to spontaneous recovery, stroke often leaves patients with predominantly unilateral motor impairments. Indeed, recovery of upper extremity function in more than half of patients after stroke with severe paresis is achieved solely by compensatory use of the less-affected limb [2]. Improving use of the more affected arm is important however, because difficulty to use this arm in daily tasks has been associated with reduced quality of life [3].
There is now definite evidence however that physical therapy interventions targeted at the more affected arm can improve both the amount of spontaneous arm use and arm and hand function after stroke [4]. Further, even after motor retraining is terminated, performance can further improve in patients with less severe strokes in the months following therapy [5],[6]. A possible interpretation of this result is that the repeated attempts to use the affected arm in daily activities are a form of motor practice that can lead to further improvements in motor performance [5].
The neural correlates of motor training after stroke have been investigated in animals with motor cortex lesions [7],[8]. Specifically, a focal infarct within the hand region of the primary motor cortex causes a loss of hand representations that extends beyond the infarction. However, several weeks of rehabilitative training can overcome this loss of representation, and yield an expansion of the hand area to its prelesion size; the larger area in turn has been correlated with higher level of performance [9]. Long-term potentiation in pyramidal neuron to pyramidal neuron synapses has been demonstrated in horizontal lateral connections [10], and may provide the basis for map formation and reorganization in the motor cortex [11], and motor skill learning [10].
Contrasting with the increase in performance due to spontaneous recovery, a concurrent decrease of spontaneous arm use has been proposed to occur following stroke. This decrease may be due both to the higher effort and attention required for successful use of the impaired hand and to the development of learned nonuse [12], in that the preference for the less affected arm is learned as a result of unsuccessful repeated attempts in using the affected arm [13]–[15]. The constraint-induced therapy (CIT) protocol, which forces the use of the affected limb by restraining the use of the less affected limb with a mitt, has been specifically developed to reverse learned nonuse [16]. Although its “active ingredients” are still not well understood [17], CIT has been shown to be effective in the recovery of arm and hand functions after stroke in multisite randomized clinical trials [4]. Because 50% of the eventual improvement in use (as measured by the questionnaire-based “motor activity log”) is seen at the end of the first day of CIT, it has been suggested that CIT is effective in reversing learned nonuse [18]. To our knowledge, however, there are no longitudinal data tracking the development of learned nonuse just after stroke and during recovery.
In summary, increase in performance after stroke due to spontaneous recovery, rehabilitation, or both does not appear to correlate simply with spontaneous arm use, and a yet-to-be clarified nonlinear mechanism seems to be at play. Here, we focus on rehabilitation in the control of reaching poststroke, a prerequisite for successful manipulation. We developed a biologically plausible model of bilateral control of reaching movements to investigate the mechanisms and conditions leading to such positive or negative changes in spontaneous choice of which arm to use. Our central hypothesis, based on the above observations, is the existence of a threshold in spontaneous arm use: if retraining after brain lesion (or spontaneous recovery) increases spontaneous arm use above this threshold, performance will keep increasing, as each attempt to use the affected arm will act as a form of motor relearning. The patient will then enter a virtuous circle of improved performance and spontaneous use of the affected arm, and therapy can be terminated. In contrast, if spontaneous use of the arm does not reach this threshold after either natural recovery or rehabilitation, or both, performance will not improve after stroke, and compensatory strategies with greater reliance on the less affected arm will either remain or even develop further.
To model spontaneous use of one arm or the other, and changes in motor performance, we simulated horizontal reaching movements towards targets distributed along a circle centered on the initial (overlapping) positions of the two arms (Figure 1A). Our computational model of bilateral arm use in arm reaching contains a left and a right motor cortex, and a single action choice module (Figure 1B). We first trained the full model (the “normal subject”) to reach with either hand, but with a bias for using the hand closer to the eventual target. Spontaneous arm use was recorded in a free choice condition, in which the action choice module can select either arm to reach targets that are randomly generated anywhere along the circle. Motor performance was evaluated by the directional error between the desired movement direction and the actual hand direction.
To simulate stroke, we partly lesion one hemisphere (i.e., remove a set of simulated neurons from the simulation). We first simulate a spontaneous recovery period in which the action choice module determines the choice of arm, and the state of motor cortex determines error in reaching, with consequent changes in synaptic weights. We then mimic CIT with a forced use condition in which only the use of the affected arm (i.e., that contralateral to the lesioned cortex) was allowed. We study in simulations the conditions that lead to successful recovery, that is, to high levels of spontaneous use and performance with the affected arm in appropriate regions of space, and low reliance on compensatory movements with the less affected arm.
Our model has two distributed interacting and adaptive systems: the motor cortex for motor execution and the action choice module for decision-making.
Strokes seem to affect only a certain range of movement directions. Outside this range, reaching is relatively spared [45]. To model this effect, we removed the neurons with preferred directions in the first quadrant of the left motor cortex (50% of the neural population coding for the right hand side workspace, as shown in Figure 2A.5), which controls the right arm (unless otherwise noted). The results would be the same had we chosen the other arm, or any other quadrant. We also tested stroke models in which neurons were affected probabilistically as a function of the range angle (with neurons being removed with 100% probability for the central angle of the simulated lesion and then with lower probability as the angles on each side of the lesion center increase); simulation results with these stroke models were qualitatively similar to those with the “hard boundary” model and thus for simplicity are not presented here. We also tested different stroke patterns, including a lesion ranging from 45° to 145°, and lesions with asymmetric bimodal distributions. Simulations (results not shown) confirmed that such lesions did not produce results qualitatively different from those presented here.
We used two measures of motor performance:
We chose these two performance measures in our model because they can be linked to actual patient performance measures. Initial directional error has been used in characterizing reaching in stroke patients (e.g., [46]). Although the population vector is normally not directly observable in patients, it can be regarded as a measure of force exerted by arm muscles on the hand [26],[47],[48], and low force generation is a characteristics of stroke [49]. Because both use and performance are stochastic, we report averages of 10 uniformly distributed samples over the affected range in all graphs (except the pie charts of Figures 2, 8, and 9).
The changes in performance and spontaneous arm use of the affected arm were recorded in four consecutive phases: (i) an acquisition phase of normal bilateral reaching behavior in 2,000 free choice trials (partially shown), (ii) an acute stroke phase of 500 free choice trials, (iii) a rehabilitation phase in a forced use condition (variable number of trials), and (iv) a chronic stroke phase consisting of 3,000 free choice trials. Values of performance and spontaneous use just after rehabilitation are called “immediate;” their long-term values at the end of the chronic phase are called “follow-up.”
In all phases, targets were randomly generated at the start of each trial, distributed uniformly across all possible angles. Unless otherwise stated, we used the following parameters: Each motor cortex had 500 neurons, with initial preferred directions θp uniformly distributed. The coefficient of variation of the signal-dependent noise ratio k was 0.15. The motor cortex learning rates were αSL = 0.005 and αUL = 0.002. The action choice module contained two networks of 20 radial basis function neurons with σreward = 0.2 (in radians, ≈11.46°), ρ = 0.2, σACM = π/10 (in radians, = 18°), αACM = 0.1, and β = 10.
The first (prelesion) phase provided a normal baseline for reaching behavior. For each desired direction, learning achieved zero mean directional error (Figure 2A.1) and a tendency of right arm use for the right-hand-side workspace, and left arm use for the left-hand-side workspace (Figure 2B.1).
Just after stroke, however, the population vectors showed directional errors in and around the affected range (Figure 2A.2). Sufficient therapy (1,000 forced use trials, Figure 2A.4) resulted in redistributing the preferred directions within the affected side of motor cortex, with the population vectors realigned to the desired directions. Although the realignment was not perfect, and a small range of preferred directions was still missing, the directional errors were much reduced. This resulted in increased rewards in these directions, thus increasing the action value for the affected arm, preparing the way for increased use of the affected arm once free choice was allowed. Lack of therapy on the contrary resulted in a still large missing range of directions (Figure 2A.3).
At the end of the “acute stroke” period, the less affected arm largely compensated for the more affected arm in the affected range (Figure 2B.2). If no therapy followed, then this behavioral compensation remained (Figure 2B.3). Sufficient therapy, however, led on the resumption of free choice trials to increased spontaneous arm use of the more affected arm (right arm) in the affected range (Figure 2B.4) and almost restored it to its prestroke levels.
We then studied the time courses of motor performance measures and spontaneous arm use (Figure 3). In the acute stroke phase, the free choice condition resulted in some spontaneous recovery in performance, as the repeated attempts to use the arm, although generated with poor performance, produced directional errors that retuned the motor cortex. However, the poor performance of these initial repeated attempts to use the affected arm caused a decrease in the action value for this arm in the affected directions, leading in turn to a reduction in spontaneous arm use. Thus, a “learned nonuse” phenomenon occurred despite improving performance. After 500 trials of natural recovery, a number of rehabilitation trials were given in the forced use condition. Rehabilitation improved performance as expected, but its lasting effects on spontaneous arm choice depended on the intensity of therapy. The increase in spontaneous arm use returned close to 0% soon after the end of therapy if only 200 trials of therapy were given. If 400 trials of therapy were given, spontaneous arm use held steady after therapy. If more therapy was given, spontaneous arm use was high after therapy and kept improving for a large number of trials thereafter.
The model thus exhibits a threshold for the intensity of rehabilitation. To precisely quantify the threshold, we computed the change in spontaneous arm use following rehabilitation by fitting a simple linear model with trials post stroke as predictor; the number of trials corresponding to a null slope corresponds to this threshold. As shown in Figure 4, with the default parameter set, there was a threshold at 420 trials of forced used trials, above which spontaneous arm use increased even after therapy was discontinued. Below this number of forced used trials, spontaneous arm use decreased to minimal levels after rehabilitation—it was “in vain.” The zero crossing in the slope in Figure 4 implies bistability of spontaneous arm use: when the number of rehabilitation trials is larger than the number of trials required to reach the threshold (420 trials), the spontaneous arm use improves in the following free choice condition until it saturates; conversely, when the number of therapy rehabilitation is less than the number of trials required to reach the threshold, the spontaneous arm use deteriorates (Figure 5C). Similar bistability is also shown in the directional error (Figure 5A) and normalized population vector (Figure 5B).
As expected, the minimal intensity of effective therapy depends on lesion size (Figure 6A). Compared to smaller lesions, large lesions require longer rehabilitation sessions to reach the threshold of spontaneous arm use above which therapy can be terminated. In our model, although directional error recovered almost perfectly for lesions sizes smaller than 50% for the right hand side workspace (follow-up test after 800 rehabilitation trials; results not shown), the long-term normalized population vector correlates almost linearly to the lesion size (same simulations conditions, see Figure 6B).
Motor performance can be judged according to two different criteria: accuracy (low bias of error) and precision (low variance of error). Figure 7 shows the effects of stroke and therapy, or the lack of it (‘no therapy’), on the accuracy and precision of the reach directional error over the affected range for the affected arm (contralateral to the lesion, Figure 7A) and for the nonaffected arm (ipsilateral to the lesion; Figure 7B). Although, stroke leads to an immediate and large deterioration of accuracy and precision for reaching movements with the affected arm (Figure 7A, thick solid line), therapy restores accuracy to near prestroke level (Figures 7A, dotted line). Because the number of available neurons is reduced after stroke, however, precision remains low after therapy compared to prestroke levels (Figure 7A). Lack of therapy (‘no therapy’ in Figure 7A, thin solid line) results in further deterioration of accuracy and precision for the affected (right) arm after stroke. In contrast, while stroke and therapy have almost no effect on performance of the nonaffected arm in our model (Figure 7B, dotted line), the increased frequency of compensatory reaching movements in the no therapy condition results in an increase of accuracy on these reaching movements (Figure 7B, thin solid line).
We then studied the organization and reorganization of the cells' preferred directions in each hemisphere before lesion, after lesion, and after therapy. Using pie histograms (Figure 8) which show the number of neurons whose preferred directions are in a certain range of directions, we observed a cortical reorganization pattern similar to that observed in animals that undergo rehabilitation or not after motor cortex lesions (see Discussion). Before lesion, more cells coded for the movements that were more often performed. After lesion, therapy or the lack of it affects the reorganization of neurons' preferred directions in both hemispheres.
Motor training with the affected arm has a profound effect on reorganization in the affected hemisphere. After sufficient therapy, the distribution of the surviving cells' preferred directions is similar to the prelesion distribution, with, however, fewer cells coding each direction, because the total number of cells is reduced (Figure 8A.4). During therapy, the directional error decreases, ensuring concordance of the supervised and unsupervised learning rules; the unsupervised learning rule is “adaptive” as it reinforces the supervised learning rule (Figure 8A.4). Conversely, motor training has almost no effect on the cell population of the nonaffected arm (Figure 8B.4).
Two patterns of reorganization are noteworthy in the affected hemisphere. First, the size of the affected range increased compared to just after the lesion; second, a large number of cells now code for movements in the fourth quadrant. If no therapy or insufficient therapy is provided, the directional error of the affected arm does not decrease (Figures 3A and 7A). This results in discordance between the supervised and unsupervised learning rules, and the unsupervised learning rule, based on desired but not actual directions, becomes “maladaptive,” further increasing the lesion size (Figure 8A.3) and largely increasing the representation of compensatory movements (Figure 8B.3) whose performance improves (decrease both in directional error bias and in directional error variability, and increase in normalized population vector). In the nonaffected hemisphere, a number of cells shift their preferred directions to the first quadrant, because the nonaffected arm must now compensate for the movements previously performed by the affected arm (Figure 8B.3).
Without the unsupervised learning term, reorganization follows different patterns: Therapy has less of an effect on reorganization, and lack of therapy does not lead to overrepresentation of compensatory movements in the affected hemisphere or in the nonaffected hemisphere (Figure 9).
To better understand the respective roles of each of the supervised, unsupervised, and reinforcement learning rates on behavior we then performed a sensitivity analysis for these three parameters on directional error for different durations of therapy (200, 400 and 800 therapy trials) followed by 3,000 free choice condition. As shown in Figure 10A, directional error decreased as the supervised learning rate increased for any amount of therapy. Figure 10B shows, however, a more complex pattern for the unsupervised learning rate. For a number of rehabilitation trials sufficient to reach threshold in the default parameter set (420 therapy trials on the threshold with 0.002 for the unsupervised learning rate), there is an optimal unsupervised learning rate for which long-term performance (after 3,000 free choice trials) is enhanced compared to either zero unsupervised learning or too large unsupervised learning. Thus, for appropriate learning rates, unsupervised learning is “adaptive,” as it enhances performance. No unsupervised learning or too large unsupervised learning rates are detrimental to performance however. A similar pattern is shown for the reinforcement learning rate, although the interpretation is more arduous as very little spontaneous use occurs with a reinforcement learning rate set at 0 (to perform the sensitivity analysis for the reinforcement learning rate, we used the default parameter set until the end of the acute-stroke phase, then the different reinforcement learning rates were tested starting with therapy condition).
We further studied the conditions under which the threshold appears by setting each of the three rates to 0 and keeping the other two to the default values. With such learning rate settings, we plotted the directional error, normalized population vector, and spontaneous hand use (Figure S1, Figure S2, and Figure S3) just after therapy and 3,000 trials after therapy as a function of the number of rehabilitation trials, as in Figure 5. Unlike for the full default parameter set (Figure 5), if one of the learning rates is set to zero, the bistable behavior disappears, as shown by the noncrossing of the curves for 0 (immediate test) and 3,000 free choice trials (follow-up test). In other words, the threshold observed in the complete model is an emergent property of the three types of learning. If supervised learning or reinforcement learning is not present, directional error worsens after 3,000 free choice trials compared to just after rehabilitation, for any number of rehabilitation trials. If unsupervised learning is not present, however, directional error improves after 3,000 free choice trials for any amount of rehabilitation trials.
We proposed a novel model of bilateral reaching that links different levels of analysis, as it combines a simplified but biologically plausible neural model of the motor cortex, a biologically plausible (but nonneural) model of reward-based decision-making, and physical therapy intervention at the behavioral level. Because our model is based on sound theoretical principles and neural mechanisms, it allows us to explore the nonlinear interactions between performance and spontaneous use in stroke recovery.
Our motor cortex model, by learning to minimize both directional errors and variability, accounts for the reversal of the loss of cortical representation after rehabilitation, and the increase of this loss together with the increase of the representation of neighboring areas without rehabilitation [7],[50].
In the lesioned cortex, during therapy, the supervised learning rule ensures that underrepresented directions are “repopulated,” decreasing average reaching errors. However, because there are fewer surviving neurons overall after stroke, stroke leads to a decrease in population vector magnitude (Figure 3B) and increased movement variability (Figure 7A)—as previously shown in [21]. The supervised learning component of our rule is consistent with monkey data showing that learning new skills, but not repetitive use, leads to motor cortical reorganization [51]. Supervised learning-like plasticity has not been reported in the cerebral cortex however, but it is thought to occur in the cerebellum [52]. A possibility is that the reduction of error due to rehabilitation, and the associated cortical reorganization, is driven by important cerebellar projections to the motor cortex. Lesion of the error signal driving cerebellar learning, presumably carried by the inferior olive [53], could be performed in animal models of stroke to test this possibility.
During therapy, the unsupervised learning rule is “adaptive” as its effect reinforces that of the supervised learning rule (compare Figures 8A.4 and 9A.4). By recruiting a greater number of neurons for often-performed actions it can counter neuronal noise and decrease directional error [21]; it is thus an adaptive process in the normal brain. After stroke, however, such unsupervised plasticity may become maladaptive. A comparison of Figures 8A.3 and 9A.3 shows that unsupervised learning further augments the effect of stroke if no therapy is given. As compensatory movements, or movements unaffected by the stroke, compete for the surviving neurons, fewer neurons code for directions around the affected area (Figure 8A.3), leading to further deterioration of performance (Figures 3A, 3B, and 7A). The representation of compensatory movements is increased and performance of these movements improves (Figure 7, decreased directional error bias). Without the unsupervised learning term, reorganization follows different patterns: Therapy has less of an effect on reorganization, and lack of therapy does not lead to overrepresentation of compensatory movements in the affected hemisphere or in the nonaffected hemisphere (Figure 8).
To our knowledge, the present computational neural model is the first developed to make specific behavioral and neuronal predictions on the efficacy of physical therapy interventions. Two previous models have been developed to account for behavior after stroke [21],[54], but these models do not address plastic changes. The model by Goodall et al. [50] predicts that focal lesions result in a two-phase map reorganization process in the intact peri-lesion cortical region, but this model does not account for the development of compensatory movements and reorganization of choice after training.
Our model is in accord with the most recent understanding and comprehensive view of the basal ganglia function in adaptive selection of alternative actions [40],[55],[56] via release of inhibition of motor cortex activity [42]. A different decision making mechanisms was however recently proposed by Cisek [57], who analyzed the time course of cortical activation before and after decision to reach one of two targets with a single arm. Unlike in our model, target choice was resolved in a distributed manner, by competition between neurons within cortical layers. Further experiments are needed to study how targets are selected when both limbs can be used, and how this selection is reorganized after lesion and therapy.
In a recent motor cortex model [58], as in our model, reorganization of preferred directions is due to a learning rule containing two terms: a supervised error correcting term, and a (unsupervised) weight decay term. Because our unsupervised learning rule is based on the activation of neighboring neurons however, it explains maladaptation and increase of lesion size in the no-therapy condition (Figure 8A.3). Furthermore, the sensitivity analysis of the three learning rates (supervised, unsupervised and reinforcement learning, Figure 10) showed that the bistability of performance and spontaneous arm use (Figures 4 and 5) requires the combination of all three types of learning (Figures S1B, S2B, and S3B)
Because of its simplicity, our model provides clear insights into a range of factors affecting recovery of arm use after stroke. However, our model does suffer from a number of limitations:
To resolve the limitations, in the future we will expand our model by adding arm and muscle models controlled by neurons grouped in adaptive motor cortical maps. We plan to investigate the tradeoff between proximal and distal regions, with cortical motor maps that change during training on tasks that require more skilled use of the hand itself. Moreover, the notion that the action choice model may correspond to the basal ganglia opens up promising lines of investigation.
In summary, despite our considerable simplifications of movement representation in the motor cortex and of the simulated lesions, our results show that our proposed mechanism of motor learning and plasticity, and the ensuing results (recovery, threshold, and neural reorganization) are general and not particular to the specifics of our model.
Our model makes the following testable behavioral and neural predictions.
In our model, neural reorganization generates bistability at the behavioral level: after therapy, spontaneous arm use will stabilize at either a low or a high value, depending on the amount of therapy. Specifically, therapy is effective and could be stopped if spontaneous arm use reaches a certain threshold, as the repeated spontaneous arm use following therapy provides a form of motor learning that further “bootstraps” performance. Below this threshold, however, motor retraining is “in vain”—there is no or little long-term spontaneous arm use after training, and the model exhibit “learned nonuse,” as has been proposed in patients with brain lesions [13].
We thus predict that a measure of spontaneous arm use may be a good indicator to determine optimal duration of the therapy. In current rehabilitation practice, all rehabilitation is concentrated in the weeks following stroke. Our model suggests that rehabilitation protocols adopt instead a spaced and adaptive train–Test A–wait–test B–train paradigm: short bouts of training (train) are followed by a spontaneous arm use test (Test A), no training for several weeks (wait), and another spontaneous arm use tests (Test B). If spontaneous arm use measured on Test B has increased since that on test Test A, the threshold is reached, and rehabilitation can be terminated. If spontaneous arm use is still low or has decreased since Test A, another bout of rehabilitation is called for. This pattern is repeated until the threshold is reached. Note that such a training paradigm will have the additional benefit of making use of the “spacing effect,” in which spaced training lead to superior retention of learned skills [63]. We plan to put this hypothesis to empirical test using a novel laboratory-based objective test of bilateral limb use.
|
10.1371/journal.pgen.1000917 | Hypomethylation of a LINE-1 Promoter Activates an Alternate Transcript of the MET Oncogene in Bladders with Cancer | It was recently shown that a large portion of the human transcriptome can originate from within repetitive elements, leading to ectopic expression of protein-coding genes. However the mechanism of transcriptional activation of repetitive elements has not been definitively elucidated. For the first time, we directly demonstrate that hypomethylation of retrotransposons can cause altered gene expression in humans. We also reveal that active LINE-1s switch from a tetranucleosome to dinucleosome structure, acquiring H2A.Z- and nucleosome-free regions upstream of TSSs, previously shown only at active single-copy genes. Hypomethylation of a specific LINE-1 promoter was also found to induce an alternate transcript of the MET oncogene in bladder tumors and across the entire urothelium of tumor-bearing bladders. These data show that, in addition to contributing to chromosomal instability, hypomethylation of LINE-1s can alter the functional transcriptome and plays a role not only in human disease but also in disease predisposition.
| A surprisingly large portion of our transcriptome originates within repetitive elements, most commonly LINE-1s. However, the mechanism of activation has not been definitively shown. We directly demonstrate for the first time the causal relationship between DNA hypomethylation and transcriptional activation of LINE-1 promoters. Hypomethylation of specific LINE-1 promoters can alter the transcriptome, including activating an alternate transcript of the MET oncogene, not only in primary bladder tumors but also in premalignant urothelium across entire bladders with tumors. Our study has important implications for tumor biology, cancer detection, and treatment, and it also answers the long-standing question of whether hypomethylation of retrotransposons induces ectopic gene expression and influences disease susceptibility in humans, a phenomenon first described in agouti mice.
| Aberrant DNA methylation is involved in the initiation and progression of carcinogenesis and includes both hypermethylation of CpG islands at gene promoters and global hypomethylation. While a small portion of hypomethylation occurs at gene promoters, resulting in overexpression of certain oncogenes [1], [2], the majority occurs at repetitive elements, such as long interspersed nuclear elements (LINE-1s or L1s) [3]. Since most of the 500,000 copies of L1 have become nonfunctional over the course of human evolution [4] and can no longer transpose, genome-wide hypomethylation at L1s during tumorigenesis is thought to contribute mainly to chromosomal instability [5]. In mice hypomethylation of transposable elements can lead to disruption of normal gene function [6]. Viable yellow agouti (Avy) mice have a retrotransposon inserted into one allele of the agouti locus and when this retrotransposon is hypomethylated, which can occur in utero by limiting the maternal intake of methyl donors, it acts as an alternate promoter for agouti. Ectopic induction of the agouti gene results in altered coat color, obesity, and an increased incidence of tumors [6]. While it is well known that repetitive elements are hypomethylated in cancer, it has never been directly demonstrated that hypomethylation of a retrotransposon leads to ectopic gene expression in humans.
A recent study has revealed that more than 30% of transcription start sites in the human genome are located within repetitive elements, with just over 7% in L1s [7]. A full length L1 sequence (6 Kb) has a sense promoter driving transcription of its two open reading frames and an antisense promoter driving transcription in the opposite direction that can act as an alternate promoter for surrounding genes [8]–[10]. Almost 500 of these retrotransposons can induce ectopic gene expression in embryonic and cancerous tissues, revealing their potential role during both development and tumorigenesis [7]. However this study did not address the potential mechanism of how repetitive elements become transcriptionally active. Since the L1 promoter is a CpG island and methylated in normal somatic tissues it seems likely that epigenetic mechanisms are involved in its transcriptional silencing. There are many layers of epigenetic regulation responsible for regulating expression of single copy genes, including DNA methylation, histone modifications, and nucleosome occupancy [11]. While it is known that unmethylated retrotransposons in Arabidopsis [12] acquire the active histone variant H2A.Z, the chromatin structure in humans of repetitive elements, particularly active ones, has been largely ignored.
Until recently it has not been possible to study the promoters of individual L1s since the sequences are too similar to design primers for one particular locus [13]–[15]. Therefore a direct correlation between the epigenetic status of a specific L1 and expression of its associated transcript has not been possible. For the first time to our knowledge, we have elucidated the role of epigenetics in the transcriptional activity of L1s by utilizing novel assays capable of examining the methylation status and chromatin structure of specific L1s and expression of alternate transcripts originating from the L1 promoters. In addition to L1s being hypomethylated and transcriptionally active in bladder tumors we also found that a specific L1 located within the MET oncogene is active across entire bladders with cancer. The clinical implication of our finding is that surgical excision of the tumor would leave behind large areas of the bladder that remain epigenetically altered and express a potential oncogene. We also provide evidence that an active L1 acquires H2A.Z and nucleosome free regions upstream of TSSs, which has only been described previously at single copy genes, and undergoes chromatin remodeling from an inactive tetranucleosomal structure to an active dinucleosomal structure.
To elucidate the mechanism of transcriptional activation of repetitive elements we used the sequence of the functional promoter of L1s to identify specific promoters potentially capable of expressing alternate transcripts of host genes. Figure S1 contains the genomic locations of the L1s, all of which are in an antisense orientation to the host gene allowing for transcripts in sense orientation to the gene's coding sequence. Interestingly, most these ESTs are from tumor cells. One such L1 is located within the MET oncogene (L1-MET) [8]. Since MET is known to be overexpressed in bladder cancer [16]–[18], we characterized two L1-MET transcripts by sequencing EST clones obtained from a bladder carcinoma cell line (GenBank accession no. BF208095) and placenta (BX334980). Both transcripts have start sites located in the L1 promoter, share the same reading frame as MET (Figure S2A), and when transiently transfected into Hela cells result in expression of truncated MET proteins (Figure S2B). Several truncated forms of the tyrosine kinase MET, which is the hepatocyte growth factor (HGF) receptor, are constitutively active and promote invasion and migration through activation of a variety of signal transduction pathways in numerous types of carcinomas, including breast, prostate, colorectal, and lung, in musculoskeletal sarcomas, and also in haematopoietic malignancies [19], [20]. Therefore hypomethylation of L1-MET could lead to expression of a transcript that encodes a truncated and potentially constitutively active MET protein.
To examine the methylation status at a specific L1 we designed bisulfite-specific PCR primers with one located in the L1 promoter and the other in the surrounding intronic region of the host gene (Figure 1A). The L1-MET promoter was highly methylated in normal cells and tissues, whereas 18 out of 20 of the bladder carcinoma cell lines showed significant hypomethylation (p<3.4×10−10) (Figure 1B). We also measured methylation of global L1s using the standard assay with two primers that anneal within the L1 promoter (Figure 1A). We found that hypomethylation of L1s was significant (p<6.4×10−5) but not as dramatic as L1-MET hypomethylation and that the methylation pattern can be quite different between global L1s and a specific L1, such as in the cell lines LD137, T24, and RT4 (Figure 1B). This result clearly shows that global L1 status does not represent the status at specific L1s.
The transcript from the L1-MET anti-sense promoter contains its own exons 1 and 2, referred to as L1-MET exon 1 and L1-MET exon 2 (Figure 1A). We designed RT–PCR primers with one primer located in either the MET exon 2 or the L1-MET exon 1 and one primer located in the shared exon 3 to examine the expression of the host gene MET and the alternate transcript from L1-MET, respectively (Figure 1A). We confirmed the transcription start site of L1-MET by 5′RACE in the T24 bladder carcinoma cell line (Figure S2C) in which the L1-MET promoter is completely unmethylated. The L1-MET transcript was lowly expressed in one bladder fibroblast cell line (LD419) and two non-tumorigenic urothelial cell lines, UROtsa [21] and NK2426 [22], and highly expressed in most bladder carcinoma cell lines (Figure 1C). L1-MET was also not expressed in normal tissues except for placenta (data not shown). Therefore L1-MET hypomethylation correlated with the expression of the alternate transcript (Figure 1C). Treatment of LD419 with the demethylating agent 5-aza-deoxycytidine lead to expression of L1-MET, suggesting that L1-MET is silenced by DNA methylation (Figure S2D). We also designed bisulfite-specific PCR primers and RT–PCR primers for two additional specific L1s from the list shown in Figure S1, which were randomly selected. One L1 was located within ACVR1C, a member of the TGF-Beta family able to induce apoptosis [23], and the other located in RAB3IP, and a protein whose exact function is unknown (Figure S3 and Figure S4). Hypomethylation of these specific L1s also correlated with expression of their associated alternate transcripts, suggesting that DNA methylation plays a role in transcriptional silencing of functional L1 promoters in general (Figure S3 and Figure S4).
The data presented thus far represents an association between hypomethylation of an L1 promoter and ectopic expression of an alternate transcript. To directly demonstrate that DNA methylation represses transcription of the bidirectional L1 promoter we utilized a luciferase promoter activity assay with a pCpGL luciferase reporter construct that has been modified to not contain any CpG sites [24]. Therefore, after insertion of the promoter sequence of interest the plasmid can be treated with the CpG methyltransferase M. SssI and the methyl donor S-adenosyl-methionine (SAM), allowing the promoter to be methylated without affecting the plasmid backbone. We created two plasmids, differing only the orientation of the L1-MET promoter, allowing us to measure either the L1 transcriptional activity or the L1-MET activity transcriptional activity (Figure 2A). Activity in both directions was inhibited in the methylated plasmid (Figure 2B). To our knowledge these data show for the first time that DNA methylation directly suppresses transcription from L1 promoter in both directions, indicating that the ectopic transcripts from L1s found in cancer [7] are a result of L1 hypomethylation. The relative activity between the two different promoters indicates that the L1-MET promoter is much weaker than the L1 promoter.
In addition to DNA methylation, epigenetic regulation of gene transcription also involves chromatin structure, specifically covalent modifications of histones, incorporation of histone variants, and nucleosome occupancy. In mice the chromatin structure of global L1s has been studied, but not in the promoter region [25]. Very few studies have addressed the chromatin structure at repetitive elements in humans. We took advantage of our ability to examine specific L1s to analyze the chromatin remodeling that occurs between the promoters of inactive and active repetitive elements in humans. Using chromatin immunoprecipitation (ChIP) we found that the level of DNA methylation at each specific L1 is inversely proportional to the level of enrichment of active histone marks (Figure 3A and Figure S5), and the chromatin structure at global L1s did not correlate with the specific L1s. Comparing the structure of the unmethylated L1-MET promoter in T24 bladder carcinoma cells to the methylated L1-MET promoter in UROtsa urothelial cells revealed a gain of the active marks H3K4me3 and acetylated H3 and the histone variant H2A.Z (Figure 3A). Therefore transcriptional activation of a repetitive element results in a similar pattern of chromatin remodeling found in active single copy genes such as p16 (Figure 3A) [12], [26], [27].
Methylase-sensitive Single Promoter Analysis (M-SPA) has previously been used to obtain single molecule resolution of nucleosome positioning at unmethylated CpG island promoters [28]. Briefly, nuclei are isolated and treated with the CpG methyltransferase M. SssI, followed by DNA extraction, bisulfite conversion, and genomic sequencing of individual clones. The resulting pattern of applied DNA methylation reveals patches of protection, indicating the location of nucleosomes on individual molecules. Previously, the main limitation of the M-SPA method was that it could not be used to assess nucleosome positioning in an endogenously methylated region. However, the enzyme M. CviPI, which methylates GpC sites [29], can be used to avoid this problem since endogenous GpC sites are not methylated in humans except in the context of a GpCpG. Therefore, by modifying our M-SPA method by using a GpC methyltransferase we have conducted the first single molecule analysis of nucleosome positioning at a methylated promoter and, in combination with our ability to study specific L1s, have shown the nucleosome occupancy at a single repetitive element in both an active and inactive state.
The endogenously methylated L1-MET promoter in the UROtsa immortalized urothelial cell line was completely occupied by nucleosomes, revealing that the methylated L1-MET promoter exists in a tetranucleosomal structure (Figure 3B). GpCpG sites were excluded from analysis since it is not possible to distinguish between endogenous CpG methylation and enzyme-induced GpC methylation at such loci. When we performed the same assay on T24 cells in which L1-MET is unmethylated we found a nucleosome occupying the region downstream of each of the two transcription start sites and no nucleosome upstream of either (Figure 3C). We were able to confirm the results in T24 cells using the CpG methyltransferase M. SssI, since L1-MET was not endogenously methylated (Figure S6). However, the number and location of CpG sites limits the resolution of this assay since the region upstream of the L1-MET start site contains only one CpG site. Therefore, the GpC methyltransferase allowed an increased resolution for this method. The unmethylated MLH1 promoter was used as a positive control for both CpG and GpC methyltransferase activity and accessibility (data not shown).
Previous work on the MLH1 bidirectional promoter has demonstrated that while each transcription start site loses the nucleosome directly upstream when active (−1 nucleosome), the nucleosome directly downstream is always maintained (+1 nucleosome) [27], [30]. The L1 promoter is a different type of bidirectional promoter that generates partially overlapping sense and antisense transcripts, commonly referred to as an antisense promoter (ASP). The L1 ASP has room for two nucleosomes between the two transcription start sites, therefore each start site has its own +1 nucleosome. These two +1 nucleosomes are maintained while the active promoter loses the −1 nucleosome at both starts sites. Therefore the inactive L1 promoter exists in a tetranucleosomal state (two +1 and two −1 nucleosomes) while the active promoter exists in a dinucleosomal state (two +1 nucleosomes). In addition, when DNA methylation levels are reduced by knocking out expression of 2 of the 3 methyltransferases responsible for maintaining DNA methylation, DNMT1 and DNMT3B [31], [32], we see acquisition of H2A.Z at L1-MET and global L1s (Figure 3D) along with induction of expression of L1-MET (data not shown) and nucleosome eviction at the L1-MET promoter (Figure 3E and 3F), revealing that a switch from an inactive tetranucleosomal structure to an active dinucleosomal structure accompanies hypomethylation.
While a single-molecule analysis of the nucleosome occupancy at the L1-MET promoter confirmed that an active L1 promoter switches from a tetranucleosomal structure to a dinucleosomal structure, we cannot generalize that other L1s exist in these states. To do so we took a cancer cell line that has a methylated and inactive L1-MET promoter, the colon cancer cell line HCT116, and performed chromatin fractionation using MNase digestion followed by sucrose gradient ultracentrifugation [33]. The fractions were run on an agarose gel and a genomic Southern using radioactively labeled input DNA was performed. Most of the DNA was present in the mononucleosome and dinucleosome fractions (Figure 4). When the same blot was probed with the L1 promoter sequence, the distribution of global L1 promoters showed enrichment in both the dinucleosome and tetranucleosome fractions, indicating that other L1s besides L1-MET could exist in an inactive tetranucleosome or active dinucleosome structure (Figure 4).
Since bladder tumors display both hypomethylation of L1s [34] and overexpression of MET [16]–[18], our next step was to determine whether hypomethylation of the specific L1 promoters and their associated alternate transcripts, including L1-MET, were present in uncultured bladder tumors. We found high levels of methylation at L1-MET and low expression in normal bladder epithelium obtained from age-matched cancer free bladders (Figure 5A and 5B) and significant hypomethylation of, and expression from, L1-MET in bladder tumors (Figure 5A and 5B). We also examined the methylation and expression of two additional specific L1 promoters located within host genes (Figure S7). Hypomethylation of the L1-ACVR1C and L1-RAB3IP promoters occurred in bladder tumors (Figure S7). Therefore we have provided the first clinical evidence that hypomethylation of functional L1 promoters results in ectopic gene expression during tumorigenesis.
Surprisingly, we also found hypomethylation and associated alternate expression of L1-MET in the corresponding histologically normal tissues from tumor-bearing bladders taken at least 5 cm away from the tumor (p<0.0001) (Figure 5A and 5B). Hypomethylation and expression of L1-MET was more prevalent in the corresponding normal tissues than L1-ACVR1C, L1-RAB3IP (Figure S7) [35]. Therefore, hypomethylation of L1-MET and activation of alternate transcripts of MET occurs not only during tumorigenesis but also in premalignant tissue. Receiver operating characteristic (ROC) curves for L1-MET revealed an extraordinary degree of both sensitivity and specificity for detecting bladder tumors (AUC of 0.97) and premalignant tissue (AUCs of 0.89) (Figure S8). Since aberrant methylation in bladder tumors can be detected in urine sediments [36] and we are able to detect hypomethylation of L1-MET in urine sediments of bladder cancer patients (Figure S9), a noninvasive urine test has the potential to be developed into an assay for tumor detection and prediction of high-risk patients.
As expected, the expression of the host gene MET was not correlated with hypomethylation of the L1-MET promoter, since the expression of MET is regulated by its endogenous promoter and not by the specific L1 promoter (Figure 5A and 5C). It has previously been shown that overexpression of MET is correlated with global L1 hypomethylation in chronic myeloid leukemia (CML) [14]. The biological mechanism behind this correlation is unclear, as MET is expressed from an entirely different promoter than L1-MET and we have shown that global L1 methylation does not correlate with specific L1 methylation. Further, we did not find overexpression of MET in bladder tumors, suggesting that it may be L1-MET that is overexpressed instead since many primers used to detect expression can amplify both products.
Since we observed hypomethylation at L1-MET in bladder tissues taken at least 5 cm from tumors we collected histologically normal tissue samples from five tumor-bearing bladders taken at various distances and directions from the tumors to determine whether distance has any effect on the level of hypomethylation (Figure 6A). When compared to the average level of methylation in normal tissues from cancer-free bladders, L1-MET was dramatically hypomethylated in normal-appearing tissues across each of the tumor-bearing bladders independent of the distance from the site of the tumor (Figure 6B). However the normal-appearing tissues were not significantly hypomethylated at L1-ACVR1C, L1-RAB3IP, and global L1 (Figure S10 and Figure 6C). Bisulfite sequencing of L1-MET in the urothelium of patients without bladder cancer revealed only fully methylated strands while in a patient with bladder cancer fully unmethylated strands were present in the tumor and the corresponding normal urothelial tissue independent of the distance from the tumor (Figure 6D and Figure S11). A plot of the distribution of DNA strands versus the percent of methylated sites reveals a biphasic distribution in the patient with bladder cancer, with the majority of strands either fully methylated or fully unmethylated (Figure S11). Our in vitro results (Figure 2 and Figure 3) suggest that these fully unmethylated strands found in tumor-bearing bladders have undergone chromatin remodeling involving a switch from a tetranucleosome to a dinucleosome structure and are transcriptionally active. To our knowledge this is the first alteration, either epigenetic or genetic, that has been found across an entire tumor-bearing organ.
The consequences of global hypomethylation at repetitive elements in cancer has long been the subject of speculation regarding the generation of genomic instability and potential activation of oncogenes. Ever since studies on viable yellow agouti (Avy) mice revealed that hypomethylation of a retrotransposon could induce ectopic expression of a gene and influence disease susceptibility [6] it has been postulated that similar events may occur in humans. While hypomethylation during tumorigenesis occurs quite frequently, a direct demonstration of the impact of hypomethylation of repetitive elements on gene expression has not been conducted. Transcriptome sequencing has recently revealed the prevalence of transcripts originating from alternate TSS within repetitive elements in humans, indicating a potential functional role of activated repeats in altering gene expression [7]. Active L1s were mostly found in embryonic and cancerous tissues, many of which result alternate transcripts of protein-coding genes. Using several specific L1s we have demonstrated the mechanism of transcriptional activation and, taken together with the results of Faulkner et al. [7], our results highlight the previously underappreciated impact of hypomethylation on ectopic gene expression, possibly contributing to tumorigenesis in a synergistic or cooperative manner (see model in Figure 7).
To elucidate the mechanism of transcriptional activation of repetitive elements, we compared the epigenetic alterations, including methylation status, histone modifications, and nucleosome positioning, that occur at a single copy of an L1 between a transcriptionally inactive and active state. Since current methods did not exist for such a study we employed several novel assays, including using primers able to amplify specific L1s, enabling methylation and ChIP assays to be performed on single copies, and a modification of the method for determining nucleosome positioning at a single molecule resolution, currently limited to unmethylated CpG islands, which allowed for the determination of nucleosome positioning in a methylated region. We were able to show that transcription from the L1 promoter is silenced by DNA methylation, providing direct evidence that one function of DNA methylation is to protect the human genome from retrotransposons.
Transcriptional activation of L1 promoters by hypomethylation results in a chromatin structure similar to that of active single copy genes such as p16, revealing that the features of active promoters, such as acquisition of active histone marks, H2A.Z, and nucleosome free regions upstream of TSSs, are not restricted to canonical gene promoters. In addition, we found that the unique structure of the L1 promoter results in two very stable nucleosome occupancy states, the inactive tetranucleosome structure and the active dinucleosome structure, and that hypomethylation could result in a switch between the two. It has been demonstrated that tetranucleosomes form a compact chromatin fiber [37]. Therefore, the widespread chromatin remodeling due to global hypomethylation of L1 promoters could contribute to chromosomal instability through the loss of many stabilizing tetranucleosome structures.
To our knowledge we have provided the first direct evidence that transcriptional activation of repetitive elements is caused by hypomethylation and chromatin remodeling at their promoters, occurs in a human diseased state, and may play a role in disease predisposition. Specifically, hypomethylation of a L1 promoter induces an alternate transcript of the MET oncogene in bladder tumors and across the entire urothelium of tumor-bearing bladders. The presence of L1-MET hypomethylation across the entire urothelium of tumor-bearing bladders has several possible explanations. Epigenetic alterations such as hypermethylation of tumor suppressor genes and hypomethylation of L1s have been found in normal epithelia adjacent to several types of tumors, including breast [38], esophageal [39], and colon [40], [41], indicating the presence of a “field defect”. Our data supports the presence of an epigenetic field defect in bladders with cancer, either due to independent events across the urothelium or clonal expansion [42]. However, another possible explanation is that the loss of L1-MET methylation occurred during early development before the bladder was fully formed. While some evidence for such abnormal epigenetic programming exists, as a recent study revealed that people who develop bladder cancer have slightly lower levels of global DNA methylation in their blood than healthy control cases [43], we did not find any evidence of a loss of methylation at global L1s or specific L1s in our patient WBC samples (Figure S9). Another possibility, which cannot be ruled out by this data, is that the presence of a tumor causes epigenetic changes across the bladder.
Whatever the underlying mechanism, the modulation of gene expression by hypomethylation of a retrotransposon such as what occurs at the agouti locus in mice is also found in humans. This leads to the activation of surrounding genes, which may contribute to tumorigenesis in a synergistic or cooperative manner. Transurethral resection of bladder tumors would leave behind large areas of epigenetically altered urothelium, possibly contributing to the high level of recurrence of bladder cancer. Fortunately, hypomethylation at specific L1s seems to provide a valuable biomarker that has the potential to significantly impact the diagnosis and treatment of bladder cancer.
The non-tumorigenic human urothelial cell lines UROtsa and NK2426 and the normal fibroblast cell line LD419 have been described previously [21], [22], [36]. Human bladder carcinoma cell lines were obtained commercially (T24, J82, HT1376, SCaBER, UM-UC-3, TCCSUP, and RT4; American Type Culture Collection, Manassas, VA) or derived in our laboratory (prefix LD). Cell culture, DNA and RNA purification were performed as previously described [36]. RNA was reverse-transcribed as previously described [36]. 5′-Rapid Amplification of cDNA Ends (RACE) to determine the 5′ end of the primary transcript of L1-MET was performed using the RLM-RACE kit (Ambion) according to the manufacturer's instruction. See Table S1 for primer sequences.
Tumor tissue samples were collected from the patients undergoing cystectomy or TURBT for bladder cancer. Normal bladder epithelium was obtained from 12 patients undergoing radical prostatectomy for prostate cancer (aged from 50 to 80) and 7 autopsy patients aged from 34 to 82, 5 of which were from non-cancer related deaths and 2 from deaths due to cancers other than bladder). All of these collections took place at Norris Cancer Hospital in IRB-approved protocols with patients' consent. Hematoxylin and eosin (H&E) sections marked with the location of the adjacent urothelium or tumor were used to guide in microdissection. DNA was bisulfite treated as previously described [44]. RNA extraction was done using a RNAeasy Micro Kit (Qiagen, Crawley, UK).
Methylation-sensitive single nucleotide primer extention (MS-SNuPE) was performed as previously described [44]. See Table S1 for primer sequences. In order to allow for a higher throughput in methylation analysis pyrosequencing was also performed as described previously [45]. Testing both methods on the same set of 66 samples yielded a correlation in the methylation levels of R = 0.91 (Figure S12). For pyrosequencing, PCR was performed on bisulfite converted DNA using a biotin-labeled 3′ primer to enable purification and denaturation of the product by Streptavidin Sepharose beads and was followed by annealing of a sequencing primer to the single-stranded PCR product. Pyrosequencing was performed using the PSQ HS96 Pyrosequencing System and the degree of methylation was expressed for each DNA locus as percentage methylated cytosines over the sum of methylated and unmethylated cytosines. See Table S1 for primer sequences. To analyze the methylation status of individual DNA molecules, we cloned bisulfite PCR fragments into the pCR2.1 vector using the TOPO-TA cloning kit (Invitrogen, Carlsbad, CA). Individual colonies were screened for the insert and the region of interest was sequenced using M13 primers. See Table S1 for primer sequences.
Expression was determined by quantitative RT–PCR as described previously [27]. See Table S1 for primer sequences.
The L1-MET and L1 promoters were cloned into the pCpGL luciferase vector [24]. The portion of the L1-MET promoter cloned was 555 bp, with 535 bp within the L1 and 20 bp within the MET gene (ch7:116364010–564). These experiments were performed as described previously [24].
ChIP was performed as described previously [27]. Briefly, chromatin was isolated from cells and crosslinked with formaldehyde. The chromatin was then sonicated to less than 500 bp in length and immunoprecipitated with an antibody to the histone modification of interest. Enrichment was determined by RT–PCR of the pulled down DNA. See Table S1 for primer sequences.
M-SPA was performed as described previously [28]. Briefly, chromatin was isolated from 250,000 cells and treated for 15 minutes with 50 U of M. SssI. DNA was isolated, bisulfite converted, and PCR fragments were cloned for sequencing of individual molecules. In order to examine endogenously methylated promoters and increase the resolution of this method, chromatin from 250,000 cells was treated with the enzyme M. CviPI, which methylates GpC sites [29], for 15 minutes with 100 U.
MNase digestion and sucrose density gradient centrifugation were performed as described previously [33]. See Table S1 for primer sequences for the LINE-1 promoter probe.
Significant differences in methylation and expression levels in normal, corresponding normal, and tumor tissues were determined using a Mann-Whitney test.
|
10.1371/journal.pcbi.0030200 | Landscape as a Model: The Importance of Geometry | In all models, but especially in those used to predict uncertain processes (e.g., climate change and nonnative species establishment), it is important to identify and remove any sources of bias that may confound results. This is critical in models designed to help support decisionmaking. The geometry used to represent virtual landscapes in spatially explicit models is a potential source of bias. The majority of spatial models use regular square geometry, although regular hexagonal landscapes have also been used. However, there are other ways in which space can be represented in spatially explicit models. For the first time, we explicitly compare the range of alternative geometries available to the modeller, and present a mechanism by which uncertainty in the representation of landscapes can be incorporated. We test how geometry can affect cell-to-cell movement across homogeneous virtual landscapes and compare regular geometries with a suite of irregular mosaics. We show that regular geometries have the potential to systematically bias the direction and distance of movement, whereas even individual instances of landscapes with irregular geometry do not. We also examine how geometry can affect the gross representation of real-world landscapes, and again show that individual instances of regular geometries will always create qualitative and quantitative errors. These can be reduced by the use of multiple randomized instances, though this still creates scale-dependent biases. In contrast, virtual landscapes formed using irregular geometries can represent complex real-world landscapes without error. We found that the potential for bias caused by regular geometries can be effectively eliminated by subdividing virtual landscapes using irregular geometry. The use of irregular geometry appears to offer spatial modellers other potential advantages, which are as yet underdeveloped. We recommend their use in all spatially explicit models, but especially for predictive models that are used in decisionmaking.
| Many different areas of science try to simulate and predict (model) how processes act across virtual landscapes. Sometimes these models are abstract, but often they are based on real-world landscapes and are used to make real-world planning or management decisions. We considered two separate issues: how movement occurs across landscapes and how uncertainty in spatial data can be represented in the model. Most studies represent the landscape using regular geometries (e.g., squares and hexagons), but we generated landscapes of irregular shapes. We tested and compared how the shapes that make up a landscape affected cell-to-cell movement across it. All of the virtual landscapes formed with regular geometries had the potential to bias the direction and distance of movement. Those formed with irregular geometry did not. We have also shown that describing whole real-world landscapes with regular geometries will lead to errors and bias, whereas virtual landscapes formed with irregular geometries are free from both. We recommend the use of multiple versions of virtual landscapes formed using irregular geometries for all spatially explicit models as a way of minimizing this source of bias and error; this is especially relevant in predictive models (e.g., climate change) that are difficult to test and are designed to help make decisions.
| The focus of this study is spatially explicit predictive models designed to support decisionmaking (e.g., population establishment and spread, climate change, and flood risk), which should have reliable, probabilistic, and mappable results. In cases in which there are few relevant validation data (e.g., nonnative species and climate change), the model cannot be calibrated statistically, and it is therefore important that biases and uncertainties are dealt with explicitly so that confidence can be placed in the results. Uncertainty may surround all components of a model (e.g., input data and processes), but bias by definition usually results from the way that processes are implemented in the model. In this study, we explored how spatial structure can be a source of bias, and present an approach that allows uncertain landscape data to be incorporated into model output with minimal bias.
There are many different landscape models in the literature (see [1] for a recent and comprehensive list), all of which allow a process (population) model to interrogate explicit locations or regions of space, and choosing the most appropriate landscape model for the study in hand is important [1]. We focus on the use of cells in a mosaic-based model [2] to represent processes in space, which requires the subdivision of space into a tessellation of discrete, internally homogeneous patches within which a process occurs. Although this is an elegant, abstract concept, the use of cells to represent uncertain spatial processes is often desirable in real-world applications. First, some information is better represented by an areal unit than by a point location (e.g., water), whereas other information is considered to conceptually occupy an area of real space defined by its boundary (e.g., an animal social group). Second, the limited understanding of many of these systems requires us to model at the scale for which most is known (e.g., the behaviour of individuals within a social group of animals is often not well-understood, whereas the size, productivity, or spatial description of the whole population may be simpler to study and is well-described). Third, the raw data used to describe the landscape (e.g., satellite and aerial photography) are subject to errors and uncertainties. By modelling processes at scales significantly larger than that of the underlying data, these problems become statistically tractable (e.g., Land Cover 2000 [3]). Although in many such models, the attribute values of cells are directly calculable from habitat or geographical data (e.g., vegetation type), here we use conceptually abstract and attribute-free cells in order to consider only geometry (specifically, shape) and neighbourhood (number and arrangement of adjacent, interacting patches) in a homogeneous landscape.
In this study, we have used population modelling concepts to demonstrate the potential for bias in cell-to-cell movement of information (e.g., individuals) resulting from the geometry of a mosaic virtual landscape. Population models predominately use raster virtual landscapes, and the description of home ranges or social groups with single squares (e.g., [4–6]) or a square arrangement (e.g., [7]) is not unusual. Cell-to-cell movement is implemented using either von Neumann (e.g., [8]) or Moore (e.g., 6,9–11]) neighbourhoods (four or eight neighbours, respectively), often with some directional component [12–14]. Some studies note that the geometry of the virtual landscape has the potential to affect simulation results [15,16], and the interaction strengths of orthogonal and diagonal neighbours in rasters are sometimes weighted using an appropriate algorithm [16,17]. Landscape permeability is sometimes defined using raster cells (e.g., [18]), although some authors have suggested using multiscaled rasters to represent patchy landscapes (e.g., [19]). Other studies use hexagonal geometry for both spatial analysis [20,21] and modelling [15,22–25] because the strengths of all neighbourhood interactions are equal. Irregular (variable shape and size) geometry has been used extensively in population modelling, but usually to parameterize or display discrete, spatially disparate habitat patches with explicit connectivity based on the distance between patches [26–28] or vector-based movement rules [29–31], and studies such as that by Ovaskainen [32] exemplify this approach. A few models use tessellated irregular shapes across a whole landscape, and implement cell-to-cell movement as part of the simulation [33,34], whereas Dunn and Majer [35] suggest that Voronoi (Dirichlet) cells are a convenient way to represent multiply scaled data, but they do not go into detail about dispersal mechanisms. However, no attempt has been made to specifically test how geometries other than rasters may affect movement in a mosaic landscape, and an explicit consideration of geometry does not appear to be an integral part of most population modelling studies.
The representation of real-world landscapes is complex, with the description of features represented as discrete objects subject to both qualitative and quantitative variability, uncertainty, or both [36]. Uncertainty within virtual landscapes is already considered in some disciplines (e.g., [37,38]). We suggest that all spatially explicit population models should consider how uncertainty in landscape representation may affect model output, just as sensitivity analyses on process model parameters have become standard practice. Clearly this is entirely dependent on the nature of the study undertaken (data, scale, structure, and discipline), so we cannot begin to describe how individual studies in diverse disciplines should address this issue. However, it seems inevitable that population modellers will adopt a probabilistic approach to spatial studies (which can be easily implemented through the use of alternative landscapes in successive runs of the model), and we provide a mechanism by which minimally biased landscapes can be created.
We created landscape mosaics with raster, hexagonal, and irregular geometries with which to model and compare the cell-to-cell exchange of information. We are not aware of any other study that directly compares the potential for systematic bias in the movement of information across the spectrum of possible geometries of mosaic virtual landscapes. We highlight how the geometry of cells in a raster virtual landscape affects both qualitative and quantitative aspects of spatial representation of irregular shapes, but leave the attributional representation of features (e.g., heterogeneous habitat [39,40]) and subsequent impacts on movement or process [41,42] to another study. We believe that these concepts are generally applicable across a broad range of spatially explicit modelling disciplines.
We created eight virtual landscapes for comparison. Three virtual landscapes used a regular geometry: two rasters with von Neumann (Figure 1A) and Moore (Figure 1B) neighbourhoods and one of equilateral hexagons (Figure 1C). Virtual landscapes with irregular geometry were created in five different ways. One was a simple tessellation around random points (hereafter, the Dirichlet landscape; Figure 1D). Three virtual landscapes were approximations to the Dirichlet landscape, but based on a raster grid, with irregular cells composed of a mean of four, nine, or 16 squares (Figure 1E, 1F, and 1G, respectively) and called the coarse-grain Dirichlet (CGD) landscapes. The final irregular virtual landscape, called the aggregate map, was derived by aggregating habitat patches from real-world coverage data [3] and thus reflected the complex structure of a real landscape (Figure 1H). Ten instances of each irregular virtual landscape were created because all were formed with random processes. Only one instance of each regular virtual landscape was used. All landscapes were based in a real-world context (Figure 2).
Cells in the raster and hexagonal virtual landscapes had a fixed number of neighbours (Table 1). Cells in the Dirichlet landscape had a mean of exactly six neighbours, though there was variation about this value within individual landscape instances. The CGD virtual landscapes all resembled pixelated versions of the Dirichlet landscape. However, both the visual and mathematical approximation improved as the resolution of the underlying raster was increased, as demonstrated by both the mean and standard deviation of the number of neighbours (Table 1). Cells in the aggregate map had approximately six neighbours, and were a range of shapes because the sequential building rules meant that growing cells were often geometrically constrained by neighbours. Of the geometries tested in this study, the mean number of neighbours of a cell was six, or its approximation, with the exception of the rasters. There was variation in the distribution of cell sizes within the irregular virtual landscapes (Table 1).
We measured and compared all possible unique cell-to-cell step lengths (measured between centre-of-mass centroids) in five landscapes: the three regular landscapes, and single instances of the Dirichlet and the CGD4 landscapes (Figure 3). In the von Neumann and hexagonal landscapes, only one step length was ever possible, with lengths 1 km and 1.074 km, respectively. In the Moore landscape, two steps were equally probable, with lengths 1 km and 1.41 km producing a mean step of 1.21 km per landscape. Step lengths in the Dirichlet landscape were gamma distributed (Figure 3) with a mean of 1.095 km, which is close to that found in the hexagonal landscape; the step lengths of each cell in the CDG4 landscape were similarly distributed with a mean of 1.18 km, though the distribution was less smooth as a result of the finite distribution of cell shapes and hence step lengths (Figure 3).
We measured the land area from a number of raster depictions of a fine-scale vector description of a real-world object (United Kingdom (UK) coastline). Rasters were created at a range of scales with a variable origin (shifted successively by 10% of the resolution west and south). A further ten rasters, at a resolution of 10 km, were created with a fixed origin but with the orientation of the grid rotated successively by 9°.
The qualitative form of a raster representation of the UK differed with a change in origin or orientation of the grid (Figure 8). Small objects such as islands appeared or disappeared, became connected or disconnected from the mainland or each other, or changed their shape radically. This would have clear effects on any model involving terrestrial movement. Although we have only shown this at one scale, these undesirable effects are fractal, and would be present as a possible bias at all scales, and would worsen at larger resolutions [44]. The mean area reported by the raster representations of the UK decreased with increasing scale (Figure 9), though as a fractal property, would show bias at all scales. The best mean estimate (99.9% of the vector original) was derived from the finest scale raster (1 km), but mean estimate of national area fell to as low as 98.5% at a 100 km resolution. In addition, the variance around these mean figures was considerable and also increased with scale. The worst performing rasters (two out of ten instances at a 100 km resolution) showed an area of only 89.6%, an alarming loss of 10.4% of the British land surface. The estimates of national area produced by iteration across multiple rotations of a raster grid at 10 km resolution had a mean of 100%; individual instances ranged from 99.3% to 100.6% of the vector original, suggesting that at this scale, the orientation of the raster had only a small quantitative effect on area.
All forms of raster representation have biases, variant with scale. In contrast, an irregular geometry (specifically a vector representation) can be subdivided into as many randomized vector cells as necessary. Any estimate of gross area, length, or geographic property is perfect (zero bias). When multiple instances of virtual landscapes are required, we recommend the use of irregular geometry to avoid introducing bias to representation of its extent.
The reduction of bias in model output should always be a priority (e.g., [45,46]). Biases can be hidden but still present in many components of a model, and their presence or interaction may produce artefacts. Sources of bias should be looked for, and where present, measured in order to decide whether they are significant in the context of the results. If they are significant, the bias should be minimized or, preferably, removed altogether. Movement in spatial models is such a fundamental process that bias in this process is likely to be critical. Some models are entirely theoretical and are used to gain insight into academic problems, in which some forms of bias may be acceptable. However, our focus is primarily on predictive models designed to support decisionmaking, where high confidence is required in the results and validation is difficult. In these models, it is essential that sources of potential bias be removed, but we believe that it is desirable in all types of models if the interpretation of results may be confounded.
We show that the geometry used to implement cell-to-cell movement has the potential to bias a diverse array of movement rules for information flow across landscapes. We tested completely random movement at one extreme, completely directed movement at the other extreme (accessibility), and a number of directed random movement rules with both forward and backward components. Other choices in this continuum were possible; however, we did not wish to test the complete library of walks ever created, but merely to demonstrate the potential for bias across the spectrum of movement types. To this end, all of the movement rules were short and simple, and the virtual landscapes homogeneous, in order to test the geometric properties of the virtual landscape and not those of the movement. We are aware that some of these methods may appear unrealistic or extreme to some disciplines. Less extreme movement rules may produce less biased results, but we would argue that some bias still exists. Inevitably the details of how cell-to-cell movement is expressed will vary between disciplines and studies, and will also be affected by scale, so we make no attempt to advise on the use or abandonment of any method of movement. However, we do wish to highlight the fundamental interaction between landscape geometry and movement, and present the use of irregular landscapes as a potential solution to some of the biases that may be encountered.
Simple random walks in virtual landscapes with regular geometries can produce enormous qualitative biases in the direction and extent of movement and hence bias the distribution of populations in space. For all our investigations (accessibility, random movement, and directed random movement), the regular geometries performed poorly at some or all scales for measures of both distance and direction. The nature and strength of the bias was, in part, a function of the length of the movement and the resolution of the landscape, with the potential for different biases to worsen after both short-distance and long-distance movement. In an ideal homogeneous virtual landscape, the distance travelled by an individual (from its origin) after random movement should be independent of the direction travelled at each step. Regular grids all restrict the direction of movement to the same few angles at every step, so the final positions of individuals cannot be independent of the grid structure. Even if the scope of the movement neighbourhood is extended to include more distant cells (e.g., to include the 16 cells adjacent to the eight immediate neighbours in the Moore neighbourhood), so that single steps may include jumps over the immediate neighbours, a regular geometry restricts the available directions to some degree. In comparison, the direction of neighbouring cells in a single, irregular virtual landscape is not set by the geometry, and when enough multiple irregular virtual landscapes are considered together, available directions assume a uniform circular distribution.
We demonstrated that, in the same way that regular landscapes restrict the available directions for movement, they severely restrict cell-to-cell step lengths, and in part, this may explain a component of the bias in movement. The hexagonal landscape is the optimum arrangement of circles packed in space using a single iteration. Dirichlet landscapes emulate circular cells over many iterations (Figure 10), and the mean step length closely approximates that of the hexagonal landscape, even in a single iteration. Although the step lengths in the CGD landscapes followed the same general distribution as the vector Dirichlet landscape, individual step lengths were more or less frequent than expected. We suggest that although the results of random walks appeared similar in all the irregular landscapes, a smooth distribution of step lengths is less likely to be a source of bias, and therefore, at fine scales in particular, we recommend the vector Dirichlet landscape over the CGD irregular landscapes.
We propose that the average number of neighbours per cell is a useful metric for quantifying the potential for landscape geometry to introduce bias in movement. A mean close to six neighbours appears to be the ideal; all the irregular landscapes have this property. Although the hexagonal landscape has six neighbours per cell, the variance is zero, and therefore directional movement must be restricted; we therefore extend the metric to include a nonzero variance. The square geometries fail in both respects.
Because we have identified the potential for bias in movement in regular geometries, we believe that studies need to show (not assume) that movement in their own spatial models produces no bias. Bias due to geometry may not be apparent in models with complex rules for cell to adjacent cell movement, e.g., dependence on heterogeneous landscape quality (e.g., habitat preference [6] and permeability [47]). Where sources of geometrical bias have been identified, individual studies have attempted to compensate with a variety of approaches [16,17]. Not only is this compensation dependent on the spatial and temporal scale of the model for which it is designed, it adds more potential for bias to the model (albeit pulling in the opposite direction) and may adversely interact with other components of the model. There is no way to be certain that biases masked at one scale will not produce artefacts when a predictive model is extrapolated in time or space. In contrast, there is no potential for bias in both the direction and distance of movement of individuals across virtual landscapes with irregular geometry.
Some of the deficiencies of cell-to-cell movement across regular landscape geometries identified here might be overcome by iteration of rasters, specifically by randomizing origin and orientation. However, this process can change the quality of the representation of the whole landscape, and it is this that we concentrate on here. The dependence on scale in the adequate representation of complex shapes with rasters has been well-discussed elsewhere [44,48–52]. Our representation of the UK coastline, and the measurement of its area, although not novel, allowed us to focus specifically on the qualitative and quantitative effects of using multiple regular geometries to represent a complete extent. We used the UK landmass as an example of a real-world object that has an irregular extent whose boundaries could not be defensibly redefined as a regular shape for a national-scale model. The representation of real-world features with regular geometries must always be an approximation [50] whose adequacy can only be measured by model output. We have shown that at any finite scale, a feature represented by a single virtual landscape with regular geometry will always show qualitative and quantitative errors. Some of these errors can be extreme, and the modeller choosing a single instance of a regular geometry with which to represent the landscape has no way of knowing how adequate it is without testing several representations. Biases in the mean output could be reduced through iteration of regular geometries (with random origin and orientation) and a careful choice of scale, but the interpretation of results would only be acceptable where the bias was quantified. In contrast, any landscape subject to a spatial modelling study can be split up into irregularly shaped cells with no detrimental qualitative or quantitative effect on the representation of the whole. If accuracy in feature representation is important, the superior alternative to the raster is the irregular and vector-based mosaic.
The problems of feature representation by regular geometries are compounded by the connection between scale and structure; the form and significance of any feature alters as soon as the scale is changed. This is especially problematic where interacting processes occur at different scales in the same landscape. Because scale has such a strong effect on the properties of model output, it is a pity that studies suggesting quantitative methods for determining the most appropriate scale have not been pursued (but see [53] for a generic approach, [54,55] for more specific applications). In the absence of quantitative rules, modellers have to rely on common sense or experience to choose scale, and therefore should clearly demonstrate that their choice is appropriate by showing that the bias produced by the model at that scale is acceptable.
The choice of resolution in many raster-based models to date is often either derived from technical data (e.g., satellite imagery) (e.g., [39,56]) or chosen to be the nearest integer measurement of an apparently relevant process (usually biological; e.g., [4,6,57]). We have shown that the potential for bias is always present in regular virtual landscapes even at high resolutions, and the impact of that bias is a function of scale. If only one raster landscape is used in a model, its origin and orientation appear arbitrary yet are usually unchallenged (this is apparent from the lack of ability to rotate rasters away from a north–south orientation in some GIS packages). If the resolution of the raster is high enough, the representation of features will be little changed by origin and orientation, but model processes may be affected. However, if the data or the process suggests modelling at a low resolution, using a raster must sorely compromise the representation of the landscape.
Using irregular mosaic landscapes solves two problems. First, the interaction of bias with scale is removed from the process model (c.f., directional bias in movement in this study). Second, the quality of the spatial representation of available data no longer depends solely on resolution; landscapes formed from vector-based, irregular cells can remain faithful to the available data at any scale and in any single instance. The modeller is thus free to set a scale appropriate to other model processes.
Lindenmayer, Fischer, and Hobbs [1] emphasize that the ability to choose one of a number of landscape models is important in fauna research. We suggest that, in the mosaic-based landscape paradigm, the automatic use of multiple landscape models should be widespread. Real landscapes comprise irregular and complex shapes [58] that can only be well-described using irregular cells. There will always be quantitative uncertainty in the location and boundary shape of cells and qualitative errors in their internal description even if irregular virtual landscapes are created deterministically from underlying habitat patch data rather than randomly (e.g., Dirichlet polygons). The only way to explore and incorporate the uncertainty associated with landscape representation is to model with multiple, alternative virtual landscapes and present results as probabilistic maps (e.g., [59]). This applies equally to regular and irregular mosaic landscapes.
The computational demands of running irregular models such as the ones in this study are not necessarily more than those of a raster model. Because the movement rules are so simple, all that is required to implement movement is a list of cell IDs and associated neighbours. All the investigations in this study were run in Python, only using the GIS for creating and displaying landscapes. We admit that preparation of an irregular landscape set requires some extra work; however, if it is accepted that multiple virtual landscapes are necessary (e.g., different random centroids for Dirichlet tessellations, and different origins and orientations for rasters), the effort in preparing irregular and regular landscapes becomes almost indistinguishable. The irregular landscapes also appear to yield greater returns, since they are scale-independent and bias-free, and can represent features within the landscape as well as the available data allow. It is worth pointing out that the ideal vector implementation of an irregular mosaic landscape can be approximated very easily by a raster aggregate; the use of models such as the CGD (Figure 1E–1G) brings the benefits of unbiased movement, although the best representation of edges and features is only achieved by using a very high resolution raster.
We have shown that all single instances of irregular mosaics have similar structure (number of neighbours) and statistical properties (e.g., step length distribution), which results in scale-free and similarly unbiased movement of information (populations). However, the structure and statistical properties of regular mosaics differ from each other and from that of the irregular mosaics, resulting in biased movement that cannot be easily compared. We suggest that two spatial models using irregular virtual landscapes of any scale may be more easily compared than those using regular geometry, which must have both the same scale and the same structure, and therefore the same bias.
We believe that the use of irregular geometry to form virtual landscapes may bring many additional benefits. We illustrate these with examples from population modelling, in which information has an integer form (e.g., individuals), but the principles should be applicable in many disciplines, including those concerned with the movement of infinitely divisible information (e.g., water). The area and shape of cells in an irregular geometry can vary across the extent of the virtual landscape so that regions requiring a detailed spatial description (complex habitat patches, linear features such as rivers, etc.) can be represented either with a greater density of irregular cells or exclusively with a single cell. This property of irregular landscapes has also been identified by Dunn and Majer [35] and is analogous to the raster approach of Tischendorf [19]. In turn, this permits an improvement in the description of cell attributes and reduces their uncertainty. This is especially important where cell attributes have the ability to affect the spatial output of the model (e.g., landscape connectivity, patch permeability, and population persistence [60]). We have validated how the aggregation of real-world features into irregular cells can provide a sufficiently irregular virtual landscape to avoid bias (i.e., aggregate map, Figure 1H).
Another useful benefit of irregular geometry is illustrated by considering a fixed point in space. A single irregular cell containing this point is obviously not circular, but cells taken from sufficient iterations of the virtual landscape will approximate a circular kernel around the point (Figure 10). This concept is useful both in describing individual behaviour (i.e., zones of perception) or group dynamics (i.e., social interaction and density dependence). Most dispersal kernels in continuous space implicitly define movement to nearest neighbours as the most frequent, with vector movement resulting in individuals moving preferentially to sites that are close [61]. By using irregular cells, such kernels can be tuned with biological and geographical realism (e.g., interaction groups are bounded by major roads or coastline, perception zones do not include impenetrable habitat, and movement cannot cross rivers; see Figure 10). Finally, we observe that the geometry and size of real-world processes and objects (such as home ranges, social group territories, habitat patches, and habitat quality) are irregular and variable ([58]; specific examples include the spatial arrangement of subpopulations of rabbits [62], badgers [63], and coyotes [64]). There is a significant body of literature in identifying habitat patches in real-world landscapes (e.g., [65–67]). Their subsequent representation with single cells of regular geometries is inappropriate. In addition, if there are few data on how cells are formed in the real landscape, the most defensible way of expressing them in a model is through multiple, randomly generated (irregular) mosaic landscapes.
This study was prompted by a desire to construct a universal framework within which uncertainties in landscape description could be included explicitly in model function and results, and in which movement could be modelled as simply as possible in an unbiased, generic, and flexible manner.
A single virtual landscape formed with regular geometry has a huge potential for bias. In contrast, this study has shown that even a single virtual landscape formed with irregular geometry has no potential to bias the direction or distance of movement of information (e.g., individuals), and although the defensible use of a random property (in this case, geometry) requires multiple instances, the variation between irregular landscapes is small.
Any representation using a regular geometry is at best a good approximation. Even when multiple instances of regular geometries with random origins and rotations are measured, the mean output still has the potential for bias, and the variation between instances is large. The representation of an extent by an irregular geometry shows no error.
We recommend the use of irregular geometry and multiple random instances in creating any virtual landscape, which eliminates bias in the movement of information and the representation of real-world extents. Both are specifically recommended for models that are designed to help make decisions, so that the probabilistic output encompasses the uncertainty in both population processes and spatial representation. As this is the first study recommending the use of irregular geometries, and we have not covered issues of internal representation (e.g., small features and heterogeneous landscape quality), it is difficult to state unequivocally that they are completely superior to regular geometries, but the results presented here suggest that they should be the first choice for modelling virtual landscapes.
Our virtual landscapes were created to have a mean cell area of 1 km2 across the study area, which was a 25,361 km2 area of southeast England. The two rasters used an arbitrarily chosen origin of (0,0) on the British National Grid (BNG); the placement of the hexagonal landscape was entirely arbitrary. The Dirichlet landscape was formed by a vector tessellation using the ArcInfo Thiessen on 25,361 random points drawn from a uniform distribution within the study area. The coarse-grain landscapes were created from three raster landscapes (origin at 0,0 BNG), with 500-m, 333-m, and 250-m resolutions. Within each raster, all squares were associated with, and then aggregated to, the nearest of 25,361 randomly chosen squares (see Figure 11 and pseudocode at the end of this section). The aggregate map was created from Land Cover 2000 [3], a fine-scale vector habitat coverage. A random habitat patch was selected and neighbours absorbed in turn until a total area of 1 km2 ± 5% was achieved. The neighbour with the next-closest centroid was always chosen with the aim of maximizing circularity; other joining rules are possible, e.g., habitat similarity, but the geometric optimization was chosen for its simplicity.
The neighbourhood of any focal cell in the hexagonal and irregular landscapes was defined as all cells in the landscape with a boundary (or point on the boundary) shared with the focal cell. The interior of a virtual landscape was defined as all cells that were more than 1 km from its boundary. The number of neighbours (Ne) and area (Ae) of interior cells were measured across all virtual landscapes and all instances on irregular landscapes. The minimum, mean, maximum, and standard deviation of Ne and Ae were calculated for each virtual landscape.
Probes were developed to quantify the effects of landscape geometry on cell-to-cell movements, and are based on the following generalized approach. In a discrete landscape, individuals move from one cell e into any neighbouring cell with probability 1/Ne, where Ne is the number of neighbours of e. We compared the performance of the probes against the simplest vector random walk. This unbiased benchmark moved an individual a fixed distance (1 km) from its starting position (x,y) at a randomly chosen angle in (0,2π). Movement probes always consisted of 10,000 independent individuals. Irregular virtual landscapes used 1,000 individuals across ten randomized instances, with figures and statistics presented as the sum of the instances (see Figure 10).
For the accessibility probe, we calculated the minimum number of steps required for an individual in any cell in the landscape to access the origin (cell containing BNG 448500, 104500; southwest corner) in all virtual landscapes. We mapped the limit of the region accessible from the origin with 100 steps and measured the mean of the minimum number of steps required to reach a range of distances from the origin at 10° angles. The random movement probe started in a cell containing the origin (BNG 524500, 179500; centre of virtual landscape). Vector random walks started at the origin. We used four sets of parameters (probability of movement, p, number of time steps, t) to simulate a variety of forced and unforced, short and long walks: (p = 1, t = 5), (p = 1, t = 100), (p = 0.5, t = 10), and (p = 0.5, t = 200). Direct comparison of population displacements can only be made between probes where the product of p and t is equal, thus unforced walks last twice as long as equivalent forced walks. The population distribution also was calculated for directed random walks, in which neighbouring cells were divided into “backward” (toward the origin) and “forward” (around or away from the origin) neighbours. Fully directed (no backward movement) and semidirected (10% possibility of movement backward) random walks were simulated for (p = 1, t = 50) and (p = 0.5, t = 100). Directed movement was implemented in the vector random walk by disallowing any choice of angle that would result in a new location closer to the origin. Notice that in a totally random walk across a landscape with mean of six neighbours, backward movement would be achieved on average in two out of six cell-to-cell movements (with two radial movements and two forward movements also possible), or with 33% chance. Therefore, a 10% chance of backward movement in the directed random walk reduces backward movement by one-third and not by one-half (if movement was only forward or backward).
For statistical tests of similarity, population density (individuals per square kilometre) was measured in the cell containing locations 0,1,2,3,... km from the origin, at angles 0°, 60°, and 90° from north. The number of individuals within a circle with area 1 km2 centred at these locations after the equivalent vector movement was used as a benchmark. Pearson's product moment correlation coefficient was calculated for the paired datasets for each cross-section. For short random walks, locations from 0,..,8 km were used; for long walks, we used 0,..,36 km.
Using a vector polygon of the British coastline (of area 245,660 km2), we created ten raster representations of the UK using a resolution of 1, 10, 50, and 100 km. At each scale, the origin of the raster (lower left corner) was initially set at BNG 0,0 with nine further representations created by moving the origin 10% of the resolution, both west and south.
Pseudocode for creating a CGD landscape with mean cell size of 1 km2 is shown below. See Figure 11 for an illustration of the process.
Let D be an X by Y grid with cells dx,y:
x,y are the coordinates of the centroid of the raster cell
dx,y.belongs_to = null
dx,y.id = (X − 1)x + (Y − 1)y
dx,y.dist = infinity
CG4: Initial raster (D) resolution 500 × 500 m (i.e., 4 cells = 1 km2)
CG9: D resolution 333.3 × 333.3 m
CG16: D resolution 250 × 250 m
# Choose the centre cells for the aggregation process
For i = 1..25361:
x = random_integer in range(1,X) inc. end points
y = random_integer in range(1,Y) inc. end points
dx,y.belongs_to = dx,y.id
dx,y.dist = 0
chosen_list.append(dx,y)
# Find which centre cell is closest to every other cell in D
For all cells c in chosen_list:
For all cells d in D but not in chosen_list:
dist = sqrt((cx - dx)2 + (cy - dy)2)
if dist < dx,y.dist:
dx,y.belongs_to = c.id
if dist = = dx,y.dist:
if random_in_[0,1] < 0.5:
dx,y.belongs_to = c.id
# Dissolve the grid D according to the .belongs_to attrib ute
|
10.1371/journal.pcbi.1007150 | Integrative modeling of the HIV-1 ribonucleoprotein complex | A coarse-grain computational method integrates biophysical and structural data to generate models of HIV-1 genomic RNA, nucleocapsid and integrase condensed into a mature ribonucleoprotein complex. Several hypotheses for the initial structure of the genomic RNA and oligomeric state of integrase are tested. In these models, integrase interaction captures features of the relative distribution of gRNA in the immature virion and increases the size of the RNP globule, and exclusion of nucleocapsid from regions with RNA secondary structure drives an asymmetric placement of the dimerized 5’UTR at the surface of the RNP globule.
| The genome of HIV-1 is composed of two strands of RNA that are packaged in the mature virion as a condensed ribonucleoprotein complex with nucleocapsid, integrase, and other proteins. We have generated models of the HIV-1 ribonucleoprotein that integrate experimental results from multiple structural and biophysical experiments, exploring several hypotheses about the state of the RNA before condensation, and the role of crosslinking by integrase. The models suggest that the 5’UTR, which shows extensive secondary structure, has a propensity to be placed on the surface of the condensed globule, due to reduced binding of nucleocapsid to double-stranded regions within the 5’UTR. This unexpected localization of the 5’UTR may have consequences for the subsequent structural transitions that occur during the process of reverse transcription.
| Electron microscopy of mature HIV-1 shows a condensed ribonucleoprotein (RNP) complex [1] packaged within the cone-shaped capsid, which is thought to include genomic RNA, nucleocapsid, integrase, transfer RNA, reverse transcriptase, and other components [2]. It is formed through a complex, multistep process where genomic RNA (gRNA) associates with a lattice of Gag polyproteins at the cell surface, which buds from the surface to form an immature virion, followed by proteolytic cleavage of Gag into capsid, nucleocapsid, integrase and other viral proteins, and finally condensation and encapsidation of the mature RNP within the capsid (Fig 1). Understanding of the supramolecular and molecular details of this RNP is important, since it must undergo large structural transitions over the course of the viral life cycle, including playing a central role in assembly of Gag proteins into new viruses and transition from single-stranded RNA to double-stranded DNA during reverse transcription.
The HIV-1 RNP has been studied by many complementary experimental methods. As revealed by SHAPE and other methods, HIV-1 genomic RNA dimerizes and has an extensive secondary structure [4,5], including a well-documented structure at the 5’ untranslated region (5’UTR) [6]. Nucleocapsid coats the gRNA at a density of about 1 nucleocapsid per 11–12 nucleotides and plays key roles as a chaperone in reverse transcription and other processes [7]. A recent study took advantage of the extreme radiation sensitivity of nucleocapsid, which causes formation of bubbles in tomograms, to localize nucleocapsid in the RNP and further show a preference for localization in the dense condensate at the large end of the capsid [8]. Studies on the mode of action of ALLINI compounds (allosteric integrase inhibitors) have revealed that integrase is also essential for maturation of the RNP, and crosslinks the RNA in the mature virion [9].
As part of our ongoing work to study the mesoscale structure of HIV virions at various stages of the viral life cycle, we have developed coarse-grain models of the mature HIV-1 RNP to explore its formation and characteristics, integrating experimental results from electron microscopy, structural biology, CLIP-Seq and SHAPE. Coarse-grain models have been instrumental in understanding of mesoscale-level protein interactions in HIV structure and maturation, including models of the structure of immature virion [10] and formation of the cone-shape capsid [11, 12]. In particular, these studies reveal the process whereby gag protein in immature virions assembles into an imperfect hexagonal lattice. CryoEM studies have further revealed that this hexagonal lattice covers roughly 2/3 of the inner surface of the immature virion, with scattered defects [13]. In the work reported here, we interpret these results with a quasisymmetrical model of gag in immature virions, where the defects correspond to missing gag hexamers at the pentagonal sites in the quasisymmetrical lattice. We then generate models of the mature RNP that explore several alternatives for how the gRNA is deposited onto this immature gag lattice, and how much of this immature virion structure is retained when the RNP matures.
Our RNP models include three elements: two copies of 9 kb gRNA, 2000 nucleocapsid proteins, and 140 integrase subunits. The coarse-grain method begins with several assumptions for the structure of the genome within the immature virion, folds the 5’UTR based on experimentally-determined base-pairing interactions, and condenses the RNA through interaction with nucleocapsid and integrase to form a mature RNP.
The coarse-grain modeling method used here builds on previous lattice-based methods for modeling bacterial nucleoids [14]. Briefly, the method begins with an initial geometric model of the gRNA built within a sphere representing the Gag polyprotein lattice in the immature virion, and assigns integrase crosslinking sites based on proximity of the experimentally-determined binding sites. The model is then condensed, driven by the weak attraction of nucleocapsid and gRNA, while being constrained by integrase crosslinks and steric bulk of nucleocapsid. Three initial gRNA configurations were tested, based on three different hypotheses for the capture and maturation of the gRNA. “Self-avoiding Gag” models deposit the two gRNA strands on the inner surface of the immature Gag lattice in a self-avoiding random walk, assuming that integrase crosslinking occurs soon after release of the gRNA. This model explores the hypothesis that local features of the RNA placement on the Gag lattice may be retained as the complex condenses. “Overlapping Gag” models similarly deposit the two gRNA strands on the Gag lattice, but allow them to overlap. “Random” models assume that the RNA is released in the immature virion and is randomly distributed within the available volume before condensation. Integrase is known to adopt multiple oligomeric states [15], so separate models with integrase tetramers and with integrase dimers were tested, as well as models without integrase.
Coarse-grain models of the mature HIV-1 RNP show a uniform, condensed form (Fig 2). Volumes decrease by roughly a third to a quarter relative to the initial models (Table 1). The three initial configurations (Random, Self-avoiding Gag, and Overlapping Gag) yield similar volumes of the final models. As seen in Fig 2, the RNP globule easily fits within the experimentally-determined structure of the capsid, and matches closely images of the intact capsid from electron microscopy (see, for example, [1]).
Models with integrase tetramers showed the greatest volume, followed by globules with integrase dimers, and models with no integrase were smallest. Control experiments varying the number of dimers and tetramers, and comparing the default square-planar model of integrase with a tetrahedral model, show that the difference in size is a direct consequence of the volume occupied by the integrase (Table 1). For example, the default square planar model for the integrase tetramer includes two steric spheres with 9.2 nm diameter that may overlap, so the total volume of 35 integrase tetramers would be in the range of 14300 to 28500 nm3. In the Self-avoiding Gag model, adding these values to the model with no integrase gives a range of 66600–80800 nm3, which is consistent with the model with integrase tetramer at 70500 nm3. As a control, we also created models using tetrahedral integrase tetramers composed of two dimers, each using the same representation used for the dimer models (6.5 nm diameter spheres). The tetrahedral tetramer model shows a volume of 61200 nm3, similar to the model with integrase dimers at 61900 nm3. Finally, we did an experiment that would be consistent with rapid exchange of nucleocapsid, by creating the condensed model of RNP without NC, and then choosing IN positions in the condensed globule. The resultant globule (69800 nm3) showed a similar volume as the globule with integrase crosslinking throughout condensation (70500 nm3).
Contact probability plots (Fig 3) reveal subtle differences between the three types of models, and these are further quantified by plotting the average value of the contact probability as a function of the separation of nucleotides within or between chains (Fig 4). Models built from the Random configurations show a random sprinkling of interactions between all regions of the gRNA, and a similar distribution of contacts between the two chains. Regions immediately adjacent to the diagonal are sparsely populated, showing the lack of organization at small scales, such as plectonemic supercoils or hairpins. Conversely, models with the Self-avoiding Gag configuration show more interaction near the diagonal. This is expected, since the contact probability of a random chain decays less rapidly with loop length when constrained in 2D compared with 3D [16]. Cross-chain interactions, on the other hand, are reduced due to the self-avoiding definition of the model. Models starting from the Overlapping Gag configurations show a reversed nature: the interaction along the diagonal is slightly reduced when compared to models starting with the Self-avoiding Gag configuration, but the cross-chain interactions are enhanced, presumably due to the many points of close proximity between chains in the initial model.
As expected, all three models show a strong interaction at the 5’UTR, due to the modeled secondary structure. A subtle band of reduced interaction extends horizontally and vertically from the diagonal at the 5’UTR, indicating that the 5’UTR forms fewer than expected interactions with the rest of the chains. As shown below, this is a consequence of a general exclusion of the 5’UTR from the body of the globule.
We also calculated contact probability plots that are averaged over the two strands, which are more indicative of results that may be obtained from a hypothetical high-resolution contact experiment (Fig 3B). The underrepresented band extending from the 5’UTR is distinguishable, but the more global features that differ between the three models are not as distinguishable between the three plots.
Given the coarse-grain nature of the model, the hairpin loops in the 5’UTR do not show a typical double helical structure, but rather end up looking more like distorted bobby pins (see the examples in Fig 5). We noticed early in this study that the secondary structure has a consequence on the global nature of the RNP: during the process of condensation, the 5’UTR is excluded from the body of the globule and often ends up on the surface. We quantified this exclusion with a simple metric, by evaluating the average radial distance of nucleotides in the 5’UTR as compared with the rest of the gRNA and the sites of integrase interaction (Table 2). The central column of this table shows that the average radius of the bulk of the RNA is fairly constant across the default model and several controls, with models with IN tetramers at just over 18 nm, and models with IN dimers or no IN slightly smaller. The right column shows that the IN-binding sites show a similar trend across the default and control models. The 5’UTR shows different behavior, however, based on the assumptions made for the nature of region.
Several control experiments help to identify the cause of the exclusion. We tested two hypotheses: the role of NC, and the role of the secondary structure itself. In the default experiments (“FullModel” in Table 2), NC is excluded from the 5’UTR (apart from one site determined experimentally), providing less attractive force for the condensation, resulting in eccentric location of the 5’UTR and larger radial average values. When hypothetical models are condensed with NC covering the 5’UTR as well as the rest of the gRNA (“FullAllNC” in Table 2), the 5’UTR instead is located in the interior of the globule, showing smaller radial average values. This is expected, since the local concentration of nucleocapsid is higher due to the close proximity of the RNA chains when they base pair.
When the two RNA chains are treated as separate chains, with no secondary structure, they show a similar behavior with respect to NC. If NC is excluded from the 5’UTR (“TwoChain” in Table 2), the 5’UTR tends to pack on the surface of the globule, and if NC is equally distributed (“TwoAllNC” in Table 2), the 5’UTR has the same properties as the rest of the chain. From these experiments, we conclude that NC is the primary cause of the exclusion of the 5’UTR under the assumptions made by our modelling method.
One of the major goals of this work is to create plausible models of the RNP to identify experimental modalities that could distinguish between different hypotheses for the effect of integrase crosslinking on the final form of the RNP. Note that our protocol is not meant to simulate the process of condensation, rather, it is designed to provide a rapid method for generating multiple models that are consistent with the available data defining the nature of the RNP. The three types of models tested here (Self-avoiding Gag, Overlapping Gag, and Random) are designed to explore different hypotheses for the initial structure of the gRNA and the possibility that IN crosslinking could trap features of this structure in the condensed RNP. Contact probably plots reveal differences in the arrangement of chains in the mature RNP, with the self-avoiding model showing stronger partitioning of the two chains in different regions of the RNP. The study also reveals that, because the two gRNA chains are identical, these differences would be difficult to quantify in a hypothetical Hi-C type experiment (Fig 3B).
Two additional morphological features are revealed in these models, which may be accessible to study by cryoEM or super-resolution microscopy. First, crosslinking by integrase leads to a larger condensed globule. Control experiments revealed that this increase in size is primarily due to the steric bulk of the incorporated integrase subunits. However, the model for integrase used here is very simple, with no constraints on the relative orientation of the gRNA strands that are bound. We might expect that the condensed globule may be larger if integrase is more rigid than modeled here, with reduced mobility in the linker to the RNA-binding C-terminal domain and consequently stronger constraints on the orientation of the gRNA binding sites. Unfortunately, there currently are no convincing models of integrase structure at this stage in viral life cycle, but based on available structures of integrase dimers and intasomes (see Methods), we might expect that the connection to the C-terminal domains is quite flexible. Our control experiments suggest that we would not expect a smaller condensed globule if integrase is found to exchange sites rapidly during the process of condensation.
The second emergent feature of the models is the exclusion of the 5’UTR from the bulk of the RNP globule. Control experiments implicate the binding propensity of NC for single-stranded nucleic acids [17] as the cause of this exclusion. Exposure of the 5’UTR might be expected to have functional consequences, for example, by allowing ready access to reverse transcriptase for the initiation of genomic DNA synthesis.
The current model includes only gRNA, nucleocapsid and integrase, and secondary structure only in the 5’UTR. There are many opportunities for future studies to explore additional functional features and their emergent effects on RNP globule structure and partitioning. These will include a more detailed study of secondary structure, starting first with the Rev response element and extending to more detailed models as defined by SHAPE data. In addition, other molecules, such as transfer RNA and reverse transcriptase, are known to interact with the gRNA and could have effects on the form and function of the RNP. We are also currently developing methods for generating full atomic models from these coarse-grain representations, for use both in simulation and educational outreach.
A lattice-based approximation of the convex hull was used to estimate the volume of RNP globules, and calculated similarly to previous work [14]. The average radial distance of all beads, or a selected subset of beads, relative to the center of gravity was used as a metric to quantify both compactness of a globule and asymmetry in placement of selected portions of the RNA chains. Contact probability plots were calculated by counting the number of contacts within a given threshold (here, 10 nm) within a given sequence bin, across all 25 instances of a particular model.
Nine sets of models of the condensed HIV-1 RNP, using the three hypotheses for the starting model, and with integrase tetramers, dimers or no integrase, are available at https://zenodo.org/record/2662964. Open-source software and input data files for generating models with the self-avoiding Gag hypothesis are available on GitHub at https://github.com/dsgoodsell/HIVnucleoid.
|
10.1371/journal.pcbi.1000562 | Discovery of Regulatory Elements is Improved by a Discriminatory Approach | A major goal in post-genome biology is the complete mapping of the gene regulatory networks for every organism. Identification of regulatory elements is a prerequisite for realizing this ambitious goal. A common problem is finding regulatory patterns in promoters of a group of co-expressed genes, but contemporary methods are challenged by the size and diversity of regulatory regions in higher metazoans. Two key issues are the small amount of information contained in a pattern compared to the large promoter regions and the repetitive characteristics of genomic DNA, which both lead to “pattern drowning”. We present a new computational method for identifying transcription factor binding sites in promoters using a discriminatory approach with a large negative set encompassing a significant sample of the promoters from the relevant genome. The sequences are described by a probabilistic model and the most discriminatory motifs are identified by maximizing the probability of the sets given the motif model and prior probabilities of motif occurrences in both sets. Due to the large number of promoters in the negative set, an enhanced suffix array is used to improve speed and performance. Using our method, we demonstrate higher accuracy than the best of contemporary methods, high robustness when extending the length of the input sequences and a strong correlation between our objective function and the correct solution. Using a large background set of real promoters instead of a simplified model leads to higher discriminatory power and markedly reduces the need for repeat masking; a common pre-processing step for other pattern finders.
| In the years following the sequencing of the human genome focus have shifted towards trying to understand how this blueprint results in the diversity of cells that we observe. Part of the answer lies in the regulation of transcription and how the proteins responsible for this recognize where they should attach to the DNA. This is a well studied problem, but most methods developed for this have a hard time dealing with the heterogeneity of the mammalian genomes. Here we present a method that greatly improves the efficiency of this search by contrasting the DNA with a large number of background DNA sequences. This enables us to handle repetitive segments of the genome that may be functional, but are usually considered intractable by most methods.
| The rapid emergence of experimental techniques that can probe for functional elements at whole-genome scales[1] necessitates computational methods to analyze data in these settings. In particular, methods that locate promoters or measure gene expression on genome-wide scales (e.g. [2],[3]) must be complemented by algorithms that can find the active regulatory elements within the larger promoters. Ab initio computational search for transcription factor binding sites (TFBS) in DNA sequences is often termed “motif discovery”. “Motif” here refers to a general pattern describing what DNA sequences the transcription factor binds[4]. Motif discovery is one of the classical problems in computational sequence analysis and can be briefly stated as: Given a set of sequences containing one or several short overrepresented sites, locate these and produce a model describing them.
There are two main avenues used to attack this problem: i) enumerative algorithms based on word counting, such as [5],[6], and ii) pattern-based approaches often using position specific weight matrices (WMs), which scores sites based on position specific weights [4]. Since the binding preferences of transcription factors (TFs) are not easily captured by a single word or consensus string, pattern-based approaches can give solutions closer to the biological reality and it has been argued that the matrix score is related to the binding energy [7],[8]. However, such approaches correspond to the problem of finding local, optimal multiple alignments, which is NP-complete [9]. Therefore, almost all pattern-based motif finders use statistical optimization methods such as Gibbs sampling or expectation maximization [10],[11].
A typical instance of motif discovery starts with a set of upstream promoter regions of co-expressed genes suspected to be co-regulated and by extension more likely to be under control by the same regulatory machinery. This set is called the “positive set” and most methods proceed from here by locating motifs that are in some way statistically overrepresented in this set. The most successful applications of motif discovery have been in organisms whose regulatory information is densely aggregated around transcription start sites, such as Saccharomyces cerevisiae (baker's yeast). In mammalian genomes, regulatory information is spread out over wider regions, which makes “pattern drowning” a significant issue; in other words, the information in the regulatory sites is too small to stand out in the large genomic region of interest. In this context, the accuracy of contemporary pattern finders is not sufficient for many biologically important problems [12].
Most methods operate with some notion of a background model describing “generic DNA” against which the over-representation is measured. The model is often a multinomial or a Markov model. The choice of model is important for obtaining good results [13],[14]. However, most such models have difficulty in capturing the complexity of the highly heterogeneous mammalian genome sequence, which has a multitude of different promoter architectures[15], numerous interspersed repeats, low complexity sequences, CpG islands, etc. [16]. Instead of simplifying the underlying DNA sequence by a general model, we take this to its extreme conclusion and use a very large set of promoters as the actual background instead of building a model describing the sequences in the promoters. For simplicity, we use the term “negative set” to describe the background set; this is strictly speaking not true as sites could occur in this set at a much lower frequency, since real promoters are sampled randomly. By contrasting the sets, it is possible to see what common features make the sequences in the positive set unique.
Discriminatory motif searching is not a new idea; several methods have been developed that take advantage of a negative set [17]–[24]. However, many of these use word-based models [19]–[21], which might not capture the diversity of binding sites. Others again use PWMs, but have binary hit models that do not distinguish between hits as long as they are over a threshold [22]. A discriminatory approach similar to ours has been combined with the use of expression data [18], but depending on the regions that are being investigated this might often not be available or even possible. We adopt an approach similar to DEME [23] to identify the most discriminative set of motifs by modeling the sequence labels (positive or negative) rather than using the conventional generative approach[10],[11]. However, there are some important differences to DEME. Firstly, DEME uses a global string-based search followed by a local gradient refinement, which may miss patterns that are not well-represented by a consensus string, whereas we use a global optimization technique (simulated annealing) for optimizing the model, which does not have this limitation, although it may have others (see below). Secondly, our method (Motif Annealer - MoAn) uses and optimizes a threshold, and uses an enhanced suffix array (ESA) to speed up pattern searches. Thirdly, in MoAn the length of the motif is also optimized. DEME is also particularly targeted towards proteins while our approach is intended for use with DNA.
Specifically, we use conditional maximum likelihood to estimate the WMs and their thresholds such that the probability of the positive and negative sets is maximized (see Methods). Thus, the resulting matrices cannot be derived from the frequency matrix for the sites found – it is rather the matrices that lead to the best discrimination. The probability of a sequence is calculated as a product of the probabilities given by the matrices matching above a threshold and a simple null model for non-matching regions. From this and prior probabilities for matches in the positive and negative sets, the probability of the set label (positive or negative) is calculated. In this probability the background model cancels. The total likelihood is a product of the class probabilities for all sequences (positive and negative).
This conditional likelihood leads to a non-trivial optimization problem which is handled using simulated annealing (see Methods), where we iteratively change the WMs and their thresholds, retaining changes that lead to higher discriminatory power using the Metropolis-Hastings algorithm [25],[26]. Given sufficient iterations, the method guarantees convergence on the optimally discriminatory motifs. To cope with the vast size of the sets we utilize a highly efficient data structure, the ESA, for searching DNA for pattern instances[27]. With reasonable cutoffs, this reduces the computation by an order of magnitude[28].
We evaluated our method by comparing its accuracy to a set of widely used motif discovery methods (MEME[29], DEME[23], Weeder[5] and NestedMICA[14]) in several different ways. In all runs, we used the same background set, which consists of 1000 experimentally defined promoters randomly sampled from the mouse genome (Text S1). The evaluation statistics are the same as used in [12] (see Methods) and we also pooled the results from all motifs (grouped by length of the input sequence; see below) and calculated the compound statistics on this. To reduce the influence of the optimization method, we ran all non-deterministic methods five times on each set selecting the best run according to their own scoring function.
In line with the recommendations of [12] we used synthetic data sets for the inter-method comparison. These were constructed by taking experimentally defined promoter regions based on strong CAGE tag clusters [2] and planting binding sites from various TFs inside these (Text S1). To decrease possible biases for the methods towards certain specific motif types, we randomly selected one TF from each of the 11 JASPAR[30] families as well as an example of a zinc-finger factor (Table S1). For a given matrix, we randomly chose sites from experimentally validated binding sequences used for constructing the JASPAR matrix instead of generating sites using the matrix. Since the accuracy of motif discovery methods normally deteriorates when sequence length is increased (“pattern drowning”), we evaluated the various methods on sets with sequence lengths varying between 200 and 1200 nucleotides (Table S3). This gave a total of 84 sets (12 motifs ×7 lengths) with 100 sequences in each. Sequences had a site from a given motif planted with a probability of 0.5. For those methods that support it, a background/negative set was provided containing 1000 sequences sampled in the same way and with the same length as the positive sequences. We used default settings for all methods except where there were obvious reasons not to (Text S2). Since DEME requires motif length as input we decided to input the correct length of the matrix. This provides DEME with an informational advantage over the other methods.
Fig. 1 (and Figs. S4, S5, S6, S7, S8) shows a significant performance gain in using MoAn compared to the other methods as measured by Matthews correlation coefficient on nucleotide level (nCC) and average site performance (ASP) – an average over the positive predictive value and the sensitivity on binding site level (see Methods for details). With both measures, MoAn performs better than any other method on all sequence lengths. In particular, the performance is not as affected by increasing the input sequence length as the other methods; at certain sequence lengths(800, 1200) MoAn has more than twice as high ASP values as the second best method. We also evaluated MoAn with the applicable subset of the evaluation set proposed by [12](Text S3 and Table S4), where the OligoDyad, AnnSpec and MoAn achieve the highest sASP values. We note that this set is challenging as none of the methods perform well overall, and the difference in performance between methods might not be significant due to this fact. In addition, this set does not evaluate how well the method can deal with increasing lengths of input sequences, which is highly relevant.
The relationship between our objective function and the correct solution was assessed by plotting the MoAn scores against the sensitivity obtained in all five runs on each of the 84 sets (not just the best from each run) (Fig. 2). There is a clear correlation (Pearson CC: 0.90) between these two measures. There is a similar correlation with other measures, such as the nCC (Fig. S1).
This finding is important, because it indicates that the raw score is an indication of quality independent of the motif analyzed. It also shows that choosing the best scoring run of several will often give the best result.
Aside from the problem with decreasing sensitivity as the length of the input sequences increase, repetitive sequences represent a severe problem for motif discovery, as these will often seem to be over-represented, and therefore it is common to mask these repeats. However, masking is always arbitrary, and some repeats are functional [31],[32], so indiscriminate repeat masking is not optimal. When using a large negative set, repeat masking is unnecessary since repeats, if commonly occurring, will feature in the negative set and therefore be avoided as potential hits in the positive. At the same time, we can avoid the reverse problem – if a type of repeat actually is over-represented in the positive set, it can still be found. To demonstrate the insensitivity to repeats on a practical level, we planted repetitive sequences in each of the positive sets with a slightly higher frequency than the real motifs and ran our predictor on these sets both with the normal background and with a background similarly spiked with repeats. Specifically, we planted 1 to 10 consecutive instances of CACTA with a probability of 60% in each sequence. Fig. 3 shows, as expected, that the results do not deviate much from the repeat-less run when repeats are planted in both the positive and negative sequences, while the method picks up the repeats instead when there are no repeats in the negative set. We also performed this test using decoy motifs instead of repeats with similar results (Text S4, Fig. S2).
Evaluation of methods on real data is difficult and often a poor indication of general performance due to lack of insight into the correct solution [12]; on the other hand, it is necessary to show that the method can be applied to real problems.
MoAn and four other methods were run on a collection of real data sets consisting of the binding sites of four human and mouse factors from the PAZAR database[33] and their associated genomic sequence. The sets were split by organism into 7 sets and the regions adjacent on the genome were merged resulting in sets ranging in size from 14 to 118. The merging means that the base sequences can have a varying number of sites and may be of different lengths. The sets were then subsequently enlarged by adding an equal number of randomly selected promoters to increase the difficulty (Text S6 and Table S5) and also padded with their cognate upstream and downstream regions of varying lengths (200–1200, as in the synthetic evaluation) to estimate the impact of noise.
Fig. 4 shows the performance over the real sets. MoAn's performance is clearly superior, but not as spectacular as in the more controlled environment with synthetic sequences.
We speculate that the reason for this is that the background and foreground of the synthetic sets are essentially sampled from the same pool (RefSeq promoters), while we have made no effort to customize the background for the PAZAR sets. If the genomic environment of the factors differ from normal promoter sequences this could lead to a reduced performance. There are also fewer sets (7 versus 12) in this evaluation leading to a higher variability.
We report additional trials using ChIP-chip data in supplementary material (Text S7, Fig. S3 and Tables S6, S7). MoAn has also been used successfully to discriminate between binding regions of human ESR1 and its paralog ESR2; the results were comparable with matrix-scanning approaches with pre-defined motifs[34].
An additional aspect of the motif finding problem is that TFs often work by forming complex interactions [35]. Examples include mutually exclusive and cooperative binding. Clusters of TFBSs are commonly termed cis-regulatory modules, and are often responsible for tissue-specific expression. We try to capture these interactions by incorporating co-occurrence of sites from different motifs into our model, with the goal of further increasing predictive power. To test whether our objective function is capable of capturing interactions between factors we constructed a set where co-occurrence of sites from different motifs occurs. We randomly chose 5 pairs of new motifs (Table S2) and planted their corresponding sites in a positive set of 100 promoters with a 40% chance of co-occurrence and 10% of single occurrence. We then spiked the background set with sites from each of the motifs (10% chance each for all sequences) to mimic a situation where it is the interactions of the two sites rather than single sites that are responsible for the regulation. MoAn was then run in co-occurrence mode and compared to two single-occurrence runs in a series. In the serial runs we masked out the predictions from the first iteration before running the second iteration. In Fig. 5 the ASP and nCC is plotted. In our experiment three of the pairs turned out to be composed of motifs with relatively low information, leading to poor performance. However, the two remaining ones show that modeling of co-occurrence can significantly improve performance. This extended model is unfortunately computationally taxing and requires more than twice the number of iterations compared to the single prediction.
In this work we have shown the value of using a large negative set instead of a pre-defined background model in motif discovery. Using raw sequences more accurately portrays the background than any general model and therefore higher discriminatory power is achieved. This method is also much less sensitive to “pattern drowning” in larger sequences, which is a bottleneck in computational analysis of mammalian regulatory regions. However, while our method takes a significant step towards routine motif discovery on large sequences, the problem cannot be considered fully solved. In particular, MoAn accuracy may be further improved by incorporating information on evolutionary constraints (phylogenetic footprinting)[36] or DNA accessibility[24],[37].
In our opinion DEME is the best runner up of the methods. It often predicts the correct motif and has a high sensitivity, but often at the cost of a large number of false positives as it predicts also in those sequences not containing a site. MoAn seems to be better at balancing the sensitivity and specificity. On the other hand DEME is also given an artificial advantage by having the correct motif length as input and it is uncertain how advantageous this is.
Weeder performed surprisingly poorly given its stellar performance in a recent evaluation[12]. This might be due to motif selection which we did according to the most redundant motif, but was in [12] done in a more complicated manner not part of the current Weeder package. This procedure led to no predictions on several of the harder sets which might give Weeder a statistical advantage (as discussed in [12]).
A concern that might be raised is that optimizing a cutoff might lead to a conservative estimate of binding sites at the expense of weaker sites. However, assessing this is hard since experiments have their own thresholds in the post-analysis and any evaluation of MoAn's threshold will be dependant upon those. Investigations where we artificially forced the cutoff to remain low, lead to a reduction in performance (data not shown). We address this potential problem indirectly by providing a matrix that can be used to search sequences at a lower threshold.
Future improvements of MoAn will focus on the optimization algorithm, which currently is not robust enough to always produce reliable results. In our current implementation we avoid this problem by running the algorithm many times to see that the solution is stable.
Evaluation is done on both site and nucleotide levels. The statistics used are similar to those in the recent large scale evaluation [12]. To get a compound statistic for all motifs at each length we used what is there described as the “combined” method for summarizing. This consists of treating all sets of a given length as one big set, summing up all the basic statistics below (nTP, nTN … sFN) before calculating the compound statistics. This removes the problem of undefined statistics in those cases where a method does not predict any sites.
nTP Number of nts part of a site correctly predicted.
nTN Number of background nts correctly predicted.
nFP Number of background nts predicted to be part of a motif.
nFN Number of nts part of a site predicted as background.
sTP Number of real sites that share over 50% of its nts with a predicted site.
sFP Number of predicted sites that share less than 50% of its nts with a real site.
sFN Number of real sites that share less than 50% of its nts with a predicted site.
Note that we are more conservative with respect to the site prediction than [12] in that we demand at least half of the nucleotides overlapped to get a single sTP.
Derived from the basic statistics:
A sequence is assumed to be described by a mixture model consisting of a background distribution and a set of WMs describing the binding affinities of the TFs. The WMs contain log-odds scores of the type:(1)where is the position in the WM, is a letter in the DNA alphabet and is the probability of having letter at position in the motif described by . The score of a matrix aligned at a position in a sequence is therefore:(2)where is the DNA letter at position in sequence .
The aim is to discriminate between two sets of sequences , where label denotes the positive set and the negative. The prior probability of binding site occurrence in a sequence contained in set is called . We assume that there is a marked difference in the site occurrence between the two sets and want to construct a score that captures how well a set of WMs describe this difference. Using two WMs as an example, and , there are four possible ways for a sequence to be generated. With prior probability it contains no sites and is only generated by the background model . Or, with prior probability , it contains a single site (one of the two) corresponding to one WM positioned at nucleotide number ( is equal to 1 or 2 corresponding to the two different matrices). This is written , where is the score of the matrix aligned to the nucleotides at position (eq. 2) and 2 is the base of the log scores contained in the WM. Note that the log scores in a WM are divided by the background model, so the background () cancels out in sites where the motif occurs. The final case, with prior probability , is the co-occurrence of two sites in a sequence, which is . However, this is only correct when the sites are not overlapping since otherwise the overlapping nucleotides would be included in the product twice. Therefore we disallow overlaps.
For efficiency reasons, we do not calculate the score in its entirety. We assume that it is the strong sites that contribute the most to the equation and introduce a cutoff for each WM on the minimum score of a site. This enables an efficient search in the ESA. This is not without biological merit since WM scores and binding energies for known TFs are correlated, and at some point the binding energies of a TF and a poor binding sequence must be too small to matter [4]. It is also a standard method to use when scanning with known matrices [38]. So we only consider sites that score above a threshold, which is called for matrix . Then the probability of a sequence from the set being generated by the WMs is(3)where is the expectation over of over all predicted sites:(4)with being the step function (1 above 0 and zero otherwise). The co-occurrence expectation is defined in a similar way with overlaps disallowed. The effective weight of no sites(5)accounts for extra weight given to no sites due to alignments not meeting the threshold. With this definition, is the probability or generative model of the sequence conditioned on the WM and threshold, .
To find the WMs that best explain the difference in occurrence between the sets we use a discriminative objective function based on the probability of the labels given the sequences and WMs, formally:(6)
This is the logistic likelihood function for binary classification, see e.g. [39]. The discriminative model can thus be viewed as logistic regression with an adaptive set of basis functions. For multiple sequences assumed to be independent, the joint probability is the product of the single sequence probabilities over all sequences in both the positive and negative set:(7)
We refer to this function as the (log likelihood) score, .
Based on the sequence density we can use Bayes theorem to calculate the probability of the label given the WMs , the thresholds , and the sequence :(8)
We observe that the prior probability of is proportional to the number of sequences in the set divided by the total number of sequences .
A very high threshold will give no matches, and the probability will then be a constant given by the priors and the size of the two sets. Matches that score above the threshold in the negative set will lower the score and matches above the threshold in the positive set will increase the score, so the game is to obtain as many high-scoring matches in the positive set as possible without introducing too many matches in the negative set.
The prior is conservative in our runs in that we are strict about promoting hits in the positive set, but only moderately strict about disallowing negative hits. For a single matrix the prior on is 0.01; : 0.99; : 0.80; : 0.20. For two matrices: : 0; : 0.1; : 0.9; : 0.80; : 0.15; and : 0.05. These priors can be set by the user if prior knowledge is available about the set (i.e. a high confidence negative set or an uncertain positive set).
In the evaluation we deliberately chose a probability of having a site (0.5) in a sequence very different from the model prior (0.99) to avoid giving our own method a big advantage. It shows that the method is not very sensitive to the choice of prior.
The objective function outlined above is optimized using simulated annealing [40]. Informally, it proceeds by iteratively proposing a candidate solution and then accepting or rejecting it depending on how good it is compared to the current solution. It sometimes accepts changes for the worse and therefore possesses the power to escape local maxima. The hope is that it will converge on a solution that is close to optimal. Formally, this translates to a walk over the search space where in the current state , the next state is either the same or the candidate solution depending on their relative scores and a temperature parameter .
The temperature parameter is lowered for each iteration using as default an exponential cooling scheme (for details see Text S5), thus incrementally constraining the neighborhood of accepted changes.
Candidate solutions are proposed by applying one of several steps outlined in the list below. In the case of multiple matrices, only one is changed at a time. We perform all steps on a integer “count” matrix which is then translated into a log-odds WM prior to searching the ESA, but notice that the “count” matrix does not represent actual letter frequencies in the selected sites. The steps are:
Note that for the extend and decrease step there is a minimum and maximum number of columns for a motif. The default for these are 5 and 15 respectively.
The matrix is initialized with random counts and the cutoff is also selected uniformly according to the last step in the list above. Termination of the optimization is only based on the number of iterations which is by default set to a rather conservative value of 30 million iterations. Time requirements for a single run is variable depending on the set size, but was for our runs comparable to NestedMICA (single threaded) and considerably faster than Weeder's “large” run and DEME.
Source code as well as data sets is freely available at the author's web site: http://moan.binf.ku.dk
|
10.1371/journal.pmed.1002463 | Treatment eligibility and retention in clinical HIV care: A regression discontinuity study in South Africa | Loss to follow-up is high among HIV patients not yet receiving antiretroviral therapy (ART). Clinical trials have demonstrated the clinical efficacy of early ART; however, these trials may miss an important real-world consequence of providing ART at diagnosis: its impact on retention in care.
We examined the effect of immediate (versus deferred) ART on retention in care using a regression discontinuity design. The analysis included all patients (N = 11,306) entering clinical HIV care with a first CD4 count between 12 August 2011 and 31 December 2012 in a public-sector HIV care and treatment program in rural South Africa. Patients were assigned to immediate versus deferred ART eligibility, as determined by a CD4 count < 350 cells/μl, per South African national guidelines. Patients referred to pre-ART care were instructed to return every 6 months for CD4 monitoring. Patients initiated on ART were instructed to return at 6 and 12 months post-initiation and annually thereafter for CD4 and viral load monitoring. We assessed retention in HIV care at 12 months, as measured by the presence of a clinic visit, lab test, or ART initiation 6 to 18 months after initial CD4 test. Differences in retention between patients presenting with CD4 counts just above versus just below the 350-cells/μl threshold were estimated using local linear regression models with a data-driven bandwidth and with the algorithm for selecting the bandwidth chosen ex ante. Among patients with CD4 counts close to the 350-cells/μl threshold, having an ART-eligible CD4 count (<350 cells/μl) was associated with higher 12-month retention than not having an ART-eligible CD4 count (50% versus 32%), an intention-to-treat risk difference of 18 percentage points (95% CI 11 to 23; p < 0.001). The decision to start ART was determined by CD4 count for one in four patients (25%) presenting close to the eligibility threshold (95% CI 20% to 31%; p < 0.001). In this subpopulation, having an ART-eligible CD4 count was associated with higher 12-month retention than not having an ART-eligible CD4 count (91% versus 21%), a complier causal risk difference of 70 percentage points (95% CI 42 to 98; p < 0.001). The major limitations of the study are the potential for limited generalizability, the potential for outcome misclassification, and the absence of data on longer-term health outcomes.
Patients who were eligible for immediate ART had dramatically higher retention in HIV care than patients who just missed the CD4-count eligibility cutoff. The clinical and population health benefits of offering immediate ART regardless of CD4 count may be larger than suggested by clinical trials.
| Starting antiretroviral therapy (ART) at HIV diagnosis has health benefits relative to deferring therapy.
Understanding the magnitude of these benefits is important for countries expanding treatment eligibility.
Clinical trials may underestimate the real-world benefits of immediate ART, because in such trials, patients assigned to deferred ART are retained in care and carefully monitored to determine future ART eligibility and to assess trial endpoints.
Although clinical trials have shown the biological benefits of early ART, the behavioral impacts on retention in care are unknown.
We assessed the association between immediate versus deferred ART eligibility and clinical retention among 11,036 patients presenting for HIV care in rural South Africa, the country with the largest HIV epidemic and largest ART program worldwide.
Using a regression discontinuity design, we compared 12-month retention in care among patients presenting just above versus just below the 350-cells/μl CD4 threshold used to determine ART eligibility during the period of study.
We found that, among these patients, immediate eligibility increased initiation of ART by 25 percentage points and 12-month retention by 18 percentage points.
Among patients starting ART because they had an eligible CD4 count, 12-month retention was 91%; among patients prevented from starting ART because their CD4 count was above 350 cells/μl, 12-month retention was 21%.
Offering ART immediately at diagnosis improves retention in care, reducing the number of patients who screen positive for HIV but are then lost to follow-up before they can start ART.
The benefits of offering ART immediately at diagnosis, regardless of CD4 count, may be greater than previously thought.
| Mass provision of HIV treatment has improved life expectancy in southern Africa [1–3], yet HIV remains the leading cause of death and disability [4]. Recent clinical trials show health benefits to antiretroviral therapy (ART) at high CD4 counts [5–7]; WHO now recommends starting HIV patients on ART at diagnosis [8], and many countries have moved to “treat all” policies [9].
Although clinical efficacy has been demonstrated under trial conditions, the effect of immediate versus deferred ART in clinical settings in the “real world” is largely unknown. In addition to the direct health benefit demonstrated in trials [5–7], starting ART immediately also may reduce the burden of disease by retaining in clinical care patients who would otherwise be lost to follow-up. High rates of attrition have been observed among patients who are not yet eligible for ART and who ostensibly are being monitored for disease progression, leading to missed opportunities for counseling and timely initiation of therapy [10–24]. However, the extent to which immediate ART mitigates attrition is unknown. Clinical trials, designed to minimize attrition in both arms, do not observe this phenomenon and may therefore underestimate the benefits of immediate ART. Observational studies have documented lower retention among pre-ART patients compared to patients on ART [11,12], yet these differences could simply reflect the selection of more highly-motivated patients onto ART, rather than a causal effect of ART on retention in care.
In this study, we assessed the association between immediate (versus deferred) ART eligibility and clinical retention in a large public-sector treatment program in rural South Africa. Using a quasi-experimental regression discontinuity design [25–28], we compared retention for patients presenting with CD4 counts just above and below the 350-cells/μl eligibility cutoff used during the study period. Regression discontinuity can be used when a treatment is assigned, at least in part, based on a threshold rule, such as the CD4 eligibility cutoff for HIV treatment [28]. CD4 measurements have high within-individual variation [29] due to laboratory instrument imprecision, sampling variability in blood draws, and random factors such as ambient temperature at the time of the blood draw. Due to random noise in measured CD4 counts, patients just above and below the cutoff are similar on both observed and unobserved factors, but are assigned to different exposures [28]. At the threshold, outcomes are observed in both counterfactual states of the world (eligible/not eligible), and comparisons have a causal interpretation [30]. Whereas most observational studies rely on strong assumptions about unobserved confounders, regression discontinuity can achieve balance by design, similar to a randomized controlled trial, and therefore enables causal inferences without strong assumptions [28]. This natural experiment provides a unique opportunity to assess the impact of immediate versus deferred eligibility for HIV treatment in a real-world clinical setting.
Ethical approval for data collection and analysis was obtained from the University of KwaZulu-Natal Biomedical Research Ethics Committee. The research in this paper consisted of secondary analysis of preexisting de-identified data and was determined to be “not human subjects research” by the Boston University Medical Campus Institutional Review Board (H-35385, “Analysis of the HIV cascade of care in rural South Africa: A secondary data analysis”).
The study population for this analysis consisted of all patients in the Hlabisa HIV Treatment and Care Programme (Hlabisa Cohort) [31] whose first CD4 count specimen was collected between 12 August 2011 and 31 December 2012. The Hlabisa HIV Treatment and Care Programme is a collaboration between the Africa Health Research Institute (https://www.ahri.org) and the South African Department of Health. The Hlabisa Cohort includes all patients receiving HIV care and treatment services at government facilities (17 clinics and 1 hospital) in Hlabisa sub-district, a poor, largely rural area where 1 in 3 adults is HIV-infected [32]. Patients initiating ART prior to their first CD4 count were excluded from the study.
Data on CD4 counts, viral loads, dates of ART initiation, and routine HIV clinic visits were obtained for all members of the study population. The Hlabisa HIV Treatment and Care Programme collects data on CD4 counts for patients who have not yet initiated ART, including patients who never initiate ART. Patients entered the study on the date when their first CD4 count specimen was collected for lab testing, typically the date of HIV diagnosis. Test results were transferred directly from the laboratory into the Hlabisa Cohort database. All patients were eligible to be followed for at least 12 months. Follow-up was closed on 1 January 2014.
The primary outcome was 12-month retention in care, which was defined as evidence of any routine clinic visit, lab result (CD4 or viral load), or date of ART initiation within the interval 6 to <18 months after a patient’s first CD4 count, regardless of receipt of ART. By South African national guidelines, all patients would be expected to have at least 2 documented lab tests within this period (Appendix A in S1 Appendices). Although guidelines delineate semi-annual laboratory monitoring, the wide interval allows for the fact that many patients were late for appointments but still retained in care. Results were robust to narrower intervals.
The primary outcome was designed to capture all clinical contact specified by the national guidelines for pre-ART and ART care. Patients on ART had more opportunities to appear and be classified as retained due to their greater frequency of scheduled clinic visits. In sensitivity analysis, we excluded routine clinic visits and counted patients as retained at 12 months only if they had a CD4 or viral load test or initiated ART during the period 6 to <18 months after their first CD4 count.
As a secondary outcome, we assessed the presence of a CD4 or viral load test or ART start date within 6-month intervals following a patient’s first CD4 count, out to 2 years (0 to <6, 6 to <12, 12 to <18, and 18 to <24 months). Because some patients do not return for lab tests precisely every 6 months, these 6-month intervals will underestimate the proportion of patients retained. However, these analyses may inform how retention evolves over time. Analyses of retention in care out to 18 and 24 months were constrained to the subpopulations observed for that amount of time, i.e., patients presenting before 2 July 2012 (18 months follow-up) and 1 January 2013 (24 months follow-up).
Per South African guidelines during the study period, patients were ART eligible if their CD4 count was <350 cells/μl and/or they had a WHO stage III/IV condition [33]. After blood was drawn for a CD4 count, all patients were instructed to return to the clinic in one week to receive their result. ART-eligible patients were enrolled in several weeks of individual and group counseling and then initiated on ART. At ART initiation, patients were instructed to return for monthly clinic visits to pick up their medication and at 6 and 12 months post-initiation (and annually thereafter) for CD4 count and viral load monitoring. Patients not yet eligible for ART were referred to pre-ART care and were instructed to return every 6 months for CD4 monitoring [31].
Based on these policies, we defined two exposures. First, we defined ART eligibility as having a CD4 count below 350 cells/μl. As our second exposure, we defined ART uptake as initiation of therapy within 6 months of a patient’s first CD4 count. Not all patients who had an eligible CD4 count went on to initiate ART: some did not return for their CD4 count results, and others did not complete the counseling sessions even after eligibility was determined. Conversely, some patients with CD4 counts at or above 350 cells/μl initiated ART on account of stage III/IV HIV illness or due to provider discretion. Thus, results for our primary exposure—an ART-eligible CD4 count—have an intention-to-treat (ITT) interpretation. We defined ART uptake at 6 months because, by national guidelines, patients who did not start ART within 6 months had another CD4 count to determine eligibility.
To determine the effect of immediate versus deferred ART eligibility on retention, we used a quasi-experimental regression discontinuity design. Regression discontinuity can be implemented when a treatment is assigned based, at least in part, on a threshold rule on a continuous assignment variable [27,28,34–36]. Though commonly used in economics [37–39], regression discontinuity has only recently made inroads in epidemiology and clinical research [26–28,40–42]. Because of random measurement error in the CD4 count laboratory assay [29], assignment to immediate versus deferred treatment is effectively random for those patients with CD4 counts near 350 cells/μl. As such, comparisons of outcomes “just above” and “just below” this threshold have a causal interpretation (see Appendix B in S1 Appendices).
Our analytic strategy, which was based on a preexisting, single, well-known clinical practice threshold, followed best practices for the conduct and reporting of regression discontinuity designs [37,38,40,43–45]. Our primary analysis tested the null hypothesis, determined a priori, that immediate (rather than deferred) ART eligibility would have no effect on retention in care. We evaluated the relationship between the value of a patient’s first CD4 count and retention in care, allowing for a discontinuity at the threshold of 350 cells/μl and different slopes on either side of the threshold. Risk differences at the threshold were estimated using local linear regression with a data-driven Imbens-Kalyanaraman bandwidth and a rectangular kernel. We assessed robustness of the results to a wide range of alternate bandwidths, following the literature (see Appendix B, pp. 3–4, in S1 Appendices) [45,46]. Results from these models are ITT effects, i.e., differences in retention for patients assigned to immediate versus deferred treatment eligibility by their CD4 count. The data-driven bandwidth selector chooses the bandwidth that minimizes the mean squared error of the difference in predictions at the threshold (i.e., the ITT effect). The goal is to identify as large a region as possible in which the conditional expectation function (relationship between CD4 count and retention) is approximately linear. The more data included, the less random error in the prediction at the threshold, but also the greater the potential for bias if the relationship is in fact nonlinear [45]. Perhaps the greatest advantage of using a data-driven bandwidth selector is that we eliminate the ability for the investigator to manipulate the results by choosing a “preferred” bandwidth. All models were estimated using a rectangular kernel, i.e., weighting observations within the window of data equally. Additionally, we estimated local logistic regression models and estimated predicted margins at the threshold. Because optimal bandwidth selectors are not currently available for logistic regression, we used the same bandwidth as for the local linear model.
We also assessed the association between immediate versus deferred eligibility and 6-month uptake of ART and used this analysis to estimate the share of patients for whom the decision to initiate ART was based on the eligibility of their CD4 count (so-called compliers) as opposed to disease stage or other factors [47]. Using an instrumental variables approach [37,48], we then estimated the effect of ART uptake on 12-month retention for compliers, using CD4 count < 350 cells/μl as an instrument for ART uptake. Under the assumption that having an eligible CD4 count affected 12-month retention only through uptake of ART, these instrumental variables estimates can be interpreted as the causal effect of ART initiation on retention among compliers (see Appendix B, pp. 5–7, in S1 Appendices). We estimated complier causal risk differences, also known as complier average causal effects (CACEs) or local average treatment effects (LATEs), using 2-stage least squares regression. We additionally estimated proportions of patients retained among patients who started ART because they were eligible (so-called treated compliers) and among patients who did not start ART because they were ineligible (so-called control compliers), and estimated complier causal relative risks as the ratio of the treated and control complier proportions [49].
The validity of the regression discontinuity design rests on the assumption that other patient characteristics that may influence retention are similar for patients with CD4 counts just above and below 350 cells/μl. To evaluate covariate balance, we assessed whether observed factors (age, sex, date of presentation, and clinic of presentation) were similar on either side of the threshold. Bias can also result if patients or providers manipulate CD4 count values in order to gain access to treatment. To assess systematic manipulation, we tested for heaping of CD4 counts on one side of the threshold [44]. All analyses were conducted using Stata/SE version 14.2. This report has been prepared according to STROBE guidelines, recommended by the Enhancing the QUAlity and Transparency Of health Research (EQUATOR) network (S1 STROBE Checklist). A de-identified analytic dataset (S1 Data) and replication code (S1 Code) are available.
A small proportion of patients (4.1%) initiated ART prior to their first CD4 count and were excluded from the study. The remaining sample included 11,306 patients who entered care with a first CD4 count between 12 August 2011 and 31 December 2012. Of these, 6,225 had CD4 counts < 350 cells/μl and 5,082 had CD4 counts ≥ 350 cells/μl. Baseline characteristics were similar just above and below the CD4 count threshold; about 70% of patients were women, and the average age was 30 years (Table 1). We found no evidence of systematic manipulation of CD4 count values around the threshold (Fig 1).
Among patients presenting with first CD4 counts close to 350 cells/μl, immediate ART eligibility (CD4 count < 350 cells/μl) was associated with higher 12-month retention in care, with a difference of 18 percentage points (95% CI 11 to 23), relative to deferred eligibility. Fifty percent of patients with immediate eligibility were retained at 12 months compared to 32% of patients with deferred eligibility, a 56% relative increase with eligibility (Fig 2; Table 2). In sensitivity analyses defining retention in terms of lab results or ART start dates (and excluding clinic visits), estimates were somewhat attenuated, with a difference in 12-month retention of 11 percentage points (95% CI 4 to 18) (Table 2; Fig C1 in S1 Appendices). A gap in retention was observed at all 6-month intervals from first CD4 count to 24 months (Table 2; Fig C2 in S1 Appendices), suggesting that loss to follow-up among patients not yet eligible for ART occurred soon after their initial clinic visit. For all retention outcomes, results were robust to a wide range of alternate bandwidths (Tables D1–D5 in S1 Appendices). Similar results were observed using a logistic rather than linear probability model (Table D6 in S1 Appendices).
Turning to uptake of ART, patients immediately eligible for ART were 25 percentage points (95% CI 20 to 31, p < 0.001) more likely to initiate ART within 6 months (Table 2) than those not yet eligible for ART, rising from 18% initiating ART among patients just above the threshold to 43% initiating ART among patients just below the threshold (Fig 3). Even among patients with an eligible CD4 count, a majority (57%) did not initiate ART within 6 months. Fig C3 in S1 Appendices shows that having an eligible CD4 count had no effect on initiation within the first 2 weeks, consistent with treatment guidelines. The gap in ART uptake apparent at 6 months persisted at 12 months. These results imply that among patients with CD4 counts close to 350 cells/μl, 18% would have initiated ART regardless of CD4 eligibility (so-called always takers), 57% would not have initiated ART regardless of CD4 eligibility (so-called never takers), and 25% of patients would have initiated ART if CD4 count < 350 cells/μl, but not if CD4 count ≥ 350 cells/μl (so-called compliers [48]) (Fig 3). Our study population was followed from the day their first CD4 count was taken, typically the day of diagnosis, regardless of whether patients returned for their results. “Never takers” therefore include patients who tested positive and never came back.
The results for ART uptake reveal that the ITT effect of CD4 eligibility on retention was substantially diluted by noncompliance—the decision to start ART was based on CD4 count for just 1 in 4 patients presenting close to the threshold. The ITT effect thus underestimates the effect of ART uptake itself on retention in care. Our instrumental variables results for compliers revealed that patients who initiated ART because they had an eligible CD4 count were 70 percentage points (95% CI 42 to 98) more likely to be retained at 12 months than patients who were not initiated because they had an ineligible CD4 count. Retention at 12 months was 91% for compliers assigned to immediate eligibility and 21% for compliers assigned to deferred eligibility. Among compliers, immediate eligibility reduced attrition by 89% (complier causal relative risk = 0.11, 95% CI 0.00–0.32) (Table 3; Table D7 in S1 Appendices).
Understanding the extent to which immediate ART initiation mitigates loss of health and life due to failure to remain in care is important for countries and funding agencies considering WHO recommendations to start patients immediately on therapy regardless of CD4 count [8,9]. Although lower retention has been observed in pre-ART patients compared to ART patients in a wide range of settings, it was hitherto unknown whether this reflected a causal relationship or selection of highly motivated patients onto ART [11,12]. If starting ART causally increased retention in care, then the real-world benefits of immediate ART would be underestimated in clinical trials that actively retain patients not yet eligible for therapy [5–7].
Using a quasi-experimental regression discontinuity design, we found that having an ART-eligible CD4 count at diagnosis significantly improved retention in care for HIV patients in rural South Africa—by 18 percentage points in the ITT analysis. The ITT effect was diluted by the fact that the decision to start ART was based on CD4 count for only a minority of patients (just 1 in 4) presenting for care. (Other patients started ART due to disease stage, while others did not start ART in spite of having an eligible CD4 count.) Among patients whose treatment decision was based on their CD4 count, immediate ART eligibility increased 12-month retention by 70 percentage points relative to deferred eligibility, from 21% to 91% retained.
The retention advantage for ART-eligible patients is perhaps surprising. Clinical guidelines specified that patients who did not start ART should return for CD4 monitoring every 6 months to reassess eligibility. One might expect that patients who are motivated to start lifelong ART would tolerate a 6-month delay without exiting care. Further, because the lifetime benefits of ART are greater the earlier a patient starts therapy, there is a strong rationale to stay in pre-ART care in order to initiate as soon as possible. Conceivably, the retention advantage could even favor pre-ART patients as some patients who initiate ART subsequently exit care after experiencing drug side effects and the inconvenience of daily therapy. With respect to the burdens imposed on patients, it might be easier to retain patients on a holding regimen of semi-annual pre-ART appointments than to retain patients on an intensive daily drug therapy. Patients who do not start ART also might be more likely to get sick and need care 12 months later, leading to higher retention in this group.
Contrary to these speculations, immediate eligibility for ART sharply increased retention in care. There are several plausible explanations for the observed results. First, starting ART may shift patients’ cognitive reference point for future care-seeking decisions [50]. Patients who have started ART may perceive large costs to defaulting therapy, compared to the more modest costs of delaying initiation among those who have not yet started therapy. These perceptions may be reinforced by clinical guidance to patients that starting ART involves a commitment to take treatment for life [31]. Second, taking daily therapy and returning to the clinic for monthly prescription refills may facilitate habit formation [51], increasing long-term retention in care. Third, auxiliary interventions targeted to ART patients including adherence counseling, support groups, appointment reminders, and outreach to patients who miss appointments may lead to differential retention between patients on ART and those not yet on ART [52,53]. Fourth, in the context of still-rampant HIV stigma, fear of HIV status disclosure may be a significant barrier to care-seeking among patients not yet on ART [54,55]. Many clinics strongly encourage patients to disclose their HIV status to close family and friends at the time of ART initiation [31]. For ART patients, HIV status disclosure may represent an up-front investment that reduces the costs of future clinic visits. In addition to the behavioral mechanisms above, a fifth possibility is that patient health and quality of life improve due to the antiretroviral drugs themselves [56] and that experiencing these benefits encourages patients to remain in care.
In addition to the benefits of immediate ART eligibility, it is also possible that patients who are told that they are not yet eligible for ART may be inadvertently discouraged from seeking care in the future. The message of deferred eligibility may falsely signal to patients that they do not need or would not benefit from ART. They may also experience anger, hopelessness, or demoralization if they wish to start therapy but are not allowed to, and these experiences may color their attitudes towards the health system and future care-seeking.
Our results highlight the challenges in retaining in care those patients who test positive but are not yet eligible for ART in a resource-limited setting [57]. And yet, the high rates of retention among patients initiated on ART because they were eligible suggest that, in fact, we already have effective techniques to improve retention among patients who have not yet started ART. Further research is needed to identify what specific aspects of initiating ART lead to improved retention in care. These findings may be important for retaining patients who do not wish to start ART on the day of diagnosis, as is now called for in 2017 WHO guidelines [58].
Our results were obtained using a regression discontinuity design, a quasi-experimental study design that enables causal inference without the strong assumptions required in most observational studies [26–28,35,40,43,59–61]. So long as values of CD4 measurements are not systematically manipulated by patients or providers, random variability in measured CD4 counts guarantees that patients will be similar (in expectation) in a small range on either side of the 350-cells/μl eligibility threshold [30,35]. We obtained CD4 counts directly from laboratory records (rather than clinical charts) and found no evidence of manipulation, which, if systematic, would lead to heaping in the density of CD4 counts on one side of the threshold [44]. Additionally, there were no systematic differences in observed baseline covariates between patients just above versus below the cutoff. Although we cannot test the assumption that unobserved factors are balanced at the cutoff, our knowledge of the assignment mechanism, the absence of systematic manipulation, and balance on observed characteristics all point to a data-generating process in which quasi-random variation guarantees balance on all factors, similar to a randomized trial.
As with all regression discontinuity designs, causal effects are theoretically identified at the threshold (i.e., in the limit, as the 350-cells/μl cutoff is approached from above and below). In finite samples, however, these causal effects must be estimated using data further from the threshold. We followed current best practice in using local linear regression to estimate the empirical relationship between measured CD4 count and the probability of retention, allowing for an intercept shift at the threshold and different slopes on either side of the threshold [46,62]. The intercept shift at the threshold—i.e., the difference in regression predictions just above and just below 350 cells/μl—estimates the causal effect at the threshold. When using a local linear model to approximate a potentially nonlinear relationship, a key choice is the bandwidth governing the window of data used in the analysis. While a larger bandwidth will increase the precision of the estimates, this may come at the cost of some bias. We used the data-driven Imbens-Kalyanaraman optimal bandwidth selector, which minimizes the mean squared error (variance plus squared bias) of the difference in predictions at the threshold [45].
In many classical applications of the regression discontinuity design, the assignment variable is associated with the outcome and with the size of the treatment effect. Interestingly, retention was not substantially correlated with first CD4 count in our sample, and slopes were similar on either side of the threshold, evidence that treatment effects may be constant, at least within a range around the 350-cells/μl cutoff. Because we did not know this ex ante, we nevertheless present results from models allowing for different slopes and constrain inferences to the area around the cutoff. By choosing a local effect estimand, our estimates do not rely on extrapolation into unobserved regions nor on assumptions about the functional form of the relationship between CD4 count and retention across the full range of the data [35].
In addition to estimating the ITT effect of ART eligibility on retention, we also estimated the effect of ART initiation itself on retention in care using the eligibility threshold as an instrumental variable. These estimates are interpretable as the effect of starting ART for so-called compliers, i.e., those patients for whom the decision to start (or not to start) ART was based on the value of their CD4 count vis-à-vis the 350-cells/μl threshold. These instrumental variables estimates have a causal interpretation under two additional assumptions. The first assumption, known as the monotonicity or “no defiers” assumption, is that having an eligible CD4 count only increases the chances that a person will start ART. Monotonicity would be violated if there are patients who would start ART if they were ineligible (CD4 ≥ 350 cells/μl) but would not start ART in a counterfactual world in which they had an ART-eligible CD4 count. It is difficult to conceive of such cases, and this assumption is likely met in our study. The second assumption, known as excludability, is that eligibility differences at the 350-cells/μl threshold affect retention only through ART uptake. This untestable assumption could be violated if eligibility led to other differences in care apart from ART initiation (e.g., screening for other conditions or pre-ART counseling) that led to increases in later engagement with care. Monotonicity and excludability assumptions are not required for a causal interpretation of the ITT result.
A key strength of regression discontinuity designs (vis-à-vis clinical trials) is the ability to assess the causal effects of interventions implemented in real-world settings and in population-representative samples [63]. We studied the complete patient population accessing public-sector HIV care and treatment in one of the poorest and highest HIV-prevalence sub-districts in South Africa. Although our analysis was limited to one sub-district of one country, potentially limiting generalizability, the model of service delivery—decentralized, nurse-led, clinic-based—is common in other HIV-endemic areas of sub-Saharan Africa. Additionally, by including the complete patient population, we avoided the sample selection bias that can result from opt-in participation in clinical trials [64].
Our retrospective analysis of a quasi-experiment avoided many of the potential pitfalls that can lead to bias when investigators knowingly or unknowingly affect outcomes. First, the CD4 count threshold we investigate was set by policy-makers in advance of the study and could not be manipulated by the investigators. Second, patients, providers, and investigators were all blinded to the CD4 count measurement (and hence eligibility status) of the patients at the time when blood was drawn for the patients’ first CD4 count. Because we obtained the CD4 results for all blood samples directly from the National Health Laboratory Service, it would have been very difficult for eligibility assignment to be manipulated. Third, the data that we analyzed were collected as part of routine laboratory and clinical monitoring of patients in the Hlabisa HIV Treatment and Care Programme, and thus were not vulnerable to Hawthorne effects or other investigator biases in collection. Fourth, our analytic approach—local linear regression using a data-driven optimal bandwidth—is a theory-driven and widely used best practice in the conduct of regression discontinuity studies [37,38,40,43–45] and was decided on a priori. By using a data-driven bandwidth selector, we eliminated an opportunity for the investigator to manipulate the results by choosing a “preferred” bandwidth. By choosing local linear regression a priori, we avoided investigator discretion in the choice of functional form. Finally, following guidelines for the conduct of regression discontinuity studies, we assessed the data for evidence of manipulation of the assignment variable and conducted tests for balance at the threshold on all baseline characteristics observed and available in the dataset. We found no evidence to suggest that patients were dissimilar just above and below the treatment threshold.
Our study has some limitations. One limitation is that, as discussed above, our local linear regression results may be biased if the relationship between earliest CD4 count and retention in care is nonlinear near the threshold. Nonlinearities were taken into consideration when choosing the bandwidth (the window of data) for the model. Additionally, visual inspection of our figures suggests that in fact the relationships were approximately linear and that any bias resulting from using a linear model would be very small relative to the size of the effect estimates. Our results were also robust to using smaller bandwidths.
A second limitation involves the generalizability of our estimates to patients presenting with different CD4 counts and to different patient populations. As with any regression discontinuity design [26,28,36,41], our results are interpretable as causal effects for patients presenting with CD4 counts close to the 350-cells/μl eligibility threshold. If treatment effects differed across CD4 counts, then our results would not be directly informative of effects at higher (or lower) CD4 counts. Although this is a limitation, it is likely that our estimates are broadly generalizable to other points in the CD4 count distribution. The probability of retention changes little with the value of the patient’s first CD4 count, except at the threshold, and prior analysis showed similar effects at the 200-cells/μl eligibility threshold used prior to August 2011 [42]. Our analysis was also limited to one sub-district of one country, and it is unknown whether the results generalize to other settings.
A third limitation regards the difficulty of measuring retention. There are competing definitions in the literature [65]. In our primary specification, we classified patients as retained if they had any routine contact with the clinic within 6 to <18 months after first CD4 count, including visits to pick up ART medication, laboratory tests (CD4 or viral load), or an ART initiation date, all of which are specified as elements of routine pre-ART or ART care. As a robustness check, we defined an alternate measure of retention based only on laboratory tests and dates of ART initiation. We note that by excluding routine visits, this measure underestimates retention among ART patients and should be interpreted as a lower bound.
A fourth limitation of our analysis is that we were unable to assess longer-term health outcomes that may result from poor retention. In previous analysis, we found large survival benefits of immediate ART eligibility for patients presenting with CD4 counts near the former eligibility threshold of 200 cells/μl [28,42]. However, it is unknown whether these survival benefits extend to patients presenting at higher CD4 counts. We observed significant gaps in retention at 18–24 months. Patients whose CD4 counts are not actively monitored for treatment eligibility may initiate long after their CD4 count falls below the eligibility threshold, or they may not initiate at all. Extended treatment delays have consequences not only for patients themselves but also for population health, with increased potential for onward transmission [66]. Further research will be needed to determine the real-world impacts of deferred ART at higher CD4 counts on long-term engagement with care, health, survival, and onward transmission.
International treatment guidelines are informed (largely) by clinical trials [8,47,58,67], which typically differ from clinical care in the “real world” in important dimensions. One of these dimensions is retention in care. Clinical trials usually seek to retain patients in all treatment arms through systematic monitoring and outreach efforts. For example, in the TEMPRANO trial, 30-month retention was 97% in both arms [6]. The parity of retention across arms in these trials stands in stark contrast to the very large gap—91% versus 21%—we observed in a non-trial setting (Fig 4). Efforts to minimize attrition improve the validity of inferences on clinical endpoints; however, trials cannot then observe the effect of the intervention on retention in care, nor any downstream health impacts that are mediated through retention [5–7,68,69]. In this and other applications, the gap between clinical efficacy (as demonstrated in clinical trials) and real-world effectiveness may turn on the nature of the relationship between the intervention and retention of patients in care—a question of patient behavior, not biology.
Prior observational studies have documented higher retention among ART patients than among pre-ART patients [11,12,23,24]; however, this association is difficult to interpret due to potential selection bias. If highly motivated patients were more likely to initiate ART—leaving a residual of less motivated patients in pre-ART care—then a policy expanding ART eligibility may simply shift patients with low motivation from pre-ART to ART, leading to low retention and poor outcomes for newly eligible patients on treatment. On the other hand, if initiating patients on ART causally increases retention in care, then immediate ART eligibility would improve retention, leading to even larger benefits than observed in clinical trials. Our results provide evidence, to our knowledge for the first time, to distinguish between these two competing hypotheses. We demonstrate a large difference in retention between pre-ART and ART patients, causally attributable to starting ART itself.
The gap in retention observed in this study would be eliminated if patients were eligible for ART regardless of CD4 count, as under test-and-treat scenarios now being implemented in many countries. We caution that some patients may have little interest in initiating therapy even if eligible [59]. Our results are not informative about the impact of immediate therapy for this group, and increasing demand for ART in such patients may be a challenge. Nor do our estimates generalize to the smaller group of patients who start ART for other reasons (e.g., disease stage) regardless of CD4 count. But for those patients currently barred from initiating due to an ineligible CD4 count, we show that a guideline change allowing immediate initiation could dramatically increase retention in care. Early WHO guidelines for HIV were designed to prioritize the sickest for ART, and there has been concern that expanding eligibility would inappropriately target resources to patients with little incentive to remain on therapy [70]. Our results suggest that expanding eligibility would target patients who are both high need (only 21% would be retained if not eligible) and high performing (91% would be retained if eligible). Countries such as South Africa [9], which have now removed CD4 thresholds, can be encouraged that such a policy will be an efficient step towards expanding HIV treatment coverage.
Clinical trials have demonstrated the biological efficacy of early ART [5–7]. Effects on patient retention, however, cannot be observed in trials that minimize attrition by design. Our study demonstrates, to our knowledge for the first time, the retention effects of early ART: denying ART eligibility to patients who would be willing to start therapy leads to very large losses from HIV care, losses that would be avoided with immediate ART. Our results thus indicate that the real-world benefits of extending ART eligibility to all patients, regardless of CD4 count, may be larger than previously thought.
|
10.1371/journal.pntd.0006159 | Barriers of attendance to dog rabies static point vaccination clinics in Blantyre, Malawi | Rabies is a devastating yet preventable disease that causes around 59,000 human deaths annually. Almost all human rabies cases are caused by bites from rabies-infected dogs. A large proportion of these cases occur in Sub Saharan Africa (SSA). Annual vaccination of at least 70% of the dog population is recommended by the World Health Organisation in order to eliminate rabies. However, achieving such high vaccination coverage has proven challenging, especially in low resource settings. Despite being logistically and economically more feasible than door-to-door approaches, static point (SP) vaccination campaigns often suffer from low attendance and therefore result in low vaccination coverage. Here, we investigated the barriers to attendance at SP offering free rabies vaccinations for dogs in Blantyre, Malawi. We analysed data for 22,924 dogs from a city-wide vaccination campaign in combination with GIS and household questionnaire data using multivariable logistic regression and distance estimation techniques. We found that distance plays a crucial role in SP attendance (i.e. for every km closer the odds of attending a SP point are 3.3 times higher) and that very few people are willing to travel more than 1.5 km to bring their dog for vaccination. Additionally, we found that dogs from areas with higher proportions of people living in poverty are more likely to be presented for vaccination (ORs 1.58-2.22). Furthermore, puppies (OR 0.26), pregnant or lactating female dogs (OR 0.60) are less likely to be presented for vaccination. Owners also reported that they did not attend an SP because they were not aware of the campaign (27%) or they could not handle their dog (19%). Our findings will inform the design of future rabies vaccination programmes in SSA which may lead to improved vaccination coverage achieved by SP alone.
| Rabies is a devastating yet preventable disease that causes around 59,000 human deaths annually of which a large proportion occurs in Sub Saharan Africa (SSA). In order to eliminate rabies, annual vaccination of at least 70% of the dog population is recommended. In SSA most rabies vaccination programmes use static point (SP) vaccination approaches. Despite being logistically and economically more feasible than door-to-door approaches, SP vaccination campaigns often result in low vaccination coverage. Here we investigated the reasons why attendance at SPs offering free rabies vaccinations for dogs is suboptimal in SSA. We analysed data from a citywide vaccination campaign in Blantyre city, Malawi in combination with household related data. Our results found that the distance from home to SP influences attendance at SPs. We also found a clear need for provision of timely and accurate information about upcoming campaigns, including information on the importance of puppies being vaccinated as well as ways to improve dog handling. Understanding the barriers to attendance at SPs and taking them into consideration, would make mass vaccination programmes more feasible thereby allowing high vaccination coverage to be achieved without the need for expensive and logistically challenging door-to-door programmes.
| Rabies has been estimated to cause around 59,000 human deaths per year [1]. Globally, rabies has been estimated to cause 3.7 million disability-adjusted life years and 8.6 billion US dollars economic losses annually [1]. Almost all human rabies cases are acquired from contact with rabies infected dogs [2]. Case fatality for patients who develop clinical signs related to rabies infection approaches 100% and successful treatment has rarely been reported [2]. Rabies disproportionately affects Sub Saharan African countries [1, 2]. Despite significant regional and international healthcare intervention initiatives, no African country has been reported rabies free to date [3].
Since 99% of all human rabies deaths are caused by bites from rabies infected dogs [2], mass dog vaccination campaigns are the single most effective strategy to eliminate rabies amongst humans and dogs [1, 4, 5]. To effectively eliminate rabies from canine and human populations, a critical requirement of mass dog vaccination programmes is to ensure that a sufficiently high proportion of dogs are vaccinated [6]. Empirical data has shown that annual vaccination coverage of 70% is sufficient to eliminate rabies from dog and human populations [6, 7]. This has been further validated by mathematical modelling [8]. For example, mathematical models have demonstrated that a cut-off of 70% would prevent a major disease outbreak at least 96.5% of the time based on rabies field data from USA, Mexico, Malaysia and Indonesia [8]. Collectively, these findings have resulted in the recommendation by the World Health Organisation (WHO) that rabies vaccination programmes should vaccinate at least 70% of all dogs annually [6, 7, 9, 10].
However, vaccinating large numbers of dogs at over 70% coverage has proved challenging despite the development of a range of mass rabies vaccination strategies [11]. Vaccination approaches which have been used include door-to-door campaigns (D2D); static point (SP) campaigns, using both fixed and temporary posts; and a combination of the two. Door-to-door programmes, which typically achieve a high vaccination coverage, are labour intensive, expensive and challenging to roll out on a large scale. Consequently, most rabies vaccination programmes in Sub Saharan Africa (SSA) have used SP vaccination approaches where the vaccination teams remain at a static location within a community and the local inhabitants present dogs to the vaccination teams. Although widely used in Africa as they are logistically and economically more feasible than door-to-door approaches, SP approaches have often failed to reach a high coverage [9, 11–14]. Consequently, many organisations have to manage a trade off when rolling out dog vaccination programmes of either utilising a door-to-door approach, which typically results in high coverage but lower numbers of dogs vaccinated, or a SP approach, which often achieves a lower coverage but facilitates the vaccination of more dogs.
This trade-off between coverage and number of dogs vaccinated would be eliminated if a higher proportion of dogs could be vaccinated through SPs. The reasons why attendance at SPs is low in many countries has surprisingly received little attention despite it being a major reason why rabies elimination programmes have been so challenging to effectively roll out in SSA. If the barriers to attendance at SPs could be understood and then overcome, mass vaccination programmes could become more feasible thereby allowing high vaccination coverage to be achieved without the need for expensive and logistically challenging door-to-door programmes.
Only a small number of studies have explored why attendance to SPs is often suboptimal. In previous small scale studies in Chad, Mali, Peru and urban Tanzania the most common reasons reported by dog owners for not attending a static vaccination point included; lack of information about the campaign [14–18], difficulty in handling dogs [13–15, 17, 18], lack of time [14, 17, 18], lack of information about rabies [15], mistrust [15], distance/location of SP [13, 15, 18], the dog being too young [16–18] or lactating [17, 18] and the lack of money to pay [14, 16]. In campaigns using a combination of SPs and D2D vaccination strategies, it is also possible that owners do not attend SPs as they expect to get their dogs vaccinated during the door-to-door campaign [16]. Furthermore, while giving out dog collars or wristbands can increase participation [18], charging the owners for vaccinations can result in lower vaccination rates [16]. Despite these studies, there is still an incomplete understanding of the barriers which limit attendance at static points.
The need to understand and overcome barriers to SP attendance is particularly important in Blantyre, Malawi where rabies is an important cause of mortality, especially in children [19]. In order to address the high incidence of rabies in this population, we have embarked on an annual mass dog vaccination campaign throughout the city. We have previously reported that although 97% of the dog population is owned, only 53% out of the 79% overall vaccination coverage we achieved in our 2015 vaccination programme was attributed to dogs vaccinated at a SP, with the remaining 26% achieved by vaccinating dogs at door-to-door [20].
In order to make the vaccination campaign financially sustainable in the longer term, we need to reduce the reliance on D2D vaccination and encourage higher attendance at SP vaccination stations. Consequently, the aim of this study was to investigate the barriers to attendance at SP vaccination clinics using a multi-faceted approach including modeling the relationship between distance to travel and attendance at SP together with dog owner questionnaires. Our study is the first large scale, city-wide study investigating the reasons for failure to attend static vaccination points in SSA.
Prior to vaccination of owned dogs, verbal informed consent was obtained from the person presenting the dog for vaccination. In the cases where an owner could not be identified, dogs were vaccinated in accordance with Government Public Health protocol, as the work was part of a non-research public health campaign.
This study was conducted in Blantyre city, the second largest city in Malawi with an estimated human population of 881,074 in 2015 [21]. The city’s dog population in 2015 was estimated to be 45,526 based on mark re-sight methods [20]. The city covers an area of 220 km2, which is divided into 25 administrative wards [22]. The campaign took place throughout the whole of Blantyre city.
The vaccination campaign has been described in detail by Gibson et al. [20]. Briefly, the city was divided in 204 working zones and their sizes were subjectively dictated according to an area that could be covered by a vaccination team in one day. Each zone was assigned a land type based on appearance in Google Satellite Maps™: a) housing category (HS) 1 (small houses—high density), b) HS 2 (small houses—medium density), c) HS 3 (small houses-low density), d) HS 4 (medium houses—ordered), e) HS 5 (large houses-medium/low density), f) industrial/commercial, g) agriculture/open space. For the purposes of the regression analysis described below these were regrouped in high (a), medium (b,d) and low (c,e,f,g) housing density areas.
Mass dog vaccination across the city was carried out between the 30th of April and the 25th of May 2016 using two approaches; static point (SP) and door-to-door (D2D). Using 8 vaccination teams working simultaneously, SP vaccinations were conducted at weekends followed by D2D vaccinations in the same area on the following Monday, Tuesday and Wednesday.
All data analysis was carried out within the R statistical software environment [24]. Specific packages used are mentioned below.
22,924 dogs recorded during the door-to-door campaign were used for this analysis. This excluded 700 dogs that had been vaccinated by someone other than MR earlier that year. 91% included in the analysis were adults and 10% were neutered. 40% of dogs were female out of which 13% were either pregnant or lactating. The vast majority of dogs were owned, out of which 24% were recorded as always roaming, 34% as roaming daily but restricted at some point during the day, less than 1% as roaming weekly and 33% as never roaming.
10,476 dogs (0.46) were reported to be vaccinated at SP locations and for 6,271 of those we could retrieve their unique SP location ID. 4,756 (0.76) of those went to their nearest SP location, while 994 (0.16) went to their second nearest SP. Fig 1 shows the location of the SPs used in this campaign, as well as household to SP paths drawn for people who presented their unique SP ID number during the door-to-door campaign. Lastly, there was no statistical evidence of a relationship between distance to the nearest SP and missingness of vaccination card evidence.
Based on the Google Maps route the distance attending dog owners were willing to travel to a SP vaccination clinic was on average 1.22 km with 75% of attending dog owners walking up to 1.5 km to the SP. Similarly, the mean straight line distance was estimated to be 0.812 km with an upper quartile of 1.016 km. S1 Fig demonstrates the difference between the two methods for calculating the distance and demonstrates why distance estimates can vary using the two methods. Fig 2 shows distance traveled to each of the 47 SP clinics, using the two distance estimation methods. It shows that there is great variation in the range of distances SPs manage to attract individuals from. This implies that there might be underlying reasons why some SPs attract individuals from much greater distances than other. In order to ensure that these differences are adjusted for, the addition of SP-level random effects was considered in the regression model.
Data used to build a multivariable logistic regression model predicting attendance to a SP included dog related data, household related data extracted from poverty, land use and land type GIS data and straight line distance from nearest SP. A summary of data used as predictor variables is presented in S1 and S2 Tables. Furthermore, Fig 3 shows how the proportion of attendance to SP decreases as distance from nearest SP increases.
Univariable analysis results are shown in Table 1. Land use data were not considered for the model as almost all dogs were located within residential areas. Similarly, ownership status was excluded as very few of the dogs seen were strays (1%). All other variables were considered for the final model.
Fig 4 shows the final multivariable logistic regression model predicting attendance to SP. Numerical results of the regression model can be found in S3 Table. While increasing distance from SP, being a puppy or pregnant/lactating decreased the odds of a dog being taken to a SP for vaccination, high proportions of poor people among a region, as well as living in a high and medium housing density area were positive predictors of attendance to SP. The model also showed that the effect of distance was increased with increasing levels of poverty i.e. there was an increased drop of attendance with distance in poorer people. Regarding dog characteristics being healthy or neutered increased the odds of a dog being taken to SP for vaccination. Lastly, compared to dogs who always roamed, dogs who were reported as never roaming were less likely to be taken to a SP, while dogs who were allowed to roam daily, but restrained for part of the day had increased odds for being taken to a SP for vaccination. The predictive ability of the model was assessed by using the model to predict whether a dog was taken to a SP or not using the test dataset. The AUC was calculated as 0.77, indicating that the model was reasonably good at predicting the outcome.
During the door-to-door survey, people who did not attend a SP clinic were asked why. Reasons quoted for not having attended a SP clinic are shown in Fig 5. The most common ones included the owners being unaware, unavailable or unable to handle their dogs, distance and the puppies being too young. This result complements the results of our model by emphasising the importance of distance and the fact that the age of the dog will influence their decision on whether to bring it for vaccination. It also provides further information on other possible reasons why dogs might not be presented at a SP clinic, which our model was unable to take into consideration. These include owner related factors such as lack of awareness, availability and difficulty in handling dogs.
Investigating further into the relationship between people being unaware of the SP vaccination campaign and distance to their nearest SP, it was found that people who said that they did not attend the SP because they were not aware of the campaign were located further to a SP than those who quoted different reasons for not attending (mean unaware = 1.10 km, mean other reason = 0.92 km, p-value = < 0.01).
This paper presents the results of the first large scale study investigating the reasons why attendance at SPs offering free rabies vaccinations for dogs is suboptimal in SSA. We were able to interrogate data from a city-wide vaccination campaign in Blantyre, Malawi using a combination of GIS and household questionnaire type data. We found that distance from household played an important part in SP attendance. Specifically, our regression model showed that for every km closer the odds that the dog was taken to a SP for vaccination were 3.3 times higher. Distance was also one of the main reasons for not attending SPs most commonly quoted by the owners (17%). This finding has been replicated in other studies, which also found distance decay in the use of health services in developing countries [33–37]. Our findings were consistent with smaller scale studies in rural Tanzania, where vaccination coverage decreased as distance from sub-village [18] or household [13] to SP increased.
In order to better inform the planning of future vaccination campaigns we also estimated the distance people were actually willing to walk to a SP, using data from 6,271 dogs for which we could retrieve vaccination cards with SP IDs. Our approach was unique in mapping both straight line distance and actual path based distance. We estimated that people were willing to travel on average 1.22 km to a SP vaccination clinic with 75% of the people walking up to 1.5 km to the SP. Similarly, the mean straight line distance was estimated to be 0.812 km with an upper quartile of 1.016 km. This information is crucial and should be used in planning efficient vaccination campaigns in urban sub Saharan settings in order to improve vaccination coverage using SPs only. In addition, our study highlighted the different uses of straight line distance versus path distance. While straight line distance is very useful when designing mass campaigns as it is much easier to estimate, path distance is more accurate and would be a more valuable tool in estimating for example the cost of travel of each SP attendant. Our study clearly demonstrates that the path distance is on average 50 per cent greater than the straight line distance in this setting.
We also found that socio-economic status influenced attendance to SP vaccinations. Our model shows that dogs from areas with higher proportions of people living in poverty are more likely to be presented for vaccination. Interestingly, the model also shows that the effect of distance described above is increased at increasing levels of poverty. In other words, there is an increased drop of attendance with distance in areas with higher proportions of people living in poverty. This is the first study to report this relationship, which highlights the importance of understanding more about which groups of people might be more inclined to bring their dog to a SP for vaccination. The only other study that has looked at this relationship has found no difference in vaccination coverage between households with high and low socio-economic status in rural Tanzania [13]. The conflicting results might arise due to the fact that our study was carried out in an urban setting. According to our experience in Blantyre, dogs are often brought to SPs by younger members of the family. Middle and high income parents might be less inclined to send their children alone to a vaccination point. Similarly, affluent people may consider their time more costly and be less willing to spend it waiting in queues in order to get their dogs vaccinated.
The signalment of the dogs was also important in influencing likelihood of attending a SP. Our model shows that young dogs, pregnant or lactating females were less likely to be brought to SP vaccination stations. Young age was also reported by 9% of the owners themselves as a reason for not bringing a dog to a SP both in our study and other studies in SSA [16–18]. Puppies less than three months old are often excluded from vaccination campaigns, either due to the misconception that they cannot mount an immune response or because it would require administration of the vaccine off-label [38]. Nevertheless, previous experimental and field studies have shown that puppies can mount a protective immune response as young as 4 weeks old [38–40]. Puppies constitute up to 30% of the dog population in SSA [11] and can therefore play a crucial role in maintaining vaccination coverage beyond the 70% threshold. In fact, WHO guidelines on mass vaccination campaigns advise vaccinating all dogs including those under three months of age [10]. This important issue needs to be addressed through improved advertising and education in order to increase vaccination coverage by ensuring that puppies as well as adult dogs are presented to static vaccination points.
Another interesting finding was the relationship between the reported dog confinement level and SP attendance. We found that compared to dogs who always roamed, dogs reported as never roaming were less likely to be taken to a SP. This might be because people believe that if dogs are not allowed to roam, they are not at risk of rabies. While this might be true for dogs that are kept in a protected area, many dogs will just be kept on a leash or in a garden where other dogs have access to and are therefore at risk of contracting rabies. This is another important issue to be raised during rabies education sessions and vaccination campaign advertisement. In comparison, dogs who were allowed to roam daily, but restrained for part of the day had increased odds of being taken to a SP for vaccination when compared to dogs who roamed all the time. This might reflect the fact that people who interact more with their dogs, are also keen to provide health care or simply be proxy for whether people were able to handle their dogs in order to bring them to the SP.
We also found that a lack of awareness of the vaccination programme was important despite high local profile within the local media, communities and schools. SP vaccination stations were advertised using posters and local radio during the weeks preceding the campaign and announced using a loud speaker in the communities around each station in the days before the actual vaccinating teams arrived at the SPs [20]. Despite these efforts, the most common reason for not attending a SP quoted by the owners (27%) was that they did not know about it. In fact, people further away from SP were less likely to be aware of the vaccination campaign. Promotion of a vaccination campaign is massively important and indeed being unaware was one of the most commonly quoted reasons for failure to attend a vaccination SP in other developing countries including Chad [16, 17], Mali [14], Tanzania [18] and Peru [15]. Timely and accurate provision of information about upcoming SP vaccination stations is likely to increase participation at SPs, and might therefore be cost-effective for future campaigns to invest a greater proportion of resources on campaign advertisement and promotion making sure they cover the area of interest homogeneously.
Another important reason for not bringing a dog to a SP identified by 19% of the owners in this study was difficulty in dog handling. This supports findings of previous studies in developing countries [13–15, 17, 18]. In settings where most dogs are owned for guarding or hunting [11], dogs may be less accustomed to being walked on a leash, making it very difficult for owners to bring them to SP vaccination stations. In order to achieve greater coverage at SP this problem cannot be ignored. Promotional campaigns and rabies education work need to include information on how to safely handle and walk dogs. Such information might need to be provided throughout the year in order for the dogs to be more likely to be able to be handled at vaccination time. Examples of programmes focusing on improving dog handling have been used in several countries in Latin America [15], but have not been previously described in SSA possibly due to economic constraints. With rapidly rising mobile phone ownership in SSA, regional mass SMS delivery through the most popular networks has the potential to greatly increase dissemination of information about time and location of up and coming SPs and therefore possibly increase turn-out.
The present study has several strengths and limitations. Data used for our regression model were sourced from an intensive vaccination campaign which aimed to cover the whole city and is therefore likely to be very representative of the dog population in Blantyre city. This provided us with data about each dog’s signalment as well as GPS locations used to estimate distance to nearest SP and extract GIS data corresponding to each location. This resulted in a detailed dataset and enabled us to extensively explore factors affecting attendance at SP locations. Our model validation showed that our model was reasonably good at predicting the outcome, but there was some unexplained variation. This might have arisen due to the fact that GIS data sources for Malawi are limited and not very detailed or due to information we did not collect such as households who did not respond, number of dogs per households and whether they had equipment to restrain dogs. Similarly, this might be due to information we were unable to include in this kind of model and indeed our household questionnaire showed that two of the most common reasons for not attending SP were being unaware of the campaign and having difficulty in handling which were not included in the initial model. Lastly, we have used google maps to calculate the path distance people were willing to travel to SP stations. This is an innovative way of estimating path distance, which has not been used in rabies relevant studies before, providing a more realistic estimate compared to straight line distance. Nevertheless, it is important to remember that the accuracy of this estimate greatly depends on the accuracy of google maps data in each region and might not be applicable in all areas.
Overall, this is first large scale study investigating the barriers to obtaining adequate rabies vaccination coverage through SPs in an urban setting in SSA. Our results suggest that future vaccination campaigns should increase efforts on improving positioning of SPs so that they become more accessible. We have also shown that there is a clear need to provide timely and accurate information about upcoming campaigns, emphasing the importance of puppies being vaccinated and identifying ways to improve dog handling. Estimates from our model could be used to estimate the impact on vaccination coverage of adapting several measures such as increasing vaccination points or increasing the proportion of puppies vaccinated, however caution should be exercised due to potential factors not accounted for by the model. In conclusion, this study has provided valuable insight into the barriers to attendance at SPs in urban settings and this should be taken into consideration when designing future mass vaccination programmes using SP vaccination stations in order to allow high vaccination coverage to be achieved without the need for expensive and logistically challenging door-to-door programmes.
|
10.1371/journal.pcbi.1003360 | Dimensionality of Carbon Nanomaterials Determines the Binding and Dynamics of Amyloidogenic Peptides: Multiscale Theoretical Simulations | Experimental studies have demonstrated that nanoparticles can affect the rate of protein self-assembly, possibly interfering with the development of protein misfolding diseases such as Alzheimer's, Parkinson's and prion disease caused by aggregation and fibril formation of amyloid-prone proteins. We employ classical molecular dynamics simulations and large-scale density functional theory calculations to investigate the effects of nanomaterials on the structure, dynamics and binding of an amyloidogenic peptide apoC-II(60-70). We show that the binding affinity of this peptide to carbonaceous nanomaterials such as C60, nanotubes and graphene decreases with increasing nanoparticle curvature. Strong binding is facilitated by the large contact area available for π-stacking between the aromatic residues of the peptide and the extended surfaces of graphene and the nanotube. The highly curved fullerene surface exhibits reduced efficiency for π-stacking but promotes increased peptide dynamics. We postulate that the increase in conformational dynamics of the amyloid peptide can be unfavorable for the formation of fibril competent structures. In contrast, extended fibril forming peptide conformations are promoted by the nanotube and graphene surfaces which can provide a template for fibril-growth.
| Investigation of the effects of nanomaterials on biological systems is crucial due to the increasing exposure to nanostructured materials with the growing developments and applications of nanotechnology in everyday life. Nanoparticles have been shown to have an effect on protein structure and interfere with protein self-assembly leading to the development of amyloid fibrils responsible for many debilitating diseases, such as Alzheimer's, Parkinson's and prion related diseases. Computational techniques enable investigation of such systems at the atomistic and electronic levels providing insight into properties not available from experiments. We employ a novel combination of computational methods, including large-scale electronic structure calculations and classical molecular dynamics to investigate the behavior of amyloidogenic apoC-II peptide in the presence of carbonaceous nanoparticles, the most prevalent form of nanoparticles found in the environment. Our results showed that carbon nanoparticles have significant effects on the peptide structure, dynamics and binding affinity. Specifically, the dimensionality and curvature of the nanomaterial can either facilitate or hinder their interaction with amyloidogenic peptides and make them adopt conformations capable of inhibiting or promoting fibril growth. These findings are important for rational design of amyloid fibril inhibitors as well as the elucidation of possible toxic effects of carbon based nanomaterials.
| The fast-developing field of nanotechnology has already had a significant impact in numerous areas of science and technology due to the ability to control the properties of nanomaterials with greater precision [1]–[3]. Despite the remarkable speed of developments in nanoscience, little is known about the effects of nanomaterials on biological matter [4]. There is a growing concern that nanomaterials, specifically those used for medical applications, may induce cytotoxic effects [5]. In addition, engineered nanomaterials, which are increasingly being used in industry and the manufacture of household goods have the ability to permeate blood-brain barriers and thus have the potential to damage cells in vivo [6]. The toxicity of nanoparticles has been associated with fibril formation, where nanoparticles can cause localization of peptides and proteins on their surfaces and promote undesirable aggregation that can favor formation of amyloid fibrils. These highly-structured protein aggregates are responsible for many degenerative diseases such as Alzheimer's, Creutzfeld-Jacob disease, and dialysis-related amyloidosis [7]–[10].
Carbonaceous nanoparticles are one of the most prevalent types of nanomaterials present in the environment. These air-borne particles are continuously injected into the atmosphere in large quantities through the process of combustion and, at the smallest scale, are in the form of clusters with nanometric dimensions. Carbon based nanomaterials, such as fullerenes, nanotubes and graphene surfaces, have been widely studied for potential applications due to their outstanding mechanical, thermal and electronic properties. There is, however, a growing volume of literature that alerts to the potential harm from both intentional (medicinal) and unintentional exposure of living organisms to such particles [6], [11], [12]. Comprehensive understanding of organic-inorganic interactions is crucial in order to minimize the potential toxicological effects associated with advances in the development and use of such nanomaterials [13], [14].
Computational modeling has been used extensively to study the dynamic, thermodynamic and mechanical properties of biological systems. Recent reviews summarize the application of computer simulations to the study of biological matter in the presence of nanomaterials, specifically the common modes by which nanomaterials interact with proteins, DNA and lipid membranes [15]–[19]. Physicochemical properties that may be important in understanding the toxic effects of nanomaterials include particle size and size distribution, shape, exposed surface area, internal structure and surface chemistry [20]. Much research has focused on the characterization of carbon-based nanomaterials such as fullerenes, carbon nanotubes and graphene surfaces [21]–[23]. At the same time, experiments involving carbonaceous nanomaterials in biological milieu are still limited and the interactions involved are not well understood [24]. Specifically, there are some contrasting findings that have recently been published on the role of carbon nanotubes in fibril formation. Linse et al. found an increase in the rate of fibrillation by β2-microglobulin in the presence of carbon nanotubes, where they suggested that a locally increased concentration of protein on the carbon nanotubes surface promotes oligomer formation [9]. Two other separate studies also suggested that carbon nanotubes act as catalyst for fibril formation [25], [26]. In contrast, Ghule et al. found that multi-walled carbon nanotubes inhibit amyloid aggregation of the human growth factor protein, hFGF-1 by encapsulating the protein structure and suppressing like-protein interactions [27]. Furthermore, recent computational studies of Aβ peptides found that carbon nanotubes drive the formation of β-barrels around the nanoparticle [28], [29]. The authors suggested that this type of aggregation would lead to; 1) blocking of the peptide structure for further peptide association; 2) reducing the population of monomers/oligomers available for fibril growth; and thus resulting in an inhibition of Aβ fibrillation [28]. In addition, they proposed that the hydrophobic and π-π interactions between the Aβ peptide and carbon nanotube inhibit β-sheet formation and destabilize fibril-seeds into random coil aggregates, which would increase the nucleation lag-time and possibly reverse the fibrillation process [29]. It is evident from the works presented above that there are contrasting views on the role of carbon nanomaterials in fibril formation. However, there is substantial evidence that suggests carbon nanomaterials can have fibril inducing and inhibiting capabilities depending on the structural architecture of the nanoparticle itself and more importantly, the affinity of the peptide/protein under investigation, which plays a crucial role in the propensity for aggregation and/or fibril formation on nanoparticles [18], [19].
While advances in experimental techniques are able to probe ever-smaller length-scales and ever-shorter timescales, atomistic modeling is a valuable complementary approach for a systematic investigation of detailed mechanisms of nanoscale phenomena at the atomistic and electronic levels [23], [30]–[40]. Herein, we present a computational study investigating the effects of curvature and shape of carbonaceous nanomaterials on the structure, dynamics and binding of an amyloidogenic apolipoprotein C-II (apoC-II) derived peptide, apoC-II(60-70). ApoC-II is a 79 amino acid protein, with an important role in lipid transport [41], [42]. Under lipid-depleted conditions, apoC-II readily forms homogeneous fibrils with a “twisted ribbon” morphology and all of the characteristics of amyloid fibrils [43]. ApoC-II amyloid fibrils are commonly associated with atherosclerotic plaques, where they have been found to co-localize with other apolipoproteins and initiate early events in heart disease [44]. Studies have shown that airway exposure to concentrated ambient particles and single-wall carbon nanotubes can promote progression of the atherosclerosis process in apolipoprotein-E knockout mice that develop plaques in blood vessels at early age [45], [46]. Similarly, a study by Vesterdal et al. demonstrated that intraperitoneal administration of pristine C60 fullerenes is associated with a moderate decrease in the vascular function of mice with atherosclerosis [47].
The apoC-II peptide derivative, apoC-II(60-70), was found to have the ability to form amyloid fibrils independently [48]. This peptide has been extensively investigated under a range of conditions and in different environments using experimental and computational techniques [33], [34], [37], [38], [49], [50]. Our previous studies using molecular dynamics simulations of the monomeric wild-type apoC-II(60-70) peptide showed that it preferentially adopts hairpin-like structures in solution. This structure was defined as an intermediate state on-pathway for the formation of fibril-seeds. Increased solvent accessible surface area and the relative orientation of the aromatic side-chains were features identified as fibril-favoring for this peptide, as they promoted hydrophobic interactions with other like-peptides. In contrast, increased flexibility and the broader distribution of angles between the aromatic residues of mutated apoC-II(60-70) resulted in slower aggregation kinetics, in other words these features were fibril-inhibiting, as demonstrated by our experiments [34], [37]. Furthermore, our research on oligomeric apoC-II(60-70) showed that extended β-sheet structures stabilize preformed dimers and tetramers of apoC-II(60-70). The results suggested that a tetrameric oligomer in anti-parallel configuration can serve as a possible seed for fibril formation of apoC-II(60-70), where side-chain-side-chain contacts contribute to the fibril stability, while the maximum exposure capacity of the whole peptide (backbone and aromatic side-chains) promotes the growth of the fibril-seed due to the increase of exposure to other peptides [34], [37]. Overall, the solution based studies on the behavior of apoC-II(60-70) in different environments provide benchmarking data for identifying the effects of nanomaterials on the structure and dynamics of this amyloidogenic peptide.
Here, we investigate the behavior of apoC-II(60-70) in the presence of three carbonaceous nanomaterials: a spherical C60 fullerene, a tubular single-wall carbon nanotube and a flat graphene surface. We study the peptide's structure, dynamics and binding, all of which can influence its fibril formation capacity and compare the results with the previously characterized peptide behavior in solution [33], [34], [37], [38]. We apply a novel combination of computational methods, including large-scale electronic structure calculations and classical all-atom molecular dynamics. This approach was recently applied for the first time to investigate the fibril inhibition mechanisms of cyclic apoC-II(60-70) and its linear analogue [39]. This combined modeling approach enables investigation of the fundamental driving forces behind the interactions of the peptide with nanomaterials and their effects on the peptide structure, dynamics and binding affinity.
To investigate the effects of carbonaceous nanomaterials on the structure and dynamics of apoC-II(60-70) (MSTYTGIFTDQ, 169 atoms) a series of simulations were performed with different starting peptide conformations and arrangements. The fullerene particle consisted of 60 carbon atoms with a radius of ∼3.5 Å. The nanotube was modeled as a (5,5) single-walled tube with 320 atoms in an open ended armchair arrangement, 6.78 Å in diameter and ∼38 Å in length, which was sufficiently long to prevent interactions between the peptide and the edges. Graphene was modeled as a periodic single sheet of 2160 carbon atoms in a hexagonal arrangement to represent an infinite graphene surface. The initial configurations were constructed by positioning the peptide 4.5–10 Å from the nanomaterial (see Table S1 and S2 in Supporting Information). The peptide together with the nanomaterial was then placed in a periodic simulation cell of at least 60 Å×60 Å×60 Å in dimension.
The molecular dynamics (MD) simulations were performed using the Gromacs 3.3 [51] simulation package, with the interactions between the particles in the system described by the united-atom Gromos forcefield and the 43A1 parameter set. The carbonaceous nanomaterials were modeled using the aromatic sp2 carbon parameters. We note that polarizable forcefields which describe the electrostatic interactions with the use of distributed multipoles [32], [40], [52] have been under development for graphitic structures, however, recent studies have shown that classical forcefields produce results comparable to experiment [16], [53], [54].
The Lennard-Jones interactions were truncated at 10 Å, with the long-range electrostatic interactions accounted for by the Particle Mesh Ewald (PME) method [55]. The LINCS algorithm was used to constrain the bond lengths to their equilibrium values [56], enabling a timestep of 2 fs to be applied for all simulations. The VMD software package was used for visualization of the dynamics and analysis of the molecular trajectories [57].
In vacuo energy minimization using steepest descent algorithm was initially performed on the peptide-nanoparticle systems to remove steric clashes. The optimized system was then solvated using the SPC water model [58] at a water density of ∼1 g/cm3. To neutralize the overall negative charge of the system, an unrestrained counterion (Na+) was included in the simulation cell. Energy minimization on the solvated system was performed to relax all of the atomic degrees of freedom. Subsequently, MD was conducted to allow the solvent to equilibrate around the solutes by keeping the peptide and nanomaterial restrained. A constant pressure of 1 bar and temperature of 300 K were maintained using the Berendsen barostat and thermostat [59]. In all simulations the geometry of the nanomaterial was restrained for ease of monitoring the peptide dynamics.
Two initial non-fibrillar conformations (native and helical) of apoC-II(60-70) peptide were simulated in the presence of each nanomaterial. To further enhance conformational sampling simulations were repeated six times with different starting orientations of apoC-II(60-70) with respect to the nanomaterial, yielding a total 800 ns of data per nanomaterial-peptide complex. The behavior and structures observed in each system exhibited distinctive trends, therefore the results from representative simulations are shown.
Using umbrella sampling together with the weighted histogram analysis method (WHAM) [60], potential of mean force (PMF) profiles were generated to evaluate the free energy of dissociation (ΔG) for apoC-II(60-70) bound to each nanomaterial in solution. This method was applied to explicitly solvated systems and therefore accounts for the entropic contributions in the determination of the dissociation energies. We determined the PMF as a function of separation distance between the center of mass of the nanomaterial and the α-carbon of the glycine residue in apoC-II(60-70). To acquire the PMF profiles, a series of simulations (windows) were performed at increasing distance between the peptide and nanomaterial, starting from typical equilibrium structures of the peptide-nanomaterial complex. The peptide was restrained at each window using Hookean functions with a force constant of 8000 kJmol−1 nm−2. In the present work, ΔG and PMF both refer to the free energy required to bring the peptide and nanomaterial from an associated form, which defines our zero of free energy, to some separation d. Adjacent windows were separated by 0.5 Å and each window was simulated for 15 ns with at least 30 windows used (until the peptide was fully dissociated from the nanomaterial), resulting in a total simulation time of at least 450 ns per nanomaterial complex. WHAM was subsequently applied on the final 5 ns of simulations to remove the biasing potentials and obtain the unbiased PMF profiles. The overlap between neighboring windows was monitored to ensure the suitability of the selected spring constant and sufficient conformational sampling (not shown).
Binding energy calculations were performed on multiple structures selected from the classical forcefield simulations of each peptide-nanomaterial system (more details in aromatic tracking section). Based on the findings of a recent methodological study, classical energy minimization was performed in solution prior to electronic structure calculations [61]. This procedure reduced any electrostatic artefacts that may arise due to the electronic structure calculations being performed in vacuum while retaining the major structural features of the system obtained during the fully solvated MD simulations. Single point electronic energy calculations performed on the resultant frames were used to calculate in vacuo binding energies between the peptide and nanomaterial. We determined the binding energy (Eb) of apoC-II(60-70) peptide on each nanomaterial, as(1)where EP+N is the total energy for the peptide-nanomaterial complex, EP is the total energy of the apoC-II(60-70) peptide and EN is the total energy of the isolated nanomaterial.
The linear-scaling DFT code ONETEP [62] was used, which combines linear scaling computational efficiency with accuracy that is comparable to traditional plane-wave DFT codes. Such efficiency opens up the possibility of performing accurate DFT calculations on thousands and tens of thousands of atoms, including proteins [63]–[65] and various nanomaterials [66], [67]. ONETEP achieves linear scaling by exploiting the ‘near-sightedness’ of the single-particle density matrix p(r,r') in non-metallic systems,(2)where K is the density kernel and φα are a set of strictly localized non-orthogonal generalized Wannier functions (NGWFs) [68]. The total energy is self-consistently minimized with respect to both the density kernel and the NGWFs. The NGWFs are expanded in a basis set of periodic sinc (psinc) functions [69], which are equivalent to a plane-wave basis, and are optimized in situ, giving plane-wave accuracy and allowing the accuracy to be systematically improved with a single kinetic energy cut-off parameter.
The PBE generalized-gradient approximation was used to describe exchange and correlation [70], and norm-conserving pseudopotentials were employed to describe the interactions between electrons and nuclei. Dispersion interactions were accounted for using a DFT+D approach [71]. Dispersion-corrected DFT has been shown to produce accurate results for weakly interacting systems, such as aromatic composites [72] and protein-ligand complexes [73]. The supercell dimensions for each system were sufficiently large to prevent interactions between periodic images. In all cases, NGWF radii of 8 bohr were used for all atoms, no truncation was applied to the density kernel, the kinetic energy cut-off for the psinc basis was 880 eV, and the Brillouin zone was sampled at the Γ-point only.
The role of aromatic residues in the adsorption of apoC-II(60-70) to each nanomaterial was investigated by tracking the placement of the peptide's aromatic rings across each graphitic surface. This technique determines the position and orientation of the aromatic rings in amino acids relative to the rings within the nanomaterials' surface at every step of the MD trajectories. The aromatic ring arrangement was categorized into three groups: no π-stacking, offset π-stacking and face-to-face π-stacking (Figure 1). The criteria to determine no π-stack register were a pair-wise contact distance over 4.5 Å between any two atoms of the aromatic ring and nanomaterial; or an angle greater than 30° between the plane normal of the aromatic ring and nanomaterial surface [74]. Face-to-face π-stacking was accounted for when the displacement between the centroids of the phenyl rings of the aromatic residues and centroid of the nearest hexagonal carbon ring was less than 0.71 Å (half the carbon-carbon bond length). A displacement greater than 0.71 Å was considered as offset π-stacking. ApoC-II(60-70) peptide having two aromatic rings in its sequence (Tyr63 and Phe67) resulted in six possible ring arrangements relative to the nanomaterials surface. The categories were defined as: (1) no π-stacking by both rings; (2) offset π-stacking by one ring and no π-stacking by the other; (3) offset π-stacking by both rings; (4) face-to-face π-stacking by one ring and no π-stacking by the other; (5) face-to-face π-stacking by one ring and offset π-stacking by the other; (6) face-to-face π-stacking by both rings. Once the aromatic arrangement was categorized, each group underwent structural clustering with RMSD cut-off of 2 Å for the entire peptide using the single linkage clustering method to determine the most frequently sampled structure within each π-stacking category. Three representative structures from each π-stacking group were selected and underwent electronic structure calculations to determine their binding energies.
Explicitly solvated molecular dynamics simulations were used to characterize the interactions between the amyloidogenic peptide apoC-II(60-70) and three exemplar carbonaceous nanomaterials. ApoC-II(60-70) showed a strong affinity to the nanomaterials, where the peptide came in contact with the nanomaterial within the first 20 ns of simulation and remained adsorbed for the entire trajectory. Below we analyze the bound states and the mechanisms responsible for the binding.
Secondary structure analysis was performed to investigate the effects of the nanomaterial curvature on the peptide's conformation. The STRuctural IDEntification (STRIDE) [53] algorithm was utilized to classify the peptide's secondary structure as a function of time. Secondary structure evolution plots depicting typical conformational trends exhibited by apoC-II(60-70) in the presence of C60, nanotube and graphene are shown in Figure 2. The initial 20 ns of simulation (equilibration) are also shown to highlight the conformational changes in the peptide induced by adsorption onto the nanomaterial surface.
The results show a structural transformation of the peptide upon adsorption to the nanomaterial surface. ApoC-II(60-70) in the presence of C60 was observed to curve around the particle, with a turn region around Gly65, as shown in the picture inset of Figure 2a. This structure allows for a large number of contacts to be made with the nanoparticle, dominated by π-interactions between the aromatic residues (Tyr63 and Phe67) and the C60 surface. Due to the presence of C60, apoC-II(60-70) is unable to form the inherent β-hairpin conformation [34], [37]. In one out of six simulations the peptide was able to dissociate within 10 ns of contact with the C60 particle. This suggests that the peptide can be weakly bound to the surface of the nanoparticle. Upon desorption apoC-II(60-70) was able to form a β-hairpin conformation (picture inset of Figure 2b). We note that the β-hairpin structure was found favorable for monomeric apoC-II(60-70) peptide in solution, and identified as an intermediate state on-pathway for fibril formation [34], [37], [75]. The results show that the presence of C60 inhibits the formation of the characteristic fibril favoring β-hairpin as well as the extended conformation suggesting that the interactions with the C60 may contribute to an increase in mobility (see Figure S1) and facilitate the formation of fibril incompetent conformations. Recent work by Andujar et al. where they showed that C60 induced significant destabilization of the amyloid-β fibrils by disrupting the hydrophobic contacts and salt-bridges between the β-sheets [76] is in line with our work. This suggests that C60 can be used as a prototype for the design of potential fibril inhibitors.
The secondary structure evolution plot of apoC-II(60-70) in the presence of a nanotube shows that the peptide exhibits different structural features compared to those in the presence of C60. The peptide tends to elongate across the surface of the nanotube, while adopting mostly turn and coil motifs. This behavior is a result of the large surface area available for contact on the nanotube (Figure 2c). The strong affinity between the nanotube and peptide is assisted by π-π interactions between the aromatic rings of the peptide and the nanotube. The curvature of the nanotube enables the peptide to arch, which facilitates the short-lived formation of a hydrogen bond between Ser61 and Thr64. In comparison to the C60 simulations, the peptide was less dynamic on the surface of the nanotube, as seen from the smaller number of conformations sampled by the peptide following the adsorption and immobilization on the nanotube surface (Figure 2c and Figure S1).
The simulations of apoC-II(60-70) in the presence of graphene exhibited similar structural features to those seen in the presence of the nanotube. Upon adsorption, the peptide elongates along the graphene surface and features predominantly turn and coil structures (Figure 2d). The large surface area available for interactions enables the peptide to freely slide on the surface, while the favorable π-π stacking interactions between the aromatic residues of the peptide and the surface define its conformational features. Linse et al. showed extended nanoparticles enhance the probability of appearance of a critical nucleus for nucleation of protein fibrils, albeit for a different combination of nanomaterials and peptides [9]. This feature was determined as fibril-favoring in our previous works on apoC-II(60-70) oligomers [34], [37]. Other studies have also shown that carbon nanotubes and graphene surfaces facilitate a change in the conformation of peptides [30], [77] and π-stacking is an efficient mode of biological recognition of π-electron-rich carbon nanoparticles [23], [30]–[32], [52], [77], [78].
A common feature in all secondary structure plots is the presence of a persistent coil motif at the C-terminal end of the peptide, where predominantly hydrophilic residues reside. This suggests that the inherent preference for interaction with the polar environment by these residues is suppressed by the attractive van der Waals forces between the large surfaces presented by the nanomaterial and the peptide, preventing the peptide dissociation from the nanomaterial. Strong hydrophobic interactions between the WW domains and carbon nanotubes have also been associated with protein function “poisoning” and disruption of the protein active site [79].
Overall, it should be noted that the adsorbed peptide may adopt both fibril initiating as well as fibril incompetent conformations. However, our analyses indicate that the extended conformation adopted on the extended nanosurfaces is in line with the fibril competent structures we found through our previous modeling and experimental studies [33], [34], [37], [38], [49], [50]. In contrast the mobility and lack of secondary structure elements needed for the fibril formation by the C60 adsorbed peptide suggests the inhibiting role of this nanoparticle in the fibril formation.
To gain a more detailed understanding of the interactions involved in adsorption of apoC-II(60-70), the contact stabilities of the peptide's residues with each nanomaterial were investigated (Figure 3). Contact stabilities were calculated as the percentage of simulation time during which a contact was maintained between each residue and the respective nanomaterial. A contact was counted when the distance between a pair of atoms was less than 4 Å, which enabled us to account for van der Waals interactions between the peptide and the nanoparticle.
High contact stabilities were found for both the aromatic tyrosine (Tyr63) and phenylalanine (Phe67) residues with all nanomaterials. The stable π-stacking arrangements between the aromatic rings of the peptide and the electron-rich carbon rings of the surface, suggest that these are the key residues that contribute to the strong interactions between the peptide and the carbonaceous nanomaterials, in line with other studies [23], [30]–[32], [52], [77], [78].
This effect is evident in all simulations, however in the C60 complex the aromatic residues dominate the interactions between the peptide and C60 surface, while the other residues exhibit less persistent contacts. Interestingly, Tyr63 exhibits higher binding affinity to C60 compared to Phe67, in accordance with a DFT study that showed Phe and Tyr bind with a similar strength to the nanotube, while Tyr exhibits a stronger binding to C60 [80], [81]. In contrast, the large contact area presented by the carbon nanotubes and graphene results in higher contact stabilities with all (not just the aromatic) residues, this effect being most evident on the graphene surface.
In our recent studies we showed that the orientation of the aromatic side chains is different in the fibril-forming and fibril-inhibiting arrangements [34], [37]. The simulations of apoC-II(60-70) in the presence of C60 exhibited structures where the aromatic rings were positioned on the same side of the peptide which enhance the π-stacking interactions with the small, highly curved C60 particle. This ring arrangement was postulated to inhibit fibril formation [34], [37]. In contrast, the aromatic rings did not show a specific facial preference in the nanotube and graphene complex simulations (Figure 2c,d). This was due to the large contact area and stronger hydrophobic interactions presented by these materials, which formed the aromatic ring stacking upon adsorption of the peptide.
A series of radial distribution functions (RDF) were calculated to determine the degree of water structuring around the peptide in solution and when bound to the nanomaterial surface. The results provide an insight into the extent of desolvation of the peptide conformation upon binding to the different nanomaterials. Typical RDFs of the peptide side-chain hydrogen atoms (H) with respect to the water oxygen atoms (O) are shown in Figure 4. For all systems, the RDF profiles show a peak at ∼2 Å, representing the first hydration shell, indicating hydrogen bonding between water and the apoC-II(60-70) side-chains. The results also show the presence of a second hydration shell at ∼4 Å. The RDFs of peptide-nanomaterial complexes exhibit an attenuation of the overall probability density, suggesting the exclusion of water due to the hydrophobic contact between the peptide and nanoparticle manifests in the RDFs through the lowering of the occurrence of water at larger separation distances in the bound state. Indeed, desolvation effects have been shown to be favorable in the self-assembly of cyclic peptides on carbon nanotubes [31].
To characterize the binding of apoC-II(60-70) peptide to each nanomaterial in the presence of solvent, the free energy of dissociation was calculated using umbrella sampling (potential of mean force, PMF) together with the weighted histogram analysis method (WHAM) [60]. This approach is applied to explicitly solvated systems and accounts for both the enthalpic and entropic contributions to the dissociation free energies. Two bound equilibrium complex structures were studied for each system to enhance sampling. The PMFs detailing the dissociation pathway of apoC-II(60-70) from each nanomaterial are presented in Figure 5.
The results demonstrate that higher degree of curvature reduces the surface area available for adsorption, and the dissociation free energy indicates that binding to C60 is weakest and binding to graphene is strongest of the systems investigated. Here, a lower value indicates a weaker binding. The size of the peptide does not allow for complete wrapping of the C60, therefore in this complex the peptide is quite mobile with terminal residues remaining free and not forming close contacts with the nanoparticle, as can be seen in the residue contact stability plot in Figure 3. Figure 5a indicates that the dissociation energy is dependent on the adsorbed peptide conformation. This suggests that C60 induces significant structural lability in apoC-II(60-70) preventing it from adopting stable conformations, in line with the peptide evolution observed through molecular dynamics trajectories (Figure 2 and Figure S1). A higher free energy of dissociation (∼1.8 kcal/mol) was obtained for the peptide that had a larger number of contacts with C60 and whose aromatic rings were continuously interacting with the C60 particle. In contrast, the system where the two aromatic rings of the peptide predominantly formed π-stacking between themselves rather than with the nanoparticle resulted in a lower dissociation energy (∼1.1 kcal/mol). The peaks and troughs are caused mostly by the transient π-stacking interactions, with peaks observed when contacts are broken, as illustrated by the insets in Figure 5. In our previous work on apoC-II(60-70) we showed that an increase in conformational flexibility and dynamics can slow down or even inhibit fibril formation [34], therefore it appears that interactions with the C60 can induce a similar, fibril inhibiting, effect.
We note that generally PMF plots for the C60-peptide system are noisier than those for the nanotube and graphene systems which is due to the transient nature of the contacts and increased mobility of apoC-II(60-70) when in contact with C60, rather than due to insufficient conformational sampling. The effect was verified by continuing the umbrella sampling simulations for a further 15 ns per window and observing that the resultant PMFs did not show significant differences (figures not shown). Similarly, the free energy differences seen between the multiple simulations of the peptide-nanotube system are due to the variety of structures sampled along each dissociation pathway. As expected, the predominantly elongated peptide conformation (Figure 5b, left) enabled a larger number of contacts between the peptide and the nanotube, which resulted in a higher free energy of dissociation (∼8.6 kcal/mol). The peptide exhibiting mostly coiled structures made fewer contacts with the nanotube (Figure 5b, right) which in turn required less energy (∼3.5 kcal/mol) to dissociate from it. The smoother PMF plots for the peptide-nanotube systems are a result of the persistent interactions between the components, in keeping with the results of our classical MD simulations.
The PMF plots representing the dissociation free energy of apoC-II(60-70) from graphene exhibited conformation independent pathways. As seen from the MD results, the π-stacking between apoC-II(60-70) and graphene contributes to the formation of elongated peptide structures and restricts the conformational flexibility of the peptide. Repeat simulations resulted in dissociation free energies of ∼15 kcal/mol irrespective of the conformations sampled along the dissociation pathway.
We note that dissociation energy peaks occur when a large number of contacts are broken, such as during the illustrated dislocation of the aromatic residues from the graphene surface (see insets of Figure 5c). In contrast, as the peptide is slowly pulled away from the surface a characteristic smooth dissociation energy profile is observed.
In addition to the classical simulation-derived dissociation free energies discussed above, we have used electronic structure calculations based on DFT to calculate in vacuo binding energies of selected frames derived from classical all-atom simulations. To investigate the role of aromatic residues (Tyr63 and Phe67) in driving the adsorption of apoC-II(60-70) onto carbon based nanomaterials, we developed an algorithm capable of tracking the position and orientation of the phenyl rings of the aromatic amino acids with respect to the aromatic rings of the nanomaterials' surface at every step of the MD trajectories (exemplar result shown in Supporting Information).
The in vacuo binding energy of three representative frames from each π-stacking arrangement was obtained by DFT calculations using the ONETEP linear-scaling code [62], comprising a total of eighteen typical structures per peptide-nanoparticle complex. This analysis provides a measure of the relative binding of apoC-II(60-70) to a nanomaterial surface with respect to the contact area. The binding energy differences between the representative structures for each π-stacking configuration versus the peptide-nanomaterial contact area are shown in Figure 6 (tabulated form available in Supporting Information). The vacuum binding energies are shown relatively to the strongest bound state (apoC-II(60-70) on graphene). In this case, higher values indicate a weaker binding (left y-axis, Figure 6).
The in vacuo binding energy results confirm the trends observed in our explicitly solvated PMF free energies showing the strength of binding between apoC-II(60-70) and the nanomaterials to follow: C60<nanotube<graphene. In all systems the aromatic rings act like “anchors“ for binding the peptide to the carbon nanomaterials via π-π interactions.
DFT binding energy calculations confirm the finding from classical MD that apoC-II(60-70) exhibits strongest binding on graphene with a face-to-face π-stacking arrangement made by the two aromatic rings of the peptide and the surface. The all-atom MD simulations show that the flat graphene surface promoted sliding of the peptide (see Figure S2 in Supporting Information) and backbone elongation to optimize the π-stacking arrangement between the aromatic rings of the peptide and the substrate. This contributes to the peptide-graphene system having the largest aromatic and total contact area, which results in the strongest binding. The DFT binding energy also confirmed that apoC-II(60-70) exhibits a weaker binding to the nanotube and the weakest binding to C60, attributed to the increased nanosurface curvature which stimulates the formation of turns and loops in apoC-II(60-70) leading to a lower contact area between the peptide and nanoparticle. We note that C60 comprises both hexagonal and pentagonal carbon rings and, therefore, has a lower probability of face-to-face π-stacking with the six-membered aromatic rings of the peptide (statistical data shown in Supporting Information). This provides a further explanation for the significantly smaller contact area and weaker binding obtained for the peptide and C60 nanoparticle, compared to the nanotube and graphene systems (Figure 6 and Table S3).
Furthermore, using our DFT calculations we were able to examine the intra-peptide electrostatic interactions which play a significant role in determining the peptide's secondary structure and consequently the binding affinity to other materials. Electron density difference (Δρ) maps showing charge accumulation (red) and depletion (blue) upon peptide adsorption on each nanomaterial are presented in Figure 7. We can see that intra-peptide interactions are more significant in the proximity of nanoparticles with high curvature which have a reduced nanoparticle-peptide contact surface area, as Figures 6 and 7 demonstrate.
The greater surface area available on the flatter “hexagonal-only” surface of graphene allows for a more efficient π-stacking and a stronger peptide binding as shown in Figures 7c. Moreover, surface adsorbed elongated peptide conformations enable polar residues such as Thr, Ser and Gln to become more solvent exposed, thus exhibiting the “snorkeling effect” [82], [83] (see inset of Figure 7b), where the hydrophobic backbone interacts with the graphitic surface, while the polar side chains are protruding to the solvent. This lowers the overall contact area between the peptide and nanomaterial, and ultimately reduces the binding affinity. Figure 7 shows small electron density differences between the aromatic groups and the graphitic surfaces. This is in agreement with the study of Poenitzsch et al. where they observed weak charge-transfer interactions between aromatic groups and carbon nanotubes using scanning tunneling spectroscopy and Raman experiments [84]. Our electron density analysis shows that, generally, a weaker binding is a result of inefficient π-stacking arrangements and intra-peptide electrostatic interactions that reduce the peptide-surface interactions, as Table S3 demonstrates. Charge redistribution can also be seen between the peptide and nanoparticle surface (Figure 7), suggesting some polarizability effects occur between the peptide and nanomaterial. Specifically, a charge depletion can be seen at Asp69 and Gln70 in all systems, while a charge accumulation develops at the closely interacting sites of the peptide and nanomaterial surface. Figure 7a shows charge accumulation at Gly65 for the C60 complex, while the nanotube and graphene exhibit charge buildup in close proximity to the predominantly hydrophilic N-terminal region of the peptide (Met60, Ser61 and Thr62).
We note that the agreement between our classical simulations and the DFT studies suggest that the classical forcefield potentials employed here are able to capture the polarization effects inherent to peptide-nanoparticle systems. Moreover, a recent study using the dispersion corrected DFTB-D method, showed that although molecular mechanics techniques with fixed-charge forcefields do not explicitly incorporate polarizability, they can predict the strength of π-π interactions between aromatic moieties and carbon nanotubes [75]. This demonstrates that molecular dynamics simulations utilizing fixed charge forcefields provide a reasonable representation of the interactions between peptides and graphitic surfaces.
Using classical forcefield and electronic structure calculations, we have shown that an amyloidogenic apoC-II(60-70) peptide exhibits a strong affinity for graphitic nanomaterials where binding is facilitated through π-π interactions between the aromatic residues of the peptide and the surface of the nanomaterial. This is generally achieved by the exclusion of water molecules from the peptide-nanomaterial interface.
The proximity of the C60 fullerene contributed to an increase in conformational lability of apoC-II(60-70), which was shown to prevent it from adopting fibril-favoring structural features. This finding is in line with the previous studies of oxidized apoC-II(60-70), where increased structural flexibility and dynamics were the key factors prohibiting this peptide to form fibrils, confirmed experimentally. Conversely, our data showed that the less curved nanotube and flat graphene nanomaterials promote elongated peptide conformations previously shown to form fibril seeds, which confirms recent findings that extended carbon nanosurfaces can act as templates able to encourage peptide fibril formation and growth.
Electronic binding energy and solution free energy calculations showed the binding affinity of apoC-II(60-70) was weakest for the C60 particle, followed by the nanotube, and strongest for the graphene. In all simulations these trends are due to the larger contact area available for peptide adsorption to the flatter graphene and nanotube than the highly curved C60. The increased curvature also results in reduced efficiency of aromatic π-stacking and higher intra-peptide electrostatic interactions which contributes to its weaker binding to the nanomaterials. The electronic structure calculations show that dimensionality that determines the electronic properties of the nanoparticle as well as size and curvature play a significant role in the contact area and binding mechanisms of the peptide. At the same time the intra-peptide interactions determined by the peptide sequence (i.e. presence of aromatic, aliphatic, polar/apolar amino acids) affect the binding mechanism of peptides to nanoparticles. The observed agreement between the classical and electronic structure calculations show that molecular dynamics simulations utilizing fixed charge forcefields provide reasonable representation of the interactions between peptides and graphitic surfaces.
In summary, our results highlight that hydrophobic nanoparticles have multiple notable effects on the peptide structure, dynamics and binding affinity. We have demonstrated that dimensionality and different degree of curvature can either facilitate or hinder the interaction of amyloidogenic peptides with the nanosurfaces and make them adopt conformations capable of inhibiting or promoting fibril development, as shown in our recent experiments. These findings can be important for rational design of amyloid fibril inhibitors as well as for clarification of possible toxic effects of carbon based nanomaterials.
|
10.1371/journal.pcbi.1003524 | Computational Analyses of Synergism in Small Molecular Network Motifs | Cellular functions and responses to stimuli are controlled by complex regulatory networks that comprise a large diversity of molecular components and their interactions. However, achieving an intuitive understanding of the dynamical properties and responses to stimuli of these networks is hampered by their large scale and complexity. To address this issue, analyses of regulatory networks often focus on reduced models that depict distinct, reoccurring connectivity patterns referred to as motifs. Previous modeling studies have begun to characterize the dynamics of small motifs, and to describe ways in which variations in parameters affect their responses to stimuli. The present study investigates how variations in pairs of parameters affect responses in a series of ten common network motifs, identifying concurrent variations that act synergistically (or antagonistically) to alter the responses of the motifs to stimuli. Synergism (or antagonism) was quantified using degrees of nonlinear blending and additive synergism. Simulations identified concurrent variations that maximized synergism, and examined the ways in which it was affected by stimulus protocols and the architecture of a motif. Only a subset of architectures exhibited synergism following paired changes in parameters. The approach was then applied to a model describing interlocked feedback loops governing the synthesis of the CREB1 and CREB2 transcription factors. The effects of motifs on synergism for this biologically realistic model were consistent with those for the abstract models of single motifs. These results have implications for the rational design of combination drug therapies with the potential for synergistic interactions.
| Cellular responses to stimuli are controlled by complex regulatory networks that comprise many molecular components. Understanding such networks is critical for understanding normal cellular functions and pathological conditions. Because the complexity of these networks often precludes intuitive insights, a useful approach is to study mathematical models of small network motifs having reduced complexity yet consisting of key regulatory components of the more complex networks. Computational studies have analyzed the behavior of small motifs, and have begun to describe the ways in which variations in parameters affect their functional properties. Here, we investigated how variations in pairs of parameters act synergistically (or antagonistically) to alter responses of ten common network motifs. Simulations identified parameter variations that maximized synergism, and examined the ways in which synergism was affected by stimulus protocols and motif architecture. The results have implications for the rational design of combination drug therapies where a goal is to identify drugs that when administered together have a greater effect than would be predicted by simple addition of single-drug effects (i.e., super-additive effects), thereby allowing for lower drug doses, minimizing undesirable effects.
| Cellular functions are regulated by complex biochemical networks that incorporate large numbers of diverse molecular components and their interactions. The large scale and complexity of these regulatory networks impedes achieving an intuitive understanding of their overall function and responses to stimuli and/or drugs. Consequently, when analyzing a complex system, it is often useful to develop and analyze reduced models that capture the key dynamical properties of the system. In analyses of biochemical networks, these reduced models are referred to as motifs [1]. Motifs depict distinct connectivity patterns that occur more frequently in a given network than in random networks of the same size. Motifs can be comprised of as few as three molecules (referred to as nodes or vertices) and their interactions (referred to as edges). Motifs are present in gene regulatory networks, protein-protein interactions, and metabolic networks of species as diverse as bacteria [1]–[2], yeast [2]–[3], and humans [2], [4]–[12]. Structurally distinct motifs appear to manifest specific dynamical features [10], [13]–[17] and modeling studies describe how the responses of distinct motifs and the robustness of these responses vary with parameters [18]–[20]. These studies are beginning to elucidate ways in which motif dynamics contribute to the functions and response properties of larger, more complex regulatory networks. Moreover, as is investigated here, small network motifs can be used to examine the ways in which combinations of parameter changes act synergistically (or antagonistically) to alter the response to stimuli. This later strategy may ultimately help guide the development of drug combination therapies that target disease-related dysfunction of a network motif.
Here, models of ten three-node motifs (Fig. 1) were developed and synergistic interactions within these motifs were investigated. These motifs are ubiquitous and are included within gene and protein networks that are associated with specific diseases [1]–[2], [4]–[6], [9], [11], [13], [15], [21]–[24]. The mechanisms of disease are usually associated with large networks of molecular pathways. However, in many clinical studies in which combination drug therapy is used for treatment of diseases, two-drug combinations are considered [25]. Therefore, for models of simple motifs or of complex pathways, simulation of concurrent paired parameter changes is of value for understanding the synergistic or antagonistic properties of many current or possible combination therapies. In the simulations, pairs of parameters were simultaneously varied, and the extent to which these ten motifs manifest synergism (or antagonism) was examined. First, a canonical model was developed. This motif constitutes a minimal representation of two convergent pathways. Element A and element B both respond to a common stimulus (S), and converge to activate a common target (T) (Fig. 1A1). Activation (e.g., increased phosphorylation, or enhanced synthesis) of T was assumed to be the output of the motif, and the target for examining the effects of combinations of parameter variations. Variations of parameter pairs in elements A and B represent combination therapies in which two drugs target two different sites of the same pathway or two pathways converging at a downstream process. Such convergence is commonly used in designing therapies. For example, Paclitaxel synergizes with Tubacin in enhancing tubulin acetylation, with the former directly increasing acetylation and the later decreasing the deacetylation of α-tubulin [25]. Aplidin and Cytarabine are synergistic in killing cancer cells because they induce apoptosis via two convergent signaling cascades [25]. In this study, synergism (or antagonism) was quantified using degrees of nonlinear blending and additive synergism (see Model Development). Then the canonical motif was modified to generate a set of similar three-node motifs that incorporated different patterns of interaction among the nodes. These interactions included a variety of positive and negative feedback loops, and autoregulatory loops. These motifs were found to greatly modify the existence and amount of synergism. For specific parameter pairs, only a subset of motif architectures exhibited synergism.
To substantiate these conclusions, the approach was applied to a model representing interlocked feedback loops that govern the synthesis of two transcription factors, cAMP-response element binding proteins (CREBs), specifically CREB1 and CREB2 [26]. CREB1 is a transcription activator and CREB2 is a transcription repressor. CREB1 and CREB2 regulate their own expression, via binding to the CRE elements in or near their genes. The feedback loops involving CREB1 and CREB2 modulate long-term memory [27]. Three of the network motifs that were simulated are included in this model: a positive auto-regulatory loop governing CREB1 synthesis; a negative auto-regulatory loop governing CREB2 synthesis; and negative feedback in which CREB2 inhibits the synthesis of CREB1. The effects of these motifs on synergism in this more biologically realistic model were consistent with the results from the more abstract three-node models.
Elements A and B converge onto a target (T) (Fig. 1A1). Both A and B are activated by stimulus S. The activities of A and B are dynamic variables that follow first-order ordinary differential equations (ODEs). kbasal_A and kbasal_B are basal activation rates of A and B. The deactivation of A and B follows Michaelis–Menten kinetics. These assumptions yield the following ODEs:(1)(2)
Two variants of this simple converging model were analyzed. In Variant M, the activation rate of T is proportional to the product of the effects of A and B. Deactivation of T follows Michaelis-Menten kinetics. The following ODE for T results:(3)
In Variant A, the activation rate of T is proportional to the sum of the effects of A and B, yielding the following ODE:(4)
Concentrations are non-dimensional. Standard parameter values are used in all simulations unless noted. These values are:
KsA = 0.1, kbasal_A = 0.1, kdA = 0.2, KA = 1, KsB = 0.1, kbasal_B = 0.1, kdB = 0.2, KB = 1, kbasal_T = 0.0001, kST = 0.01, KTA = 2.5, KTB = 2.5, kdT = 0.01, KT = 0.5.
The parameters were adjusted by trial and error so that the dynamics of A, B and T display properties of common biochemical responses; e.g., 1) the activation of A and B was rapid, whereas their deactivation was relatively slow; 2) the peak level of T was well below saturation when stimuli were weak; and 3) the basal activation rate of T, kbasal_T, was much smaller than the activation rate induced by A and B, kST. For simplicity, the strengths of both pathways were initially balanced. Therefore, the parameters of dynamics of A and B shared the same values. The extent to which the synergism is dependent on these values is discussed in the Results.
In the three-node model, it was assumed that 1) both A and B have excitatory effects on T; and 2) the standard parameter values governing A and B dynamics are identical. There are then nine biochemical variations of the canonical motif that involve adding a single feedback or feedforward interaction, or autoregulation, involving A or B. These nine motifs were represented by extensions of the canonical model. For simplicity and consistency, these extensions were all constructed by adding regulation, within the motif, of a given pair of parameters, kdA and kdB. After construction, synergism was examined with paired parameter variations for each of these motifs.
CREB1 is assumed to bind to cAMP response elements (CREs) near the promoters of both the creb1 and creb2 genes, activating expression of both genes. CREB2 is assumed to bind competitively to the same CREs and to repress transcription of both genes. In this minimal model, differential equations for mRNAs are not included. Thus the model consists of two ODEs, for the levels of CREB1 and CREB2, with gene regulation represented by activation and repression of CREB1 and CREB2 formation.(14)(15)Concentrations are non-dimensional. Standard parameter values from Song et al. [26] are used. These values are:
Vx = 0.4 min−1, Vy = 0.01 min−1, Kx = 5, Ky = 10, kdx = 0.04 min−1, kdy = 0.01 min−1, rbas,x = 0.003 min−1, and rbas,y = 0.002 min−1
The parameters involved in the positive auto-regulatory loop in which CREB1 enhances its own synthesis are Vx (the maximum induced synthesis rate of CREB1) and Kx (the dissociation constant of CREB1 from a CRE). The parameters involved in the negative auto-regulatory loop in which CREB2 represses its own synthesis are Vy (the maximum induced synthesis rate of CREB2) and Ky (the dissociation constant of CREB2 from a CRE). All four parameters are involved in the negative feedback loop in which CREB2 represses the synthesis of CREB1, because in this loop, CREB1 first activates CREB2 synthesis (parameters Vy, Kx) and CREB2 then represses CREB1 synthesis (parameters Vx, Ky).
Measures to assess synergistic drug actions are diverse and include additive synergism [32]–[34], Bliss independence [32], [35]–[37], the Chou-Talalay Combination Index [37], the isobolographic approach [32], Loewe additivity [33]–[35], and nonlinear blending [38], but there is no agreement on which is preferable. Given the diversity of methods for measuring synergism, it is useful to adopt more than one method in studies of combination drug treatments. In the present study nonlinear blending synergism and additive synergism were selected to assess the effects of combinations of parameter variations on the output of the motifs. These methods were chosen because: 1) Additive synergism is a straightforward way to calculate synergism and can be easily conceptualized, and 2) Nonlinear blending allows for synergism to be assessed by considering the shape of a curve constructed over a range of concurrent drug dosages, as opposed to assessing at a single dose. In this way, nonlinear blending is closely related to several of the more complex methods for calculating synergism, such as isobolograms.
In nonlinear blending [38], a fixed total amount of drug 1 is selected, which gives a substantial, but not saturated, response. Then the response is quantified for mixtures of drugs 1 and 2, holding the total drug amount the same as with drug 1 alone, but varying the percentage of drug 1 in the mixture from 0 to 100%. The resulting dose-effect curve, with percentage of drug 1 on the x axis, will be concave up if antagonism is present between the drugs and concave down if synergism is present. Weak nonlinear blending synergism is present if the curve is concave down, but still has its maximum at an end point [38]. Weak nonlinear blending may not be useful, however, because the maximal response is still obtained at one endpoint, using only one parameter change. However a stronger, more useful form of synergism may be seen, termed strong nonlinear blending synergism [38]. In this case, the curve is concave down with the maximum response at a point removed from either end point (Fig. 2A). Thus, for a given total drug amount, the best response is obtained for a mixture of drugs 1 and 2.
For each nonlinear blending (NB) curve, a corresponding additive effect (AE) curve can be constructed. The same combination of drugs is used as for the NB curve. However, the additive effect is simply calculated as the sum of the response from with the altered value of drug 1 alone and the response with drug 2 alone (Fig. 2B). Additive synergism is then present if the response to the combination of drugs 1 and 2, shown by the NB curve, is greater than the sum of the responses to the individual drugs, shown by the AE curve.
In the simulations, the effects of pairs of drugs were mimicked by varying pairs of parameters. In the canonical model, 14 parameters in Eq. 1–4 determine the dynamics of A, B and T. The peak level of T was considered to represent the response to the stimulus S (Fig. 1A2). The peak level of T in the absence of parameter changes was regarded as the control peak level. The percentage increase of the peak level with parameter changes over the control peak level was taken as the simulated response to the parameter changes. Individual parameters were varied in the direction that increased the peak level of T. For 14 parameters, 91 distinct combinations of two parameters are possible. To simulate a reasonable range of drug effect strengths, each individual parameter was varied within the range bounded by 90% changes of its standard value, either increased (0%–+90%) or decreased (0%–−90%), according to which direction increased the peak level of T. When two parameters were modified simultaneously, the sum of absolute values of individual parameter percent changes was maintained at 90%. The value of 90% was chosen so that the maximal degree of reduction, or inhibition, of biochemical processes governed by these parameters – synthesis, degradation, or activation/deactivation of A, B, or T – was 90%. For example, kdA/kST is one pair that was selected. For simulations with this pair, kdA was decreased, because a decrease in the degradation rate of A tends to increase the peak level of the target T, and kST was increased, because an increase in the maximum induced synthesis rate of T also increases the peak level of T. In a series of combined parameter changes, the decrease in kdA varied from 0% to 90% of its standard value, and the increase in kST concurrently varied from 90% to 0% of its standard value, such that |% decrease in kdA|+|% increase in kST| = 90%. To simulate dose-effect curves for individual parameters, as well as to construct NB curves, 30 points were used, evenly spaced between 0 and 90% parameter variations from standard values, evenly spaced between 0 and 90% parameter variations from standard values.
Synergism was quantified by defining the degrees of NB and additive synergism. The left end point of an NB curve corresponds to a 90% change in parameter 2 (i.e., 0% change in parameter 1). The right end point corresponds to a 90% change of parameter 1. The greater value of these two endpoints was considered as the maximal single-parameter effect (Fig. 2A). Then:(16)
Based on Eq. 16, the degree of weak NB synergism (i.e., NB curve is concave down and the peak response is obtained from either end point of NB curve) is always zero. Therefore, this degree was only calculated for parameter pairs that exhibit strong NB synergism, with degree >0. All other pairs that exhibit weak NB synergism were assigned a degree of 0. A negative degree corresponds to antagonism (i.e., NB curve is concave up).
The degree of additive synergism was calculated using the maximal difference between the NB curve and AE curve, as follows (see also Fig. 2B).(17)A positive degree >1 corresponds to additive synergism, whereas a negative degree corresponds to antagonism.
Fourth-order Runge-Kutta integration was used for integration of ODEs, with a time step of 3 s. The model was programmed in XPP-Aut version 6.1 (http://www.math.pitt.edu/~bard/xpp/xpp.html). The XPP-Aut code is provided as Supplemental Material.
Perhaps the simplest model for studying synergism is a three-node model of two pathways that converge onto a single target. Here, each pathway represents stimulus-induced activation of an enzyme, which in turn modifies the activity or level of a target effector molecule. For generality, the activity of each pathway is simply represented as the level of an element (A or B). A and B converge to induce activation of the target, T (Fig. 1A1). To characterize the responses of this motif (and the additional motifs described below), a stimulus (S) is modeled as a brief (10 min) square-wave pulse, rising from a basal value of zero to an elevated value. To examine whether the strength of S affects synergism, the latter value is varied from weak (1) to strong (40) (non-dimensional units). After the pulse, S returns to zero. S concurrently activates A and B. The activities of A and B are dynamic variables A and B that follow first-order ODEs (see Model Development). Two variants of this model were analyzed. In Variant M, the effect of the two pathways on the expression of T is multiplicative. The activation rate of T is proportional to the product of the effects of A and B. Thus, Variant M is equivalent to a logical AND gate. In Variant A, the activation rate of T is proportional to the sum of the effects of A and B. Each element is thus able to separately activate T. Thus, Variant A is similar to a logical OR gate.
For Variants M and A, after systematic simulation of the effects of modifying 91 pairs of parameters, the degrees of NB and additive synergism for each parameter pair were evaluated (Fig. 3). Strong stimuli might induce a ceiling effect, which could preclude the synergistic effect of combined parameter changes. Therefore, in these simulations S = 1. In Variant M, 21 parameter pairs had a degree of NB synergism >0, and 74 pairs had a degree of additive synergism >1. In Variant A, only 7 parameter pairs had a degree of NB synergism >0, and 34 pairs had a degree of additive synergism >1. The histograms of the degrees of NB synergism and additive synergism for the 91 parameter pairs show that: 1) None of the parameter pairs in Variant A had a degree of NB synergism >20, whereas 7 parameter pairs in Variant M had a degree of NB synergism >20 (Fig. 3A), and 2) No parameter pairs in Variant A had a degree of additive synergism >20, whereas 19 parameter pairs in Variant M had a degree of additive synergism >20 (Fig. 3B). Thus the multiplicative effects of A and B on T in Variant M produced stronger NB and additive synergism than did the additive effects of A and B on T in Variant A.
In Variant M, all parameter pairs displaying NB synergism also produced additive synergism, whereas parameter pairs displaying additive synergism did not necessarily produce NB synergism. For example, with Variant M, the kbasal_A/KTA pair produced antagonism in the NB curve (i.e., the NB curve is concave up). However, this pair still produced additive synergism (Fig. 3C). In contrast, with Variant A, the KB/KT pair produced an AE curve that, in some portions, was above the NB curve, even though NB synergism was present (red box in Fig. 3D). Therefore, additive synergism may be observed in the absence of NB synergism, and NB synergism may be observed in the absence of additive synergism. When the AE curve is above the NB curve, the effect of the combined parameter variations (the NB curve) is less than the sum of the effects of the individual parameter variations (the AE curve). This corresponds to additive antagonism, and yields a negative degree (Eq. 17).
Because Variant M produced, on average, stronger NB and additive synergism than did Variant A, subsequent simulations were performed with Variant M.
In Variant M, for simplicity, the standard parameter values of elements A and B were identical, thus the time courses of A and B following a stimulus S were identical (Fig. 4A1). To test whether a difference in the dynamics of these two elements would affect synergism, parameters for element B (ksB, kbasal_B, and kdB) were all reduced by 50%. This modification made the activation and deactivation of B somewhat slower than that of A. Because of the relatively slower kinetics, the peak amplitude of B was reduced by ∼50% (Fig. 4A2), and the subsequent decay of B was slowed. NB and additive synergism were simulated after parameter modification. The histograms of NB and additive synergism degrees were compared between the models with fast B dynamics (original Variant M) and with slower B dynamics (Fig. 4B). Only slight shifts were observed in the distribution of NB and additive synergism degrees with slower B dynamics (Fig. 4B). In this case, similar to the original model Variant M, 9 parameter pairs had a degree of NB synergism >10, and 36 pairs had a degree of additive synergism >10 (Fig. 4B). In another test, the activation of B by S was delayed by 60 min compared to activation of A (Fig. 4A2), and the effect on synergism was assessed. With delayed B activation, 7 parameter pairs still had a degree of NB synergism >10, and 29 pairs had a degree of additive synergism >10 (Fig. 4B). Therefore, NB and additive synergism in Variant M were robust to moderate variations in dynamics.
In further simulations, the percent reduction of ksB, kbasal_B, and kdB was increased to 70% and then to 90%, which made the activation and deactivation of B much slower than A. These changes led to a decrease in the number of parameter pairs showing synergism. When parameters for B were reduced by 70%, 7 parameter pairs had a degree of NB synergism >10 and none had a degree >40. 33 pairs had a degree of additive synergism >10 but none had a degree >40. When parameters for B were reduced by 90%, only 3 parameter pairs had a degree of NB synergism >10 and none had a degree >20. 28 pairs had a degree of additive synergism >10 but none had a degree >40. Although 50%, 70% and 90% were arbitrarily selected, they effectively represent the range of substantial variation of kinetics of B. These results suggest that the use of combined parameter changes might be less effective in inducing synergy if two converging pathways have very different dynamics. In the simulations, the highest degrees of synergism were produced when the basal parameter values governing the dynamics of both pathways were identical. Therefore, further analyses of effects of stimulus strength and network motifs on synergism were performed for the optimal initial condition (identical basal dynamics of elements A and B).
To further test the robustness of Variant M, parameter sensitivity analysis was performed for the dynamics of T. Six parameters that affect the dynamics of T (Eq. 3) were altered by either −25% or +25% from their standard values. NB and additive synergism for the 91 parameter pairs were simulated after each of these 12 parameter modifications. The histograms of NB and additive synergism degrees were compared for the cases of standard and varied parameters. Both NB and additive synergism were robust to these moderate variations in T dynamics. For example, after reducing or increasing kST by 25% or reducing KT by 25%, 9 parameter pairs exhibited a degree of NB synergism >10, and 36 parameter pairs had a degree of additive synergism >10 (Figs. 4C–D). These pair numbers were the same as with standard parameter values. After increasing KT by 25%, 7 parameter pairs exhibited a degree of NB synergism >10, and 38 parameter pairs had a degree of additive synergism >10 (Fig. 4D).
Initially, the effects of stimulus strength and network motifs on synergism for all 91 parameter pairs were examined. However, the most dramatic changes in the dynamics of the motifs were related to alterations in the deactivation rate constants of elements A and B, kdA and kdB. Therefore, the detailed analyses of effects of stimulus strength and network motifs on synergism concentrated on variations to the three parameter pairs (kdA, kdB), (kdA, kST), and (kdA, KT), each of which includes at least one of these deactivation rate constants. Each of these pairs consistently displayed substantially different degrees of synergism. For example, with S = 1, throughout almost all parameter variations, kdA and kdB maintained degrees of both NB and additive synergism >40, kdA and kST maintained degrees of NB synergism >5 and additive synergism >20, and kdA and KT maintained a degree of additive synergism >10 and a degree of NB synergism = 0. For the kdA/KT pair the NB curve remained concave down with a maximal at an endpoint (weak nonlinear blending).
Decreasing kdA or kdB prolongs activities of A or B. When a decrease in kdA was paired with a decrease in kdB, strong NB synergism resulted for S = 1 (Fig. 5A1). kST is the activation rate of T induced by A and B. When a decrease in kdA was paired with an increase in kST, moderate NB synergism resulted for S = 1 (Fig. 5B1). KT is the dissociation constant for deactivation of T. A decrease in kdA was paired with an increase in KT, which directly elevated the peak level of T. This pair of parameter changes produced no strong NB synergism (degree of 0), and yielded only weak additive synergism, when S = 1 (Fig. 5C1).
The AE curves of these pairs under different strengths of stimulus were simulated: S = 1 (Figs. 5A1, B1, C1) and S = 10 (Figs. 5A2, B2, C2). For S = 1, additive synergism is also evident for the kdA/kdB pair. The NB curve was well above the AE curve (Fig. 5A1). For the kdA/kST pair, additive synergism was strong, with the NB curve lying significantly above the AE curve (Fig. 5B1). For kdA/KT, only slight additive synergism remained (Fig. 5C1). Interestingly kdA/kdB, which produced the strongest NB synergism, also produced the greatest increase of T (Figs. 5A1, B1, C1). This result indicates that the lack of strong NB synergism with the other two pairs is not due to saturation of the peak level of T.
When S = 10, for kdA/kdB, both NB synergism and additive synergism were still evident (Fig. 5A2). For kdA/kST, the NB curve was only slightly above the AE curve, and NB synergism was almost absent (Fig. 5B2). For kdA/KT, additive synergism was reversed, because the AE curve was slightly above the NB curve (antagonism) (Fig. 5C2). Because S = 10 induced a higher control peak level of T than S = 1 in the absence of any parameter changes, the additional percentage increases of the T peak due to parameter variations were smaller when S = 10 (Figs. 5A2, B2, C2) than when S = 1 (Figs. 5A1, B1, C1), which indicated some saturation of the peak level of T for the higher stimulus, S = 10. However the kdA/kdB pair, which produced the strongest NB synergism, yielded a greater increase of T than did kdA/KT (Figs. 5A2, C2). This result indicates that the saturation of peak T is not the only reason for reduction of synergism.
The results of Fig. 5 indicate that increasing the strength of the stimulus tends to reduce NB synergism and additive synergism. To further test this hypothesis, the degrees of NB synergism and additive synergism were measured under a series of stimuli ranging from S = 1 to S = 40. Fig. 6 illustrates dose-effect curves of synergism vs. stimulus strength. Although, for all three pairs of parameters, NB synergism and additive synergism gradually decreased with increasing S when S>10, the dose-effect curves were not monotonic, displaying a type of inverted U-shaped curve. Each pair required a different stimulus strength to maximize each type of synergism. For each pair, maximal NB synergism and additive synergism could occur at distinct stimulus strengths.
Three network motifs are included in the model of Song et al. (2007) (Fig. 8A1): a positive auto-regulatory loop in which CREB1 enhances its own synthesis; a negative auto-regulatory loop in which CREB2 represses its own synthesis; and a negative feedback loop in which CREB2 represses the synthesis of CREB1. There are 8 parameters in Eqs. 14 and 15, leading to 28 possible distinct combinations of two parameters. The effects of these three motifs on NB synergism and additive synergism were examined by varying all 28 parameter pairs.
In Song et al. [26], when the standard parameter values are used, the model has two stable steady states: LOW and HIGH states. Vx is transiently increased from 0.4 min−1 to 3.7 min−1 to simulate the application of a neurotransmitter, 5-HT [26]. A standard 5-HT protocol that is widely used to induce long-term synaptic facilitation (LTF) (5 pulses of 5 min 5-HT with interstimulus interval of 20 min) [39] is simulated, by increasing Vx for 5 min for each of the 5 pulses. After the 5-HT stimulus, CREB1 and CREB2 were switched from the LOW state to the HIGH state (Fig. 8A2). The ratio of [CREB1]/[CREB2] after they converge to the HIGH state was considered to represent the response to 5-HT. An increase in this ratio corresponds to an increase in the level of the transcription activator CREB1 and/or a decrease in the level of the transcription repressor CREB2. Therefore an increase in this ratio corresponds to increased induction of genes necessary for LTF. LTF is a cellular correlate of long-term memory (LTM), and an increase of [CREB1]/[CREB2] corresponds qualitatively to improved formation of some forms of LTM. The [CREB1]/[CREB2] ratio in the absence of parameter changes was regarded as the control, and the percentage increase of the ratio over the control was taken as the simulated response to each of the 28 variations in parameter pairs. As was done previously for the three-node motifs (Methods), for each parameter pair, individual parameters were varied in the direction that acted to increase the response, i.e., the [CREB1]/[CREB2] ratio.
After systematic, concurrent variation of each of the 28 pairs of parameters, only one pair of parameters, Vx/Kdy, was found to exhibit degrees of NB and additive synergism that exceeded 20 (Fig. 8B). Vx is the maximum induced synthesis rate of CREB1, whereas Kdy is the degradation rate of CREB2. In these simulations, Vx and Kdy were both increased, because for both parameters, an increase acts to elevate [CREB1]/[CREB2]. Increasing Vx enhances the strength of the positive auto-regulatory loop of CREB1. Increasing Kdy suppresses the negative feedback between CREB1 and CREB2 and suppresses the negative auto-regulatory loop of CREB2 by accelerating the degradation of CREB2. Thus increasing positive auto-regulation while simultaneously suppressing negative auto-regulation was found to be the only way to obtain substantial NB synergism and additive synergism. These simulation results are consistent with the earlier results from the three-node motifs, that positive auto-regulation acts to enhance NB synergism, and that negative auto-regulation acts to suppress synergism.
The effects of varying parameter combinations in the context of common motifs in biochemical pathways were simulated, and for each motif, parameter combinations were identified that have a relatively greater likelihood of exhibiting synergism. For every motif (Fig. 1), simulations quantified degrees of NB synergism and of additive synergism for an extensive set of 91 parameter pairs. Three of these pairs, which each exhibited substantially different amounts of synergism, were then used to analyze parameter combination effects in detail. By concurrent simulation of NB and AE curves, we were able to visualize which pairs produced synergism (e.g., Fig. 5A), and which produced antagonism (e.g., Fig. 3C). Simulations demonstrated that NB synergism can exist without additive synergism, and vice versa (Figs. 3C, 3D).
For the basic motif of two converging signaling pathways (Fig. 1A1), the model variant with a multiplicative effect of the pathways on the regulation of a target (Variant M) produced stronger synergism than did the model with a simply additive effect of the two pathways (Variant A). Thus, combined parameter variations may be less efficacious if they occur within signaling pathways that have additive effects on a downstream target. However, and as discussed below, a multiplicative mechanism does not guarantee synergistic effects in other motifs.
The synergism produced by Variant M was robust to modest variations of parameters governing the dynamics of signal pathways. In contrast, synergism was much more sensitive to the strength of stimulus and to the presence of feedback or feed forward interactions between elements and their target. NB and additive synergism were decreased when the stimulus became strong (Fig. 6). These results are not surprising, because increasing stimulus intensity will saturate the signaling cascade. This ceiling effect will prevent the combined parameter changes from further enhancing the activation of their target. However, the relationships between synergism and stimulus strength were non-monotonic (Fig. 6). Each parameter pair exhibited maximal synergism at a different stimulus strength. For example, kdA/kdB produced maximal NB synergism when S = 4, but kdA/kST produced maximal NB synergism when S = 2 (Fig. 6). When S = 8, a simulated parameter combination was still effective (i.e., synergism was observed) for kdA/kdB, but not for kdA/kST (Fig. 6). These simulations suggest that empirically, if the physiological stimulus strength activating a given pathway varies, the optimal choice of combined drug therapy to target that pathway might also change.
Mutual excitation and mutual inhibition feedback loops are two motifs commonly observed in signaling pathways [29]. Variation of a pair of parameters yields synergistic effects in mutual inhibition motifs (Figs. 1B3, 7, S2), but has opposite effects on motifs with a mutual excitation loop (Fig. 1B2), where both NB and additive synergism (Figs. 7, S1) were eliminated. This elimination may be due to the self-reinforcing effect of the positive feedback, enhancing the ability of combined parameter changes to saturate the response. Similar results were obtained for paired parameter changes in a positive feed forward motif (Fig. 1B1). These results suggest that it may not be advantageous to design a combined drug therapy for which both drugs activate a positive feedback loop similar to that of Fig. 1B2, or a feed forward loop similar to Fig. 1B1. In contrast, mutual inhibition tends to prevent activation of the target from saturating, and this effect may explain the enhanced synergism. Thus, a promising strategy for combination therapy might be to target drugs to elements of mutual inhibition loops.
The effects of other network motifs on synergism were also examined. A positive auto-regulation loop (Fig. 1C1) enhanced NB synergism only if: 1) the upstream stimulus activating the signaling pathway was weak, and 2) one or both of the parameters were part of the positive auto-regulation loop (Figs. 7, S3). Although positive auto-regulation surprisingly reversed additive synergism for some parameter pairs (Fig. S3), these simulations nevertheless suggest that given conditions 1) and 2), a positive auto-regulation loop may be a worthwhile target of a combination drug therapy. In contrast, a negative auto-regulation loop had a diverse, pair-specific effect on synergism (Fig. 7). When the upstream stimulus was weak, this auto-regulation tended to reduce synergism. Only for kdA/kST and for a strong upstream stimulus was a substantial enhancement (degree >10) seen. These simulations suggest that a signal pathway with negative auto-regulation may not be a good target for combined drug therapies that affect pathway parameters in a similar manner to that modeled here. Simulations with the more biologically realistic model of [26], which contains multiple motifs, supported the above suggestions. With this model, the only way to obtain substantial NB synergism and additive synergism was to suppress negative auto-regulation while substantially enhancing positive auto-regulation.
The effects of a negative feedback loop between the target and an upstream element were similar to those of a mutual inhibition loop between elements (Figs. 7, S2, S4). In the negative feedback and the mutual inhibition loops:
For kdA/kdB, both NB synergism and additive synergism were enhanced regardless of stimulus strength, compared to the model without feedback or inhibition.
For kdA/kST, additive synergism was enhanced when the stimulus was sufficiently strong (S = 10).
For kdA/KT, both forms of synergism were enhanced when the stimulus was sufficiently strong (S = 10).
These results further support the suggestion that some pathways with negative interactions, either mutual inhibition between elements, or inhibition between a downstream target and an upstream element, may be good targets for combined drug therapy, especially when both drugs target the parameters involved in the negative interactions. However, a simple negative feedback loop between A and B (Fig. 1D1) may not be a good target for combined drug therapies that alter parameters in a way similar to that modeled here, because with this negative feedback, the degrees of NB synergism for all three parameter pairs were zero regardless of stimuli (Fig. 7).
Although elements A and B are representative of enzymes acting on a common target (T), the motifs of Figs. 1A1–1D2 are found in signaling pathways that also include regulation of gene expression [1], [29], [40]. Thus A, B, and/or T could alternatively refer to genes, or A and B could refer to transcription factors inducing gene T. Drug combinations targeting elements of motifs with such elements might include antisense RNA or siRNA to affect translation of a regulatory protein, or compounds that affect the activity of transcription factors.
Indeed, numerous feedback loops and feed-forward motifs have been identified in signaling pathways that include gene regulation and that are associated with cancer and other disorders. Cui et al. [4] mapped interactions among ∼1,600 genes associated with oncogenesis. Over 850 three-node FFLs similar to that in Fig. 1B1 were identified, and ∼200 feedback loops similar to those in Figs. 1B2 and 1B3 were found. One well-known example is the p53-mdm2 negative feedback loop [41]. Aberrant activation of the Ras→Raf→MAP kinase pathway is also commonly implicated in cancer [42], and a strong negative feedback loop within which activation of MAP kinase leads to inhibition of upstream Raf kinase has been identified [43]. Considerable effort is also being directed toward developing pharmacotherapies to improve learning and memory in individuals with cognitive deficits due to molecular lesions (e.g., Rubinstein-Taybi syndrome [44]; or neurofibromatosis type 1 [45]). Long-term synaptic potentiation (LTP) is a correlate of learning and memory, and within signaling pathways important for LTP, numerous three-node FFLs and feedback loops have been identified [46]. Interestingly, over the entire human gene network, three-node FFLs are the most common regulatory motif [5].
A series of sensitivity analyses were performed. The results indicated that although synergism and antagonism could be dependent on initial setting of individual parameter space, synergism may be less likely if parameters are concurrently changed in two converging pathways with substantially different dynamics. In the simulations, the highest degrees of synergism were produced when the basal parameter values governing the dynamics of two converging pathways were identical. This finding that choosing two target pathways with similar dynamics tends to favor synergism might be useful in designing some combination drug therapies.
Combination drug therapies are commonly used for complex diseases and neurological disorders such as Alzheimer's disease, depression, traumatic brain injury, cancer, type 2 diabetes, and infections [47]–[54]. A potential benefit of combination therapies is synergism [50], [55]–[56]. With synergism, drugs administered together have a greater effect than would be predicted by simple addition of single-drug effects (i.e., super-additive effects). Thus synergism allows for lower drug doses, minimizing undesirable effects. However, the molecular mechanisms underlying the synergism produced by combined drugs are not well understood for many combination drug therapies. In Axelrod et al. [57], 420 drug combinations were screened in 14 different cell lines. 84 combinations were found to generate synergism in multiple cell lines. The mechanistic analysis did help to suggest possible mechanisms involved in the induction of synergism. For example, the analysis implied that a combination of Ro31-8220 and lapatinib might produce synergism by their compensatory crosstalk between the p70S6 kinase and EGF receptor pathways [57]. However, none of the 84 combinations were synergistic in more than half of the 14 cell lines, and no pattern of lineage specificity was observed. Moreover, the authors found that even compounds from the same family or with similar structures cannot substitute for each other to induce synergism, which reflects the diversity of complex signaling networks.
In the current study, concurrent variation of parameter pairs to increase a response can represent the effect of a pair of drugs targeted to affect the corresponding interactions in a specific pathway motif. Therefore, alterations in parameter pairs that result in both NB synergism and additive synergism, and for which both types of synergism are robust to moderate variations in other parameters, may help to suggest, or to prioritize, drug combinations for empirical investigation.
This study did not focus on examining how motifs affect a parameter pair for varying initial values of all model parameters. Instead, the aim was to investigate how network motifs affect the degrees of synergism associated with variation of different types of parameter pairs, when the initial values of model parameters favor the induction of synergism. In the analysis of effects of stimulus strength and network motifs on synergism, we examined whether the motifs enhanced or eliminated the synergism for the parameter pairs that already exhibited synergism in the canonical model (kdA/kdB and kdA/kST), and whether the motifs helped to generate synergism for the parameter pair that did not exhibit synergism in the canonical model (kdA/KT).
The network motifs examined in the present study are obvious simplifications of the processes taking place in realistic biological networks and the analysis of these motifs is not expected to yield a complete understanding of these networks and the ways in which combined parameter variations affect them. However, models of simplified motifs have their own advantages. They have relatively few parameters and a relatively simple mathematical structure, so it is possible to simulate and analyze the dynamics characteristic of a specific motif, which can be otherwise obscured within a complex network. Moreover, once the key dynamical elements of basic motifs are characterized, it is easier to incorporate these elements into more complicated and biologically realistic models. Under some circumstances, the dynamics characteristic of a single motif may dominate a local biological network [29], [58]. Indeed, Tyson and Novak [58] review several cases for which a single motif appears to dominate the dynamics of a cellular response. In these circumstances, the analysis presented here could be particularly useful for understanding cellular responses to pairs of drugs affecting that motif.
In summary, models such as those studied here provide insights into the dynamical properties of network motifs. Moreover, the stimulus protocols and parameter manipulations used here may help to predict the extent of synergism and consequently may prove useful in prioritizing empirical investigations of combination therapies. If combined drugs target signaling pathways that contain mutual excitation or positive feed forward interactions, a single drug might be as efficient as combined drugs. Whereas, if combined drugs target signaling pathways that contain a negative feedback loop between the target and an upstream element, or a mutual inhibition loop between elements, the combination may be more likely to exhibit empirical synergism. The usefulness of this approach is dependent on knowledge of the motifs that participate in signaling pathways affected by specific disorders. Often this information is currently limited. However, as progress is made in understanding motifs affected by diseases, the use of such computational models in the initial stage of designing combination drug therapies may become a common and efficient methodology.
|
10.1371/journal.pntd.0002504 | Diagnostic Accuracy of Loopamp Trypanosoma brucei Detection Kit for Diagnosis of Human African Trypanosomiasis in Clinical Samples | Molecular methods have great potential for sensitive parasite detection in the diagnosis of human African trypanosomiasis (HAT), but the requirements in terms of laboratory infrastructure limit their use to reference centres. A recently developed assay detects the Trypanozoon repetitive insertion mobile element (RIME) DNA under isothermal amplification conditions and has been transformed into a ready-to-use kit format, the Loopamp Trypanosoma brucei. In this study, we have evaluated the diagnostic performance of the Loopamp Trypanosoma brucei assay (hereafter called LAMP) in confirmed T.b. gambiense HAT patients, HAT suspects and healthy endemic controls from the Democratic Republic of the Congo (DRC).
142 T.b. gambiense HAT patients, 111 healthy endemic controls and 97 HAT suspects with unconfirmed status were included in this retrospective evaluation. Reference standard tests were parasite detection in blood, lymph or cerebrospinal fluid. Archived DNA from blood of all study participants was analysed in duplicate with LAMP. Sensitivity of LAMP in parasitologically confirmed cases was 87.3% (95% CI 80.9–91.8%) in the first run and 93.0% (95% CI 87.5–96.1%) in the second run. Specificity in healthy controls was 92.8% (95% CI 86.4–96.3%) in the first run and 96.4% (95% CI 91.1–98.6%) in the second run. Reproducibility was excellent with a kappa value of 0.81.
In this laboratory-based study, the Loopamp Trypanosoma brucei Detection Kit showed good diagnostic accuracy and excellent reproducibility. Further studies are needed to assess the feasibility of its routine use for diagnosis of HAT under field conditions.
| Diagnosis and effective treatment are cornerstones in the control of human African trypanosomiasis (HAT). Molecular tools such as the polymerase chain reaction (PCR) detect the parasite's DNA and are generally very sensitive and specific. However, PCR is not applicable in field settings because it requires a laboratory infrastructure and sophisticated equipment. A recently developed loop-mediated isothermal amplification (LAMP) has emerged as a simpler alternative to conventional molecular methods for the diagnosis of HAT. The test has been transformed into a diagnostic kit for qualitative detection of the parasite's DNA in clinical specimens, the Loopamp Trypanosoma brucei Detection Kit. In this study, we evaluated this kit in laboratory conditions on DNA extracted from blood samples of 142 patients, 97 suspects and 111 healthy endemic controls in the Democratic Republic of the Congo. The test showed good diagnostic accuracy and excellent reproducibility. Given the practical advantages of LAMP over conventional nucleic acid methods these are promising results. Further studies are needed to assess the test's accuracy and feasibility in field conditions.
| Human African trypanosomiasis (HAT) is a protozoan disease caused by the Trypanosoma brucei species, which are cyclically transmitted by tsetse flies. Two subspecies are pathogenic to man: Trypanosoma brucei (T.b.) gambiense in central and western Africa, and Trypanosoma brucei rhodesiense in east and southern Africa [1]. Currently, less than 10 000 cases per year are reported by the World Health Organization, of which over 70% occur in the Democratic Republic of Congo (DRC) [2].
Diagnostic algorithms for T.b. gambiense HAT generally start using the Card Agglutination Test for Trypanosomiasis (CATT) as initial screening for the presence of antibodies. Those testing positive in CATT are then subjected to parasitological tests for confirmation of the infection [3]. Parasitological confirmation relies on the microscopic search for parasites either in lymph, blood or cerebrospinal fluid (CSF). The most sensitive method is based on the mini-anion exchange centrifugation technique (mAECT), yielding an analytical sensitivity of <50 parasites per mL of blood [4], [5]. However, given the low parasitemia associated with T.b. gambiense infection, some truly infected individuals remain negative in the mAECT.
Because of the limited sensitivity of parasitological confirmation tests, molecular methods have been developed [6], [7] and they generally show high sensitivity and specificity [7]. They can be performed on various specimen types such as whole blood, blood stored on filter paper and CSF. However, the need for laboratory instruments for nucleic acid extraction, amplification and visualization are obstacles to their application in clinical settings in HAT endemic areas [8]. Isothermal reactions such as nucleic acid sequence-based amplification (NASBA) and loop mediated isothermal amplification (LAMP) have recently been developed for the diagnosis of HAT [7], [9]. In contrast to PCR, they do not require thermocyclers and amplification can be conducted in a heating block or a hot water bath. A potential advantage of NASBA is that it targets RNA and thus might have greater utility as a test of cure compared with DNA-targeting molecular tests [6]. However, NASBA is not yet ready to be used under field conditions due to the complexity of RNA purification [6]. Instead of RNA, LAMP amplifies DNA that is less prone to damage during transport and storage of samples and during extraction. Sets of specific inner and outer primers are needed for autocycling strand displacement DNA synthesis by the Bst DNA polymerase at a temperature between 60–65°C. The results can be interpreted by several detection formats, such as turbidity, fluorescent DNA intercalating dyes, fluorescent hybridisation probes and oligochromatography [9], [10].
There are published reports on two LAMP assays for Trypanozoon DNA. One assay targets the single copy paraflagellar rod protein A (PfrA) gene and the second is based on the repetitive insertion mobile element (RIME) [9]. Recently, the latter has been transformed into a commercially available kit, the Loopamp Trypanosoma brucei kit (Eiken Chemical Co LTD, Japan in collaboration with FIND, Geneva, Switzerland) [9]. Ready-to-use reaction tubes are provided with the reagents dried down in the caps of the tubes, together with negative and positive controls. LAMP showed great promise with purified DNA and with trypanosome-spiked blood but has not been yet evaluated on specimens from HAT patients and controls. We here present the data from the first diagnostic evaluation of the commercial LAMP kit on DNA extracted from blood of 142 gambiense HAT patients, 97 gambiense HAT suspects and 111 healthy endemic controls from the DRC.
All samples analysed in this study were collected within the framework of two earlier diagnostic studies for HAT, PARAHAT and HAT-PolyB. Both studies were approved by the ethical committees of the University Hospital in Antwerp (registration numbers ITG09415684 and B30020108363, respectively) and the Ministry of Health of the D.R Congo (registration numbers M-D/226/2010 and M-D/179/2010, respectively). The ethical committees approved extended use of the samples in further HAT diagnostic studies. Written informed consent was obtained from all study participants and all samples were anonymized.
For this retrospective evaluation we used DNA extracts from blood of study participants recruited consecutively in 2010 in Bandundu, the most HAT endemic province in DRC [11]. From all participants testing positive on CATT whole blood, the CATT was repeated with sequential plasma dilutions and the end titer was recorded [3]. For diagnostic purposes all were subjected to parasitological confirmation, irrespective of CATT results. Trypanosomes were detected by examination of lymph node aspirate (in subjects with swollen cervical nodes) or by blood examination (all study subjects) with the capillary centrifugation technique (CTC) [12], mAECT on whole blood [8], and mAECT on buffy coat [5]. For patients with parasites detected in the lymph or blood, or with a plasma CATT end titer ≥1∶8 a lumbar puncture was done. Parasite detection in CSF was performed with the single modified centrifugation technique [4]. DNA was extracted from blood with the Maxwell® 16 Blood DNA Purification robot (Promega Corporation, Madison, WI,USA) from 200 µL blood stabilised in an equal volume of DNA stabilising GE buffer (6 M guanidium, 0.2 M EDTA, pH = 7.5). Final DNA extraction volumes were 300 µL and extracts were stored at −20°C. Time between DNA extraction and LAMP testing was 1.5 to 2 years. All blood samples were also analysed with a Trypanozoon-specific 18S rDNA PCR in duplicate [13]. This PCR amplifies a 120 bp DNA sequence of the Trypanozoon 18S rRNA gene and the amplified product is visualized using conventional electrophoresis in agarose gels and ethidium bromide staining. All PCR testing was done in duplicate at the Institute of Tropical Medicine in Antwerp.
Participants were considered as HAT patients if parasites were detected by any parasitological method in any blood or lymph or CSF sample. Healthy endemic controls were recruited during active screening in the villages [14]. Healthy endemic controls are individuals presenting themselves for CATT screening but with no clinical symptoms of HAT, no previous history of HAT and negative results in CATT whole blood, trypanolysis and mAECT. Individuals with suggestive clinical findings, a positive CATT (cut-off titer ≥1∶4) and positive trypanolysis test that were not confirmed as cases on parasitological testing and who had no previous history of HAT, were classified as HAT suspects. Altogether, frozen DNA from blood of 350 study participants were tested by LAMP: 142 from confirmed HAT patients, 97 from HAT suspects and 111 from healthy endemic controls. In the confirmed HAT patient group, standard tests showed parasites in the blood in 131 cases while in 5 and 6 cases parasites were only found in the CSF and lymph respectively.
The Loopamp Trypanosoma brucei Detection Kit (Eiken Chemical,Taito-ku, Tokyo, Japan) was applied in duplicate on the DNA extracts by one of the authors (PM), a trained clinical microbiologist, who was blinded to the disease status of the samples. The test was performed at ITM Antwerp according to the product insert. Briefly, the dried reagents in the tube were reconstituted in a 25 µl reaction solution, containing 3 µl template DNA and 22 µl negative control buffer, and immediately placed in the LAMP incubator (LF-160 incubator, Eiken Chemical co,Taito-ku, Tokyo, Japan). LAMP amplified Trypanosoma brucei DNA was visualised using the provided UV-LED device. Amplified DNA emits green fluorescence while there is no fluorescence in negative samples. The provided positive and negative controls were taken in each run (14 tests) to validate the test results.
Sensitivity and specificity values and their 95% confidence intervals were calculated for the LAMP in the confirmed HAT patients and in healthy endemic controls, respectively. The sensitivity was defined as the proportion of confirmed HAT patients who are positive by the index tests and specificity as the proportion of healthy endemic controls who are negative by the index test. Each DNA extract was tested in duplicate by LAMP. Agreement between LAMP and PCR and reproducibility of LAMP were assessed on all specimens (patients, suspects, controls) with Cohen's Kappa and interpreted following the grading system described by Landis and Koch (1977) [15]. Data were analysed in Stata, version 11.1 (StataCorp, College Station, Lakeway, Texas, USA).
Of the 142 HAT patients, 132 and 124 were LAMP positive in respectively the first and second run, corresponding with sensitivities of 93.0% (95% CI: 87.5%–96.1%) and 87.3% (95% CI : 80.9–91.8), respectively (table 1). Of the 11 patients with trypanosomes detected only in lymph or in CSF, 7 were positive in both LAMP runs on blood. Of the 97 HAT suspects, 6 were positive in both replicates of LAMP, 8 and 20 were positive in the first and second replicate, respectively. Of 111 healthy endemic controls, 4 tested positive twice with LAMP and 4 tested positive only once. Specificity estimates range from 92.8% (95% CI 86.4–96.3%) to 96.4% (95% CI 91.1%–98.6%). Sensitivities and specificities of PCR were in the same range as LAMP with overlapping confidence intervals (table 1).
Assessed on all participants (patients, suspects and healthy controls), agreement between the two LAMP replicates was excellent with a kappa value of 0.81 (95% CI: 0.71–0.92) (table 2), which is in the same range as the PCR replicates (kappa value = 0.82, 95% CI: 0.72–0.92). Agreement between the first replicate of LAMP and the 18S PCR was also excellent with a kappa value of 0.82 (95% CI: 0.72–0.93). Kappa values of LAMP replicates were lower in the subgroups but in the same range as for PCR and with overlapping confidence intervals (table 2).
In this diagnostic accuracy study, the LAMP showed a sensitivity of 87.3% and 93.0% in the two testing runs. Specificity was 92.8% and 96.4%, with a lowest lower limit of the 95% confidence interval of 86.4%. Agreement between LAMP replicates as well as between LAMP and PCR was excellent with kappa values above 0.8.
The sensitivity of the commercial LAMP kit tested here was equivalent to that of the 18S PCR test, which showed a sensitivity between 87.3% (95% CI: 80.9–91.8) and 90.1% (95% CI: 84.1–94.0) on the same DNA extracts. This is in line with the observation that both tests showed identical analytical sensitivities of 100 parasites per mL of blood in a head-to-head comparison using experimentally prepared blood samples (data not shown). While the LAMP detects the RIME DNA elements (500 copies per haploid genome) [16], the PCR targets the 18S rRNA gene (10–100 copies) [17]. In the 11 confirmed HAT patients with parasites only detected in the lymph or CSF, 7 were positive in both LAMP runs. In contrast, we also observed 5 false negative LAMP results in mAECT positive patients. In the HAT suspects, who could not be confirmed by the parasitological methods, we observed particularly poor agreement between the two LAMP repetitions (kappa = 0.35). These discordances are probably due to the fact that the target DNA concentration in such samples is at the detection limit of the test. If LAMP is to be used to confirm non-confirmed HAT suspects, testing multiple samples from the same patient may increase its sensitivity.
The specificity of the LAMP kit was in the same range as the 18S PCR, which showed a specificity between 96.4% (95% CI 91.1–98.6%) and 97.3% (95% CI 92.3–99.1%) on the same samples. The LAMP was twice positive in 4 of the 111 healthy endemic controls. Three of these LAMP positive controls were at least one time also positive by PCR. Some positive healthy endemic controls may actually be infected with T.b. gambiense because the parasitological confirmation algorithm using mAECT is not 100% sensitive, and this may lead to an underestimation of the specificity of the index tests. Another possible reason may be the absence of the LiTat 1.3 variable surface glycoprotein (VSG), which is the antigen used in the CATT and the trypanolysis test, in some strains of T.b. gambiense [18], [19]. In addition, low antibody titers may be present in early or latent infections [20]. However, false positive LAMP results due to non-specific amplification reactions cannot be excluded. Since the LAMP detects the RIME DNA of all Trypanozoon, a transient human infection with T.b. brucei could also have led to a positive test result [21]. The recently developed LAMP assay that targets the T.b. gambiense specific glycoprotein (TgsGP) gene [16] can exclude an infection with other Trypanozoon and thus may be more specific. However, in the same publication the authors showed that the diagnostic sensitivity of the TgsGP LAMP is lower than the sensitivity of the RIME LAMP.
Reproducibility of LAMP was excellent and as good as that of PCR, with kappa values of 0.81 and 0.82 respectively when all samples were considered. Within the sub groups lower kappa values were observed, which is due to the fact that in these more homogenous groups the expected agreements were much higher. Values observed within the groups were in the same ranges for LAMP and PCR. The LAMP-amplified DNA is visualised by a UV-LED device attached to the LF-160 incubator. This single-tube and easy read-out avoids the risk for sample contamination due to post-amplification manipulations. Another advantage is that Loopamp Trypanosoma brucei Detection Kit is thermostable at 30°C which greatly enhances the feasibility of use in peripheral health facilities in tropical countries. Although in the present study LAMP was performed on DNA extracted with the Maxwell® DNA Purification robot, simplified DNA extraction methods that are compatible with LAMP are currently under development. The requirement of electrical power supply to operate the incubator for the amplification step constitutes a potential drawback for use in remote health facilities, even if it can be circumvented by using an alternative power source such as an electrical generator and/or a photovoltaic panel.
In recent years there has been a sharp decline in HAT prevalence in most of the endemic countries and the classical case finding approach by mobile screening units is becoming less cost-effective. There is thus an urgent need to consider alternative ways of surveillance and case detection, and the LAMP technology could play a role [22]. Though still more complicated than the parasitological methods, LAMP is feasible for use at the level of a district hospital laboratory and could be useful as part of a testing algorithm for samples collected at more peripheral levels. LAMP can be applied on samples collected elsewhere without the need to be processed the same day. Either CATT or one of the newly developed rapid tests [23] can be used to screen suspects for HAT at health center or at village level; LAMP can then be used in a more centrally located laboratory as a second step in the diagnostic algorithm. LAMP data for serologically positive individuals who remained negative in parasitological testing may be particularly informative. However, future research should determine if HAT suspects with positive LAMP testing need further diagnostic work-up before being put on treatment and if the detection of LAMP positive individuals from the same geographical origin should be a trigger for intensified surveillance efforts. Although we feel that LAMP is best suited for use in central laboratories, the feasibility and cost-effectiveness of including LAMP in the screening process by the mobile teams may be determined in specific evaluation studies but should also take into account the test specimen preparation prior to the LAMP itself.
In conclusion, the study shows that the LAMP has similar diagnostic accuracy as the 18S rDNA PCR and can replace PCR for accurate and simplified detection of Trypanozoon DNA in clinical specimens. LAMP may have an important role to play in disease surveillance. However, one should note that the specificity of LAMP is not 100%, that HAT treatment is complex and toxic, and that the positive predictive value of tests in low incidence settings is low. Based on this study we cannot yet recommend initiating treatment of patients based on LAMP results; further evidence from prospective clinical studies under field conditions is needed, as well as cost-effectiveness analysis of competing algorithms. Feasibility studies of LAMP are currently conducted in the D.R. Congo (http://www.finddiagnostics.org/programs/hat-ond/hat/molecular_diagnosis.html).
|
10.1371/journal.pcbi.1003269 | Catalysis of Protein Folding by Chaperones Accelerates Evolutionary Dynamics in Adapting Cell Populations | Although molecular chaperones are essential components of protein homeostatic machinery, their mechanism of action and impact on adaptation and evolutionary dynamics remain controversial. Here we developed a physics-based ab initio multi-scale model of a living cell for population dynamics simulations to elucidate the effect of chaperones on adaptive evolution. The 6-loci genomes of model cells encode model proteins, whose folding and interactions in cellular milieu can be evaluated exactly from their genome sequences. A genotype-phenotype relationship that is based on a simple yet non-trivially postulated protein-protein interaction (PPI) network determines the cell division rate. Model proteins can exist in native and molten globule states and participate in functional and all possible promiscuous non-functional PPIs. We find that an active chaperone mechanism, whereby chaperones directly catalyze protein folding, has a significant impact on the cellular fitness and the rate of evolutionary dynamics, while passive chaperones, which just maintain misfolded proteins in soluble complexes have a negligible effect on the fitness. We find that by partially releasing the constraint on protein stability, active chaperones promote a deeper exploration of sequence space to strengthen functional PPIs, and diminish the non-functional PPIs. A key experimentally testable prediction emerging from our analysis is that down-regulation of chaperones that catalyze protein folding significantly slows down the adaptation dynamics.
| Molecular chaperones or heat-shock proteins are essential components of protein homeostatic machinery in all three domains of life, whose role is not only to prevent protein aggregation but also catalyze the protein folding process by decreasing the energetic barrier for folding. Importantly, chaperones have often been implicated as phenotypic capacitors since they buffer the deleterious effects of mutations, promote genetic diversity, and thus speed up adaptive evolution. Here we explore computationally the consequences of chaperone activity in cytoplasm via long-time evolutionary dynamics simulations. We use a 6-loci multi scale model of cell populations, where the fitness of each cell is determined from its genome, based on statistical mechanical principles of protein folding and protein-protein interactions. We find that by catalyzing protein folding chaperones buffer the deleterious effect of mutations on folding stability and thus open up a sequence space for efficient and simultaneous optimization of multiple molecular traits determining the cellular fitness. As a result, chaperones dramatically accelerate adaptation dynamics.
| Evolutionary selection of protein sequences is a complex task whereby several traits such as translation efficiency, structural integrity (i.e. folding stability and kinetics), molecular function, as well as interactions with other proteins in the cellular milieu should be simultaneously optimized. Imposing simultaneous and often contradictive (pleiotropic) constraints on protein sequence evolution severely limits the repertoire of possible solutions in sequence space and thus slows down the evolutionary dynamics. It is widely accepted that strong selective pressure against protein misfolding plays a key role in determining the rate of protein evolution and sustainable mutational loads [1]–[5]. However, other constraints such as the need to avoid protein sequestration to non-functional protein-protein interactions (NF-PPIs) in the cytoplasm are also emerging as important determinants of the rates and outcomes of evolutionary dynamics of proteins [6]–[10].
From de novo folding of nascent polypeptides to refolding of mature misfolded proteins, chaperones or heat-shock proteins assist in maintaining the necessary abundance of folded proteins, compensating for the selective costs of erroneous protein synthesis, misfolding, and sequestration of proteins in NF-PPIs. In three domains of life, chaperones are essential components of protein homeostatic machinery. Chaperonins, like GroEL, effectively catalyze the folding process by increasing the rate at which misfolded proteins are converted into their folded conformations [11]–[13]; this process can lead to diminished aggregation and NF-PPIs due to the limited presence of aggregation-prone misfolded species in the cytoplasm. Lindquist and others posited that chaperones may act as phenotypic capacitors by buffering the fitness effects of deleterious mutations [14], leading to a greater genetic diversity and speeding up adaptive evolution [15], [16]. A recent in vivo study from our lab [12] also showed that the chaperone action in dynamic cellular milieu can be pleiotropic, i.e. it extends beyond the immediate effect of protein folding by reducing the participation of destabilized proteins in NF-PPIs and affecting their accessibility to ATP-dependent proteases .
Apparently, chaperones play a key role in sculpting the fitness landscape of organisms. However, understanding the evolutionary implications of this fact requires a multi-scale modeling that realistically represents the mechanism of chaperone action and reaches across the necessary length and time scales. Recently, we developed a multi-scale evolutionary model for population dynamics simulations [7], where the fitness (rate of division) of each cell is derived explicitly from its genomic sequence by using the physical principles of protein folding and interactions. The model provided insights into the co-evolution of molecular properties of proteins, their abundances in the cytoplasm, and their functional and NF-PPIs. Here we significantly extend this ab inito model to explicitly account for chaperone activity in the cytoplasm of model cells. The model elucidates not only the immediate pleiotropic effect of chaperone action on cellular fitness but also its long-term evolutionary consequences. We find that the chaperone activity provides a significant acceleration of adaptive evolution by minimizing the detrimental effect of protein misfolding and therefore opens new paths in sequence space for efficient and simultaneous optimization of multiple molecular traits, determining the fitness of model cells.
Our ab initio 6-loci model cells contain explicit genomes that encode six essential, birth rate controlling, proteins that are modeled as 27-mer lattice proteins as introduced in [7]. The advantage of this coarse-grained protein model is that a crucial conformational subset, consisting of all maximally compact conformations, can be enumerated [17], making the calculations of binding and folding stabilities exact within a selected representative conformational ensemble. At the initial stage of the simulations, each protein in the model is assigned a conformation, which is deemed folded and thus functional, and each protein complex in the functional PPI network is assumed to be functional only in one specific docking mode out of 144 possible ones [7]. The model of Ref. [7] considered NF-PPIs only between folded proteins. Here we also take into account the misfolded compact Molten Globule (MG) states of proteins [18] by modeling the ensemble of unfolded states as maximally compact yet non-native conformations (see Methods). As shown in Fig. 1A, we allow all proteins in their folded and MG states to interact with each other in the cytoplasm of model cells to form functional and non-functional protein complexes. Experimental studies show that GroEL and several other chaperones do not interact strongly with proteins in their native state, see e.g. [19]–[22]. Therefore, here we only consider interactions between the model chaperone and proteins in their MG state. As shown in Fig. 1B, the interaction surface of the chaperone is modeled as a 2D (3×3) lattice fragment, consisting of nine amino acid residues that are found in the apical domain of the chaperonin GroEL and that have been shown to be essential for substrate binding [23].
We assume that functional protein complexes constitute the same prototypical PPI network as in [7]: the first protein is active in monomeric form, the second and third proteins are functional as a heterodimer, and finally, the fourth, fifth and sixth proteins form a “date triangle” where they function in various combinations of pairwise complexes between them (Fig. 1A). We then postulate, as in [7], that the division rate of an individual cell is a product of the functional concentrations of proteins for the postulated prototypical PPI network:(1)Here is a parameter used to scale the rate and thus the time, is the postulated “optimal” total concentration of proteins, which reflects the assumption that protein synthesis comes at a cost, are the total concentrations of individual proteins, and is a control parameter that defines a fitness penalty for deviation from the optimal total concentration of all proteins. Overall, the role of the denominator in Eq. [1] is to penalize the deviations from the optimal protein levels and to avoid a fitness gain by a mere overexpression of proteins. Hence, the cell division rate in our model is determined by a fitness function, which stems from an intuitive physical-biological assumption that a subset of gene products acts in concert to promote healthy cell divisions.
In what follows, we define the functional concentrations of monomer and dimers in Eq. [1] as(2)where is the Boltzmann probability that proteins i and j interact with each other in a specific docking conformation (see Methods), is the concentration of the monomeric protein product of gene 1 in its native folded form, is the concentration of the binary complex formed by the folded states of proteins i and j.
We employ a simplified two-step kinetic model to describe the catalytic activity of active chaperones, as illustrated in Fig. 1C (see Methods for the technical aspects of the formulation). In this active model, the chaperone acts as a catalyst to accelerate the rate of protein folding. As a control, we also consider a passive model of chaperone action, whereby the role of chaperone is simply to bind and release proteins in their MG states. It is noteworthy that, in contrast to a conventional catalyst, which decreases the activation barrier for both forward and backward reactions, an active chaperone increases the rate of conversion of misfolded proteins into their folded form without increasing the rate of reverse reaction of unfolding. Such “one-way” catalysis, which requires consumption of ATP, increases the concentration of folded species, which is equivalent, under steady state conditions, to an effective increase of thermodynamic stability of a protein as outlined in [12]. We model binding of MG proteins to the chaperone with a pre-equilibrium assumption since the association/dissociation of chaperone with an MG protein is a fast process as compared to subsequent kinetic steps in which the actual protein folding occurs. It has been shown that these later kinetic steps, which lead to folding, are rate limiting as they almost always require ATP hydrolysis [24].
Examples of active chaperones with catalytic folding activity include the chaperonins, GroEL in prokaryotes [24] and TRiC in eukaryotes [25]. While the applicability of our model is not limited to the GroEL-like chaperonins, the catalytic activity of this class of chaperones has been well established, see e.g. [13], [24]. Therefore, our model directly applies to this class of chaperones, which forms a good experimental system to test our predictions. Henceforth, unless otherwise indicated, we refer to the chaperones with catalytic activity simply as chaperones.
We explored the effect of chaperones on evolutionary dynamics by running long time evolutionary simulations (200,000 generations) of model cell populations. Our simulations start from monoclonal populations of model cells, whose sequences have been designed by using the method reported in [26] to provide high stabilities for all 6 proteins in their folded states without regard for their functional and NF-PPIs (see details in Methods). In our model, the acceleration of protein folding rate due to chaperone action is determined by the parameter x, which is the ratio of the rate at which a folded protein is released by the chaperone to the rate at which spontaneous protein folding occurs (defined in Methods).
To determine the effects of chaperone buffering on adaptive evolution, we tested two models – an active and a passive model. In the active model, the chaperone acts as a catalyst and accelerates protein folding. However, in the passive model, the chaperone assumes a simple role by merely binding and releasing proteins in their MG states. While for the passive model we set for the active model we assume a modest throughout this work, consistent with the estimates of the dynamic model given in [12]. In both cases we keep the chaperone concentration fixed at . To highlight the role of chaperones we always, in parallel, run control simulations for cells without chaperones, i.e. setting .
To determine broadly the effect of chaperones on adaptation dynamics we ran evolutionary simulations at three different temperatures, i.e. T = 0.85 (low), T = 1.05 (medium), and T = 1.25 (high). Throughout this work, all temperatures are in units calibrated to Miyazawa-Jernigan (MJ) potentials [27].
The effect of chaperones on the evolution of fitness is presented in Fig. 2 as fitness ratio, i.e. the ratio of birth rate in the model with chaperone to that without chaperone. Fig. 2A shows the time evolution of for the active model . The chaperones provide dramatic fitness benefit during the adaptive evolution, especially at early stages. The effect of chaperones is more pronounced at higher temperatures, where proteins in the MG state are more prevalent. While the fitness ratio reaches its peak of 100 at intermediate adaptation times for T = 0.85, it peaks at 250 for T = 1.05 and dramatically over 1000 for T = 1.25. After the initial fast adaptation period, the relative fitness effect of chaperones abates. Up to this point, however, the cells already have gained a considerable fitness advantage, and in the long time limit, we see gradually declining fitness ratios as the organisms become more and more fit. Nevertheless, the evolutionary dynamics with chaperones always leads to a higher long-time fitness than the evolutionary dynamics without chaperones, although the final fitness ratio is not as dramatic as those observed at intermediate evolutionary times.
Fig. 2B shows the time evolution of for the passive chaperone model . It is clear that the chaperones in the passive model do not provide a noticeable fitness gain but rather a small fitness loss (due to the sequestration of proteins by chaperones) at the initial stage of adaptation for all three temperatures. In light of these results, we conclude that the active folding of proteins by chaperones is necessary to provide a fitness benefit to cells in the evolutionary dynamics. In the following, unless otherwise indicated, we present the data only for the active model at the low temperature T = 0.85 as representative of our general results.
Now, we turn to a detailed account of the evolutionary dynamics of the physicochemical properties of proteins, i.e. their stabilities and functional interaction probabilities for the functional heterodimers and date triangles (see Methods for the definitions of these quantities). We present the time evolution of for the monomer in Fig. 3A, for the heterodimer proteins in Fig. 3B, and for the date triangle proteins in Fig. 3C. The chaperones provide a noticeable increase in stability for the monomeric proteins, as seen in Fig. 3A. Interestingly, at the initial stage of adaptation within 500 generations, the monomer loses its stability considerably by accumulating destabilizing mutations in the presence of chaperones. However, subsequent mutations bring about a rapid turnaround, resulting in a very stable monomer, which persists throughout the rest of the evolutionary dynamics. The non-monotonic dependence of stability of the monomeric protein on evolutionary time is an indication of a chaperone-enhanced epistatic behavior. The chaperone buffering relaxes significantly the stability constraint and allows the accumulation of more mutations in the locus encoding natively monomeric protein. This effect is mainly responsible for the initial sharp drop in the stability of the monomer. The resulting enhanced genetic diversity provides a path to a faster optimization of collective properties of all proteins in the cytoplasm such as NF-PPIs, as we show below.
The evolutionary dynamics of stability for the heterodimer and date triangle proteins show quite a different trend, as seen in Figs 3B and 3C. Initially, both the heterodimer and date triangle proteins lose their stability, but later on, the stability of date triangle proteins is slowly restored. However, in striking contrast to the monomer, the stability of heterodimer and date triangle proteins in the presence of chaperones shows a downward trend with evolutionary time, as compared to that of the chaperone-free cytoplasm of model cells.
The evolution of strengths of functional interactions, reflected in the parameter for the heterodimer and date triangle proteins, is given in Fig. 3D and 3E, respectively. The chaperones significantly increase for both the heterodimer and date triangle complexes. for the heterodimer increases rapidly within first 1000 generations, in the presence of chaperones. The rate of increase of for the date triangle is slower than that for the heterodimer; nevertheless, the chaperones provide a significant increase in for the date triangle complexes as well. Hence, our results show that the chaperones shift the balance between the strengths of functional interaction and stability of proteins in favor of the former at the expense of the latter. Indeed, it is more advantageous for faster adaptation that the heterodimer and date triangle proteins primarily develop strong interaction surfaces to contribute to the fitness. High stability of proteins establishes later on once the strong functional interaction between them is ensured.
While the effect of chaperones on protein stabilities and interactions is significant, it cannot fully account for the huge overall fitness increase, which transiently reaches up to a factor of 100 at the low temperature T = 0.85 (see Fig. 2A). Therefore, there must be another factor affecting fitness, where the effect of chaperones appears even more pronounced. To that end, we turn to the analysis of NF-PPIs, which affect fitness through modulation of concentrations of proteins in their functional form. We find that at the early stages of adaptation the chaperones dramatically decrease the concentrations of protein complexes engaged in NF-PPIs, releasing more proteins to become functional. This can be seen in Fig. 4, where we plot the time evolution of the fraction of protein material wasted in NF-PPIs in the absence and presence of chaperones. Specifically, we present the time evolution of for the monomeric protein and for the heterodimers and for the date triangles, where is the total concentration protein i and is the total concentration of protein i involved in NF-PPIs. Fig. 4 shows that the vast majority of proteins are lost to NF-PPIs at the beginning of evolutionary runs, where the sequences are optimized for stability only without regard for functional PPIs. Apparently, at the very early stage of adaptation, cell resources are mostly wasted unproductively to NF-PPIs. Both Fig. 4A and 4B show that the chaperones give rise to a rapid increase in the functional concentrations of monomeric and heterodimer proteins within the first 5,000 generations. As shown in Fig. 4C, the rate of decrease of NF-PPIs for the date triangle proteins is slower than that observed for the monomer and heterodimer proteins; nevertheless, with chaperones, it still occurs at the early stage of adaptation within 10,000 generations.
We find therefore that, while the chaperones interact directly with proteins to affect its molecular properties, their greatest impact on cellular fitness occurs indirectly through the optimization of a collective property of all proteins in the cytoplasm of model cells, namely, their NF-PPIs.
Our results indicate that the chaperones significantly accelerate the rate of adaptive evolution. Customarily, a well-known parameter , where and are the non-synonymous and synonymous substitution rates, respectively, represents a quantitative measure of evolutionary rate. A straightforward approach to calculate and at any time step in simulation is to compare the genome of the dominant clone in the population to the initial starting genome. However, we find that this approach is problematic for our model in the long time limit when multiple substitutions at a single site become frequent. Here, we employ a slightly different approach. Following Wilke [28], we define the evolutionary rate as where and are the cumulative non-synonymous and synonymous substitution counts summed over short time intervals of 100 generations, and and are arithmetic means of weights for non-synonymous and synonymous sites, which account for different degeneracies of codons in the genetic code, calculated over time frames of 100 generations, see Methods for details.
We summarize our results, averaged over multiple evolutionary runs, in Fig. 5A to 5C for to highlight the type and magnitude of selection acting on different proteins at different stages of adaptation. In Figs., from 5D to 5F, we present the cumulative weighted non-synonymous substitutions for different types of proteins in our system. Further, in Fig. S1, we provide the synonymous substitution rates . The evolutionary dynamics of , and for individual trajectories are also given in Fig. S2.
We found that at the very early stage of adaptation, after 500 generation, the chaperones induce a strong positive selection pressure on the monomer, which lasts, in average for about 10000 generations, after which the monomer falls under purifying selection. However, without the chaperones, the monomer evolves under positive selection only for a short time between 1500 to 6000 generations. On the other hand, without the chaperones, the net selection on both the heterodimer and date triangle genes is negative, apparently due to the dominance of the stability constraint. In the presence of chaperone buffering, however, these loci evolve under positive selection for about 8000 to 10000 generations before they revert back to purifying selection. An important generic effect apparent in the time evolution of is that the chaperone buffering relaxes the negative selection pressure on all proteins and promotes the fixation of a greater number of beneficial mutations. Therefore, after the initial stage of fast adaptation, when all genes evolve under positive selection, we observe that the chaperones bring all genes closer to neutral regime in the adapted populations.
Next, we evaluated the effect of chaperones on the polymorphism in evolving populations of model cells. To that end we determined the average sequence entropy for each protein locus in our model. This quantity is determined from the alignment of gene sequences between all model organisms within the population (see Methods). These results are presented in Fig. 6.
Overall, we find that chaperones greatly enhance polymorphism in evolving populations. For all protein types, the sequence entropy rapidly increases within a few hundred generations with chaperone. For the monomer and date triangle proteins, the entropy stays approximately at the same level for the duration of an evolutionary run after the initial fast increase. For the heterodimer proteins, however, the entropy gradually increases reaching a level, which is almost two times higher than that for the monomer and date triangle proteins. A greater degree of polymorphism observed for the heterodimer proteins helps these proteins evolve faster than other loci in the model, as we note below.
The enhanced neutrality due to chaperone buffering also increases the rate of protein evolution considerably. Indeed, as seen in Figs., from 5D to 5F, the chaperones increase the net number of non-synonymous mutations for all loci. Initially, the monomer still evolves with the chaperones faster than the heterodimer and date triangle genes. However, the rate of evolution of heterodimer is the fastest as a result of more phenotypic diversity of this gene in population as indicated by the entropy plot (see Fig. 6B). Apparently, the evolutionary rates of the heterodimer and date triangle loci are slower that that of the monomer throughout the evolutionary dynamics with or without chaperones. Finally, Fig. S1 shows that the rate of synonymous substitutions is approximately the same for all protein types, as could be expected. However, we also see that the rate of synonymous substitutions is slightly faster with chaperones as compared to that of chaperone-free evolution. Such slightly faster evolution of synonymous substitutions might be due to hitchhiking of neutral mutations with beneficial ones that should more pronounced with chaperone evolution.
Our ab intio cell model, while much simpler than real biological systems, captures the essence of biological complexity that stems from the fact that the main effects, epistatic effects, and pleiotropic effects on different parameters often act in antagonistic directions. The pleiotropic concept of optimization of antagonistic traits in evolutionary biology, which gives rise to a complex fitness landscape, has its analog in the concept of frustration in physics, where competing interactions lead to a complex energy landscape with many suboptimal minima equal to or close to global minimum [29], [30]. In our model, the molecular traits, whose optimization might be antagonistic, include protein stability, abundances, functional, and NF-PPIs. An earlier study showed how antagonistic constraints result in a peculiar co-evolution of protein abundances and functional PPIs [7]. Here we introduced a new essential component of the cellular milieu – the chaperone activity, which enhances the conversion of proteins from the MG state to their native conformation. The chaperone action in our model partially relaxes an essential constraint on protein sequences to maintain high stability of proteins. The resulting chaperone buffering dramatically affects evolutionary dynamics by opening up sequence space, to provide a dramatic acceleration of the adaption process. We find that only the active chaperone model has a strong effect on evolutionary dynamics, while the passive chaperone model, where an MG protein is bound to the chaperone to prevent its sequestration to NF-PPIs has no effect on fitness (Fig. 2B). However, an important caveat here is that our model does not consider an irreversible aggregation and other elements of protein quality control such as proteolytic activity. Hence, the passive model might also be efficient when all the kinetic aspects of protein quality control are taken into account.
Mechanistically, our main finding is that the chaperones act pleiotropically to affect fitness in a number of ways. Firstly, the relaxation of the stability constraint allows achieving stronger functional PPIs at the expense of lower thermodynamic stability for the proteins participating in the functional PPIs (Fig. 3). Secondly, a more dramatic effect of the chaperone on cellular fitness stems from faster and greater decrease of the NF-PPIs in the course of the evolutionary dynamics (Fig. 4). The NF-PPIs are a collective property of all proteins in the cellular milieu. There is evidence that proteins in their MG state are largely responsible for NF-PPIs [31]. The active chaperone in our model converts the proteins in their MG states into their folded conformations, leading to a drop in NF-PPIs with an ensuing increase in fitness due to a diminished sequestration of functional proteins.
Recent experimental and theoretical studies with the chaperonin GroEL corroborate some of our findings. Tokuriki and Tawfik performed a series of random mutagenesis experiments on a number of non-endogenous enzymes expressed in E. Coli to investigate the impact of overexpression of GroEL on enzyme evolution [16]. They found that GroEL indeed helps folding of destabilized proteins and potentially facilitates the evolution of enzymes to gain new functions. The acceleration of adaptive evolution by GroEL is also found in a recent study [15] in which the evolutionary rates of GroEL clients and non-clients [32] were compared. It was found that the GroEL obligatory proteins evolve 35% faster than the proteins that fold spontaneously without the GroEL assistance [15]. The importance of GroEL for adaptive evolution is highlighted by the case of the endosymbiotic bacterium Buchnera, which often undergoes population bottlenecks through maternal transmission and thus quickly accumulates random mutations that destabilize proteins [33]. Remarkably, the expression level of GroEL in Buchnera is almost 8 times greater than that of E. coli under the normal conditions.
Quayle and Bullock define evolvability as the number of generations that it takes for a population to reach its phenotypic target that maximizes fitness [34]. Our study highlights the dual role of chaperones not only as a catalyst of protein folding but also as a catalyst on the fitness landscape, which lowers the genetic “barriers” between phenotypes and thus promotes evolvability. A key prediction emerging from our analysis is that the catalytic activity of chaperone gives rise to a dramatic acceleration of adaptive evolution. Hence, we predict that the depletion of active chaperones through down-regulation of their expression should directly affect the rate at which organisms adapt to new environments, which can be directly experimentally testable. This work is currently in progress in our lab.
Our proteins consist of 27 amino acid residues that fold into 3×3×3 cubic lattice conformations [17]. We use the MJ potentials to model intra- and inter-molecular interactions [27]. While the 27-mer lattice model has 103,346 maximally compact conformations [17], we employ a uniform subset of randomly selected 10,000 conformations as our conformational ensemble to speed up calculations [7]. We calculate the Boltzmann probability of folding to a native state, i.e. for each protein as follows(3)where is the energy of the native conformation and is the temperature in units corresponding to the calibration with MJ potentials.
We model the functional protein-protein and protein-chaperone interactions using a rigid docking scheme. The six faces of a cubic lattice provide six possible interaction surfaces and there are four rotational degrees of freedom to dock two interaction surfaces of two lattice proteins. Hence, in total, there exist 6×6×4 = 144 docking modes for a binary protein complex. The Boltzmann probability of interaction between the dimer proteins i and j are calculated as(4)where is the interaction energy of the functional binding mode, are the interaction energies for 144 docking modes and is the temperature. Because 27-mers are quite small (as compared to real protein sizes), in order to represent both interior and interaction surface of proteins, we employed two different lattice conformations to model our proteins: one for interior part that determines stability and one for interaction surface that determines PPI, as has been done in previous studies, see e.g. Ref. [7]. Hence, the lattice conformations that we used to represent protein surfaces in order to calculate are randomly chosen conformations but they are not the same lattice conformation that we used to represent the stability energetics of native folds. This approach provides a less tight coupling between interior and exterior of proteins that would be the case for small 27-mers representing therefore a more realistic description of protein geometry and energetics.
In the absence of chaperones, the folded state and the unfolded ensemble of states (which also includes compact MG states) for any protein “i” are at equilibrium, satisfying detailed balance with the corresponding folding and unfolding rates :(5)
The active chaperone changes this picture dramatically. In general, the operation of chaperones requires input energy by ATP hydrolysis. The energy flux due to the ATP hydrolysis by chaperone causes the violation of detailed balance between the folded and unfolded forms of a protein. Therefore, following the findings in [19]–[22], we assume that the chaperon interacts with a protein in its misfolded MG conformation to form a pre-equilibrium dimer complex , from which the protein is released in its folded form ,(6)where are the pre-equilibrium constants for the chaperone-unfolded protein complex, and are the rate constants for the chaperone assisted folding. While the native state is uniquely defined by a single conformation in our model, the unfolded states constitute an ensemble of conformations, which we take into account as a representative ensemble via a mean field approximation (see below).
The steady state solution of Eqs. 5 and 6 leads to the following,(7)By introducing the ratio of the rate constant for chaperone assisted folding to the rate constant for unassisted folding, i.e. and also by making use of the following two equilibrium relations,(8)we arrive at the equation,(9)where is the ratio of the rate of release of native proteins from the chaperone complex to the rate of spontaneous folding.
Our simulations start from initial sequences designed to be stable in their respective native conformations ; the PPIs of initial sequences were not optimized. We randomly assigned the functional docking modes for the heterodimers and date triangles. In order to keep protein folds fixed throughout the simulations, we discarded the cells encoding proteins whose assigned native folds were no longer the lowest energy configurations. A constant population size of M = 1000 is maintained throughout the simulations.
We use a variant of the Gillespie algorithm to generate stochastic evolutionary trajectories in our simulations. Using two uniformly distributed random numbers, , the algorithm decides when and which cell undergoes a cell division. Given the normalized birth rates as where for each cell i in a population of size we define the cumulative probabilities as . Note that and . The waiting time dt for the next cell division to occur at time t+dt is determined by . The cell “i” divides when the second random number falls in the interval .
Upon cell division, a mother cell gives birth to a daughter cell. A newborn cell replaces a randomly chosen cell in the population in order to maintain constant population size. Also, whenever a mutation changes the lowest energy protein fold or hits a stop codon, such cells are discarded from the population. Upon semi-conservative replication, both the mother and daughter cells are subject to either a mutational event with constant probability m = 0.001 per gene per replication (whereby a nucleotide is randomly changed) or the expression level of one randomly chosen protein in a cell can change with a constant rate er = 0.01 per cell division such that the concentration of a protein i in a daughter cell is derived from that of a mother cell as follows where is a Gaussian random number with zero mean and variance 0.1. At the beginning of our simulations, we set the concentrations of each protein and chaperone equally at .
Six proteins encoded in our cell model make four functional interactions in total. In addition, we allow all possible non-functional interactions between all proteins in their folded as well as MG conformations (see Fig. 1A). More specifically, we consider the equilibrium reactions, forming homo- as well as heterodimers between the folded proteins,(10)between the folded and unfolded proteins,(11)and between the unfolded proteins,(12)Next, we discuss how we calculate the equilibrium constants for protein-protein and protein-chaperone interactions. We use the index set for different molecular species, the subscripts i and j for different proteins, and the superscripts n and m for different protein conformations. Given the two conformations n and m of lattice proteins i and j, the binding constant can be written as(13)where are the interactions energies and is the temperature.
In order to reduce computational cost in calculating the binding equilibrium constants, we make use of a mean field approximation in which we choose representative MG conformations randomly out of 10,000 conformations and assume that each of these conformations is equally likely to occur in the MG ensemble. In what follows, we calculate the binding equilibrium constants for the dimers formed by the folded proteins exactly as(14)and the dimers formed by the folded and unfolded proteins as(15)The binding equilibrium constants for the heterodimers, i.e. , formed by the unfolded proteins are calculated by(16)and the homodimers formed by the unfolded proteins are given by(17)To model the protein-chaperone interactions, we use a 3×3 square lattice face to mimic an interaction surface for the chaperone (See Fig. 1B). For the protein-chaperone interactions, there are 1×6×4 = 24 docking modes. Hence, the binding constant for an unfolded protein Ui with conformations n and the chaperone is of the form,(18)By using our mean field approximation, the pre-equilibrium constant for the association of chaperone with an unfolded protein is given by(19)The conservation of mass for each protein and chaperone in our system can be written as(20)
Due to the non-linear nature of the law of mass action (LMA) equations, a direct integration of these coupled equations may only be possible for small systems. However, by using iterative algorithms, the LMA equations can easily be solved even for large systems. Previous iterative algorithms were developed to solve the LMA equations involving only equilibrium reactions between different molecular species. The LMA equations in our system involve not only different molecular species but also equilibrium reactions between conformational isomers of the same molecular species and therefore may not be solved by the existing algorithms, see e.g. [35]. Here, we present a straightforward generalization of the existing iterative algorithms.
We start our iterative algorithm by initializing the concentrations of monomeric unfolded proteins and chaperone, i.e. and . Our algorithm consists of iterations of three sets of equations to find equilibrium concentrations of all chemical species in our system. First, by substituting and into the right hand side of Eq. [9] we determine . Second, by using the new value of along with and we calculate the two quantities:(21)Third, we find the new values of and by using,(22)By updating the old values of and with the new values, i.e. and , and substituting them back into Eq. [9], we continue our iterations until the error threshold is achieved where is defined by(23)
In order to determine the degree of polymorphism in a population, we calculated the sequence entropy for each protein by using the formula(24)where is the frequency of amino acid type “i” in the jth position in the multiple alignment (among all cells in the population) of sequences of a protein “k”.
In calculation of we used the following formula , where and are the normalized cumulative non-synonymous and synonymous substitutions rates, and and are the arithmetic mean of non-synonymous and synonymous substitutions reflecting the instant composition and degeneracies in the genetic code [28]. In order to calculate the quantities , , and we partitioned the overall simulation time into the time intervals of length 100 generations. Given the two DNA sequences, say, DNA-1 and DNA-2 that are 100 generations apart, we first count the number of synonymous and non-synonymous substitutions between them, and second determine the number of synonymous and non-synonymous sites at the frames in this time interval, where is non-degenerate, is 2-fold and 4-fold degenerate sites, respectively [36]. By using the above quantities, we next calculate the cumulative sums and , over all time intervals and the arithmetic averages and , and finally determine
|
10.1371/journal.pgen.1005528 | Ty3 Retrotransposon Hijacks Mating Yeast RNA Processing Bodies to Infect New Genomes | Retrotransposition of the budding yeast long terminal repeat retrotransposon Ty3 is activated during mating. In this study, proteins that associate with Ty3 Gag3 capsid protein during virus-like particle (VLP) assembly were identified by mass spectrometry and screened for roles in mating-stimulated retrotransposition. Components of RNA processing bodies including DEAD box helicases Dhh1/DDX6 and Ded1/DDX3, Sm-like protein Lsm1, decapping protein Dcp2, and 5’ to 3’ exonuclease Xrn1 were among the proteins identified. These proteins associated with Ty3 proteins and RNA, and were required for formation of Ty3 VLP retrosome assembly factories and for retrotransposition. Specifically, Dhh1/DDX6 was required for normal levels of Ty3 genomic RNA, and Lsm1 and Xrn1 were required for association of Ty3 protein and RNA into retrosomes. This role for components of RNA processing bodies in promoting VLP assembly and retrotransposition during mating in a yeast that lacks RNA interference, contrasts with roles proposed for orthologous components in animal germ cell ribonucleoprotein granules in turnover and epigenetic suppression of retrotransposon RNAs.
| Cells undergoing changes in gene expression programs such as nutritional deprivation and other stresses exhibit formation of ribonucleoprotein (RNP) complexes. In Saccharomyces cerevisiae, the majority of investigations to date involve analysis of P-body (PB) and stress-granule RNP formation following nutritional stress. Few studies have investigated RNP formation induced by the mating-MAP-kinase pathway. In this study, we examined how this process influences genome stability via control of retrotransposon activation. During the mating response, expression of the retrotransposon Ty3 is induced and Ty3 virus-like particles form in RNP clusters called retrosomes. We show that mating retrosomes contain PB components that are essential for Ty3 expression, re-localization of Ty3 RNA and protein from polysomes into foci, and retrotransposition. In animal germ cell lineages, PB components are found in perinuclear complexes with RNA interference suppressors of retrotransposition. We speculate that when RNA interference is relaxed and retrotransposition is observed, some members of these complexes play positive roles as we observe in budding yeast.
| RNA processing bodies (PB) are ribonucleoprotein (RNP) granules that contain proteins associated with cytoplasmic deadenylation, decapping and 5’ to 3’ degradation of RNAs in eukaryotic cells [reviewed in [1–5]]. In previous work we showed that artificial overexpression of the long-terminal-repeat retrotransposon Ty3 in the budding yeast, Saccharomyces cerevisiae, causes formation of foci of Ty3 RNA and protein and virus-like particles (VLPs) and that components of PB are required for retrotransposition [6, 7]. Formation of these foci is disrupted by the artificial overexpression of Ty3 RNA and Gag3 assembly mutants [8–12]. These and other data support the interpretation that these foci represent VLP assembly factories or retrosomes [8, 13]. Ty3 is naturally induced and retrotransposes in mating cells [14, 15], and mating cells form foci containing PB components [16]. The present investigation was undertaken to identify proteins that associate with Ty3 capsid proteins during VLP assembly, to determine whether Ty3 expression causes the formation of RNP granules reported in mating cells, and to determine the roles of these in mating cell retrotransposition.
RNP formation is associated with cells undergoing changes in gene expression programs including cells undergoing nutritional deprivation and other stresses [17–20]. Yeast PB contain components of deadenylation-dependent degradation [1]. These include CCR4-Not1-Pop2 deadenylation complex proteins; translational repressor and decapping enhancer DEAD box helicase Dhh1/DDX6 [21]; Ded1, DEAD box helicase [22]; translational repressor, decapping enhancer, and scaffolding factor Pat1; scaffolding factor Edc3 [23]; decapping proteins Dcp1 and Dcp2; and 5’ to 3’ exonuclease Xrn1 [24]. Although it was originally thought that yeast RNPs that contained deadenylation dependent degradation factors were PB dedicated to RNA turnover, a more complex view is emerging. In addition to concentrating factors associated with RNA degradation, RNPs contain RNAs that can resume translation [25, 26] and RNPs dominated by PB components can sponsor formation of stress granules (SG) the function of which is to sequester non-translating RNAs [19, 27–29]. SG and PB contain overlapping components with SG generally lacking deadenylation and decapping factors and containing poly(A) RNA binding proteins, translation initiation factors and small ribosomal proteins in addition to non-translating RNAs [1, 30].
The finding that overexpressed Ty3 proteins and RNA concentrate with PB factors was surprising not only because it suggested that Ty3 poly(A) RNA was somehow resistant to degradation, but because in animal cells PB components include proteins implicated in retroelement restriction [31–33]. These include APOBEC cytidine deaminases [34–37]; DEAD box helicase MOV10 [38, 39]; and some fraction of Argonaute/PIWI components of RNAi [40–42]. Despite the absence of RNA interference (RNAi) in budding yeast, Dicer and Argonaute components introduced from related yeasts, suppress Ty1 retrotransposition [43].
The association between Ty3 overexpression and formation of RNPs in growing cells was especially interesting because the regulation of formation of PB and their differentiation from SG is incompletely understood. In yeast, Edc3 and Pat1 are important for assembly of PB with Edc3 considered to play a scaffolding role and Pat1 interacting with several other components [44–46]. Proteins with P/Q/N rich or prion-like domains are present in PB and at least some of these, including yeast Lsm4 subunit of the Lsm1-7 also foster PB formation [45, 47]. In addition to these structural interactions, RNPs are dynamic in response to conditions signaled through modifications of protein and RNA components. In animal cells O-linked glycan modification of proteins enhances inclusion in PB [48], but this modification has not been reported in budding yeast. YTH and Tudor [49] domain proteins are found in PB, resulting in enrichment of RNAs containing M6A [50] or proteins, for example Sm domain proteins or PIWI proteins, containing di-methyl arginine [51], respectively. CAMP-dependent PKA phosphorylation of PB component Pat1 inhibits its association with other PB components, and reversal leads to PB, but not SG formation [52]. In addition, the mitogen-activated protein (MAP) kinase signaling pathway is implicated in signal transduction from the environment to cellular components of PB. For example, phosphorylation of Dcp2 by Ste20 kinase, an upstream kinase in MAP-kinase signaling, is required for Dcp2 accumulation in PB [53].
Two major classes of long terminal repeat (LTR) retrotransposons, Ty1/Copia and Ty3/Gypsy, are expressed in S. cerevisiae during periods when expression programs are changing in response to environmental signals. Surprisingly, genetic screens for host factors discovered that PB components enhanced retrotransposition of both classes of elements [7, 54]. Ty3, the subject of this investigation, is 5.4 kb in length, including LTRs of 340 bp. In the S. cerevisiae reference strain BY4741 there are two endogenous full-length Ty3 elements and thirty-eight additional Ty3 LTRs, most likely formed by recombination between the LTRs of full-length elements (reviewed in [8]). In animal cells retrotransposons are de-repressed and retrotranspose during transient de-methylation in germ cell lineages [33, 40]. In a somewhat parallel cycle, Ty3 proliferates into new genetic backgrounds by retrotransposing in haploid mating cells [14, 15]. Exposure of mating type MATa and MATα cells to pheromone from the opposite mating type stimulates the mating MAP- kinase pathway [reviewed in [55]]. This pathway initiates changes to enable haploid cells of the two types to form diploid zygotes by inter-cellular fusion. Among these changes, cells arrest in G1 to synchronize cell cycles and induce expression of Ty3 fifty to eighty-fold [14, 56].
Ty3 5.2-kb genomic RNA (gRNA) encodes precursor proteins in two overlapping open reading frames, GAG3 and POL3 [57]. RNA is translated into the 290-aa structural precursor polyprotein, Gag3, containing capsid (CA) and nucleocapsid (NC) domains and into a 1547-aa Gag3-Pol3 precursor polyprotein. In addition to the Gag3 domain, this polyprotein contains catalytic protein domains representing protease (PR), reverse transcriptase (RT), and integrase (IN). Interactions between Gag3 domains facilitate formation of virus-like particles (VLPs) [58]. VLP formation is associated with activation of Ty3 PR, resulting in processing of polyprotein into mature domains [59]. After haploid cells mate and the cell cycle resumes, cDNA synthesis occurs [60]. The cDNA, together with a subset of VLP proteins including IN, is translocated through the nuclear pore complex and integrated at chromosomal targets. Expression of Ty3 from a high-copy- number plasmid under the GAL1-10 promoter causes formation of clusters of assembling VLPs referred to as retrosomes [6].
The current study was undertaken to identify the proteins associated with Ty3 components during VLP formation, to determine how the RNP foci observed in mating cells [16] relate to PB and SG, whether Ty3 expression causes these mating cell RNPs to form, and to test the possible roles of these RNPs in Ty3 retrotransposition. Proteins that interact with Ty3 structural protein Gag3 were identified by mass spectrometry. Viable mutants lacking these were screened for differences from WT in pheromone-induced retrotransposition. We found that activation of the mating MAP-kinase pathway stimulated formation of foci that concentrated multiple PB, but not SG components and Ty3 RNA and protein. Expression of full-length Ty3 enhanced formation of these foci in one mutant background and formation was strongly correlated with retrotransposition. Individual PB components were found to be essential for Ty3 expression, formation of retrosomes, and re-localization of Ty3 RNA and protein from polysomes into foci.
Genetic screens are a powerful way to identify host factors, but are limited with respect to identifying redundant and essential functions and informing as to whether interactions are direct or indirect. To circumvent these limitations, proteins that physically interact in a complex with Ty3 Gag3 were identified using a mass spectrometry (MS) approach. To increase the sensitivity of our analysis, the strain BY4741 (Open Biosystems) was cured of two particle types encoded by killer dsRNA, L-A and L-BC, to yield strain yVB1586 (S1 Text and S1 and S3 Tables). YVB1586 cells in logarithmic phase were induced to express Ty3 and hyper-active Ty3 variant, K15A (S3 Table) for 2 h. Cells expressing Ty3 were broken in a mechanical mixer in liquid N2 (Retsch, Inc.), and the solubilized fraction was separated by chromatography in parallel over control IgG and anti-VLP IgG matrices. Proteins were processed and subjected to liquid chromatography and in-line mass spectrometry (LC MS/MS) Peptides were analyzed by 1DLC-MS/MS using LTQ-Orbitrap XL MS (ThermoElectron) and data were searched and analyzed as described (Materials and Methods)[61]. This analysis identified a set of 154 proteins, of which 106 are not essential (S4 Table). Twenty-four percent of the proteins were designated by Gene Ontology Slim Process (http://www.yeastgenome.org/) as mRNA-binding proteins. This represented significant enrichment for RNA-related proteins over the 2.7% representation in the genome overall (S5 Table, S1 Fig).
To identify Gag3-interacting host factors that specifically restrict or enhance retrotransposition in mating cells, BY4741 (MATa) or gene knockout collection derivative strains were treated with the pheromone, α-factor, to induce expression of Ty3 under the native promoter. Ty3 was carried on a plasmid and tagged with his3AI (Ty3-his3AI)[62]. Ty3 transcripts are spliced, but HIS3 transcripts expressed in the opposite orientation are not. Therefore cells that have undergone retrotransposition can be directly identified by the His+ phenotype. Eight of these strains had increased transposition of more than twofold; transposition went down by more than two-fold in 30 strains. The screen identified fourteen genes encoding translation initiation factors, and components of PB and SG RNPs. The functions of these genes and the effect of deletion on retrotransposition are described in Table 1. Five strains lacking PB (Dcp2, Lsm1, Stm1) or PB and SG (Xrn1, Dhh1/DDX6) proteins and three strains lacking SG (Eap1, Pub1, and eIF4G1) proteins were decreased for Ty3 retrotransposition by greater than two-fold. This was not due to changes in the relative efficiency of splicing of the Ty3-his3AI-reporter in the mutant strains tested (S2 Fig). Six strains lacking PB (Edc3, Pat1, and Sbp1) and SG (Pbp1, Pbp4, and eIF4G2) proteins did not show significant differences from WT. Strains that failed to support WT transposition levels were analyzed further to understand their roles in Ty3 retrotransposition.
To establish the timeframe of assembly of Ty3 VLPs and cDNA synthesis in cells undergoing pheromone induction, haploid yeast strain MATa BY4741 (Table 1), that contains two endogenous Ty3 elements (yGRWTy3-1 and yILWTy3-1), was exposed to pheromone from MATα cells (α-factor) for 2 h (Fig 1). Levels of the 5.2-kb Ty3 gRNA increase dramatically between 0 and 2 h of induction (Fig 1A) [14]. Gag3 polyprotein precursor was detected by 2 h and continued to accumulate up to the 8 h sampling. The Gag3-Pol3 fusion product of frameshifting is made at about one-twentieth the level of Gag [63] and was not detected under these experimental conditions. Gag3-Pol3 contains the PR domain. Upon VLP formation the Ty3 aspartyl protease, PR, is activated to convert Gag3 into CA, spacer and NC, and Gag3-Pol3 additionally into PR, RT, and IN species [59]. Appearance of these forms is consistent with substantial formation of VLPs between 4 and 6 h [9] (Fig 1B). Ty3 VLP maturation occurs in G1, but cDNA synthesis is delayed until after cells undergo fusion to form diploid zygotes and resume the cell cycle in S phase [60]. Between 6 and 8 h, haploid cells recover from pheromone arrest and cDNA synthesis is substantial (Fig 1C).
To understand the relative distribution of cytoplasmic Ty3 components between individual and multimerized forms, cell extracts were analyzed by sedimentation. Cells were pulsed with pheromone for 1.5 h and washed to allow recovery from pheromone arrest. Four hours after initiation of pheromone treatment, VLP formation was assessed by velocity gradient sedimentation. Western, northern, and Southern analysis of gradient fractions indicated that processed VLP components CA, RT, and cDNA concentrated within a few high-density fractions (Fig 1D). The majority of Ty3 RNA was not present in this fraction. The most likely explanation is incomplete packaging, since previous results indicate that after induction for 6 h, only 20–23.4% Ty3 RNA is packaged [9, 10]. The appearance of Ty3 cDNA near the top of the gradient suggests that mature VLPs are unstable under gradient conditions.
To visualize VLPs, BY4741 cells treated with pheromone for 6 h were examined by transmission electron microscopy (TEM). In induced cells (Fig 1E) VLPs accumulated in clusters that were not apparent in uninduced cells. Based on previous analysis of WT, PR and RT mutant VLPs [6, 12], these were a mixture of immature forms with less dense centers and mature forms with dense centers. Overall, mating-cell VLPs were more heterogeneous in morphology than VLPs from the GAL1-induced WT Ty3 element. The inactive YILTy3-1 contains a frameshift in the IN-coding domain [64] and this could contribute to the heterogeneity of mating-cell VLPs.
Having identified proteins that interact with Gag3 and are required for pheromone-induced retrotransposition, we asked whether they are present in pheromone-induced VLP clusters. Strains derived from BY4741 expressing Dcp2, DEAD box helicase Ded1/DDX3, DEAD box helicase Dhh1/DDX6, Edc3, Lsm1, Pat1; Eap1, eIF4G1/Tif4631, Pub1, and Xrn1 fused to green fluorescent protein (GFP) under their respective native promoters (Open Biosystems, Inc.) were transformed with low-copy- number Ty3 plasmid pVB3734 expressing Gag3 fused to mCherry (mCh). Cells in logarithmic phase were either treated with pheromone or left untreated, and were visualized after 4 h by confocal fluorescence microscopy (Fig 2). In the absence of induction, Gag3-mCh fluorescence was minimal and GFP reporters showed patterns varying from diffuse to punctate. Induced cells formed one or a few clearly distinct Ty3 Gag3-mCh foci. PB reporters Dcp2-, Ded1-, Dhh1/DDX6-, Lsm-, Pat1, and Xrn1-GFP fusions formed or enlarged foci that overlapped with these Ty3 foci. Stm1 has genetic interactions with PB factors, but has not been reported to localize to PB foci [65]. The pattern of Stm1-GFP was unafffected by Ty3 expression. Eap1-GFP SG protein reporter formed punctate foci, but these foci seemed qualitatively unaffected by pheromone induction. PB-SG reporter eIF4G1-GFP was robustly expressed and upon pheromone treatment did not form foci in pheromone induced cells. Under conditions in which Gag3-mCh foci formation occurred, SG protein Pub1-GFP was also strongly expressed, but foci were not detected in uninduced or induced cells. Overall, these results indicated that mating cells form foci containing multiple components of PB, but do not concentrate RNP proteins more typically associated with SG into these foci. In cells that formed these pheromone-dependent foci, PB-GFP reporters co-localized withTy3-mCh.
The regulation of PB formation is not completely understood. However, PB component Edc3 is thought to provide a physical scaffold for PB formation in glucose deprived cells, and Pat1, that interacts with Edc3 and multiple additional PB components, contributes significantly to PB assembly [29, 44]. Deficiencies in either of these components reveal an additional contribution of a prion-like domain in Lsm4 [45] We previously reported that galactose-induced, high-copy-number Ty3 expression can drive visible PB formation in logarithmic phase cells [6]. Because evidence presented here and in previous studies of foci of mutant Gag3 suggests that Gag3 physically associates with PB proteins [12] and is multimeric [58, 66] we sought to understand whether Ty3 foci could form in the absence of PB structural proteins. A low-copy-number Ty3 reporter created by fusion of GFP in-frame to POL3 (Ty3-GFP, pTD3548) was used to monitor Ty3 focus formation (Fig 3, S7 Table). We used deletion mutants to test effects of the loss of PB/SG components Eap1, Dhh1/DDX6, Dcp2, Edc3, Lsm1, Pat1, Pub1, eIF4G1, or Xrn1 and PB-related factor Stm1 for ability on Ty3 foci formation. By 4 h, the induced WT strain showed Ty3-GFP foci in approximately 94±5% of cells. In contrast, strains lacking proteins that both localized to foci and were required for retrotransposition showed significantly reduced percentages of cells with Ty3-GFP foci. In order of effect, these were: dhh1Δ (8±3%, p˂0.0001) < lsm1Δ (15±8%, p ˂0.0001) < xrn1Δ (38±5, p ˂0.0001) <dcp2Δ (52±6%, p ˂0.0001). The strain lacking Eap1, which was required for retrotransposition, but did not concentrate in foci, also showed reduced formation of Gag3 foci (48±6%, p ˂0.0001). Decreased percentages of cells with Ty3 foci might reflect reduced amounts of Ty3 protein, disruption of foci formation or a combination of these effects. These possibilities were investigated further as described below. Lack of Stm1 or Pub1, which were required for retrotransposition, but did not concentrate in foci, did not reduce the number of cells with Ty3 foci (pub1Δ, 90±6%, p = 0.5066; stm1Δ, 91±2%, p = 0.778). In the case of pub1Δ, foci appeared larger, although this was not quantified. Notably, lack of Edc3 and Pat1, proteins identified as contributing to assembly of glucose-deprivation PB in fermenting yeast [45] showed lesser effects on mating-cell PB formation (edc3Δ = 76±%, p = 0.0025 and pat1Δ = 93±4%, p = 0.857). This result is in line with the lack of effect of deletion of these proteins on Ty3 retrotransposition frequency (Table 1). Together these results reinforce the interpretation that mating cell Ty3-PB foci are more closely related to PB than SG, but differ significantly from glucose-deprivation PB in that their formation is independent of Edc3 and Pat1 PB assembly factors.
Evidence that mating cell PB contain Ty3 proteins and are resistant to effects of deletions of scaffolding proteins, together with previous results showing that Ty3 expression at high levels actually drives formation of PB [6], suggested that expression of endogenous retrotransposons contributes to formation of mating cell PB foci. Although many retrotransposons including Ty1 are present in high-copy number, full-length Ty3 is represented by only two full-length copies in the Ty3(WT) strain, BY4741 [67], enabling a direct test of its role in mating-cell PB formation. The two endogenous full-length Ty3 elements were deleted from BY4741, creating the Ty3Δ strain (yVB1672). PB formation in these strains was monitored by C-terminal fusion of the chromosomal locus of the PB component DHH1 with GFP. The Ty3(WT) strain was treated with pheromone for 4 h or left untreated, and examined by fluorescence microscopy. As expected, induced cells showed 66% Dhh1-GFP foci compared to 8% in uninduced cells (Fig 4A).
G1 arrest during mating enables synchronous resumption in S phase after zygote formation [68]. To investigate the effect of pheromone-induced cell cycle arrest on PB formation, strains were constructed that are defective in cell cycle arrest, but are otherwise responsive to pheromone. Pheromone stimulation of the mating MAP-kinase pathway activates downstream MAP-kinase Fus3, that phosphorylates Far1, activating its inhibition of Cdc28-Cln, and culminating in G1 arrest. The far1Δ cells retain significant pheromone-mediated transcriptional induction, but fail to undergo efficient G1 arrest in preparation for mating. To assess the contribution of Far1-mediated cell cycle arrest on Ty3 focus formation, FAR1 and far1Δ strains containing Ty3-mCherry (pTD3655) were induced with pheromone and Ty3 focus was measured (Fig 4B). No significant difference was observed between FAR1 and far1Δ strains in the proportion of cells with Ty3-mCherry foci despite reduced Ty3 protein in far1Δ cells (Fig 4C).
To test whether Ty3 expression during pheromone-induced cell cycle arrest affected PB formation, Dhh1-GFP focus formation after pheromone induction was compared in Ty3(WT) and Ty3Δ strains with and without Far1. The results showed that PB formation as measured by Dhh1-GFP foci formation was indistinguishable between Ty3(WT) and Ty3Δ cells [Ty3(WT) = 71±3% versus Ty3Δ = 72±2%] showing that PB focus formation after pheromone treatment was not solely dependent on Ty3 expression. However, deletion of FAR1 from Ty3(WT) cells resulted in reduced Dhh1-GFP foci formation (FAR1 = 71±3% versus far1Δ = 44±1%), indicating that cell-cycle arrest contributes to PB formation. Interestingly, in far1Δ cells Dhh1-GFP foci formation was reduced significantly in the absence of Ty3 [Ty3(WT) far1Δ = 44±1% versus Ty3Δ far1Δ = 32±3%, p = 0.0015]. This result suggested that Ty3 expression although reduced in far1Δ cells nonetheless plays a positive role in PB formation, possibly more apparent when foci formation is attenuated (Fig 4D). Consistent with decreased expression of Ty3 in the far1Δ background, retrotransposition was decreased by about twofold (Fig 4E). Thus, MAP kinase-controlled cell-cycle arrest contributes to mating cell PB formation and in this genetic background, induction of Ty3 expression during mating enhances, but is not essential for mating PB formation.
Yeast PB segregate preferentially into daughter buds [69]. Furthermore, maternal RNA granules in metazoans, affect segregation of important RNAs in subsequent cell divisions. The observation that Ty3 contributes to PB formation suggested that mating might be influenced, either positively or negatively, in Ty3-populated strains. We therefore constructed derivatives of Ty3(WT) and Ty3Δ strains with appropriate nutritional markers to allow selection of diploid cells (see Materials and Methods). These strains were mixed and concentrated onto filters to promote mating for 1.5 or 2.5 h. Mated cells were spread onto double drop-out medium to select for diploid prototrophs with complementing nutritional markers. Mating efficiency at 1.5 h and 2.5 h was similar in Ty3(WT) and Ty3Δ strains (Fig 5A and 5B). To test whether a higher level of Ty3 expression produced an effect, the experiment was performed in the Ty3Δ background in the presence and absence of a high-copy-number plasmid from which Ty3 was expressed under the native promoter. In this context, Ty3 expression slightly reduced average mating efficiency at 3.5 h, but given the variation within the experiment this difference was not significant (Ty3Δ+pTy3 = 42.5±5.0% (mean ± SD) versus Ty3Δ+vector = 49.5±3.0%, p = 0.052) (Fig 5C).
A significant subset of Ty3 Gag3-interacting factors affected Ty3 retrotransposition and assembly focus formation (Table 1). These were further investigated to understand the basis of their effects on retrotransposition. WT, dhh1Δ, eap1Δ, lsm1Δ, pub1Δ, stm1Δ eIF4G1Δ and xrn1Δ cells were induced with pheromone or left untreated and sampled at 2, 6 or 8 h after initiation of treatment. Samples representing each time point were examined for Ty3 RNA, protein, packaging, and cDNA retrotransposition intermediates. Full-length Ty3 and control SNR17a RNAs were measured by quantitative northern blot analysis of RNA extracted from cells induced for 8 h and normalized values were compared (Fig 6A and S3A Fig). Dhh1Δ and eIF4G1Δ cells had the lowest levels of Ty3 RNA; eap1Δ and lsm1Δ cells were modestly reduced; stm1Δ cells had no significant difference from WT; pub1Δ mutant strain was indeterminate, due to variability among independent samples; and xrn1Δ was significantly elevated for Ty3 RNA. These qualitative differences in Ty3 RNA levels among mutants that were reduced for Ty3 retrotransposition suggested that these factors might act at different points in the Ty3 replication cycle with Dhh1/DDX6 and eIF4G1 acting earliest, on the level of Ty3 RNA. Although the difference between dhh1Δ and WT was greatest at 8 h where WT expression was highest, even at 2 h, dhh1Δ strain was reduced in Ty3 RNA (S4A Fig).
Gag3 expression was measured by immunoblot analysis of Gag3 and CA-containing processing intermediates in extracts of cells collected at 0, 2 and 6 h. The Gag3 signal was normalized to loading control Pgk1 (Fig 6B and S3B Fig). Amounts of Gag3 were very low for dhh1Δ cells and reduced for eap1Δ cells. Eap1 is one of two eIF4E binding-proteins with a role in translation repression proposed to occur through competition with eIF4G, which binds to eIF4E during cap-dependent translation initiation [70]. Eap1 deletion would be predicted to positively affect both RNA and proteins levels, which is contrary to our observation. There was a significant decrease Ty3 RNA and protein that could account for the decrease in retrotransposition as the packaging assay indicated that RNA could be packaged with similar efficiency to WT. Low amounts of RNA correlated qualitatively with reduced Gag3. Levels of Ty3 Gag3 in stm1Δ, pubΔ, and eIF4G1Δ cells were not distinguishable from WT. Despite WT levels of RNA, amounts of Gag3 were modestly increased in lsm1Δ cells. Consistent with increased RNA levels, Gag3 levels were higher in the xrn1Δ cells than in WT yeast.
As Gag3 and Gag3-Pol3 assemble into VLPs with RNA, Ty3 PR is activated to process precursors into mature forms [59] and fails to process efficiently in assembly mutants [11, 12]. Although the extent of the phenotype was variable, the CA:Gag3 ratio was decreased in xrn1Δ, potentially indicating a deficiency either in assembly of RNA, Gag3 and Gag3-Pol3 into VLPs or in post-assembly processing. In addition to maturation of polyprotein precursors, increasing Ty3 gRNA resistance to nuclease digestion, as gRNA is packaged into VLPs serves as a measure of assembly [12]. In parallel with preceding experiments, cells were induced with pheromone for 8 h and native whole cell extracts were treated with nuclease or left untreated to evaluate RNA protection (Fig 6C and S3C Fig). Lsm1Δ, eIF4G1Δ, and xrn1Δ cells were significantly decreased for Ty3 RNA protection, indicating that RNA was not efficiently or correctly packaged.
Reverse transcription of gRNA into cDNA occurs subsequent to packaging, maturation of Ty3 RT, and recovery from G1 arrest [71]. CDNA was measured at 8 h by Southern blot analysis in which samples were normalized to a genomic fragment of ARG4 (Fig 6D and S3D Fig). This assay showed that dhh1Δ, eap1Δ, lsm1Δ and xrn1Δ cells had reduced levels of Ty3 cDNA, consistent with reduced RNA and protein for mutants lacking Dhh1 and Eap1 but suggesting a later role for Lsm1 and Xrn1. Xrn1 and Lsm1 were implicated in restricting protein accumulation and promoting RNA protection. Pub1Δ cDNA levels were not significantly different than those of WT; stm1Δ cells had elevated cDNA, indicating that the defects in retrotransposition must occur at a later stage, such as nuclear entry or integration.
The lsm1Δ and xrn1Δ strains showed increased amounts of Gag3, but this contrasted with decreased processing in xrn1Δ cells and decreased packaging of Ty3 gRNA in both mutants. In addition, the Ty3-GFP reporter showed deficient focus formation in lsm1Δ and xrn1Δ cells (lsm1Δ = 15±8% and xrn1Δ = 38±5%) (Fig 3). These properties suggested a defect in transitioning the gRNA from translation into assembly phases or in assembly itself. Because of programmed frameshifting between GAG3 and POL3, Gag3 is about 20-fold more abundant than Gag3-Pol3 [63]. To increase the sensitivity of analysis, cells were induced with pheromone for Ty3 expression for 8 h and Gag3 was analyzed by immunofluorescence using an anti-VLP antibody that reacts with Gag3 (see Materials and Methods). RNA was detected by fluorescence in situ hybridization (FISH) using three pooled oligonucleotide probes complementary to Ty3 internal regions. WT cells showed strong co-localization of Gag3 and Ty3 RNA in cytoplasmic foci (Fig 7A). Analysis of lsm1Δ and xrn1Δ strains showed a punctate cytoplasmic signal interspersed with fewer intense foci than in WT cells (lsm1Δ = 55±3% and xrn1Δ = 35±6%, compared to WT = 71±5%). However, the overall faintness of the RNA signal prevented a determination as to whether the Ty3 RNA in the mutants also formed fewer foci. RNA foci that were observed were co-localized with Gag3 puncta. These observations suggested that Gag3 retained interaction with Ty3 gRNA, but failed to coalesce normally into assembly foci. This interpretation was confirmed by comparison of this phenotype to that of a Ty3 Gag3-NCΔ mutant, lacking most of the NC-coding region so that it fails to interact with gRNA. This mutant failed to form Gag3 foci and did not have detectable co-localization of Gag3 and gRNA [11](Fig 7B). We conclude that despite an abundance of Ty3 RNA and Gag3, lsm1Δ and xrn1Δ strains fail to efficiently form Ty3-PB foci so that protein and RNA remain diffuse.
PB components are associated with both deadenylation and translational suppression, and the fluorescence experiment implicated Xrn1 and Lsm1 in localization of Ty3 RNA and Gag3 to foci. In addition, lsm1Δ and xrn1Δ cells had elevated levels of Gag3, suggesting defects in those PB functions. If Xrn1, deletions of which displayed the more severe phenotype with respect to protein localization and retrotransposition, were required to transition Ty3 RNA from polysomes into assembly, then Gag3 and Gag3-Pol3 polysomes in the mutant should persist relative to those in WT. To test this hypothesis, extracts from WT and xrn1Δ cells were fractionated by velocity sedimentation over sucrose density gradients. WT cells treated with pheromone for 2 h or left untreated and processed for polysome analysis did not show gross differences in the RNA A260 profile (S5A and S5B Fig). WT and xrn1Δ cells were induced with pheromone for 2 h and processed for polysome gradient analysis in the absence (Fig 7C) and presence (Fig 7D) of EDTA. As anticipated, EDTA dissociated ribosomes into 40 and 60S subunits. Alternate gradient fractions were processed for northern blot analysis using Ty3 and ACT1 control probes and immunoblot analysis using anti-CA (Fig 7C and 7D). In the polysome gradients of WT and xrn1Δ extracts not treated with EDTA, Ty3 Gag3 and Ty3 RNA, were concentrated in fractions 17 to 21. In the WT and xrn1Δ EDTA-treated samples, most Ty3 and ACT1 RNA shifted to the top of the gradient as expected for RNAs not associated with polysomes. For extracts from WT cells the pattern of Ty3 Gag3 and CA persisted in fractions 17 to 21 as expected for VLPs that are EDTA resistant (Fig 7D). However, in extracts from xrn1Δ cells, Gag3 and CA shifted to gradient fractions 6 and 7 and only a small amount of precursor Gag3 remained in the fraction normally associated with assembled VLPs (Fig 7D). These results indicated that compared to Ty3 protein in WT cells, most Ty3 protein in xrn1Δ mutant cells is in an EDTA-sensitive form, suggesting that Gag3 associates co-translationally with its RNA template; furthermore, this form accumulates in the xrn1Δ mutant, consistent with delayed transitioning into assembly.
Three important insights into the events between Ty3 retrotransposon transcription and VLP formation resulted from this investigation. First, activation of the mating MAP-kinase pathway stimulates formation of cytoplasmic foci that concentrate PB components together with retrotransposon RNA and protein. Second, PB components Dhh1/DDX6, and Lsm1 and Xrn1 are required for Ty3 expression and assembly of RNA into VLPs, respectively (see S7 Table, Comparison of Ty3 transposition frequency and retrosomes formation in mutant strains). Third, Ty3 expression enhances formation of mating PB foci in cells with attenuated cell cycle arrest. These findings raise questions about how PB proteins facilitate packaging of retrotransposon RNA into VLPs, how any RNA avoids degradation during sequestration in PB and whether retroelements enhance their own survival by promoting formation of perinuclear RNPs that facilitate nuclear entry or transfer during or after mating.
Analysis of RNP foci induced by the MAP-kinase mating pathway identified PB proteins Dhh1, Dcp2, Ded1, Lsm1, Xrn1, Pat1, and Edc3. Because Ty3 proteins formed only one or two foci per cell and individually co-localized with each of these PB proteins, we concluded that individual foci generally contain Ty3 proteins and all or most of these PB proteins. We previously established that induction of WT Ty3 expression under the GAL1-10 UAS, causes formation of foci of PB proteins, but expression of Gag3 assembly mutants disrupts foci formation in various ways [6, 10–12]. These foci were designated retrosomes because of their apparent role in retrotransposon VLP assembly. Haploid cells undergoing pheromone-stimulated MAP kinase signaling were previously reported to form foci containing Dhh1/DDX6 and Dcp2 [16]. Because Ty3 is expressed during pheromone induction and because both PB and SG contain these components, we tested for Ty3 RNA and proteins and additional PB and SG components in these foci. Ty3 and PB components localized to these foci, but two SG proteins tested, Eap1 and Pub1, did not form microscopically visible foci (reviewed in [4, 19, 72, 73]). We concluded that mating cell retrosomes are more closely related to PB than to SG RNPs.
Dhh1/DDX6, Lsm1, Xrn1, and Pub1 were among the proteins required for WT levels of Ty3 retrotransposition. One of the most surprising aspects of Ty3 retrosomes is that Ty3 components including RNA assemble in association with factors implicated in RNA degradation. Further underscoring this unexpected relationship, in the absence of Dhh1, Ty3 RNA levels are dramatically reduced. In yeast, Dhh1/DDX6 is implicated in RNA expression, translation repression, tRNA retrograde translocation into the nucleus during starvation, and RNA turnover [21, 74–76]. However, because homologs of Dhh1 protect maternal RNAs in other organisms [21, 77], one attractive possibility is that Dhh1 shields Ty3 RNA from degradation. Another possibility is that Dhh1 affects Ty3 RNA levels through its role in promoting translation of Ste12 a downstream mating MAP kinase target transcription factor [16] or via its action in transcription elongation [78]. However, these last two roles in transcription seem inconsistent with the observation that expression under the heterologous GAL1 promoter or expression of a truncated form lacking POL3 do not rescue the low levels of Ty3 RNA in dhh1Δ cells (S4A and S4B Fig).
Ty3 RNA and Gag3 levels were modestly elevated in xrn1Δ compared to WT cells, which is consistent with absence of a major exonuclease [24]. However, polysome analysis showed that the xrn1Δ mutant lacked the high molecular weight, EDTA-resistant complexes associated with VLP formation in WT cells. Instead, upon EDTA treatment, Gag3 and RNA shifted upward in gradients, consistent with a greater proportion of Gag3 and Ty3 RNA in xrn1Δ cells remaining associated with polysomes rather than exiting translation for assembly. Although Gag3 levels were elevated and there were Gag3 foci in these cells, the pattern was relatively dispersed. Xrn1, has previously been implicated in co-translational degradation of polysomal RNAs indicating that at least a fraction is associated with polysomes [79], but our findings suggest that Xrn1 also interacts with RNAs that are not targeted for degradation. Xrn1 has been previously implicated in mRNA localization via microtubule trafficking [80]. One straightforward explanation for our results is that Xrn1 has a role in trafficking Ty3 RNA from polysomes to PB retrosomes. The lsm1Δ cells displayed a similar, although less severe, phenotype to that of the xrn1Δ cells. Other studies have suggested that the Lsm1-7 heptameric complex and Pat1 interact with the 3’ ends of RNAs and associate with other PB proteins, including decapping proteins and Dhh1 bound to the 5’ end (summarized in [1]). Based on these Lsm interactions and the mutant phenotype, we propose that Lsm1-7 also participates in releasing the translating Ty3 RNA from polysomes, thereby promoting its localization into PB retrosomes.
Unlike Dhh1, Lsm1 and Xrn1, poly(A) binding-protein Pub1 did not localize to PB retrosomes, and pub1Δ cells had enlarged Ty3 Gag3 foci. This result is consistent with a model in which the association of Pub1 with translating poly(A) RNAs acts in opposition to PB suppression of translation and subsequent sequestration of Ty3 RNA. While on the surface, the decrease in transposition observed in this mutant might appear to be at odds with larger retrosomes, these retrosomes might exceed an optimal size for assembly foci so that enmeshed VLPs are unable to access nuclear pores.
Because Ty3 RNA and proteins associate with PB proteins in mating cells, we asked whether Ty3 expression promotes formation of mating cell PB. However, mating cell PB formed in a Ty3Δ strain, indicating that full-length Ty3 is not essential for mating PB formation. In mating, the MAP-kinase Fus3 phosphorylates Far1, which inhibits Cdc28-Cln thereby arresting cells in G1 [68]. Far1Δ cells, which fail to arrest, but initiate the mating transcriptional program, showed reduced PB formation compared to WT cells. This attenuated PB phenotype allowed us to further test for a role of Ty3 in PB formation.
Because mating cells undergo dramatic changes in transcription and translation, we postulated that PB formation during mating is associated with a specific role in the haploid to diploid transition. For example some cellular RNAs are known to exit and re-enter translation as cells transition from pheromone arrest to recovery [81, 82]. PB protein functions are known to affect mating, for example, PB factors Xrn1 and Dhh1, promote mating cell nuclear fusion and Ste12 expression, respectively [16, 83]. Although Ty3Δ cells formed mating PB, some observations indicate that Ty3 expression could contribute to mating PB formation. First, as previously reported, expression of Ty3 under the GAL1 promoter in non-mating cells, causes formation of one or two large PB per cell [6]. Second, Ty3 expression increases PB formation in far1Δ cells, in which PB formation is attenuated. Third, Edc3 and Pat1, which provide structural support for PB formation in glucose-deprived cells [44–46], are dispensable for Ty3 Gag3 mating PB formation. In mating cell PB, Gag3 may compensate for the deficiency of Edc3 or Pat1 by acting as a scaffold for PB protein recruitment in a manner similar to the contribution observed for the prion-like domain of Lsm4 [44–46]. Finally, a role for persisting transcripts from almost forty solo Ty3 LTRs some of which are induced by pheromone [14] cannot be excluded by our experiments with Ty3Δ cells. Production and storage or degradation of such non-coding transcripts might promote PB formation in Ty3Δ cells. PB formation was recently correlated with larger daughter buds [69]. The tolerance of genomes for retrotransposons is still not completely understood. If Ty3 or its LTR transcripts do enhance mating PB formation, it might, over an evolutionary-scale timeframe contribute to Ty3 genomic retention.
Similar to the S. cerevisiae Ty3/Gypsy LTR retroelement described here, the Ty1/Copia LTR retroelement requires PB proteins for retrotransposition [84, 85]. However, there are significant differences in the lifestyles of these two elements and some of these likely affect their interactions with PB components. First, Ty1 transcripts are among the most abundant in haploid cells but Ty3 transcripts are present in relatively low amounts except under induction. Ty1 transcripts form poly(A) T body foci, which are discrete from PB [86]. Ty1 retrotransposition is low due to repressive transcriptional and post-transcriptional copy number control (CNC) [87–89]. In addition to the full-length Ty1 RNA, subgenomic anti-sense and sense transcripts with roles in CNC are produced. The former has been implicated in epigenetic repression of Ty1 transcription [88, 90], while the latter has recently been shown to encode a truncated Gag protein with both specific and non-specific nucleic acid binding activities that interfere with Ty1 VLP formation. The 5’ to 3’ exonuclease, Xrn1, negatively controls the half-lives of these subgenomic RNAs. In contrast to Ty1, full-length Ty3 copy number, transcription and retrotransposition are relatively low except under conditions of mating pheromone induction [14, 15, 67]. However, these observations do not exclude a similar role for Xrn1 in the Ty3 lifecycle to that in the Ty1 lifecycle. For example, Ty3 expresses a subgenomic transcript of about 3.1 kb which could support expression of truncated Gag3 species. In addition, a species that reacts with anti-CA antibodies, but which is smaller than CA is observed (S3B Fig). However, there do not appear to be relatively higher amounts of the subgenomic sense transcript and the shorter protein in the Xrn1 mutant. In addition, the nucleic acid binding activity for Ty3 upon which PB association of Ty3 RNA depends is lodged in a discrete Gag3-derived NC zinc knuckle species [11, 91] which is not found in Ty1. Overall, the low level of Ty3 copy number compared to that of Ty1 (2 versus 32 in BY4741 for example), and low level of RNA under non-inducing conditions suggests that post-transcriptional CNC if it exists for Ty3 in BY4741 is not as strong as for Ty1.
Second, Ty3 and Ty1 apparently differ in their mode of utilization of PB factors and in the fold-impact of PB factor loss. As discussed in this work and previous work, Ty3 retrosomes are coincident with PB. In contrast, Ty1 retrosomes form in association with the ER through which Gag proteins transit [92], and Ty1 retrosomes associate transiently with or overlap partially, but are not coincident with PB. Nonetheless, deletion of several genes encoding PB factors, also reduced Ty1 retrotransposition compared to that observed in WT cells [xrn1Δ (0.5%) < dhh1Δ (1.6%) < upf1, 3Δ (2.5,3.4%) < dcp2Δ (5.7%), lsm1Δ (10%) < pat1Δ (16%) < ccr4Δ (20%) < edc3, 2Δ (50,86%)] (for example, [85]). Loss of these PB factors generally had the greatest effect at a step between Gag production and cDNA levels [84, 85]. Consistent with this defect, and similar to what was observed for Ty3, the xrn1Δ strain failed to package genomic RNA [85], and xrn1Δ, lsm1Δ, and pat1Δ showed reduced RNA foci [84]. Ty3 and Ty1 differ qualitatively and quantitatively with respect to dependence upon PB factors. For example, DHH1 deletion causes almost complete loss of Ty3, but not Ty1 transcripts, and deletion of several PB genes, for example PAT1, affected Ty1, but not Ty3 retrotransposition. In addition, under pheromone induction conditions utilized in our experiments, Ty3 retrotransposition frequency was several orders of magnitude greater than that of native Ty1 retrotransposition and with the exception of dhh1Δ and xrn1Δ, the Ty3 frequency was much less sensitive to loss of PB function than was that of Ty1 perhaps explaining some of the apparent differences in factor requirements.
The observation that Ty1 and Ty3 expression in response to MAP-kinase activation [55] and common PB factors [7, 54, 84, 85, 93], suggests that they could be in an evolutionary arms race with each other as well as their host. However, further considerations show distinctions. For example, as described here, Ty3 responds to mating MAP-kinase signaling which further destabilizes Ty1 protein [94, 95]. In diploid cells, where Ty3 is repressed, nutritional deprivation MAP-kinase signaling activates Tec1/Ste12 induction of Ty1 expression and retrotransposition [96, 97]. Together, these observations suggest that that despite shared host factors, Ty1 and Ty3 have privatized cellular programs within which they retrotranspose, thus potentially minimizing direct competition.
Comparison of the roles of PB proteins supporting Ty1 and Ty3 replication in yeast to PB roles in the replication of viruses and retroelements in other systems produces a mixed picture. The plant positive-strand RNA virus, Brome Mosaic Virus, depends upon PB proteins for replication in a yeast model [98]. DEAD box helicase Dhh1/DDX6 was implicated in formation of HIV intracellular intermediates in human immunodeficiency virus (HIV) core assembly and in prototypic foamy virus RNA packaging [99, 100]. However, other studies report no correlation between virus assembly and PB formation or DDX6 activity [35]. In human, DDX3 is required for WT nuclear export of HIV-1 RNA [101]. In our study, PB protein Ded1/DDX3 was identified by mass spectrometry, and co-localizes with Ty3 Gag3 in pheromone-treated cells, but its role remains to be tested. Although not identified in this study, several PB components are also classified in animal cells as restriction factors for retroviruses and retroelements [37, 38]. Collectively, these observations show that highly-conserved PB components associate with retroelement proteins and/or RNAs and may be used in different manners by retroviruses and retroelements.
Ty3 [102] and Ty1 [103] retrosomes associate with the nuclear envelope. The concentration of retrotransposons in perinuclear RNPs has obvious advantages for retrotransposition. Accumulating Gag associated with perinuclear clusters would have access to newly-exported gRNAs for packaging. Intriguingly, PB factors Dhh1 and Pat1 are required in parallel pathways for retrograde transport of tRNAs into the nucleus in nutritionally starved cells [75]. Multiple steps occur during VLP assembly and PB factors may promote more than one aspect of this process, including tRNA capture. Indeed, although its relationship to retrograde tRNA transport is not clear, tRNA association has been shown to promote nuclear entry of HIV cores and cDNA [104].
It is noteworthy that Ty3 activation in mating yeast displays many parallel features to retrotransposon activation in germ cell lineages of higher organisms. Animal retroelements are transcriptionally repressed in somatic tissues by epigenetic mechanisms. Repression is relieved in germ cell lineages by transient de-methylation. During phases of transcriptional activation, retrotransposition is suppressed by post-transcriptional degradation of repeated sequence/retroelement RNAs in specialized perinuclear RNP granules. In D. melanogaster, C. elegans and M. musculus, these RNPs contain PIWI Argonaute proteins, piRNAs, and PB components DEAD box helicases DDX3 (ScDed1) and DDX6 (ScDhh1), decapping enzyme subunits Dcp1 and Dcp2, and 5’ to 3’ exonuclease Xrn1 [reviewed in [40, 42, 105]]. Germ cell granule suppressors of retroelements (e.g. APOBEC3, RNAi Argonaute and Dicer components of PB) are not endogenous to budding yeast. However when expressed, two of these factors, APOBEC3G [85] and heterologous yeast Argo/Dicer, suppress Ty1 retrotransposition [43]. This suggests that localization of retroelement RNAs to these RNPs is conserved.
We show here that yeast cells undergoing mating MAP-kinase signaling form PB containing Ty3 RNA and proteins, and that components of yeast mating cell PB promote Ty3 retrotransposition at multiple stages. We speculate that in animal germ cells retrotransposition might also be promoted by PB functions associated with perinuclear granules, but is suppressed by components of RNAi. This work poses the following questions for future investigations: Is MAP kinase pathway signaling the underlying regulator of PB formation? How are RNAs recognized for inclusion in mating cell PB, but protected from degradation? In animal cells when suppression of retroelements in germ cell lineages is incomplete do PB components promote the spread of elements into new genomes?
Yeast and bacterial culture methods were as previously described [106, 107] except where noted. Bacterial strains HB101 or DH5α were used for plasmid preparation. All S. cerevisiae strains were derivatives of BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0). (S1 Table). BY4741 derivatives included strains deleted for Ty3 elements and killer RNA, and strains tagged with fluorescent reporter proteins. Manipulation of strains and plasmids utilized standard molecular genetics. For all α-factor induction, yeast cultures were grown at 24°C to OD600 of 0.2 and α-factor was added to 6 μM final concentration. A complete description of procedures is provided (S1 Text, Supplemental Materials and Methods; S2 Text, Plasmid construction and sequences; and S2 Table, Primers used in this study and S3 Table Plasmids used in this study.
Polyclonal rabbit antibodies (Berkeley Antibody Company) were raised against WT Ty3 VLPs purified as previously described [57] and affinity purified using recombinant Gag3 (this work). Polyclonal rabbit antibodies against Ty3 CA [60] were used for immunoblot analysis and immunofluorescence. A complete description is provided (S1 Text, Supplemental Materials and Methods).
Gag3-associated proteins were purified by α-VLP and control IgG affinity chromatography of whole cell extracts. Co-purifying proteins were digested with LysC and trypsin as described [108]. Peptides were analyzed by 1DLC-MS/MS using LTQ-Orbitrap XL MS (ThermoElectron) as described [108]. Database searching was performed using Batch-Tag and search Compare within the developmental version (v.5.2.2) of Protein Prospector as described. Proteins were identified by at least two peptides with a false-positive rate of 0.1%. A total of 154 proteins were identified of which 106 are not essential (S4 Table). A complete description is provided (S1 Text, Supplemental Materials and Methods).
GO term enrichment analysis was performed using the Gene Ontology Term Finder (http://www.yeastgenome.org/cgi-bin/GO/goTermFinder.pl) and the SGD Gene Ontology Slim Mapper(http://www.yeastgenome.org/cgi-bin/GO/goSlimMapper.pl). Genes were sorted according to biological process, molecular function, and cellular component. For Gene Ontology Term Finder, threshold P-value of 0.01 were used to identify specific enrichment.
Gag3-associated RNA interaction network was visualized by Cytoscape [109].
Gag3-interacting P-body proteins (boxes) were queried for protein-protein physical interactions using SGD:Yeastmine (http://yeastmine.yeastgenome.org/yeastmine/begin.do). Cytoscape v3.1.0 (http://www.cytoscape.org/) was used to graphically map interactions (S1 Fig). Additional details including gene names of all interacting partners used to create the network map (S6 Table).
BY4741 or derivatives were transformed with a plasmid expressing Ty3 marked with a HIS3 gene inactivated by the presence of an antisense artificial intron (Ty3-his3AI) [62, 110] under the native promoter (pDM3193). Strains were grown in SD-ura at 24°C to OD600 = 0.2, induced with α-factor and plated on synthetic dextrose medium (SD)-his, to select for cells that had acquired a chromosomal copy of Ty3-HIS3 in which the artificial intron was spliced out, and plated onto rich medium (1% yeast extract, 2% peptone, 2% dextrose) for total cell counts. At least four independent transformants were assayed for each strain. Transposition frequency was calculated as the frequency of His+ colonies and p values determined using the student t-test.
Cells were grown at 24°C to OD600 = 0.2 and induced with 6 μM α-factor (GenScript USA Inc.) for 4 h or left uninduced as control. Fluorescent images of live cells were visualized using either a Zeiss Axioplan2 or a Zeiss LSM510 META inverted confocal scanning Plan-Apochromat fluorescence microscope. FISH was performed with an adaptation of previous methods [84, 111, 112]. Ty3 mRNA was detected with an equal parts mixture of antisense oligonucleotides complementary to coding sequence in Ty3 Gag3, RT, and IN. Ty3 protein was visualized using α-VLP followed by Alexa-fluor 568-conjugated donkey anti-rabbit IgG (Life Technologies). Cells were stained with 3 μg/ml DAPI (4’, 6-diamidino-2-phenylindole). FISH was imaged using a Zeiss LSM780 Plan-Apochromat inverted confocal laser scanning microscope. Images were processed for publication with Adobe Photoshop CS3 (Adobe Systems Inc.). For all experiments, images from at least three replicate samples consisting of at least 100 cells each were analyzed, and single plane images are shown. A complete description is provided (S1 Text, Supplemental Methods and Materials).
Quantitative mating assays were done essentially as described [113]. In general: Cell cultures were grown at 23°C to OD600 = 0.2–0.3 and 106 cells of each mating type were mixed, collected onto 25mm-GF/C filter disc (Whatman), and incubated at 30°C for the indicated times. The cells were eluted from the filter in 1 ml medium. For mating efficiency, 100 μl was plated on medium selective for complementation of auxotrophic markers in MATa and MATα cells upon cell fusion. To determine the number of unmated cells, eluted cells were diluted 1–1,000 and plated onto the appropriate medium. Cells were incubated at 30°C for several days and colonies were counted. Mating efficiency (%) was calculated as the ratio of diploid cells to unmated cells. For specific experiments: For mating of Ty3(WT) BY4741 (MATa) x BY4742(MATα) and Ty3Δ yVB1913 (MATa) x yDM1456(MATα), cells were grown in YPD medium, mated on YPD medium for 1.5 and 2.5 h. Cells were eluted in SD medium lacking methionine and lysine, plated the same medium for mated cells, and on SD lacking methionine or lysine for unmated cells. For the effect of high level of Ty3 expression on mating, Ty3Δ strains yVB1672 (MATa) and yVB1926 (MATα) were transformed with either a high-copy-number plasmid containing Ty3 (pDM3194) or a plasmid control (pYES2.0). Cells were grown in SD medium lacking uracil, mated on the same medium for 2.0 or 3.5 h. Cells were eluted in SD medium lacking histidine and methionine and plated on the same medium for mated cells, and on SD lacking histidine or methionine for unmated cells.
RNA, protein and packaging, and cDNA levels were measured essentially as described [10]. Statistical significance was determined using the program R [114] to perform one way ANOVA, with planned contrast between WT and individual mutant samples compared.
For polysome profilings, cells were grown at 24°C in YPD to OD600 = 0.2, induced with 6 μM α-factor (GenScript USA Inc.) for 2 h. Polysomes were prepared essentially as described [115]. Extracts from 10–12 OD260 cells were fractionated on a 4–47% linear sucrose gradient. A complete description is provided (S1 Text, Supplemental Materials and Methods).
|
10.1371/journal.pgen.1002020 | Towards Establishment of a Rice Stress Response Interactome | Rice (Oryza sativa) is a staple food for more than half the world and a model for studies of monocotyledonous species, which include cereal crops and candidate bioenergy grasses. A major limitation of crop production is imposed by a suite of abiotic and biotic stresses resulting in 30%–60% yield losses globally each year. To elucidate stress response signaling networks, we constructed an interactome of 100 proteins by yeast two-hybrid (Y2H) assays around key regulators of the rice biotic and abiotic stress responses. We validated the interactome using protein–protein interaction (PPI) assays, co-expression of transcripts, and phenotypic analyses. Using this interactome-guided prediction and phenotype validation, we identified ten novel regulators of stress tolerance, including two from protein classes not previously known to function in stress responses. Several lines of evidence support cross-talk between biotic and abiotic stress responses. The combination of focused interactome and systems analyses described here represents significant progress toward elucidating the molecular basis of traits of agronomic importance.
| A major limitation of crop production is imposed by a suite of abiotic and biotic stresses resulting in 30%–60% yield losses globally each year. In this paper, we used a yeast-based approach to identify rice proteins that govern the rice stress response. We validated the role of these new proteins using additional analyses to evaluate the function of these genes in rice and assessed whether they serve to positively or negatively regulate the stress response. This approach allowed us to identify ten genes that control resistance to bacterial disease and tolerance to submergence. The combination of approaches described here represents significant progress toward elucidating the molecular basis of traits of agronomic importance.
| A major limitation of crop production is imposed by a suite of abiotic and biotic stresses resulting in 30%–60% yield losses globally each year [1]. The burgeoning field of systems biology provides new methodologies to make sense of plant stress responses, which are often controlled by highly complex signal transduction pathways that may involve tens or even thousands of proteins [2]. Complementary to large-scale approaches to delineate organisms' entire interactomes [3], we have developed a focused, high-quality Y2H-based interactome around the following key proteins that control the rice responses to disease and flooding: XA21 [4], NH1 (NPR1 homolog1/OsNPR1) [5], [6], SUB1A and SUB1C (submergence tolerance 1A, 1C) [7] (Figure 1A, Table S1). XA21 is a host sensor (also called a pattern recognition receptor (PRR)) of conserved microbial signatures that confers resistance to the Gram-negative bacterium Xanthomonas oryzae pv. oryzae (Xoo) [4], [8], [9]. Overexpression of Nh1 in rice also enhances resistance to Xoo [5]; whereas reduced expression of Nh1 impairs benzothiadiazole-induced resistance to Pyricularia oryzae [10]. SUB1A and SUB1C are ethylene response transcription factors that regulate response to prolonged foliar submergence [7]. Much remains to be learned about the signaling pathways controlled by these pivotal stress response proteins.
To identify components of these signaling pathways, we carried out yeast two hybrid screening to construct a rice response interactome. We then validated the robustness of the interactome using bimolecular fluorescence complementation [11], yeast mating-based split ubiquitin system assays [12], and phenotypic analysis. Transgenic analysis of genes encoding key proteins coupled with correlation analysis of transcriptomics data and protein-protein interactions revealed ten interactome members that function as positive or negative regulators of biotic or abiotic stress tolerance in rice. Fourteen additional members of the interactome have previously been reported to function in stress tolerance. The high-quality interactome and systems-level analyses described here represent significant progress toward elucidating the molecular basis of traits of agronomic importance.
We initially reconstructed four separate sub-interactomes for NH1, the intracellular kinase domain of XA21 (termed XA21K668 [13]), SUB1A, and SUB1C by screening a rice cDNA library pool. Subsequent rounds of screening with identified interactors, targeted assays with additional proteins identified based on sequence homology, and inclusion of connections from the rice kinase interactome [14] revealed that the NH1-, XA21-, and SUB1-anchored interactomes form a single rice stress interactome (Figure 1A, Table S1).
The four sub-interactomes were constructed by using a high throughput yeast two hybrid (Y2H) approach to identify components of the XA21-, NH1-, and SUB1- signaling pathways. We identified a total of 8 unique XA21 binding proteins (XBs, Table S1). Five of these XBs, XB2, XB10 (hence forth called OsWRKY62), XB11, XB12 and XB22, were chosen for further screening as baits in the Y2H to identify XB interacting proteins (XBIPs). Using Arabidopsis NPR1 as bait, six interacting proteins (NRR, NRRH1, rTGA2.1, rTGA2.2, rTGA2.3, and rLG2) were isolated by the same approach as described above. With NRR as bait, we isolated an additional six proteins (NH1, NH2, NRRIP-1, NRRIP-2, and NRRIP-3). With rTGA2.1 as bait, 4 interacting proteins were identified (TGA2.1IP-1, TGA2.1IP-2, GRNL1 and GRNL2). GRNL1 was used as bait to isolate nine interacting proteins (rTGA2.1, rTGA2.2, GIP-1, GIP-6, GIP-9, GIP-11, GIP-13, GIP-18, GIP-20, and GIP-23). Using SUB1A and SUB1C as baits, we identified 20 SUB1A binding proteins (SABs) and 9 SUB1C binding proteins (SCBs) (Table S1). Two proteins, SAB8 (SCB5) and SAB18 (SCB9), were identified using both SUB1A and SUB1C as baits. All identified proteins were repeatedly confirmed through secondary screenings were further characterized.
Additional proteins were incorporated into the XA21 and NH1/NRR interaction based on literature curation and subsequent experimentation. For example, ten interactors identified through our previous rice kinase Y2H screen [14], were incorporated into the the rice stress response interactome (Figure 1A, Table S1). We also demonstrated, through Y2H and co-immunoprecipitation assays, that OsRac1 (rice small GTPase, previously shown to play an important role in the rice defense response) interacts with RAR1 (required for Mla12 resistance), HSP90 (heat shock protein 90), OsRBOHB (rice respiratory burst oxidase homologB), and OsMPK1 [15], [16], [17]. We also showed that OsMPK12 (blast- and wound-induced MAP kinase (BWMK1)), which was previously demonstrated to be induced upon infection by Magnaporthe grisea), interacts with XB22IP-2 (hereafter, called OsEREBP1 (rice ethylene-responsive element-binding protein 1, AP2)) [18]. We tested additional interactions based on of the presence of predicted protein motifs. For example, a tetratricopeptide repeat domain found in XB22 is also found in SGT1 (Suppressor of G-two allele of Skp1). XB12 shows sequence similarity with p23, a protein that modulates Hsp90-mediated folding of key molecules involved in diverse signal transduction pathways [19]. We therefore tested the protein interactions of these two XBs with components of the HSP90/SGT1/RAR1 chaperone complex [20]. Positive interactions were incorporated into the rice stress response interactome. Similarly, because NH1 interacts with NRR, we tested two predicted paralogs (NRRH1 and NRRH2) with NH1.
While a genetic interaction between the NH1 and XA21 signaling pathways has previously been demonstrated [21], signaling components shared between submergence tolerance and Xoo-resistance have not yet been described. The current network is composed of 100 proteins and shows significant enrichment (by q<0.05, Fisher exact test with multiple hypothesis adjustment [22]) for several gene ontology (GO) terms related to both abiotic and biotic stress responses (Figure 1B, Table S2). Among molecular functions, the rice stress response interactome is particularly rich in transcription factors (diamond nodes in Figure 1A, p-value = 7.1×10−5, Fisher exact test), including 5 WRKY proteins, 4 TGA proteins, and 4 AP2 factors.
Validation of subsets of protein-protein interactions (PPIs) with two additional in vivo assays provides evidence that the interactome is of high quality. Using a mating-based split ubiquitin system that measures interactions with transmembrane proteins [12], we confirmed that 80% (8 out of 10 tested) of the XA21-binding (XB) proteins are able to interact with the full-length, membrane-spanning XA21 (the initial screen was conducted with the truncated XA21K668 protein) (Figure 1A, Figure S1). To assess whether the observed Y2H protein-protein interactions occur in plant cells, we examined 30 candidate proteins pairs using bimolecular fluorescence complementation (BiFC) in rice protoplasts. To rule out false-positive interactions, we tested the interaction of each protein with negative control vectors consisting of half of the yellow fluorescent protein. We found that 14 of the 30 tested showed interactions as detected by fluorescence only in the presence of the interacting rice protein but not in the presence of the negative control. Four proteins fluoresced in the presence of the negative control but displayed greatly enhanced fluorescence intensity in the presence of the interacting rice protein indicating that the interaction could be reproduced in vivo. Together these results indicate that 60% (18/30) of the tested pairs of interactome members interact in rice protoplasts as revealed by BiFC assays (Figure 1A, Figure S2, Table S3).
Components showing a large number of interactions with other interactome members (high degree) have been hypothesized to be essential for survival of the organism [23] although this finding has been disputed [24]. To identify such key hub proteins, we identified components in the rice stress interactome that displayed high degrees of interactions and then subjected them to pair-wise PPI assays. We tested a 24×20 matrix of 27 biotic stress (XA21) interactome components, a 14×14 matrix of 16 abiotic stress (SUB1) interactome components, and a 24×16 matrix of biotic-abiotic interactome components (Text S1, Table S4). An interaction was considered significant and reproducible if we observed it was replicated in two to three independent assays (Table S4).
Pair-wise PPI assays among interactome members revealed large numbers of possible interactions within and between the biotic and abiotic sub-interactomes (average degree 11±8, Figure 1C, Table S4). These interactomes have a high percentage (21.8%) of interactions beween their components (232 interactions out of 1060 tested) (Table S4). The biotic stress response interactome exhibits the highest level of interactions at 27.5% (132/480). The abiotic stress response interactome and the union between the biotic-abiotic stress response interactomes are even more highly connected [18.9% (37/196) and 16.4% (63/384), respectively]. The high number of interactions observed in the stress response interactome suggests that a large fraction of the components are capable of interacting with each other. These results also suggest that these components serve as members of large and/or changing complexes in vivo [25].
While the high number of interactions we observed is an order of magnitude greater than observed for studies of large-scale interactomes [3], it is comparable to smaller scale, more focused studies, such as that carried out for Arabidopsis MADS box transcription factors. In the MADS-box factor study, an average of only 5.4% of the components showed interactions (272/4998). However, when transcription factors predicted to function in the same biological process were examined, they displayed an increased number of interactions. For example, MADS-box factors predicted to be involved in floral development showed >15% interactions [26].
Consistent with their demonstrated key roles in response to stress, XA21, SUB1A, and SUB1C exhibit a high degree of interactions. In the matrix-based PPI tests, each of these interacted with over 10 additional proteins not initially identified as interactors in the original screen (Table S4). Other proteins with published roles in biotic stress signaling, including XB15 [13], XB3 [27], OsWRKY62 [28], and XB24 [29] are also among those with an above average degree of interaction. Such hubs may have a higher chance of engaging in essential functions because they participate in more interactions [30].
Coexpression network analysis and stress-specific transcriptomics of the interactome components support the validity of the interactome as an integrated module and highlights specific nodes that may function in cross-talk between the abiotic and biotic stress responses (Figure 2). The interactome is highly enriched for genes with correlated or anticorrelated expression compared with the whole genome (Figure 2A and 2C). For this analysis, we built rice biotic and abiotic stress gene transcript coexpression networks for the interactome members based on Pearson's correlation coefficients (PCC) calculated from publically available Affymetrix microarray data (Table S5). We define a correlated or anticorrelated interaction by PCC > |0.5|, a criterion under which 15% of interactome gene pairs interact, compared with ∼5.5% of pairs in the whole rice genome, and no pairs when the expression profiles are randomized (Figure 2A and 2C, Table S5). In both the coexpression networks derived from the abiotic and biotic microarray datasets, many components of the SUB1A (abiotic stress) and the XA21/NH1 (biotic stress) sub-interactomes display highly correlated or anticorrelated expression (Figure 2B and 2D, Table S5). This result further supports cross talk between the abiotic and biotic response networks. Contrasting the networks built from the different array sets, reveals that only a fraction of edges are conserved between the biotic and abiotic gene expression networks. This suggests that the expression of interactome members, and thus their availability to form PPIs with each other, varies depending on the stress regime, consistent with a model of dynamic complex formation [31] (Figure 2B and 2D).
We also generated microarray data to monitor transcriptional responses of Xa21-expressing and Nh1- and Nrr-overexpressing rice (NRR binds NH1 and is a negative regulator of resistance [21]) before and after Xoo infection. Analysis of this dataset as well as a previously reported Sub1a-specific response dataset [32], reveals that interactome members are significantly enriched among differentially expressed genes (p<0.05, Fisher exact test, Figure 2E, Figure 3, Table S6, Figure S3).
The interactome includes fourteen components that have previously been shown to regulate resistance to Xoo, further supporting the high quality of the interactome (Figure 1A, Table S7). We measured the Xoo and/or submergence response phenotypes of mutant rice lines for twenty additional interactome members, focusing primarily on genes encoding proteins with a high degree of PPIs (Table S7). Note that because of this bias in our experimental design, we are unable to test for correlation between a high degree of PPIs and a functional role in rice stress tolerance. Our phenotypic results show that nine out of seventeen genes (53%) that we assayed for a role in resistance to Xoo showed altered defense response phenotypes. Only one out of nine genotypes assayed showed altered tolerance to submergence, possibly due to the absence of SUB1A in the genotypes we examined (Table 1, Figure 3A–3H, Figures S4, S5, S6, S7, S8, S9, S10, S11, S12, S13).
Importantly, our phenotypic analysis revealed roles for two protein classes that, to our knowledge, were previously unknown to function in the plant stress response. on sequence similarities, SAB18 is a SANT-domain transcription factor, and, SCB3, is an enzyme involved in lysine biosynthesis (Table 1). SAB18 is a negative regulator of submergence tolerance suggesting that it may modulate the antagonistic activities of its two binding partners, SUB1A and SUB1C (Figure 3G and 3H, Figure S13). SCB3 serves as a positive regulator of resistance to Xoo (Figure S8). This result together with an earlier report showing that lysine levels increase in the Xoo-challenged Xa21 rice compared to mock treated controls [33], suggests that lysine plays an important, although undefined, role in the rice innate immune response.
The remaining eight proteins that we demonstrate to be involved in rice innate immunity have similarity to known stress-response factors (Table 1, Table S7, Text S1). Though many of these proteins were identified due to association with XA21 or an XB, modification of the expression of four of these genes gives altered resistance phenotypes in the absence of XA21 (Table 1), suggesting that they function in multiple biotic stress-response signaling pathways. Of particular significance, knockdown or knockout experiments show a role for three proteins, (RAR1, WAK 25 (wall associated kinase 25), and SnRK1 (sucrose non-fermenting-related protein kinase 1)), in XA21-mediated immunity.
The chaperone complex, HSP90/RAR1/SGT1 has been long known to play a positive role in intracellular NBS-LRR-mediated immunity [34]. RAR1 and HSP90 have also been shown to play a role in Arabidopsis FLS2-mediated signaling [35] and maturation of the rice chitin extracellular receptor OsCERK1 [36], respectively. Our observation that RAR1 serves as a positive regulator of XA21-mediated immunity (Figure 3A and 3B, Figure S6) further affirms that this complex contributes to host sensor-mediated immunity.
Wak25 (LOC_Os03g12470), compromises XA21-mediated immunity (Figure S10), indicating that WAK25 is a positive regulator of this process. WAKs have previously been shown to function as positive regulators of plant defense responses [37]. Although we do not yet know how WAK25 serves to regulate XA21-mediated immunity, there is precedence for interaction of PRRs with other receptor kinases. For example, the Arabidopsis FLS2 PRR interacts with the BRI1-associated kinase (BAK1) to transduce the immune response [38].
We also found that OsMPK5, previously demonstrated to serve as a negative regulator of resistance to the fungus, Magnaporthe grisea, and the bacteria, Burkholderia glumae [39], also negatively regulates resistance to Xoo (Figure S4). In contrast, the Arabidopsis protein with highest similarity to OsMPK5, AtMPK3, acts downstream of the Arabidopsis host sensor FLS2 and is a positive regulator of camalexin-mediated resistance to Botrytis cinera [40], [41]. The opposite regulatory roles for these Arabidopsis and rice predicted MPK orthologs underlines the limitations of extrapolating function between plant species.
OsMPK12 -and OsEREBP1 - are also positive regulators of resistance to Xoo (Figure S5, Figure S12). OsMPK12 was previously shown to phosphorylate OsEREBP1 [18]. OsEREBP1, as phosphorylated by OsMPK12, exhibits enhanced binding to the GCC box element of pathogenicity-related (PR) gene promoters. Overexpression of OsMPK12 in tobacco enhances expression of PR genes and increases resistance to Pseudomonas syringae and Phytophthora parasitica infection [18]. Thus, our results together with previously published studies indicate that OsMPK12 and OsEREBP1 are positive regulators of resistance to many pathogens.
We have also demonstrated a negative regulatory function for OsWRKY76 (Figure 3E and 3F, Figure S11), as has previously been shown for OsWRKY62 [28]. These two OsWRKYs are in the same WRKY subgroup (IIA) and are orthologs of barley HvWRKY1 and HvWRKY2, which serve as negative regulators of resistance to Blumeria graminis [42]. Along with our observation that the OsWRKY IIA proteins interact with members of the XA21 and SUB1 sub-interactomes [28], [43], these data are consistent with the WRKYIIA proteins playing a key role in fine-tuning grass defense responses.
SAB23 is a plant homeobox domain- (PHD) containing protein, which is known to function in development [44] and has been linked to response to pathogen stress [45] (Table 1). SAB23 serves as a negative regulator of resistance to Xoo (Figure 3C and 3D, Figure S7). This result supports previous observations that components regulating XA21-mediated resistance are also involved in developmental regulation [21], [46], [47]
SnRK1A, a well-known regulator of sugar sensing [48], was identified as a positive regulator in XA21-mediated immunity (Figure S9). Arabidopsis SnRK1 has been identified as a key regulator in sugar sensing and abscisic acid (ABA) signaling [49]. Though ABA has typically been found to act as a positive regulator of abiotic stress responses and a negative regulator of biotic stress responses [50], several positive regulators of the rice biotic stress response including SnRK1A and OsMPK12 participate in ABA signaling. Genes with ABA-related GO annotations are also up-regulated in Nh1-overexpressing and Sub1a-expressing transgenic rice (q = 1.3×10−2 and q = 5.3×10−10, respectively, Fisher exact test, multiple hypothesis adjustment) (Table S9). Together these observations support the hypothesis that ABA also has important functions in resistance to Xoo and tolerance to submergence in rice.
Comparable to analyses that show a correlation between essentiality and network degree centrality for essential genes [51] and negative regulators of growth (i.e., tumor suppressors) [52], we found that the rice interactome proteins with a validated role in the stress response have a significantly higher degree centrality in the abiotic co-expression network compared with those for which we were unable to measure a phenotype (Figure 3i, p = 3.7×10−2, Wilcoxon signed rank test, Table S8). Thus, interactome members that serve as central hubs as measured by co-expression analysis are more likely to function in the stress response than those members that do not serve as central hubs. This observation indicates the power of using the “guilt-by-association principle” to guide experiments based on co-expression maps [53], [54].
Here, we constructed a rice stress response interactome composed of 100 proteins governing the rice response to biotic and abiotic stress. Integration of protein-protein interaction assays, co-expression studies, and phenotypic analyses allowed us to efficiently identify ten novel proteins regulating the rice stress response.
The XA21 kinase fragment K668 was cloned into the Y2H bait vector pMC86. SUB1A and SUB1C were also cloned into pMC86. Sequence information is provided in Table S1. The Y2H screening experiments for SUB1A and SUB1C were conducted in the same manner as those for XA21. Bait constructs were transformed into yeast strains HF7c MATa, plated on selective medium, and screened as described (Clontech's Matchmaker Pretransformed Libraries User Manual). Colonies from the HF7c baits were grown to approximately 2×108 cfu/mL in 50 mL synthetic dextrose (SD: 6.7 g Difco yeast nitrogen base w/o amino acids, 2% glucose, 1X drop out solution [supplemented with appropriate amino acids], pH 5.8) lacking Tryptophan (Trp) media for use in the primary screens. Cells of HF7c baits were pelleted, washed once with sterile H2O and resuspended in 50 mL rich yeast media, YPAD (20 g Difco peptone, 10 g yeast extract, 40 mg Adenine hemisulfate, 2% glucose, pH 5.8). Target yeast (Y187) were transformed with cDNAs from a Hybrizap (Stratagene) Y2H library derived from seven-week-old IRBB21 (Indica cultivar containing Xa21) leaf mRNA. One aliquot of the Y187 target yeast was mixed with the Hf7c bait yeast in 50 mL YPAD and poured into a tissue culture flask. Yeast strains were allowed to mate for 20 to 24 hrs at 28°C with slight shaking. Yeast were then isolated and washed twice with sterile water and plated on SD medium lacking Histidine (His), Tryptophan (Trp), Leucine (Leu) and supplemented with 2 mM 3-amino-1, 2, 4-triazole (3-AT). Putative positive diploids from the primary screens were isolated and plasmids extracted. Confirmation of interacting proteins through plasmid re-transformation eliminates many false positives; a step often dispensed of in high throughput Y2H studies due to the encumbrance of bacterial transformation and plasmid propagation [14]. Yeast plasmids were transformed into E. coli DH5α to amplify plasmids. Amplified plasmids were then re-transformed into the yeast strain AH109 (Clonetech) to confirm interactions. Transformed yeast for the secondary screens were first plated on selective medium lacking Leu and Trp. Once yeast colonies appeared, they were then streaked on selective medium lacking His, Leu, and Trp, plus 2 mM 3-AT and medium lacking Ade, Leu, and Trp. Prey plasmids were isolated and sequenced only after confirmation in secondary screens. The PPI datasets were submitted directly to DIP and assigned the International Molecular Exchange identifier IM-15311[55].
For mating based-split ubitquitin assays, we followed protocols and used vectors and yeast strains as described previously [12]. In brief, using Gateway LR Clonase (Invitrogen) we constructed the bait by transferring XA21cDNA from pENT/D into pMetYC_Gate and the preys through transfer of the corresponding cDNA from pENT/D into pNX_Gate32-3HA. Primers for these constructs are described in Table S10. For identification of positive interaction via yeast mating, the bait and prey constructs were transformed to yeast strain THY.AP5 and THY.AP5, respectively by using the yeast transformation kit, Frozen-EZ yeast transformation II (Zymo Research). Positive interactions were selected by colony growth in minimal SD/Ade-/Leu-/Trp-/His- media (Figure S1).
We conducted BiFC assays as described in Ding et al. [14]. As negative controls, we included the both empty vectors (735 (YC)-EV and 736 (YN)-EV) for each pair-wise test. The BiFC assays are summarized in Table S3 and Figure S2.
We calculated Pearson correlation coefficient (PCC) scores to measure tendency of coexpression between genes based on two sets of publicly available Affymetrix microarray data—219 rice abiotic and 179 rice biotic category data—for 37,993 genes which have Affymetrix probe set matched, of which 34,016 have unique Affymetrix probe set available and only these genes were included in this database (Table S5). The raw Affymetrix data was downloaded from NCBI Gene Expression Omnibus [56] and EBI ArrayExpress [57]. We processed raw Affymetrix data using the MAS 5.0 R-package. The trimmed mean target intensity of each array was arbitrarily set to 500, and the data were then log2 transformed. The Rice Multiple-platform Microarray Element Search was used to map the Affymetrix probesets to rice genes [58]. Distributions of PCC scores of 578,527,120 pairs of rice genes with processed microarrays or with randomized microarrays (by random shuffling of arrays) are summarized in Figure 2A and 2C and Table S5.
We grew TaiPei309 (TP309), Xa21::Xa21 106-17-3-37, LiaoGeng (LG), Ubi::Nh1 LG 11, and Ubi::Nrr 64 LG plants for six weeks in the greenhouse. We then transferred the plants to a growth chamber set for a 14-h daytime period, a 28/26°C temperature cycle and 90% humidity. We employed the scissors dip method with multiple cuts to inoculate the plants using a suspension (OD600 of 0.5) of PXO99 Xoo. One and two days after inoculation, mock-inoculated and inoculated leaves were harvested for gene expression profiling using the NSF45K array. The replicate mRNAs for the comparisons of Ubi::Xa21 TP309 vs TP309, Ubi::Nh1 LG vs. LG, and Ubi::Nrr LG vs. LG were labeled with either Cy3 or Cy5 dyes, resulting in one technical replicate and three biological replicates per genotype pair. Gene expression data were processed as previously described [58]. The microarray data have been deposited to NCBI GEO and have the accession number GSE22112.
|
10.1371/journal.pbio.2005594 | The mammalian decidual cell evolved from a cellular stress response | Among animal species, cell types vary greatly in terms of number and kind. The number of cell types found within an organism differs considerably between species, and cell type diversity is a significant contributor to differences in organismal structure and function. These observations suggest that cell type origination is a significant source of evolutionary novelty. The molecular mechanisms that result in the evolution of novel cell types, however, are poorly understood. Here, we show that a novel cell type of eutherians mammals, the decidual stromal cell (DSC), evolved by rewiring an ancestral cellular stress response. We isolated the precursor cell type of DSCs, endometrial stromal fibroblasts (ESFs), from the opossum Monodelphis domestica. We show that, in opossum ESFs, the majority of decidual core regulatory genes respond to decidualizing signals but do not regulate decidual effector genes. Rather, in opossum ESFs, decidual transcription factors function in apoptotic and oxidative stress response. We propose that rewiring of cellular stress responses was an important mechanism for the evolution of the eutherian decidual cell type.
| Animals consist of a wide variety of cells that serve different functions depending on their location in the body. Cells with similar functions, or cell types, in different animal species are related both by an evolutionary line of descent—similar to the relatedness of species themselves—and by a developmental line of descent in the embryo. Networks of interacting genes, or gene regulatory networks, control gene expression in the cell, thereby specifying cell type identity. Understanding how new cell types arise by changing gene regulatory networks is critical both to comprehending fundamental aspects of human biology and to manipulating cell types in the laboratory. We approached this question by studying endometrial stromal fibroblast (ESF) cells from the uterus of humans and opossums, two distantly related mammals. We showed that the distantly related cell type in opossum expresses a similar set of regulatory genes as the human cell, but in response to pregnancy-related signals, the opossum cells induce a stress response. In the human cells, these signals induce differentiation into decidual cells, a specialized cell type present in humans and closely related mammals. These results suggest that a gene regulatory network that modulated an ancestral, pregnancy-related stress response was hijacked and repurposed to function in differentiation and specification of the decidual cell type.
| Multicellular organisms consist of numerous specialized cells, or cell types, that play an important role in the structural and functional diversity of organisms. Evolutionary diversification of cell types in metazoans has been a significant source of novelty and was essential to the elaboration of increasingly complex body plans. One model that can explain the evolution of novel cell types is the “sister cell type” model, which suggests that cell types originate by differentiation from an ancestral cell type [1,2]. According to this model, novel cell types have arisen from ancestral cell types through modification of developmental programs leading to two derived cell types, termed “sister cell types.” The origination of novel cell types may allow for organisms to manage an imposed physiological or environmental challenge that may have induced stress or morbidity in the ancestral condition. However, while it is clear that cell types have diversified prodigiously in evolution, the molecular mechanisms leading to the origination of a novel cell type are not well understood [3].
The evolution of mammalian pregnancy offers an opportunity to investigate cell type origination. Intensive selective pressures during the evolution of mammalian pregnancy led to the evolution of many functional specializations of the uterus that accommodate the implantation of the embryo and development of the placenta, including proper control of an ancestral implantation-induced inflammatory response [4,5]. These novelties include the origin of specialized cell types such as the decidual stromal cell (DSC), the uterine natural killer cell, and a specialized form of resident macrophages [6].
During the menstrual cycle and pregnancy, human decidual stromal cells (HsDSCs) differentiate from endometrial stromal fibroblasts (HsESFs) on exposure to progesterone and signals from the embryo [7]. Responding to these signals, genes critical to human DSC differentiation drive and install a complex decidualization gene regulatory network (GRN). Numerous transcription factors have been shown to transcriptionally and post-translationally interact to regulate effector gene sets conferring DSC cell type identity. Phylogenetic cell type studies make clear that eutherian endometrial stromal fibroblast (ESF) and DSC are sister cell types [8]. While ESFs are found in the oviduct of numerous amniotes, DSCs are exclusive to eutherians [9]. Moreover, it is clear that DSCs evolved from an ancestral ESF cell type, hereafter referred to as paleo-ESF, i.e., ESF that cannot give rise to DSC and are not derived from cells that can. This cell type existed prior to the stem lineage of eutherian mammals, having diverged 65 to 80 million years ago [10], i.e., DSC evolved after the most recent common ancestor of marsupials and eutherians and prior to the most recent common ancestor of eutherian mammals [11]. Hence, the evolutionary origin of DSC is an outstanding model to investigate the molecular mechanisms that led to the origin of a novel cell type.
To characterize the molecular changes that gave rise to the origin of the decidual stromal cell type, we isolated ESFs of the marsupial grey short-tailed opossum Monodelphis domestica, hereafter called MdESFs, which we use as a proxy for paleo-ESF. In humans and other eutherians, neo-ESF differentiates into DSC in utero when exposed to progesterone and estrogen, as well as ligands upstream of cyclic AMP (cAMP)/protein kinase A (PKA) signaling such as prostaglandin E2 (PGE2) [12–14] and relaxin (RLN) [15]. We assayed the response of MdESF to the stimuli that differentiate HsESF to HsDSC in vitro in order to identify the ancestral gene regulatory program from which the core network of DSC evolved. We found, surprisingly, that core components of the decidual GRN are responsive to progesterone and cAMP in opossum ESF, but rather than undergoing DSC differentiation, these genes regulate a cellular stress response.
We utilized an established protocol to isolate MdESF by Percoll column gradient [16]. We validated by immunostaining and western blotting that cells isolated by this procedure are positive for the mesenchymal marker vimentin and negative for the epithelial marker cytokeratin (S1 Fig). Relative to other layers in the column, these cells expressed higher levels of the ESF markers HOXA11, HOXA10, and PGR. We also show that these cell preparations have low levels of CD45, a marker of white blood cells, compared to RNA isolated from opossum spleen (S1 Table).
We assayed the response of MdESF to treatment with eutherian ESF differentiation media containing the cAMP analogue 8-br-cAMP and the progesterone analogue medroxyprogesterone acetate (MPA) (Fig 1A), hereafter referred to as decidualizing stimuli or 8-br-cAMP/MPA. RNA sequencing (RNAseq) of both stimulated and unstimulated MdESF revealed endogenous expression of numerous core regulatory genes critical to eutherian decidualization (Fig 1B). From a curated list of 28 transcription factor (TF) genes with documented roles in decidualization (Table 1), 22 are expressed in stimulated MdESF and 13 are significantly up-regulated (p < 0.05) (Fig 1B and S2A Fig). Seven decidualization TF genes are down-regulated, though still expressed, and two TFs are unchanged in expression. Most notably, the up-regulated gene set contains numerous TFs with well-characterized roles in decidualization: FOXO1 [17,18], PGR [19], CEBPB [17,20], HOXA10 [21,22], HOXA11 [23,24], GATA2 [25], ZBTB16 [26,27], KLF9 [28–30], HAND2 [31], STAT3 [32,33], and MEIS1 [34] (Fig 1B and S2A Fig). In contrast to this conserved transcriptional regulatory response, classical markers of decidualization, e.g., PRL, IGFBP1, CGA, and SST, are neither expressed in unstimulated MdESF nor induced in response to decidualizing stimuli (Fig 1C). We conclude that a substantial part of the DSC core GRN is also in place in opossum ESF and is responsive to progesterone and cAMP but does not control a decidual phenotype.
We conducted experiments to determine if the observed up-regulation of regulatory genes critical to eutherian decidualization is specific to both the progesterone receptor as well as specific to the M. domestica ESF cell type. As MPA can also stimulate the glucocorticoid receptor (GR), we knocked down GR with small interfering RNA (siRNA) and subsequently assayed the transcriptional response of MPA-responsive regulatory genes in MPA-stimulated MdESF. We observed no significant change in RNA abundance for five decidualization regulatory genes (S2B Fig), suggesting that the observed up-regulation of these factors in response to MPA is not associated with stimulation of GR.
It could also be argued that the observed response is a more general feature of M. domestica fibroblasts rather than specific to MdESF. Thus, we sought to determine if up-regulation of decidualization regulatory genes is specific to the M. domestica ESF cell type or whether a similar response also occurs in skin fibroblasts from M. domestica. We isolated skin fibroblasts from M. domestica, stimulated them with either 8-br-cAMP/MPA or PGE2/MPA, and assayed six decidualization TFs by qPCR. Our results showed a strong up-regulation of ZBTB16 in response to 3-day treatment of either 8-br-cAMP/MPA or PGE2/MPA (S2C Fig). Conversely, five other TFs that were up-regulated in MdESF in response to either treatment were unchanged in RNA abundance or were down-regulated. This result suggests that the induction of decidual regulatory genes is specific to endometrial fibroblasts in the opossum and, interestingly, that ZBTB16 may be a more general inducible factor in fibroblasts with elevated intracellular levels of cyclic AMP.
Gene ontology (GO) enrichment analysis of differentially expressed genes after 8-br-cAMP/MPA treatment revealed up-regulation of genes associated with oxidative stress, mitochondrial stress, and apoptosis, as well as down-regulation of genes associated with mitosis, DNA replication, and cytoskeletal organization (Fig 1E, Fig 2A). Outwardly, stimulated MdESF exhibited a rapid morphological response suggestive of cytoplasmic architectural remodeling (Fig 1D, Fig 2B, S1 Movie). The extent of this morphological response was dependent on both 8-br-cAMP concentration and duration of treatment (Fig 2B, S3A Fig). Remarkably, this morphological effect was reversible insofar as the cells reverted back to their normal morphology within 19 hours after withdrawal of decidualizing stimuli (Fig 2C, S3B Fig). GO treemaps, which represent the function of genes and degree of their differential expression in response to 8-br-cAMP/MPA, supported the hypothesis that stimulated MdESF undergo a cellular stress response, as GO terms associated with endoplasmic reticulum (ER) stress, apoptosis, reactive oxygen species (ROS) metabolism, and protein folding response were significantly up-regulated (S2D Fig). In line with this observation, stimulated MdESF exhibited elevated levels of intracellular ROS relative to unstimulated cells or cells stimulated with MPA alone (Fig 1D, S3C Fig). These data indicate that treating MdESF with decidualizing stimuli results in a rapid morphological response that is associated with increased intracellular ROS and the induction of genes counteracting oxidative stress, suggesting that, rather than leading to decidual differentiation, MdESF exposed to decidualizing stimuli undergo a classical cellular stress response.
Next, we considered whether stress induced by treatment with decidualizing stimuli could be an artifact of treating cells with extracellular 8-br-cAMP, rather than a natural ligand activating intracellular cAMP signaling. To address this, we sought a physiologically relevant signal that increases intracellular cAMP in these cells. PGE2 signaling is of particular interest given that (1) PGE2 is able to induce decidualization via cAMP signaling in human and rodent ESFs [12,14], (2) the PGE2 receptor PTGER4 is widely expressed in ESFs in mammals [16], and (3) the recent finding that prostaglandin synthase (PTGS, also known as “COX2”) and prostaglandin E synthase (PTGES) are both expressed in the opossum uterus after embryo attachment [5]. Furthermore, PGE2 is likely a key component of the inflammatory signaling from which the eutherian implantation reaction is derived [5,6,47]. In our 8-br-cAMP/MPA stimulated cells, we see a particularly striking effect on lipid metabolism, a critical pathway in the production of phospholipid-derived prostaglandins (S2D Fig). Indeed, prostaglandin Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway genes were enriched in lipid metabolism, e.g., 33% of genes listed in fatty-acid derivative metabolic process are involved in prostaglandin metabolism (S2D Fig). Furthermore, transcriptomic analyses of stimulated MdESF suggested that 8-br-cAMP/MPA treatment negatively regulates the predominant PGE2 receptor, PTGER4, as well as genes for synthesis of prostaglandins, e.g., PTGS2 and PTGES (Fig 3A), and positively regulates catabolic enzymes that function to degrade prostaglandins, e.g., HPGD and PTGR1 (Fig 3A). These data suggest stimulated MdESFs compensate for the effect of 8-br-cAMP by modulating the prostaglandin synthesis and signaling pathways, further suggesting that PGE2 is likely the natural ligand of MdESF activating the cAMP/PKA pathway.
A survey of RNAs present in unstimulated MdESF showed that only two prostaglandin signaling receptors, PTGER4 and TBXA2R, are expressed in these cells (Fig 3B). In order to test whether PGE2 could be the natural ligand inducing intracellular cAMP signaling in these cells, we assayed by RNAseq the response of MdESF to PGE2 with and without MPA. KEGG pathway genes involved in prostaglandin synthesis and inactivation exhibited similar differential regulation in response to PGE2/MPA as do decidualizing stimuli, suggesting similar regulatory responses by MdESF (Fig 3C). Interestingly, a survey of prostaglandin signaling components revealed a strong up-regulation of the prostacyclin receptor (PTGIR) across all treatment groups (Fig 3B). Uterine tissue from pregnant and nonpregnant M. domestica showed substantially higher amounts of PGE2 in pregnant females versus nonpregnant females, suggesting that PGE2 increases in utero during gestation (Fig 3D). Contrary to treatment with decidualizing stimuli, MdESF treated with PGE2 and with or without MPA did not exhibit a readily apparent dendritic phenotype (Fig 3E, S4A Fig). Nevertheless, PGE2/MPA-treated MdESF do show elevated levels of intracellular ROS (S4B Fig).
We next investigated the effect of PGE2/MPA on the expression of core decidualization regulatory genes. Remarkably, all 22 decidual TF regulatory genes expressed in 8-br-cAMP/MPA–stimulated MdESF cells are also expressed in PGE2/MPA-stimulated cells (Fig 3F, S2A Fig), showing a marked nonparametric correlation in expression levels (Spearman’s rho = 0.863, p = 3.46 × 10−9), and 12 of the 13 up-regulated genes under 8-br-cAMP/MPA are also up-regulated with PGE2/MPA. However, this up-regulation was not seen when MdESFs were treated with PGE2 alone (S4C Fig), indicating a synergistic effect of elevated intracellular cAMP and components responding to MPA. Similar to the response to MPA alone and 8-br-cAMP/MPA, previously characterized marker genes of human decidualization did not respond to 2-day treatment with either PGE2 or PGE2/MPA (Table 2). We conclude that responding to PGE2 signaling is a physiological part of MdESF biology and that PGE2/MPA also regulates the expression of the same TF network as 8-br-cAMP/MPA treatment. Moreover, the results suggest that this PGE2/MPA-induced TF network is homologous to that activated during the differentiation of human DSCs.
We next asked if, as seen in cells treated with decidualizing stimuli, PGE2/MPA treatment also induced a stress response in MdESF. We surveyed differential expression of genes involved in oxidative stress response, apoptosis, and ROS-associated ER stress (unfolded protein response [UPR]) (Fig 4). MdESF treated with either 8-br-cAMP/MPA or PGE2/MPA significantly up-regulated genes associated with counteracting oxidative stress, including GCLM, GPX3, GPX4, SOD1, SOD3, and CAT (Fig 4A). Both treatments up-regulated the apoptotic genes BCL2L11 (BIM) and GADD45A (Fig 4B). In contrast to treatment with PGE2/MPA, cAMP/MPA induced a distinct stress response in genes associated with UPR and TNF-related apoptosis-inducing ligand (TRAIL)-related apoptosis, up-regulating ERN1 (IRE1), HSPA5, HSP90B1, HSP90AA1, CALR, and TNFSF10 (Fig 4B and 4C). Lastly, to determine if clusters of genes associated with oxidative stress and apoptosis were differentially expressed in both cAMP/MPA and PGE/MPA, we analyzed GO term clusters shared between the treatments. This analysis also suggested a shared stress response in MdESF treated with decidualizing stimuli or PGE2/MPA, in which GO terms associated with stress and inflammation, e.g., “regulation of reactive oxygen species metabolism,” “protein folding,” and “leukocyte degranulation,” were shared between these treatments (S2D Fig, S5 Fig). Similarly, PGE2 alone and PGE2/MPA shared GO terms specifically related to “hypoxia,” “autophagy,” and “regulation of cell death” (S5 Fig, S6 Fig). These results suggest that PGE2/MPA, signals that are present in the pregnant opossum uterus, induce a stress response similar to treatment with 8-br-cAMP/MPA, including elevated levels of intracellular ROS (one-tailed t test on log transformed data, p = 0.0186), but without the dendritic morphological response (Fig 3E, S4B Fig). We conclude that the in vitro 8-br-cAMP/MPA–induced stress reaction is mimicked with the more physiological PGE2/MPA treatment. Furthermore, these data are consistent with a conserved role for PGE2 during pregnancy of therian mammals.
In many cells, forkhead box class O (FOXO) TF family members generically function in stress response, counteracting oxidative stress, and apoptosis, as well as regulating gluconeogenesis and glycolysis [48]. FOXO1 also is an early acting TF in the differentiation of human DSCs [49]. Therefore, we sought to compare how FOXO1 mRNA and FOXO1 protein stability and subcellular localization are regulated in opossum ESFs. As observed in human ESFs, FOXO1 RNA is present in unstimulated MdESF, but FOXO1 protein is absent [16,50], likely due to protein kinase B (Akt)-dependent proteasomic degradation [48] (Fig 5A). In response to MPA treatment, FOXO1 accumulates in the cytoplasm (Fig 5B), suggesting MPA alone can counteract the proteasomic FOXO1 degradation. Decidualizing stimuli and treatment with PGE2/MPA resulted in nuclear translocation of FOXO1 as well as cytoplasmic loading (Fig 5C–5E), suggesting cAMP/PKA signaling controls FOXO1 nuclear localization. In response to induction of oxidative stress, FOXO1 behaved similarly to cAMP or PGE2 treatment, suggesting that post-translational modifications of FOXO1 act as a sensor of oxidative stress in MdESF (Fig 5F). Moreover, immunofluorescence on uterine sections from pregnant M. domestica females in the late stages of gestation (11.5 days post coitus [d.p.c.]) found nuclear FOXO1 in uterine stromal cells near the luminal epithelium, suggesting that FOXO1 activation is part of the physiological role of MdESF during pregnancy (Fig 5H). Strikingly similar results were obtained for FOXO1 in human DSCs (S7A Fig), suggesting post-translational regulatory control of FOXO1 as found in human DSCs in response to decidualizing stimuli has been inherited from the ancestral paleo-ESF and ancestrally was part of a PGE2-induced cell stress response (Fig 5G).
In order to assess the functional role of FOXO1 activation in opossum ESF, we assayed oxidative stress and apoptosis by treatment with 2′,7′ dichlorodihydrofluorescein diacetate (H2DCFDA) (to detect ROS) and propidium iodide (PI) (to detect early stages of apoptosis) in stimulated and unstimulated MdESF (Fig 6). Apoptosis, i.e., PI staining, was markedly elevated in MdESF treated with decidualizing stimuli (Fig 6A and 6B; 4-fold increase, one-tailed t test p = 2.0 10−5). That was also the case for PGE2/MPA but to a lesser degree (1.8-fold increase, one-tailed t test p = 0.015). To determine if FOXO TFs function in this stress response, we transfected MdESF with siRNAs targeting FOXO1 and FOXO3 RNA transcripts and subsequently treated them with 8-br-cAMP/MPA or PGE2/MPA. We confirmed depletion of FOXO1 RNA (as well as FOXO3 RNA) by qPCR and depletion of FOXO1 protein by western blot (S7B Fig, S7C Fig). siRNA-mediated knockdown (KD) of FOXO1 and FOXO3 increased signals for apoptosis (Fig 6A and 6B) (ANOVA on log transformed fluorescence values, KD effects FOXO1 = 2.4-fold, FOXO3 = 1.8-fold, overall ANOVA p = 5.91 10−8). Surprisingly, we did not find a significant interaction effect of FOXO1 and FOXO3 KD, suggesting they function additively in protecting against apoptosis (Fig 6A and 6B). However, there is no significant effect of FOXO KD on ROS levels in our data, suggesting that ROS production in response to decidualizing stimuli is not regulated by FOXO proteins.
FOXO1 KD resulted in marginally more cells positive for apoptosis than did FOXO3 perturbation (Fig 6B; 1.31-fold, p = 0.045). These results suggest that the ancestral function of FOXO genes in paleo-ESF was similar to the highly conserved, pan-metazoan roles of FOXO genes in classical stress response [51]. Therefore, we hypothesized that the regulatory linkages downstream of FOXO1 have diverged since the eutherian–metatherian split. From a list of genes in human DSCs that are positively regulated by FOXO1 (i.e., decrease in expression after FOXO1 KD and treatment with decidualizing stimuli), we selected seven genes that are also strongly up-regulated in MdESF stimulated with 8-br-cAMP/MPA or PGE2/MPA. Of the seven genes, one, IGFBP3, significantly decreased in expression and all other genes either did not respond or increased in expression in response to FOXO1 KD in MdESF (Fig 6C). This result suggests that FOXO1 transcriptionally regulates distinct sets of genes in MdESF and HsDSC. If we assume that the reproductive mode in MdESF is representative of the ancestral paleo-ESF, these data also suggest that the evolution of mammalian DSCs proceeded through modifying the target gene set of a largely conserved core GRN that includes FOXO1.
Here we show that, whereas human ESF respond to decidualizing stimuli with a compensated physiological phenotype, the opossum ESF exhibit a classic stress phenotype. This difference was also found, though to a lesser degree, when we used a more physiological signal, PGE2, instead of extracellular 8-br-cAMP. The responses of human and opossum ESFs were remarkably similar at the level of DSC regulatory gene expression, both in terms of transcriptional as well as post-translational regulation as in the case of FOXO1. While post-translational activation of FOXO1 is necessary in human cells for the expression of DSC effector genes, e.g., PRL, IGFBP1, etc., in opossum ESF, the functional role of activated FOXO1 is to protect the cells against apoptosis.
We propose that the signaling pathway and large parts of the TF network are homologous and to some degree conserved between eutherian DSC and marsupial MdESF, suggesting that these components were also present in paleo-ESF, i.e., prior to the evolutionary origin of DSCs. Furthermore, we hypothesize that there were at least two distinct molecular changes that led to the evolution of DSCs. On one hand, we find a small number of decidual TFs that are not expressed in stimulated MdESF, viz. members of the 5’ HoxD cluster HoxD12, HoxD11, and HoxD9, as well as FOXM1, TFAP2C, PRRX2, and E2F8. This suggests that one aspect of the evolution of the core GRN is the recruitment of additional TF genes through the evolution of cAMP response elements in existing or novel cis-regulatory elements. For example, in this list we find E2F8, which functions in regulating polyploidization [35], a derived feature of DSCs. On the other hand, we also find that DSC effector genes of the conserved TF network are different. Thus, another element by which the ancestral ESF cell type evolved into DSCs was the rewiring of gene regulation downstream of FOXO1 and other decidual regulatory genes, e.g., CEBPB, PGR, HOXA10, HOXA11, and GATA2 (Fig 7). We tested this model by comparing the effect of FOXO1 KD on effector gene expression and found that in fact the regulatory role of human FOXO1 in these cells is extensively different from that in opossum ESF. Exactly how this “downstream reprogramming” was effected in evolution is not known and needs to be the subject of further investigation, although previous work suggests that recruitment of transposable elements may have been a key factor [52,53].
An alternative explanation for our results is that the decidual cell type may have evolved in the stem therian lineage, i.e., before the most recent common ancestor of eutherians and marsupials, but has been lost in marsupials. Marsupial reproduction shares many plesiomorphic reproductive traits with the most basal branching mammals, the monotremes [5]. Although opossums do not lay eggs, along with the platypus, they have a relatively short gestation period and give birth to highly altricial young [54,55]. Furthermore, DSCs are necessary for maintaining pregnancy in species with the kind of invasive placentation that is only found in eutherian mammals [9]. For these reasons, it is likely that the DSC is truly a specific trait of eutherian mammals and not one with an older origin that was subsequently lost in the marsupial group.
Our model of stress-derived decidual differentiation explains a number of otherwise puzzling facts. First, there is evidence that during physiological decidualization in humans the stromal cells produce intracellular ROS that mediate the decidualization process [56–59]. Here, we show that genes responsive to decidualization stimuli in a distantly related ESF cell type are also intimately linked to stress-related decidualization in eutherians [56], e.g., GPX3, SOD1, SOD3, and CAT, and more interestingly these genes are also associated with a stress coping mechanism. Furthermore, we show that apoptosis-related genes known to play a role in decidualization, e.g., GADD45A, TRAIL, and BCL2L11 [56,60], are also up-regulated in MdESF. In the context of these stress-related genes, we also show that FOXO1 acts as an oxidative stress sentinel that counteracts apoptosis. Moreover, decidualization is associated with ER stress [61], which we also observe through the expression of genes specifically associated with ER stress and UPR in opossum cells treated with 8-br-cAMP/MPA. Finally, a subpopulation of endometrial stromal cells undergoes cellular senescence in humans, a senescent phenotype that plays a critical role in implantation [62]. Thus, peculiar features of human decidualization, e.g., redox signaling, ER stress, and cell senescence, are readily explained by our model, wherein decidual differentiation mechanisms arose in evolution from a pregnancy-related stress response that consequently activates many of the same regulatory genes and physiological processes as are activated in decidual cells during differentiation.
Our results point to a model for the origin of a novel cell type, namely the modification of an ancestral cellular stress response. Very few studies have addressed the molecular mechanisms for the origin of novel cell types. However, two previous studies are of particular interest in regards to this. The study by Nagao and colleagues [63] investigated the leucophore pigment cell type in the perciform lineage. Experimental evidence suggests that the origin of the leucophore in the perciform lineage was achieved by a change in the functional interaction between TFs already expressed in a precursor cell type. Similarly, changes in the functional interactions of FOXD3 during the evolution of the neural crest cell type, a vertebrate novelty, is also especially noteworthy [64]. These results are broadly consistent with the model we propose for the evolution of the decidual cell type, in which similar TFs are expressed ancestrally and changes in the functional interactions between those TFs occurred that were critical for the evolution of the novel cell type. In fact, we have shown previously that, coincidental with the origin of decidual cells, the function of TF proteins, namely HOXA11 and CEBPB interacting with FOXO1, has changed [65,66]. We note these genes are among those already responsive to decidualizing signals in opossum ESF. Hence, transcription factor protein evolution may be a common feature of the origin of novel cell types.
In evolution, structural and developmental changes can result in cells being exposed to drastically altered or novel environments. In mammals, the evolution of pregnancy and in particular the evolution of extended gestation resulted in the endometrium functioning under exposure to a range of new stimuli and with new requirements for the success of reproduction. As these developmental changes occurred, it is reasonable to expect that exposure of cell types to novel tissue and developmental environments can be a source of cellular stress. These anciently conserved pathways can be a rich source of hardwired, modular components of stress GRNs. In this case, evolution can mitigate cellular stress in a variety of ways, including decreasing the stress-inducing stimulus, but also through co-option of the stress pathways into normal physiological function. Our data suggest that the decidual stromal cell type has evolved from a physiological stress response that is likely directly related to the invasion of trophoblast into maternal tissues, the condition seen in crown eutherians today.
Whereas the evidence presented here pertains to the special case of the evolution of mammalian decidual cells, the recent discovery that ROS are important physiological signaling molecules during the differentiation of other cell types, e.g., neurons [67,68] and mesenchymal cells [69] (and functioning in tumorigenesis), may indicate a role for stress responses in the evolutionary derivation of novel cell types. Some cell types with novel physiological functions could therefore be understood as fulfilling physiological needs for which the ancestral body plan cannot compensate. The evolutionary rewiring of stress responses and functional changes in the interactions of transcription factors could be a few means within a suite of mechanisms involved in the evolutionary process of cell type origination to address physiological challenges.
All animal procedures were conducted under protocols approved by the Institutional Animal Care and Use Committee (#11313) of Yale University.
Opossum uterine tissue was collected from a M. domestica colony housed at Yale University. For ESF isolation, uterine tissue was harvested from a nonpregnant M. domestica female. For immunohistochemistry and ELISA, tissue from specific stages of the reproductive cycle were collected by following a standard breeding protocol outlined in Kin and colleagues [16]. Once collected tissue was stored for immunohistochemistry analysis in 4% paraformaldehyde in PBS 24 h, then washed in 50% ethanol for 1 h, then twice in 70% ethanol for 1 h then stored in 70% ethanol at −20°C. Tissue was stored for western blot and ELISA analysis by snap freezing in liquid nitrogen and then stored at −80°C.
M. domestica ESFs were isolated as previously described [16]. Briefly, primary ESFs were harvested by enzymatic digestion and centrifugation combined with Percoll density gradient. The uterus of an adult female grey short-tailed opossum M. domestica was dissected, cut in half longitudinally, and cut into 2–3–mm fragments. These were digested with 0.25% trypsin-EDTA for 35 min at 37°C and digested in Dissociation Buffer (1 mg/ml collagenase, 1 mg/ml Dispase, 400 μg/ml DNaseI) for 45 min at room temperature. Cell clumps were subsequently homogenized by passage through a 22-gauge syringe. Passage through a 40-μm nylon mesh filter removed remaining fragments. This lysate was used to generate a density gradient by centrifugation at 20,000 g for 30 min. A single cell suspension was generated from this lysate and was subsequently layered onto a Percoll gradient (1.09 g/cc Percoll, GE Healthcare Life Sciences) and centrifuged at 400 g for 20 min to allow for cells to settle to their respective density layers. Using a 25-gauge needle, each 1-ml layer was removed working from low to high density and washed into 5 ml of 50-mM NaCl solution. As with previous iterations of this protocol in this laboratory, layers 6 through 8 generally contained a fairly homogeneous population of cells that outwardly exhibited fibroblast characteristics. Cells in these layers were pelleted, resuspended in growth media, and cultured in 24-well plates. To facilitate enrichment of fibroblasts versus epithelial cells, media was exchanged in each well after two hours in order to remove floating cells that had not yet attached. To validate our cell line, we conducted comparative qPCR and immunofluorescence for proteins that mark either epithelial or mesenchymal (fibroblast) cells. Transcription factors indicative of ESFs were enriched in RNA from Percoll layer 8 relative to layer 3 or RNA isolated from spleen (S1 Table). Immunofluorescence on cells from Percoll layer 8 showed enrichment of the mesenchymal protein VIMENTIN and no epithelial contamination as judged by expression of CYTOKERATIN (S1 Fig). Therefore, cells from layer 8 were used in these experiments and were propagated in a T75 culture flask by sub-passaging over a period of 2 months at 33°C. Once confluent, cells were subpassaged using Accutase (AT104, Innovative Cell Technologies) or a cell scraper. The experiments detailed here were conducted with passages 12–20. To isolate M. domestica skin fibroblasts from the animal, ethanol was applied to the skin, hair was removed, and a small sample was excised. The sample was transferred to PBS, and the subcutaneous tissue was removed by scraping the dermal side with a razor blade and forceps. The sample was cut into strips approximately 1.0 cm2 with a scalpel and was incubated in 0.3% trypsin-PBS for 30 min at 37°C. The epidermis was subsequently removed, and the sample was washed in ABAM-PBS with gentle shaking. The sample was transferred to a tissue culture dish and sliced into squares approximately 3 mm in size. Five to 10 of these squares were transferred to a 35-mm tissue culture dish, and a sterile 22-mm glass coverslip was placed over the samples. The samples were grown to confluency in growth media, which was refreshed every 3 days.
MdESF were cultured in growth media with no antibiotic-antimycotic, containing (per liter) the following: 15.56 g DMEM/F-12 (D2906, Sigma Aldrich), 1.2 g sodium bicarbonate, 10 ml sodium pyruvate (11360, Thermo Fisher), 1 ml ITS (354350, VWR), and 100 ml charcoal-stripped fetal bovine serum (100–119, Gemini). Media was replaced every 4 days unless otherwise stated. Over the duration of these experiments, MdESF were found to be mycoplasma free (S8 Fig), as shown by periodic PCR assays for mycoplasma contamination (30-1012K, Universal Mycoplasma Detection Kit, ATCC). For decidualizing stimuli, MdESF were cultured in growth media supplemented with cAMP-analgoue 8-bromoadenosine 3′-5′-cyclic monophosphate sodium sale (B7880, Sigma Aldrich) and progesterone-analgoue medroxyprogesterone 17-acetate (MPA) (M1629, Sigma Aldrich), at final concentrations of 0.5 mM and 1 μM, respectively. Growth media was supplemented with prostaglandin E2 (14010, Cayman Chemicals) at a final concentration of 10 μM. For RNA sequencing of unstimulated and stimulated MdESF, cells were grown to 70% confluency in T75 flasks and treated with the respective stimuli for two days prior to harvesting with Buffer RLT and subsequent processing with RNeasy Mini Kit (74104, Qiagen) following the manufacturer’s protocol. Illumina sequencing libraries were generated from RNA by Poly-A selection and sequenced by the Yale Center for Genome Analysis on the Illumina Hiseq 2500 system. For transcriptomic and GO enrichment analyses, see below under sub-heading Transcriptomic Analyses. Human ESF and DSC transcriptomic data used in these analyses were reported previously [8], and FOXO1 KD RNAseq data in HsDSC have been deposited under GEO GSE115832.
MdESF or HsESF were grown in an 8-well chamber slide (12-565-18, Fisher) to 70% confluency. After treatment, cells were fixed with 4% paraformaldehyde in PBS for 15 min at room temperature. Cells were washed 2 times in ice-cold PBS, subsequently incubated for 10 min in 0.25% Triton X-100 in PBS, and finally washed 3 times for 5 min/wash in PBS. A blocking solution was applied with 1% bovine serum albumin (BSA) and 0.25% Triton X-100 in PBS for 30 min at room temperature. Cells were then incubated in blocking solution at 4°C overnight in the following primary antibodies: 1:200 rabbit anti-cytokeratin (ab9377, Abcam); 1:200 mouse anti-vimentin (sc-6260, Santa Cruz); mouse anti-FKHR (FOXO1) (sc-374427, Santa Cruz). Cells were subsequently washed the next day 3 times for 5 min each in PBS, and secondary antibody incubation was for 1 h at room temperature in the dark. Secondary antibodies used in this study were as follows: 1:200 Alexa Fluor 555 goat anti-mouse IgG (A21422, Thermo Fisher); 1:200 Alexa Fluor 488 goat anti-rabbit IgG (A11008, Thermo Fisher). Cells were then washed 3 times for 5 min each in PBS, and nuclei were stained with DAPI (10236276001, Roche). Finally, cells were washed one time for 5 min in PBS and observed with an Eclipse E600 microscope (Nikon) equipped with a Spot Insight camera. Lastly, it should be noted that we tested two different antibodies targeting human FOXO3 in M. domestica. These antibodies were anti-FOXO3a/FKHRL1 (EMD Millipore 07–702) and anti-FKHRL1 D-12 (Santa Cruz, Sc-48348). Both of these antibodies produced nonspecific bands on western blot and pervasive signal in MdESF. Therefore, we were not able to assess the protein localization dynamics of FOXO3 in MdESF.
MdESF or HsESF were cultured in T75 flasks to 80% confluency. Cells were rounded up with Tryple and homogenized in RIPA buffer (89900, Thermo Fisher) supplemented with HALT Protease Inhibitor Cocktail (PI87785, Thermo Fisher) for 15 min. Suspensions were centrifuged for 15 min at 13,000 RPM at 4°C. Protein concentrations were determined with Pierce BCA Protein Assay Kit (23225, Thermo Fisher). Cell lysates were diluted to achieve a solution with 30–60 μg total protein, combined with an equal volume of 2× NuPage LDS Sample Buffer (NP007, Thermo Fisher) with 2× NuPage Sample Reducing Agent (NP0004, Thermo Fisher), heated to 70°C, loaded in a NuPage 4%–12% Bis-Tris gel (NP0321BOX, Thermo Fisher), and electrophoresed at 130 volts for 60–90 min. Proteins were transferred to polyvinylidene difluoride membranes with the iBlot Gel Transfer System (Thermo Fisher). Membranes were subsequently incubated for 1 h at room temperature in blocking buffer (3% BSA in PBST) and incubated with primary antibodies, listed above, overnight at 4°C. Primary antibody dilutions were the following: 1:200 FOXO1, VIMENTIN, and CYTOKERATIN. After primary incubation, membranes were then washed 3 times for 5 min each in PBST and subsequently incubated with the corresponding HRP-conjugated secondary antibody, 1:5,000 of either goat anti-mouse (sc-2005, Santa Cruz) or goat anti-rabbit (sc-2054, Santa Cruz). Signal was detected by incubating membranes in Clarity Western ECL substrate (1705060, Bio-Rad) in the dark for 5 min and visualized with a Bio-Rad Gel Doc System. Uncropped western blots are provided in S9 Fig and S10 Fig.
Uterine tissue was dehydrated through a graded ethanol series, cleared in toluene, and then embedded in paraffin. We cut 7-μm cross sections on a microtome and mounted on Shandon polysine precleaned microscope slides (6776215Cs, Thermo Fisher). Sections were stored in the dark at room temperature until they were stained. We localized the expression of FOXO1 using immunohistochemistry. Slides were deparafinized in three successive washes of xylene (3 min each), then three successive washes of ethanol (3 min). Antigen retrieval was performed in citrate buffer (12 mM sodium citrate, pH 6.0, 98°C, 1 h). Endogenous peroxidase activity was blocked with Dako Peroxidase Block (Dako, 30 min). Slides were then incubated in primary antibody overnight. We used a goat polyclonal antibody generated against the N-terminal of human FOXO1 protein (1:5,000 dilution, FKHR antibody sc-9809, SantaCruz). On day two, slides were incubated in a donkey anti-goat IgG-HRP secondary (1 h, 1:200 dilution, sc-2056 SantaCruz). Slides were then rinsed in PBS (5 min), PBS-BSA (0.1%, 5 min), then incubated for 5 minutes in TSA Plus Cyanine 3 system (1:50 NEL744001KT, PerkinElmer Inc.). Slides were again washed in PBS (5 min) and PBS-BSA (0.1%, 5 min), counter stained in DAPI (1 time, 2 min, 10236276001; Roche), washed in deionized water for 5 min, and mounted in glycerol (50%).
Snap frozen uterine tissue from adult female M. domestica was homogenized in extraction buffer (0.1 M phosphate, 1 mM EDTA, pH 7.4) containing indomethacin (10 μM final concentration) using a mechanical homogenizer (TissueRuptor, QIAGEN). Cellular debris was removed by centrifugation (>13,000 RPM, 4 oC, 10 min). Tissue lysates were aliquoted into single use tubes and frozen at −80°C. Protein concentration was measured using Pierce BCA Protein Assay Kit (23225, Thermo Fisher). After determining protein concentration, 1 mg of protein for each sample was used in the first round of ELISA against PGE2 following the manufacturer’s protocol (514010, Cayman Chemical). Each sample was run in duplicate at two different dilutions. A dilution series of PGE2 standard provided by the manufacturer was included in each run, and PGE2 values in ng per mg protein were calculated from these standards. Due to the sensitivity of this assay, some samples contained excess PGE2 and therefore required additional dilutions in order to fall within the calculable range of the standard. For these samples with excess PGE2, an additional ELISA plate was run with two additional dilutions. Samples were incubated in the provided PGE2 monoclonal antibody ELISA plate overnight at 4°C. The following day, the wells were washed 3 times and the staining reaction was allowed to proceed for 45 min with shaking at 400 RPM on an orbital shaker in the dark. The plate was read at 405 nm on a Viktor X multilight plate reader (Perkin Elmer).
MdESF were cultured in T25 culture flasks to 70% confluency and subsequently transfected with siRNAs as above. After two days, cells were treated either with 8-br-cAMP/MPA or with PGE2/MPA for four days. The change of media was accompanied by an additional round of siRNA transfection per above. At the time of RNA harvest, media was removed, cells were washed 2 times in PBS, and cells were lysed directly in the flask with Buffer RLT Plus plus beta-mercaptoethanol. RNA was harvested according to the manufacturer’s protocol (74034, RNeasy Plus Micro Kit, Qiagen). Reverse transcription of 3 μg of RNA was carried out with iScript cDNA Synthesis Kit (1708891, Bio-Rad) with an extended transcription step of three hours at 42°C. For qPCR, all reactions were carried out with Power SYBR Green PCR Master Mix (4368708, Thermo Fisher) in triplicate with 40 ng of cDNA for template in each technical replicate reaction. Fold change was calculated by finding the ddCt values relative to the expression of TATA Binding Protein. All qPCR primer sets were validated by analysis of melting curves for 2 different sets of primers for the same gene. Primer sets used in this study are listed in S2 Table.
MdESF at 70% confluency in 6-well culture plates were transfected with siRNAs targeting FOXO1 (Mission custom siRNA, V30002, Sigma Aldrich), FOXO3 (Mission custom siRNA, V30002, Sigma Aldrich), or negative control scrambled siRNA (Silencer Negative Control No. 1, AM4611, Thermo Fisher). In preparation for transfection, siRNAs in 37.5 μl of OptiMem I Reduced Serum Media (31985, Thermo Fisher) were mixed with an equal volume of OptiMem containing 1.5 μl of Lipofectamine RNAiMax (13778, Thermo Fisher), incubated at room temperature for 15 min, and subsequently added dropwise to cells in 3 ml growth media. Final concentration of siRNAs was 25 nM. In experiments involving stimulated media, an additional round of siRNA transfection was prepared and added dropwise after growth media with decidualizing stimuli was added. Two siRNAs were transfected for each gene. Custom siRNAs were synthesized by Sigma to target the mRNAs of M. domestica FOXO1 and FOXO3, using the NCBI Reference Sequences XM_001368275.4 (FOXO1) and XM_001368456.2 (FOXO3). Sense and antisense sequences are listed in S3 Table. siRNA-mediated KD of human FOXO1 in human decidual cells was carried out as previously described [70].
We confirmed depletion of both RNA and protein by qPCR and western blot. For FOXO1, qPCR analyses showed that KD efficiency for these pooled siRNAs was >90% (S7B Fig). This depletion in RNA led to a corresponding depletion in FOXO1 protein, as confirmed by western blot (S10 Fig). We also desired to conduct a protein-level analysis for M. domestica FOXO3. However, for M. domestica FOXO3, we were able to confirm RNA depletion only (S7B Fig), as we did not find a commercially available antibody that showed high specificity for FOXO3 in M. domestica.
Cells were transfected as above with RNAi reagents and subsequently incubated in growth media in the presence of decidualizing stimuli for four days. For FACS analyses, conditioned media from each well was decanted into separate 15-ml conical centrifuge tubes, and cells were then washed in 2 ml PBS. PBS was decanted into the same 15-ml conical tube as the conditioned media, and 250 μl warm Tryple Express (12604, Thermo Fisher) was added to each well. To facilitate detachment, cells were placed in a 33°C incubator for 10 min. The detachment reaction was stopped by adding 700 μl growth media to each well. Cells were then transferred to their separate 15-ml conical tubes with conditioned media and PBS and subsequently pelleted by centrifugation at 400 RPM for 5 min. Supernatant was removed, and each cell pellet was resuspended in 1 ml pre-warmed Hank’s Balanced Salt Solution (HBSS) without Phenol Red, Ca2+, Mg2+ (10–547, Lonza) containing freshly resuspended H2DCFDA (Image-IT LIVE Green ROS Detection Kit, I36007, Thermo Fisher) to a final concentration of 25 μM. Cells were incubated at 33°C for 40 min. Just prior to FACS analysis, cells were placed on ice, and 0.5 ml HBSS containing 1 μg/ml propidium iodide was added to each tube. We utilized a doublet discrimination gating strategy on a BD Aria FACS instrument, wherein on average 93% of all cells were included in the analyses (S11 Fig). Fluorescent signal detected in scrambled siRNA negative control cells were utilized to set the quadrant boundaries. Data for three replicates of each experiment were collected and mean percentages for each quadrant were calculated.
Raw sequencing reads were mapped to opossum M. domestica genome assembly monDom5 with Ensembl annotation v86, using Tophat2 v2.1.1 [71] and Bowtie2 v2.2.9 [72]. Read counts for all genes were calculated using HTSeq v0.6.1p1 [73] with Python (v2.7.13). Transcripts per million were calculated to estimate relative mRNA abundance [74]. We used Bioconductor package edgeR v3.16.5 [75] to assay for differential gene expression between unstimulated and stimulated MdESF. Genes that met the following criteria were considered to be significantly differentially expressed: (1) change in expression of at least 1.5-fold; (2) resulted in an adjusted P-value smaller than 10−6; and (3) expressed in at least one condition under comparison (TPM ≥ 3) [76]. GO enrichment analyses were performed using Gorilla [77], and the results were visualized using REViGO [78].
The data from the experiment testing the effect of 8-br-cAMP/MPA and PGE2/MPA, as well as FOXO1 and FOXO3 KDs on the presence of ROS and apoptotic cells, were analyzed as a two-factor ANOVA. The two factors were “stimulation” and “treatment,” where “stimulation” had the levels, “growth media,” 8-br-cAMP/MPA, and PGE2/MPA and “treatment” had the levels random siRNA, FOXO1 KD, and FOXO3 KD. The response variable was the fraction of cells showing either ROS or PI fluorescence. The analysis was performed with raw frequencies as well as with log-transformed response variables. For the PGE2 ELISA, the concentration of PGE2 for each sample was determined by standard curve. The values were subsequently log transformed and used in a one-tailed t test.
|
10.1371/journal.ppat.1002142 | Novel Chikungunya Vaccine Candidate with an IRES-Based Attenuation and Host Range Alteration Mechanism | Chikungunya virus (CHIKV) is a reemerging mosquito-borne pathogen that has recently caused devastating urban epidemics of severe and sometimes chronic arthralgia. As with most other mosquito-borne viral diseases, control relies on reducing mosquito populations and their contact with people, which has been ineffective in most locations. Therefore, vaccines remain the best strategy to prevent most vector-borne diseases. Ideally, vaccines for diseases of resource-limited countries should combine low cost and single dose efficacy, yet induce rapid and long-lived immunity with negligible risk of serious adverse reactions. To develop such a vaccine to protect against chikungunya fever, we employed a rational attenuation mechanism that also prevents the infection of mosquito vectors. The internal ribosome entry site (IRES) from encephalomyocarditis virus replaced the subgenomic promoter in a cDNA CHIKV clone, thus altering the levels and host-specific mechanism of structural protein gene expression. Testing in both normal outbred and interferon response-defective mice indicated that the new vaccine candidate is highly attenuated, immunogenic and efficacious after a single dose. Furthermore, it is incapable of replicating in mosquito cells or infecting mosquitoes in vivo. This IRES-based attenuation platform technology may be useful for the predictable attenuation of any alphavirus.
| Chikungunya virus (CHIKV) is a mosquito-borne alphavirus that has reemerged since 2004 to cause millions of cases of severe and often persistent arthralgia. Because no licensed vaccine exists to prevent this disease, we utilized an attenuation approach to produce a live CHIKV vaccine candidate that elicits a robust, protective immune response yet causes no detectable disease in mice. It is also incapable of infecting mosquito vectors, an important safety feature for a live virus vaccine that may be used in nonendemic locations to immunize travelers or laboratory personnel. This vaccine approach, which exploits the attenuating effect of altering the expression of the alphavirus structural proteins with a picornavirus IRES, may be broadly applicable to other alphaviruses that cause important febrile diseases as well as encephalitis.
| Chikungunya (CHIK) virus (CHIKV) is a reemerging arboviral pathogen that has recently caused explosive urban outbreaks involving millions of persons in Africa and Asia. The virus was first isolated from a human in Tanzania in 1953 during a major epidemic [1], and derives its name from a Makonde word meaning “that which bends up,” which describes the posture observed in afflicted persons. CHIKV typically causes a febrile illness and severe joint pain, which is clinically similar to dengue fever. These 2 viruses also share similar endemic distributions in the Eastern Hemisphere, resulting in many CHIKV cases being misdiagnosed when laboratory testing is not available [2]. Large CHIK outbreaks were described during the 1950's and 60's in India and Southeast Asia [3], [4]. However, it was not until 2005 that CHIKV gained widespread public attention due to massive outbreaks on islands of the Indian Ocean [5] and later in India [6] and Southeast Asia [7]. In total, several million persons have been affected [8], [9]. On the Island of Reunion alone, ca. 300,000 persons or one-third of the population was affected [10]. Another factor driving the resurgence of interest in CHIK is the detection of occasional fatal cases, which were not documented before. Previously, individuals who became severely ill typically presented with hemorrhagic manifestations and occasionally shock [11], [12], [13]. However, the recent outbreaks have been linked to thousands of deaths in Reunion and India due to neurologic disease [14], [15], [16].
CHIKV exists in two transmission cycles: an enzootic or sylvatic cycle and an endemic/epidemic urban cycle. The African sylvatic cycle likely involves several arboreal Aedes mosquitoes as vectors and nonhuman primates as reservoir/amplifying hosts [17]. African outbreaks occur from direct enzootic spillover or when CHIKV is introduced into an urban areas inhabited by the anthropophilic mosquito vector, Aedes aegypti. [17], [18]. More permanent endemic/epidemic transmission cycles were established when the virus was introduced into Asia ca. 1950, and into the Indian Ocean region, India and then Southeast Asia since 2005 [19]. A mutation in the E1 envelope glycoprotein gene that results in an A226V amino acid substitution dramatically increased the infectivity of some epidemic strains for an alternative urban vector, Ae. albopictus [8], [20]. The nearly ubiquitous distribution of Ae. aegypti, and the expanding distribution of Ae. albopictus into tropical and temperate regions of both hemispheres has raised concern that CHIKV may spread outside of its previous endemic region into the Western Hemisphere and Europe. The latter scenario was realized in 2007 during a small epidemic in Italy [21] and during autochthonous transmission in southern France during 2010 (ProMED archive 20100926.3495).
CHIKV belongs to the family Togaviridae, genus Alphavirus, whose members are enveloped virions that contain a positive sense, single stranded, RNA genome of ∼12 kb. The genome encodes 4 non-structural proteins (nsP1-4) and 3 major structural proteins (Capsid, E1, and E2 envelope glycoproteins)(Fig. 1A). During replication, two distinct RNA's are produced: the genomic and subgenomic RNAs. A negative sense template RNA is also produced. The nonstructural polyprotein open reading frame (ORF) is translated via a cap-dependant mechanism from the genomic RNA, whereas the structural protein gene ORF is translated from the subgenomic RNA, also in a cap-dependent manner. The subgenomic RNA is transcribed late during infection from its promoter, which is found in the 3′ end of the nsP4 gene [22].
There is no licensed vaccine or therapeutic CHIK, so outbreaks can only be controlled by preventing the exposure of people to infected mosquito vectors. Scientists at the Walter Reed Army Institute of Research produced an investigational vaccine called 181/clone 25 (hereafter called 181/25) during the 1980s. This live-attenuated strain was generated via serial plaque-to-plaque passages of a wild-type Thai CHIKV strain using MRC-5 cells [23]. The virus is attenuated in both rodents and non-human primates and is highly immunogenic in humans. However, during phase II trials, strain 181/25 caused mild, transient arthralgia in 5 of 59 vaccinees [24]. Also, strain 181/25 can be transmitted experimentally by the natural mosquito vector, Ae. aegypti [25].
To be effective in resource-limited nations that are endemic for CHIK as well as to combat an epidemic, an ideal CHIK vaccine would induce rapid and long-lived immunity after a single dose, have a low risk of reactogenicity and reversion to virulence, and be inexpensive. Vaccines against arboviral diseases should also have a low risk of transmission from vaccinated persons via mosquitoes in the event that viremia occurs, especially those used in non-endemic regions. Although replication-defective vaccine candidates have been described that emphasize safety [26], [27], [28], none has been shown to induce rapid or long-lived immunity after a single dose, and some may be expensive to produce. In contrast, live-attenuated vaccines like the yellow fever 17D vaccine [29] have been spectacularly successful in preventing disease in developing tropical regions.
To generate a safer and more effective live-attenuated CHIK vaccine that meets the criteria outlined above, we previously produced and tested a series of chimeric alphaviruses containing either Venezuelan equine encephalitis virus (VEEV), eastern equine encephalitis or Sindbis virus non-structural protein genes along with the CHIKV structural protein genes [30]. These vaccines produce robust neutralizing antibody (Ab) responses and provide complete protection against disease after CHIKV challenge. However, some residual ability to infect potential mosquito vectors remains, and attenuation is dependent on an intact murine interferon response (SCW, unpublished). To overcome these limitations, we developed a new attenuation strategy and conducted proof-of-principle studies with another alphavirus, VEEV vaccine strain TC-83. Both attenuation and elimination of mosquito infectability relied on the inactivation of the subgenomic promoter, and addition of a encephalomyocarditis virus (EMCV) internal ribosome entry sequence (IRES) to drive translation of the structural protein genes [31]. Chimeric alphaviruses incorporating the IRES element have also been generated as vaccine candidates [32]. The EMCV IRES also mediates inefficient translation in arthropod cells [33], rendering these mutants unable to infect mosquitoes. However, starting with the attenuated TC-83 strain, the IRES-based attenuation resulted in inadequate immunogenicity and the lack of a neutralizing Ab response.
Here, we implemented this IRES-based vaccine design for CHIKV using a cDNA clone generated from the wild-type La Reunion strain [34]. Testing of this novel vaccine candidate in several murine models indicated that it is highly attenuated, even in the absence of an intact murine IFN response, is immunogenic and efficacious in preventing CHIK disease, and is unable to infect mosquitoes.
The CHIKV/IRES vaccine candidate was generated in cDNA form using standard recombinant DNA techniques using the IRES-based attenuation strategy tested previously in TC-83 [31]. The IRES element was amplified from the original TC-83/IRES construct including the first 4 codons from the EMCV sequence that were previously shown to have no effect on viral replication [31]. The IRES sequence was placed directly downstream from the subgenomic promoter of the La Reunion (LR) CHIKV infectious cDNA clone (Fig. 1B) [34]. The subgenomic promoter was inactivated using 13 synonymous mutations to preserve the wild-type amino acid sequence of nsP4 (Fig. 1C). The resultant virus, rescued by electroporation of in vitro-transcribed RNA into Vero cells, contained a non-functional subgenomic promoter as indicated by the absence of subgenomic RNA within infected cells (Fig. 2A).
Titers of CHIKV/IRES collected 30 h after electroporation were 6×106 plaque forming units (PFU)/ml, in comparison to titers of 1.1x107 for wild-type (wt) CHIKV strain LR. To assess replication kinetics, viruses derived from the electroporation were compared after infection of Vero cells. The CHIKV/IRES replicated more slowly than 181-25 or wt-CHIKV, requiring 48 hours at 37°C to reach a peak titer of 2.5×107 PFU/ml. Strain 181-25 replicated almost to peak titer within 24 hours and reached 7.9×107 PFU/ml. The wt-CHIKV also replicated close to its peak titer by 24 hours and reached 4.2×107 (Fig. 2B). Unlike wt-CHIKV, which produced visible plaques within 48 hours of Vero cell infection, the CHIKV/IRES plaques were not readily visible before 3 days of incubation at 37°C. CHIKV/IRES plaques were 0.5–2 mm in diameter, whereas vaccine strain 181/25 produced 2–4 mm and wt CHIKV produced ca. 6 mm plaques under 0.4% agarose at 3 days post infection (Fig. 2C).
To assess phenotypic and genetic stability, CHIKV/IRES was passaged 10 times in Vero cells at 37°C using a multiplicity of infection of 0.1 PFU/cell. The plaque morphology remained heterogeneous but consistent after the 10 passages (Fig. 2C). Sequencing of reverse transcription-polymerase chain reaction (RT-PCR) amplicons covering the entire genome revealed no consensus mutations aside from the presence of adenine insertions within a poly-A track of the IRES element itself. Plaque purified clones were sequenced through the IRES to determine the frequency of these mutations; 8 of 10 plaque clones examined had 7 As like the original cDNA clone and the 10th passage consensus sequence. However, 3 biological clones had up to 17 As in this region. These differences in sequence showed no obvious correlation with plaque size (data not shown).
CHIKV/IRES was also blind-passaged 5 times in C6/36 Ae. albopictus cells and the presence of virus was detected by the ability to produce cytopathic effects (CPE) on Vero cells and by RT-PCR amplification. Virus was detected only after the first passage, which presumably reflected residual virus that could not be washed from the cells after inoculation, and was not detected thereafter (data not shown). In contrast, the wt-CHIKV strain replicated in the mosquito cells throughout the passages, with titers ranging from 3–5×107 PFU/ml.
Infant outbred CD1 mice develop CHIK disease similar in many ways to that seen in humans [35]. We therefore used this model to evaluate the attenuation of our CHIKV/IRES vaccine candidate. Cohorts (N = 3) of 6-day-old CD1 mice were injected subcutaneously (SC) with 105 PFU (A high dose to increase sensitivity to detect virulence) of strains 181/25, wt LR, or CHIKV/IRES, and were sacrificed on days 2, 4, 6, and 8 to compare viral loads. Blood, brain, and leg tissue (including the knee) were collected and titrated for infectious virus. The CHIKV/IRES strain produced no detectable virus in any tissue measured throughout the sampling period. In contrast, both vaccine strain 181/25 and wt CHIKV produced measurable and significantly higher viremia through day 4 (p<0.05)(Fig. 3A). Surprisingly, vaccine strain 181/25 produced higher viral titers in leg tissue than wt strain LR, and both wt-CHIKV and 181/25 leg titers were significantly higher than those of CHIKV/IRES (p<0.05)(Fig. 3B). The wt CHIKV strain produced significantly higher brain titers than either vaccine strain on day 2 (p<0.05)(Fig. 3C). These results indicated that CHIKV/IRES is strongly attenuated in the baby mouse model.
Another murine model for CHIKV pathogenesis is the A129 mouse, which lacks functional type I interferon receptors. This model has the advantage of producing disease in adult animals, thus permitting efficacy testing using wt-CHIKV challenge [36]. Cohorts of 10-week-old homozygous A129 mice were injected intradermally in the footpad with 104 PFU (more than 100 LD50 for wt-CHIKV) of either CHIKV/IRES (N = 7) or 181/25 (N = 4), and negative controls were sham (PBS)-infected (N = 6). Mice infected with the CHIKV/IRES vaccine showed no visible signs of illness (weight loss, temperature change, ruffling of fur or hunched posture) during 14 days of observation. Mice receiving strain 181/25 exhibited significant hyperthermia from day 4–5, and also showed significant weight loss on day 6 post vaccination (p<0.05), compared to the more constant temperatures and weight increases observed in the mice receiving CHIKV/IRES (Figs. 4A and B). Both CHIKV/IRES and 181/25 produced viremia in A129 mice, but mean titers were consistently lower for CHIKV/IRES (Fig. 4C). These data suggested greater attenuation of CHIKV/IRES compared with181/25.
Another sign of disease monitored in A129 mice was swelling of the feet. For this measurement, mice were vaccinated as described above and subsequently challenged with 100 PFU of wt-CHIKV one month post-vaccination in the same foot as the vaccination site. Two days after vaccination or challenge, the vertical heights of the hind feet were measured using a caliper at the balls. PBS and 181/25 vaccination produced small and similar amounts of swelling (ca. 0.05 mm), while CHIKV/IRES vaccination produced slightly greater but still minimal swelling of 0.1 mm (Fig. 5). Sham-vaccinated mice that were challenged showed a strong inflammatory response with a mean increase of 0.8 mm in footpad thickness. In contrast, both vaccines protected significantly against swelling (p<0.001) with no significant difference between 181/25 and CHIKV/IRES.
Attenuation of the 2 vaccine candidates was also compared by infection of 3-week-old A129 mice. Cohorts of 5 were injected intradermally with 104 PFU (>100 LD50 for wt-CHIKV) of either 181/25 or CHIKV/IRES. The mice were monitored for weight and survival. There was no significant difference between the weight changes of the two cohorts (Fig. 6A). All animals that received the 181/25 vaccine died or had to be euthanized by day 8 (Fig. 6B). In contrast, none of the animals inoculated with the CHIKV/IRES vaccine showed any signs of illness and all survived to the end of the study 14 days after infection.
All A129 mice that received vaccine candidates 181/25 (N = 4) or CHIKV/IRES (N = 7) at a dose of 104 PFU seroconverted. All titers measured 35 days after vaccination exceeded 320, except for one mouse immunized with strain 181/25 that had a PRNT80 titer of 160. None of the animals that received CHIKV/IRES or 181/25 showed a significant temperature change (data not shown) or any other signs of illness (as described above) after challenge with 100 PFU of wt CHIKV, and all survived until day 14 after challenge, when the study was terminated. Mice vaccinated with 181/25 exhibited stable or slightly increasing weight after challenge, while the CHIKV/IRES-vaccinated mice lost some weight on days 8 and 9 post challenge, then recovered. In sharp contrast, sham-vaccinated animals rapidly lost weight before succumbing to infection (Fig. 7A and B). Both vaccines were significantly (Kaplan-Meier, p<0.05) and equally efficacious in preventing fatal CHIK in the A129 model.
The ability of the CHIKV/IRES vaccine candidate to protect against disease was also measured histopathologically in A129 mice after wt-CHIKV challenge. Because unprotected mice die before muscle or joint lesions develop (SCW, RS, unpublished), we examined the spleen, where earlier lesions occur. Cohorts of three 8–10-week-old A129 mice were vaccinated intradermally in the footpad with either 104 PFU of CHIKV/IRES or were sham-vaccinated with PBS. One mouse from each cohort was sacrificed 4 days post vaccination, and the remaining 2 mice were challenged with 100 PFU of wt-CHIKV at 26 days post-vaccination, then sacrificed 4 days post-challenge. The spleens of the sham-vaccinated mice challenged with CHIK-LR exhibited severe necrosis with markedly reduced numbers of small lymphocytes in the mantle and marginal zones. Only the central portion of the remnant lymphoid follicle remained. In addition, monocytoid cells with abundant eosinophilic cytoplasm in the interfollicular region were observed (Fig. 8C). In contrast, the spleens of animals receiving the vaccine as well as CHIKV/IRES-vaccinated mice challenged with wt-CHIKV (Fig. 8B & D) exhibited normal splenic architecture with intact lymphoid follicles and appropriate quantities of white and red pulp. The key histopathologic finding was the absence of any necrosis in the CHIK/IRES-vaccinated animals, when compared to the sham-vaccinated mice.
To evaluate the duration of immunity and protection after vaccination, cohorts of six A129 mice were immunized with CHIKV/IRES as described above, bled 21, 42, 56 and 92 days later, then challenged 94 days after vaccination. Similar to the results described above, no significant weight loss, footpad swelling, or other signs of disease were noted after vaccination compared to sham-vaccination (data not shown). Antibody PRNT80 titers prior to challenge were all ≥640. After challenge with 100 PFU of wt-CHIKV, vaccinated animals were significantly protected against foot swelling, fever and mortality (6/6 sham-vaccinated mice died by day 5, whereas all CHIKV/IRES-vaccinated mice survived until day 14 when the study was terminated)(Fig. 9A and B). The sham-vaccinated group experienced significant hyperthermia on day 2, followed by significant hypothermia on day 3 as the animals became moribund (Fig. 9C). There were no significant differences in weight change between the two cohorts (Fig. 9D).
To test the immunogenicity and efficacy of the CHIKV/IRES vaccine candidate compared with strain 181/25 in immunocompetent mice, cohorts (N = 9-10) of 3-week-old C57BL/6 mice were vaccinated SC with 105 PFU, or with PBS as negative controls. Although 14-day-old and adult C57BL/6 mice develop lesions in the leg after footpad inoculation with wt CHIKV [37], [38], we used a more stringent, lethal intranasal (IN) challenge C57BL/6 model with the neuroadapted Ross CHIKV strain for efficacy testing [30]. Three weeks after infection, all mice were bled and Ab titers were measured using an 80% plaque reduction neutralization test (PRNT80). The mean Ab titers in response to strains CHIKV/IRES and 181/25 were nearly equal, with all animals exhibiting PRNT80 titers ≥20 (p>0.1; Table 1). The mice were then challenged IN with 106 PFU of the Ross CHIKV strain. All vaccinated animals survived without any signs of disease (weight loss, temperature change, ruffling of fur or hunched posture) through day 14. One of ten sham-vaccinated mice died on day 9 and 6 died on day 10 after challenge. These results demonstrated the immunogenicity and significant efficacy of the CHIKV/IRES vaccine candidate in immunocompetent mice.
To confirm that neutralizing antibodies mediated protection of A129 mice from CHIKV challenge, pooled serum collected 21 days after immunization of A129 mice was inoculated intraperitoneally into naïve 6–7-week-old A129 mice (N = 5) either undiluted or at dilutions of 1:2 or 1:4; undiluted normal mouse serum was used as a negative control. Following challenge with 100 PFU of wt CHIKV, mortality was monitored for 15 days. All mice that received immune serum exhibited increased survival compared with those that received normal mouse serum (Kaplan Meier, p<0.001) and greater dilutions of the immune serum resulted in reduced survival (Fig. 10). These data indicate that neutralizing antibodies protected against fatal CHIK and indicate a correlation between Ab levels and protection.
To confirm that the CHIKV/IRES strain was incapable of replicating in mosquitoes, cohorts of 20 adult female Ae. albopictus, a highly susceptible urban vector [20], were inoculated intrathoracically with ca. 1.0 µl of a 104 PFU/ml suspension of either CHIKV/IRES or the wt LR strain. Intrathoracic infection was used rather than oral exposure because mosquitoes are uniformly susceptible to small CHIKV doses delivered via this route, whereas the oral portal of entry is less permissive even after large doses. After 7 days of incubation at 27°C, mosquitoes were triturated and serial 10-fold dilutions were tested for virus by inoculation of Vero cells followed by examination for cytopathic effects (CPE) through day 7. Mosquitoes inoculated with CHIKV/IRES as well as PBS-inoculated negative control mosquitoes produced no detectable CPE. In contrast, all 20 mosquitoes receiving wt CHIKV produced extensive CPE on the Vero cells. To ensure that temperature sensitive or host-restricted mutants were not generated following mosquito infection, RT-PCR targeting the 5′ end of the capsid gene was also used to detect viral RNA. No amplicons were detected from the CHIKV/IRES-infected mosquitoes by gel electrophoresis, whereas all mosquitoes injected with wt CHIKV produced strong bands of the expected size (Fig. S1).
Nearly 80 years after the introduction of the first vaccine against an arboviral disease, yellow fever [39], vaccination remains the most effective method to protect against arboviruses and many other infectious agents. In the case of CHIK, the 181/25 live-attenuated vaccine developed during the 1980s showed promise in preclinical studies [23] but was mildly reactogenic in human trials [24]. More recent vaccine development has focused on inactivated [26], DNA [27] or virus-like particle approaches [28]. However, in our opinion, the requirements for multiple doses administered over several weeks and/or the higher cost of such vaccines, as well as the probability that boosters will be required to maintain immunity, will limit their usefulness in the developing nations of Africa and Asia where CHIKV is endemic. We have therefore focused on live-attenuated vaccines to prevent both endemic and epidemic CHIK.
The maturity of reverse genetic technology has provided unprecedented opportunities for manipulation of the alphaviral genome to improve attenuation strategies [40]. Thus, unlike traditional attenuation approaches that rely on cell culture passages, which typically result in attenuation that depends only on small numbers of attenuating point mutations [41], alternative genetic strategies such as viral chimeras offer the promise of more stable attenuation [30], [42], [43], [44]. In addition to the risk of reactogenicity, attenuation based on small numbers of mutations can also result in residual alphavirus infectivity for mosquito vectors. This risk, which was underscored by the isolation of the TC-83 VEEV vaccine strain from mosquitoes in Louisiana during an equine vaccination campaign designed to control the 1971 epidemic [45], is especially high when a vaccine that relies on a small number of point mutations is used in a nonendemic location that could support a local transmission cycle.
To overcome the aforementioned limitations, we exploited the finding that the EMCV IRES sequence functions inefficiently for translation in insect cells [33], yet can replace the alphavirus subgenomic promoter to mediate translation of the structural polyprotein open reading frame from the genomic RNA in mammalian cells [31], [32]. The resultant CHIKV strain replicated efficiently in Vero cells, an acceptable vaccine substrate, and exhibited a stable plaque morphology and consensus genome sequence after 10 passages in this cell line. The CHIKV/IRES vaccine candidate was unable to replicate in mosquito cells or in the mosquito vector, Ae. albopictus, an important safety feature for an live arbovirus vaccine that may be administered to travelers or laboratory workers in nonendemic locations.
Attenuation, immunogenicity and efficacy of the CHIKV/IRES vaccine candidate was assessed alongside that of the 181/25 CHIKV strain, which is highly immunogenic in humans and other animals yet inadequately attenuated. The goal was to equal the immunogenicity of the 181/25 vaccine strain but to achieve greater attenuation. Using infant and adult immunocompetent [35] and interferon type I receptor-deficient mouse models [36], we demonstrated that CHIKV/IRES met both goals. As measured by survival and weight gain or maintenance, CHIKV/IRES was similarly or better attenuated than 181/25 in multiple mouse models, yet generated comparable neutralizing Ab titers and nearly complete protection against disease or mortality after CHIKV challenge. Immunity and protection were maintained for at least 3 months. Viremia after CHIKV/IRES vaccination was never detected in infant CD-1 mice, and was transiently present at a very low level in immunocompromised A129 mice, an important attenuation phenotype considering that viremia could potentially lead to mosquito infection. However, even in the unlikely event that vaccination of an immunocompromised human led to viremia, the mosquito-incompetent phenotype discussed above should prevent transmission. The only measure of efficacy for which strain 181/25 exhibited a slight superiority was in the prevention of footpad swelling post challenge; CHIKV/IRES-vaccinated A129 mice challenged with wt-CHIKV exhibited a greater mean of 0.15mm swelling versus only 0.09 mm for strain 181/25. However, both vaccines provided significant protection compared with sham-vaccination.
Splenic histopathology was used as a second measure of protection. Mice challenged with wt-CHIKV after sham vaccination developed severe necrosis along with a monocytoid infitrate. in contrast, the CHIKV/IRES vaccine induced no splenic histopathology and protected against splenic lesions upon challenge.
Previous attempts to use the EMCV IRES to generate an alphavirus vaccine used the VEEV live-attenuated vaccine strain TC-83 [30]. Although these studies succeeded in eliminating the ability of TC-83 to infect mosquito vectors, immunogenicity was reduced to the point where most vaccinated mice did not develop detectable neutralizing antibodies (although significant protection against challenge was still detected). In contrast, our CHIKV vaccine started with the genetic backbone of a virulent wt alphavirus (LR) (Fig. 1A), and robust immunogenicity was maintained despite strong attenuation. These results suggest that the IRES attenuation level may be optimal when applied to other wild-type alphavirus backbones. The application of this platform for attenuation is now being applied to Venezuelan, western, and eastern equine encephalitis viruses to test this hypothesis.
In summary, a novel CHIK vaccine candidate, CHIKV/IRES, was generated by manipulation of the structural protein expression of a wt-CHIKV strain via the EMCV IRES. This vaccine candidate exhibits a high degree of murine attenuation that is not dependent on an intact interferon type I response, yet is highly immunogenic and protects against CHIKV challenge. This promising vaccine candidate is being tested in nonhuman primates to determine if it is suitable for evaluation in humans.
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 Committees of the University of Texas Medical Branch or the University of Wisconsin.
Vero African green monkey kidney cells were obtained from the American Type Cell Culture (Bethesda, MD). The cells were maintained at 37°C in Eagles minimum essential media (MEM) supplemented with 10% fetal bovine serum (FBS), penicillin and streptomycin. C6/36 Ae. albopictus cells were also maintained in MEM containing 10% FBS at 32°C and supplemented with 10% tryptose phosphate.
The CHIKV cDNA clone containing the EMCV IRES with the subgenomic promoter ablated using 13 synonymous mutations (CHIKV/IRES) was produced using standard recombinant DNA techniques in which the infectious clone of La Reunion strain (LR) described previously was used as a template [34]. This CHIKV clone, a gift from Stephen Higgs, contains an SP6 bacteriophage promoter for transcription of RNA that is identical to genomic viral RNA. The IRES sequence was PCR amplified from a cDNA clone described previously [31]. The inactivation of the subgenomic promoter was done using site-specific mutagenesis. An intermediate construct encoding the 3′ end of the nsP4 gene through the subgenomic promoter was produced using PCR with high fidelity Phusion DNA polymerase from Finnzymes (Espoo, Finland). The resultant amplicon was cloned into a shuttle vector, prS2, and was sequenced using the BigDye kit (Applied Biosystems, Foster City, CA). The 5′ end of capsid gene from the LR strain was amplified using PCR with an overhang complementary to the IRES sequence. The IRES-containing and capsid fragments were then joined using fusion PCR, and this fragment was cloned back into the shuttle vector and resequenced. The IRES/Capsid fragment and the mutated subgenomic fragment were finally ligated together through the SpeI site introduced into both fragments. The completed insert was then cloned into the LR backbone and this final construct was completely sequenced.
Large-scale plasmid purification was done using CsCl preparations. The purified DNA was then linearized using NotI restriction endonuclease (New England BioLabs, Ipswich, MA), and a small sample was analyzed on a 1.2% agarose gel to verify linearization. The remaining DNA was transcribed using an Ambion SP6 In vitro transcription kit. The RNA was quantified and used to electroporate Vero cells using a BTX ECM 830 electroporator. Briefly, two T-150 flasks containing 90% confluent Vero cells were trypsinized and washed 3 times in RNAse-free DPBS. The cells were resuspended in 700 µl of DPBS and 10 µg of RNA was added. The solution was placed in a 4mm cuvette and was pulsed 2 times at 250v for 10 msec at 1 sec intervals. The cells were then left at room temperature for 10 minutes before being plated in T-75 flasks. The virus was harvested at 24 hours post-electroporation and centrifuged at 771×g. Supernatant was collected and titered by plaque assay on Vero cells.
The CHIKV/IRES vaccine candidate was passaged in Vero and C6/36 cells to assess phenotypic and genetic stability. T-25 flasks were grown to 90-95% confluency, then were infected at a multiplicity (MOI) of 0.1 Vero PFU/cell. Following 30 h of incubation at 37°C or 32°C, respectively, the medium was diluted and used to infect another flask with a MOI of 0.1. Following 10 serial passages, consensus sequences were determined for both passaged populations and plaque-purified biological clones by RT-PCR amplification and amplicon sequencing. We also selected 10 well-isolated, random plaques, harvested virus using a plastic micropipette tip. The agar plug containing the plaque was placed in 300 µl of MEM containing 2% FBS and RNA was extracted using TRIzol LS (Invitrogen, Carlsbad, CA). RT-PCR and sequencing were performed as described above. Vero plaque sizes were measured and compared to assess stability.
Replication kinetics was measured in 35 mm 6-well plates with duplicates for each virus tested. The wells were seeded to a confluency of 95% using Vero cells. Media was removed and they were infected at an MOI of .1 for one hour. Then 2.1 ml of DMEM containing 5% FBS was added. A 0 time point was immediately removed (100 µl). At each of the remaining time points 12, 24, 36 and 48 100 µl was removed and replaced. The samples were tittered as described above.
Depending on containment requirements and sensitivity needs, virus stocks and experimental samples were titered by plaque assay as previously described [46] or were estimated using quantitative real-time PCR with dilutions of virus to generate standard curves from which PFU titers could be extrapolated. This assay used primers (5′-GAYCCCGACTCAACCATCCT-3′) and (5′-CATMGGGCARACGCACTGGTA-3′) and the probe (5′-AGYGCGCCAGCAAGGAGGAKGATGT-3′) which contained the dye FAM. Ab titers were measured using plaque reduction neutralization tests with 80% reduction endpoints [46].
Vero cells were infected on 35 mm2 6 well plates at an MOI of 20. The media was removed 18 hours after infection and replaced with .8 ml of complete media with 1 µg/ml of actinomycin D from Sigma, and 20 µCI of [5,6-3H] uridine from Moravak Biochemicals (Brea, CA.). The cells were then incubated for 4 hours and RNA is removed by TRIzol extraction. The RNA was placed into a sodium phosphate buffer containing DMSO and glyoxal at 50°C for 1 hour. The RNA was loaded into a 1% agarose gel and run at 150 v for 3–4 hours. The gel was then washed twice in methanol for 30 minutes. Then a 2.5% PPO and methanol solution was placed with the gel overnight. The gel was washed with DI water to precipitate the PPO and the gel was then dried. The gel is then placed with X-OMAT AR film (Kodak), at −80°C for 8 hours.
Five-to-seven-day-old CD1 outbred mice [35] were obtained from Charles River (Wilmington, MA). These animals were infected subcutaneously (SC) with 105 PFU and were serially sacrificed on days 2, 4, 6, 8, and 10. Blood, brain, and hind femoral tissues were collected for assays of virus content. C57BL/6 mice were obtained from Jackson labs (Bar Harbor, ME) and used in challenge experiments as described previously [30]. Briefly, the animals were infected SC at 3 weeks of age with 105 PFU in the hind leg and observed for signs of illness for 21 days. Then, they were challenged intranasally (IN) with 106.5 PFU of the neuroadapted Ross CHIKV strain. The animals were observed daily for illness and were sacrificed when they became moribund.
A129 mice were bred at the University of Wisconsin from a breeding pair obtained from B & K ltd. Grimston, England. Animals 3 or 10 weeks of age, were infected with 1x104 PFU of vaccine strains ID in the left rear footpad. Footpad measurements were taken 48 hours post vaccination with a caliper as the vertical height of the hind feet at the balls. The animals were maintained for 38 days and bled on days 21 and 35. These animals were then challenged with 100 PFU of wt CHIKV and were monitored for morbidity and mortality. All animals were euthanized by CO2 overdose if they became moribund. A129 animals were used for a longitudinal study of protection in which they were challenged with 100 PFU ID of wt-CHIKV 94 days after being vaccinated.
Tissues were fixed in 10% neutral buffered formalin (RICCA Chemical Company, Arlington, TX.). Bone tissue was decalcified overnight using fixative/decalcifier (VWR International, Radnar, PA.). Tissue was then embedded in paraffin wax and 5 um sections were cut for analysis. Sections for hematoxylin and eosin staining were deparaffinized in Xylene for 15 minutes. Sections were then rehydrated in ethanol and ethanol/water mixtures as follows: 100% ethanol for 9 minutes, 95% ethanol/5% deionized water for 3 minutes, 80% ethanol/20% deionized water for 5 minutes. Sections were then stained with hematoxylin (Richard-Allan Scientific) for 3 minutes and then rinsed with deionized water. Sections were then rinsed in tap water for 5 minutes and placed in Clarifier I (Richard-Allan Scientific, Kalamazoo, MI.) for 5 minutes. Sections were then rinsed in tap water for 2 minutes and then in deionized water for 2 minutes. Sections were then stained in eosin (Richard-Allan Scientific) for 30 seconds. They were then dehydrated as follows: 95% ethanol/5% deionized water for 15 minutes, 100% ethanol for 15 minutes and then Xylene (Richard-Allan Scientific) for 15 minutes. Cover slips were applied to slides using Permount (Fisher Scientific) and dried overnight. Deparaffinizing and hematoxylin-eosin staining was performed on the Varistain Gemini ES (Shandon, Thermo Fisher Scientific).
All animal studies were approved by the UTMB and/or the Univ. Wisconsin Institutional Animal Care and Use Committee.
An Ae. albopictus colony established in 2003 from mosquitoes collected in Galveston, TX was used for these experiments. This species was selected because it is highly susceptible to the LR CHIKV strain [20]. Adult female mosquitoes collected 3–4 days post-eclosion were anesthetized using a chill table (Bioquip, Rancho Dominguez, CA) and were then injected intrathoracically with ca. 1.0 µL of a 104 Vero PFU/ml virus stock. The mosquitoes were incubated for 7 days at 27°C with 10% sucrose provided ad libitum. The mosquitoes were then frozen and triturated in MEM containing 2% FBS and fungicide using a Tissuelyser II (Qiagen, Venlo, Netherlands) for 2 min. Following centrifugation for 10 minutes at 10,000×G, the supernatant was plated on Vero cells using 96 well plates. The cells were infected for 1 hour at 37°C and then covered with 2% FBS containing MEM and allowed to incubate for 48 hr to measure CPE.
RNA was collected through Qiagen RNeasy columns (Qiagen, Venlo, Netherlands) or TRIzol LS (Invitrogen) using the manufacturer's protocols. 130 µl of sample were taken from the mosquito homogenates and the RNA was collected. The RNA was then amplified via RT-PCR using a Titan single step RT-PCR kit (Roche, Basel, Switzerland). The primers used to amplify annealed to the 5′end of capsid, 5′-TGGCCTTTAAGCGGTC-3′ and 5′-TATGGTCTTGTGGCTTTATAGAC-3′.
Student's T-tests were performed using Excel (Microsoft, Redmond, WA). ANOVA tests were performed using SPSS v18 (IBM Corporation, Somers, NY). Kaplan-Meier tests were performed using Prism 5 (GraphPad Software, La Jolla, CA). P-values <0.05 were considered significant. Negative data points were counted at one-half of the corresponding limit of detection for statistical analyses.
|
10.1371/journal.pntd.0002952 | Variability in Dengue Titer Estimates from Plaque Reduction Neutralization Tests Poses a Challenge to Epidemiological Studies and Vaccine Development | Accurate determination of neutralization antibody titers supports epidemiological studies of dengue virus transmission and vaccine trials. Neutralization titers measured using the plaque reduction neutralization test (PRNT) are believed to provide a key measure of immunity to dengue viruses, however, the assay's variability is poorly understood, making it difficult to interpret the significance of any assay reading. In addition there is limited standardization of the neutralization evaluation point or statistical model used to estimate titers across laboratories, with little understanding of the optimum approach.
We used repeated assays on the same two pools of serum using five different viruses (2,319 assays) to characterize the variability in the technique under identical experimental conditions. We also assessed the performance of multiple statistical models to interpolate continuous values of neutralization titer from discrete measurements from serial dilutions. We found that the variance in plaque reductions for individual dilutions was 0.016, equivalent to a 95% confidence interval of 0.45–0.95 for an observed plaque reduction of 0.7. We identified PRNT75 as the optimum evaluation point with a variance of 0.025 (log10 scale), indicating a titer reading of 1∶500 had 95% confidence intervals of 1∶240–1∶1000 (2.70±0.31 on a log10 scale). The choice of statistical model was not important for the calculation of relative titers, however, cloglog regression out-performed alternatives where absolute titers are of interest. Finally, we estimated that only 0.7% of assays would falsely detect a four-fold difference in titers between acute and convalescent sera where no true difference exists.
Estimating and reporting assay uncertainty will aid the interpretation of individual titers. Laboratories should perform a small number of repeat assays to generate their own variability estimates. These could be used to calculate confidence intervals for all reported titers and allow benchmarking of assay performance.
| Plaque Reduction Neutralization Tests (PRNTs) remain the most popular approach to characterize an individual's ability to neutralize dengue viruses and are widely used in both epidemiological studies and vaccine trials. However, the underlying variability in the assay is poorly understood, hindering the interpretation of PRNT titer estimates. Further, there is little standardization of its use across laboratories limiting our ability to compare results across settings. Here we used many repeated experiments on the same serum under identical laboratory conditions to estimate the variance in titer measurements. We also identified both the optimum PRNT evaluation point and statistical model to calculate titers. By providing an estimate of the variability in the assay, laboratories will be able to provide a confidence bound on individual PRNT readings. In addition by providing recommended statistical approaches that could be used across laboratories, our findings will help the standardization of the assay across settings.
| Dengue remains a substantial public health problem in tropical and subtropical regions [1]. All four serotypes of the mosquito-borne virus are capable of producing significant morbidity and death [2]. As part of efforts to monitor and control the disease, public health agencies and vaccine developers use serological methods to perform surveillance and assess vaccine trial outcomes. A standard for characterizing serotype-specific neutralizing dengue antibody levels is the Plaque Reduction Neutralization Test (PRNT) [3]. PRNT readouts are known to vary substantially, even on samples from the same individual, however, the extent of the underlying variability in estimates remains unclear [4]. There are many potential sources of variation including within experiment and between experiment sources. In addition, different laboratories use different cell lines, different viral strains with varying viral passage number, and parametric models to calculate PRNT with the impact of the alternative approaches poorly understood [5]–[7]. Laboratories also use PRNT evaluation points that range between PRNT50 to PRNT90, and may perform varying numbers of serial dilutions [6], [8]–[10]. Understanding and characterizing the variability of the assay may greatly increase the accuracy and quantifiability of the assay, important both in epidemiological and vaccine studies.
After infection by one of the four dengue virus serotypes, individuals develop antibodies against the infecting virus [2]. The PRNT assay is used to measure neutralizing antibodies produced in response to this exposure. When an in vitro monolayer of cells is exposed to the virus without the presence of neutralizing antibodies, the viral particles enter and kill the cells. Where viral particles have spread between neighboring cells, a ‘plaque’ of dead cells is created that can be observed and counted. The presence of neutralizing antibodies from an individual's serum reduces the number of plaques formed by inhibiting the virus. In most cases, for a given concentration of antibodies, the addition of lower dilutions of serum result in fewer plaques formed than higher serum dilutions. PRNT50 is the estimated serum dilution that produces a 50% reduction in the number of plaques formed compared to the number formed on monolayers in the absence of antibody [3]. PRNT50 is believed to give an indication of an individual's ability to neutralize the dengue virus if exposed in vivo and to indicate whether an individual has been exposed in the past.
Absolute titer estimates from a single serum sample are used in vaccine studies as a potential marker of protection following immunization [11], [12]. In addition, epidemiological cohort studies may estimate the risk of serotype-specific disease by absolute titer at baseline [13]. Comparisons between titers from two samples taken from the same individual are also routinely undertaken. For example, four-fold differences in titers from acute and convalescent sera are typically taken to signify seroconversion [14]. Cohort studies also use large serotype-specific titer differences between study visits as evidence of infection, allowing the detection of asymptomatic infections that cannot be identified through symptomatic disease surveillance [15].
An individual's ability to successfully neutralize a strain of dengue may depend on the age of the individual, gender, nutrition, genetic factors as well as the history and time of previous infections by other flaviviruses [2], [16]. In comparing single PRNT estimates between individuals, it is not possible to separate differences due to these host factors from differences due to assay variability. Understanding the variability of the assay instead requires a large number of repeated experiments on the same serum. This necessitates large pools of serum that are rarely available. However, as part of each experiment, laboratories often use high titer and low titer serum controls to ensure consistency of experimental conditions between assays. Control sera lots can come from pooled human sera that are maintained and remain unchanged for several years. In each experiment, PRNTs are calculated for each control serum (as well as the test serum under investigation). Using the plaque counts from the control sera from a large number of assays, we can estimate the variability in the PRNT within identical experiments.
The Armed Forces Research Institute of Medical Sciences (AFRIMS) in Bangkok, Thailand developed the dengue PRNT assay in the 1960s and has been performing it since for surveillance of dengue immunity in the population and supporting vaccine trials and cohort studies [3], [4], [13], [17]. Data for the current study comes from control assays of PRNTs performed at AFRIMS between 2007 and 2013. Briefly, in each assay, a monolayer of continuous Macaca mulatta kidney cells (LLC-MK2) was infected with virus, predetermined to be in the range of 30–50 plaque-forming units in the presence of 4-fold serial dilutions of heat-inactivated serum (range of 1∶10 to 1∶163840). The well for each dilution was 4.5 cm2 in size (12 wells per plate). For each dilution, the number of viral plaques was counted and compared to the number of plaques in a control where no serum was added. Each dilution and control was performed in duplicate. During the study period there were changes to the number and cell lines used to passage the virus, and the number of passages that the virus went through. In addition, the DENV-4 viral strain was changed in 2009 (Table 1). Three technicians conducted over 95% of all assays in the study period.
Two serum pools (a high titer and a low titer pool) were collected and created in 2006 and used throughout the study period. The high titer pool was obtained by pooling residual blood samples from multiple Thai individuals that tested positive for dengue virus using IgG ELISA. A portion of the pool was then diluted with human sera from PRNT-negative blood donors to create a low titer pool.
Five viruses were used during the study period, one each for DENV-1, DENV-2 and DENV-3 and two for DENV-4 (Table 1). Around every two years, viral stocks were generated in batches by passaging virus through C6/36 mosquito cell lines (between one and eight passages) and up to three passages in either suckling mice (SM) or LLC-MK2 cells.
Basic regressions were used to interpolate the titer at which defined reductions (PRNT “evaluation points”) occur from the observed reductions (e.g., a 50% reduction for a PRNT evaluation point of PRNT50). We calculated PRNTs over the range PRNT40 to PRNT90 using either (a) probit regression, (b) logistic regression, (c) complementary log-log (cloglog) regression or (d) four-parameter non-linear regression [8]. To explore the reduction in variance from repeat dilutions, PRNTs were calculated using both individual set of dilutions and by using the average plaque reductions across repeat dilutions.
As PRNTs can be resource intensive, laboratories may perform two dilutions that they expect will contain the PRNT evaluation point of interest and use straight line interpolation on the log-transformed dilutions [6]. To estimate the variability of this approach, we initially identified the expected PRNT titer using all assays from a viral strain and serum pool and identified the two sequential dilutions that contained this value. For each experiment we then only used the values from the two sequential dilutions to calculate PRNT using straight-line interpolation. We did not calculate PRNTs in experiments where the two dilutions did not contain the PRNT evaluation point of interest.
Finally, some laboratories perform Single Dilution Neutralization Tests (SDNTs) to identify exposure using a single dilution [18], [19]. The test is scored as positive if greater than 70% plaque reduction is observed (using one or more reference dengue viruses) at a 1∶30 dilution, although the optimal plaque reduction or dilution remains unclear [18], [19]. To provide insight into the risk of incorrectly categorizing individuals as positive or negative, we calculated the variance in the neutralization proportions for all experiments from each individual dilution for each viral strain and serum pool.
We assumed that the ability of each of the two serum pools to neutralize a particular viral strain was constant within any year, reflected in a single ‘true’ PRNT titer for each virus for both the high titer and the lower titer pools (i.e., one for each row in Table 1). We considered PRNT estimates from a flexible non-parametric spline, fitted to the plaque reductions from all experiments within each year from a single serum pool as the best, unbiased estimate of the ‘true’ PRNT for that pool.
We explored whether there existed any systematic differences (bias) in PRNT estimates calculated using the different models. For each experiment, we calculated PRNT titers using each of the models (probit, logit, cloglog regression and non-linear regression). Bias was suggested when there was a systematic difference between the PRNT estimates using the model and the estimate of the ‘true’ titer. In addition we calculated the mean squared error (MSE) in the estimates. We reported an average MSE, bias and variance for each PRNT evaluation point and model, weighted by the number of experiments using each virus and serum pool. We used bootstrapping to generate 95% confidence intervals for the bias, variance and MSE estimates from each PRNT evaluation point and model. Over 1,000 resamples, we recalculated the bias, variance and MSE of the assays. Ninety-five per cent confidence intervals were calculated from the 2.5% and 97.5% quantiles from the resultant distributions. Finally, to explore whether time difference between assays was associated with observed variability, we estimated the variance in titers for assays performed across different time periods.
We can divide variability in titers into ‘plate-specific’ and ‘non-plate- specific’ sources of variance, where ‘plate-specific’ is taken to mean experimental factors that are identical in assays performed on the same plate. In particular this will include the preparation of diluted virus solution that is added to wells on the same plate (a new viral solution is made for each separate plate). Non-plate-specific factors are all other sources of variability that will be present across all assays, irrespective of whether they are conducted on the same plate or not. This will include variability in the ability of the serum to neutralize the virus and variability in plaque counting. We estimated non-plate-specific variability by calculating the variance in titer estimates across assays performed on the same plate. Plate-specific variance was calculated by subtracting the non-plate-specific variance from the total variance in titers from assays performed on difference plates.
Absolute antibody titers may be used as a marker of immunity. More common, however, is the comparison of titers from convalescent and acute sera taken from the same individual for detection of seroconversion. Increasing the number of repeat dilutions for each serum will reduce the uncertainty of both absolute and relative titer estimates. Using our estimates of the variability in the assay, we calculated the expected variance from performing different numbers of repeat dilutions (varied between 0 and 3 repeats). We considered scenarios where repeated dilutions were conducted on the same plate (and would therefore not reduce plate-specific sources of variability) and where repeats were conducted on different plates where both plate-specific and non-plate-specific variance would decrease. In addition we used our variance estimates to calculate the proportion of paired samples that would result in a greater than four-fold difference in titers where no true difference exists (i.e., a false positive result).
Plaque density may be associated with differential levels of plaque overlap, which would bias titer estimates. We can use the number of plaques in the reference well (where no sera is added) as a marker of plaque density. Heterogeneities in the passaging of the virus may also be associated with changing PRNT estimates. To quantify systematic differences in titer estimates by the number of passages and the type of cell (C6/36, SM and LLC-MK2 cells), plaque density and the age of the virus stock, we constructed a multilevel model with a random intercept for each viral strain and serum pool combination (listed in Table 1).
All experiments were conducted using pooled residual sera from public health service testing and, as per Walter Read Armed Institute of Research (WRAIR) policy, did not require ethics review. WRAIR is the parent organization of AFRIMS.
Detailed methods can be found in the supplementary materials.
Between 2007 and 2013, a total of 2,319 assays were performed using five different viruses on two different control sera (Table 1). An average of 4.4 dilution steps were performed in each assay (range: 3–5) with each dilution performed twice (20,286 individual dilutions in all).
There existed substantial variability in the plaque reduction proportions (Figure 1) with consistent variability in plaque reductions for each dilution (Figure 2). On average, the variance for an individual dilution performed on the same serum using the same virus was 0.016, equivalent to a 95% confidence interval of 0.45–0.95 for an observed plaque reduction of 0.7. Performing a repeated set of dilutions reduced the variance to 0.011, equivalent to a 95% confidence interval of 0.50–0.90 for the same observed plaque reduction of 0.7.
The variability in plaque reduction proportions led to heterogeneity in PRNT titer estimates. For each set of dilutions performed on each serum and using each virus, we used four different statistical models to calculate titers at PRNT evaluation points varying from PRNT40 to PRNT90. In each case we calculated the bias, variance and mean squared error in the estimated titers. Four-parameter non-linear regression could only be used on 63% of the assays, as the remaining assays had either insufficient dilutions or the resultant curve fluctuated too much to allow model fit (Table 2). All assays could be used for the other statistical models. When comparing the models, we only used the assays where estimates existed for all four approaches. We found that the probit and logit models consistently over-estimated titers (Figure 3A). For example, for a PRNT evaluation point of PRNT50, the probit model overestimated the titer by an average of 0.14 (log10 scale) and logit by 0.12 (log10 scale). The cloglog and four-parameter non-linear regression were largely unbiased. The variance in titer estimates was very similar across models (Figure 3B). However, the variance varied widely by PRNT evaluation point. Variance was lowest at PRNT75 for the probit, logit and four-parameter non-linear regression and lowest at PRNT80 for the cloglog model. Overall, the lowest mean-squared error existed from using the cloglog model at a PRNT evaluation point of PRNT75 (Figure 3C).
Where only two dilutions were used, only 50% of the experiments could be used as the two sequential dilutions did not contain the PRNT evaluation point in the remainder and would have required extrapolation. Where it could be estimated, the standard deviation of PRNT50 using two dilutions was estimated at 0.13 (log10 scale), however, this only represents the variability of the subset of the experiments where the two dilutions had reductions in plaques that were closest to the best estimate of the unbiased PRNT and cannot be compared to the variability observed in the other approaches.
To calculate the variance in non-plate-specific factors, we calculated the variance in titers calculated from single sets of dilutions performed on the same plate using a cloglog model and a PRNT evaluation point of PRNT75 (the model and evaluation point with the lowest mean squared error) from all assays. We found a variance of 0.014 (log10 scale) for titer estimates calculated on the same plate compared to a variance of 0.032 (log10 scale) for titer estimates from different plates (all estimated from single sets of dilutions). These findings indicate that the variance from non-plate-specific factors was 0.014 (log10 scale) and the variance from plate-specific factors was 0.018 (the difference in the two variance estimates, log10 scale). We found no difference in the variance in titers between assays by the time separation between them: titers from assays performed within the same month were as variable as from assays performed over a year apart (Figure S1). When a single titer estimate was calculated from two sets of dilutions from the same plate (i.e., the common practice of performing a repeat set of dilutions on the same plate), the variance in titer estimates was 0.025 (log10 scale). This is identical to what we would expect from a reduction in non-plate-specific variance only (i.e., 0.018+0.014/2). Had repeats been performed on different plates, we estimate that the variance would have reduced further to 0.016 (i.e., 0.018/2+0.014/2, log10 scale) reflecting a reduction in both non-plate-specific and plate-specific variance.
Laboratories may wish to calculate their own lab-specific variability measures by performing repeated assays on the same serum. To estimate the number of assays laboratories would need to perform to get a reliable variance estimate, we estimated the precision in the variance estimate using different numbers of assays (between 2 and 30) on the same serum using the same viral strain. We found that with only 20 assays, the width of the 95% confidence interval for the variance would be only 0.013, only slightly higher than the width calculated from all samples (Figure S2).
We found that the titers estimated for each of the ten virus- serum pool combinations varied similarly (Figure 4B). Therefore a single variance estimate appears appropriate when calculating confidence intervals for a titer. We found that, for example, for a titer estimate of 1∶500 where no repeat dilutions have been performed and the titer was estimated using a cloglog model with a PRNT evaluation point of PRNT75, a 95% confidence interval would be 1∶200–1∶1100 (2.70±0.35 on a log10 scale). Performing a repeated set of dilutions on the same plate (the most common practice) results in a confidence interval of 1∶240–1∶1000 (2.70±0.31 on a log10 scale). Finally, performing repeated dilutions on a separate plate reduces the confidence intervals further to 1∶280–1∶880 (2.70±0.25 on a log10 scale). Note that performing a single repeat on a different plate reduces the width of the confidence intervals substantially greater than performing even additional repeats on the same plate (Figure 4A). We tested the coverage of our uncertainty estimates by calculating individual confidence intervals for each assay (2,319 confidence intervals in all). We found that 96% of these intervals contained the true titer estimates, suggesting excellent coverage.
Researchers are often interested in the ratio of titers between acute and convalescent serum samples taken from the same sick individual. We found that while the probit and logit models produced substantially biased results, the extent of the bias did not appear to differ by the magnitude of the titer (Figure S3). Therefore, while individual PRNT estimates may be biased using these models, ratios of PRNTs would not be as both the numerator and the denominator would be similarly biased. As the different statistical models had similar levels of variance (Figure 3B), the choice of model when calculating relative titers is less important. The PRNT evaluation point, however, is crucial to minimizing variability, with a PRNT75 evaluation point up to 50% less variable than alternatives (Figure 3B). Four-fold differences or greater in titers are often used as evidence of seroconversion. Using our variability estimates, we calculated the probability of detecting a greater than four-fold difference in titers where there was no difference in true titers (i.e., a false positive result). We found that when a single set of dilutions is performed and a PRNT evaluation point of PRNT75 is used, only 1.7% of assays would falsely detect a greater than four-fold difference in titers compared to 6.4% with a PRNT evaluation point of PRNT50 (Figure 5A). Performing a single set of repeat dilutions on the same plate (but with the acute and convalescent samples still performed on separate plates) reduced the probability further, to 0.7%. Finally, if both the acute and convalescent samples are on the same plate, the probability of falsely detecting a greater than four-fold difference in titer was 0.03%. Figure 5B sets out the 95% confidence intervals for different observed ratios: for example the 95% ratio for an observed two-fold difference in titers using a PRNT evaluation point of PRNT75 is 0.7–5.5 (equivalent to 0.3±0.4 on a log10 scale).
To estimate the effects of experimental conditions on titers we built a multilevel model incorporating the number of viral passages, the cell type, reference well plaque count and the age of the virus stock used in the experiments. We found that passaging the virus in SM increased titers compared to LLC-MK2 cells (effect size of 1.15, 95% confidence interval of 1.09–1.21). The total number of passages and the age of the viral stock at the time of the experiment did not affect the titers. Higher plaque counts in the reference well was associated with a very small reduction in titers (effect size of 0.99, 95% confidence interval of 0.98–99). Less than 0.7% of the variability in PRNT50 estimates could be explained by the model covariates, leaving over 99% of variability unexplained (Table 3).
Using repeated assays on the same serum sample with the same viral strain, we estimated the extent to which measured PRNTs vary. We found a consistent level of variability in titer estimates across the viruses and serum pools used during the study period. A measure of the variability of titers provides information on the potential misclassification of individuals falling above or below any specified PRNT evaluation point, information routinely used in calculating sample sizes for a wide range of studies. By characterizing the variability in measured titers, these findings will aid in the determination of an individual's immunity, the design and interpretation of results from immunogenicity trials, epidemiologic studies and allow the benchmarking of assays across laboratories. Based on our findings, we set out a number of recommendations for laboratories performing PRNTs (Box 1).
There have been a number of efforts to standardize the assay [7], [20]–[24]. A comprehensive effort by Roehrig et al., set out guidelines for PRNTs, which were supported by the World Health Organization [20], [21]. Nevertheless, heterogeneities in approaches between laboratories persist, particularly in PRNT evaluation points, in large part because there had been no comprehensive study of the variability in titers. The WHO recommends using a PRNT50 titer for vaccinee sera and PRNT90 titers for epidemiological studies. However, both vaccine and epidemiological studies regularly use alternative PRNT evaluation points [6], [10]. The stated benefit of the higher evaluation point is to decrease variance while the stated benefit of the lower evaluation point is of increased accuracy (i.e. reduced bias) [21]. We found that the evaluation point with the lowest variance was actually between PRNT75 and PRNT80 with little difference between the different models used to calculate the titers. Bias, by contrast, differed substantially by model with four-parameter non-linear regression and cloglog largely unbiased across PRNT evaluation points whereas probit and logit regression consistently over-estimated titers. However, the choice of model only appeared important where absolute titers were of interest as biases cancelled out in the calculation of relative measures. Overall, we found that a PRNT evaluation point of PRNT75 minimized the MSE between the model PRNT estimates and our best estimate of the unbiased PRNTs and should be used, where possible, across study types. Lower PRNT evaluation points may remain preferable in the estimation of low titers where the estimate at PRNT75 can be regularly below the limit of detection (typically 1∶10). However, the biological meaningfulness of low PRNT titers is currently under scrutiny. A recent dengue vaccine trial observed infections in vaccinated individuals despite the apparent presence of detectable titers [25]. These findings suggest that ‘some’ titers (e.g., greater than 1∶10) versus ‘no’ titers (e.g., less than 1∶10) is insufficient to differentiate between individuals with and without protective immunity. Further research is urgently required to understand if vaccines need to elicit higher titers than those generated by the trial vaccine or if there exist qualitatively different markers of protection (such as T-cell responses).
We found that around half of the variance in titers could be explained through plate-specific factors, experimental conditions that differ between plates but not within plates. In particular, creating the viral preparations for each plate may be an important contributor. Performing repeat dilutions on the same plate cannot reduce plate-specific variance, as both repeats will be perfectly correlated for these factors. This explains why there was only a small reduction in variance in titers calculated from repeat dilutions compared to the variance from only single sets of dilutions (0.025 versus 0.032). Performing repeats on separate plates would reduce this further (we estimate to 0.016) and should be considered where precise absolute titers are required. Where relative titers are calculated, we estimated that the probability of observing four-fold differences in titers where it does not truly exist is less than one per cent when a PRNT evaluation point of PRNT75 is used (it is twice as high when a PRNT50 evaluation point is used). Our findings indicate that the risk of a false positive detection of a significant difference in titers is low even where a cut-off of a three-fold difference in titers is used, especially where both the acute and convalescent sera are placed on the same plate. It remains unclear whether such titer differences are correlated with immune protection or past exposure.
Laboratories may perform only two dilutions and use linear interpolation to obtain PRNT estimates. We found that we could only use half of the assays for this analysis, as the remaining experiments would require unwise extrapolation outside the results from the two dilutions. In these situations, laboratories need to repeat the assays at wider dilution ranges. The substantial number of experiments that could not be included in the analysis suggests that performing only two dilutions may only have minimal benefits. Single dilution neutralization tests only require a single dilution, however, individual plaque reduction estimates had wide confidence intervals: individuals with a true plaque reduction of 75% (and therefore should be scored as ‘positive’ in a SDNT using an evaluation point of 70%) had 95% confidence intervals of 54%–96%, suggesting that many such individuals would be wrongly characterized as negative [18], [19]. This provides no insight into the sensitivity or specificity of detecting exposure from using an evaluation point of 70%.
The DENV-4 strain used in the assays was changed in 2009 resulting in a 11.4-fold increase in mean titers in the high serum pool and a 12.1-fold increase in the low serum pool, confirming previous findings of the importance of viral strain on titer estimates [4]. The two DENV-4 strains come from two different genotypes (the earlier strain was genotype 2 whereas the later one was genotype 1). These findings highlight the possibility of vastly different immunological response even within a single serotype. Alongside the effect of viral strain, it has been suggested that the number and cell type of viral passages could produce systematic differences in PRNT estimates [4], [7]. We found a small increase in titers in experiments using viruses passaged through SM compared to LLC-MK2 cells supporting similar previous findings [4]. The total number of viral passages did not appear to impact PRNT estimates, however, only small numbers of passages were conducted (maximum of eight). Increasing this substantially or only using mammalian cell lines such as Vero cells as recommended by the WHO, may nevertheless impact estimates. The presence of many overlapping plaques in a well may lead to under-estimates in true plaque counts. However, we found only negligible difference in titer estimates by the number of plaques in the reference well, suggesting plaque overlap did not affect our results. These findings suggest that the under-estimate in true plaque counts was consistent across dilutions; alternatively, the wells were sufficiently large and the plaque counts sufficiently small to avoid substantial overlap. Laboratories using smaller wells may nevertheless experience titer differences from differential levels of plaque overlap by dilution. Overall, aside from viral strain, experimental factors varied in our assays explained less than one per cent of the observed variability in titer estimates. Experimental factors that were held constant throughout our experiments, such as incubation time and plaquing cell line, may nevertheless impact titer estimates.
Our findings show that the assay is inherently variable. There are many potential sources of variability in each experiment: (a) the number of viral particles pipetted into each plate, (b) the extent of viral–antibody interaction (c) the spatial arrangement of cells in the monolayer and (d) the number of non-overlapping plaques successfully generated and counted. While technicians can minimize differences through effective mixing and careful dilutions, there may be a limit to the extent that variability in these factors can be reduced. The use of automated counting methods that allow faster and more accurate particle counting may help [26]. A related approach, the flow reduction neutralization test that relies on immunofocus rather than cell death, may produce less variable titer estimates [27]. In addition, flow-based methods in laboratories with access to flow cytometry equipment show some encouraging results, especially as these methods can use human cells and allow for high-throughput of samples [26], [28]–[30]. Further work is needed to quantify the variability of these alternative approaches.
The serum pools come from pooled human sera that contain a wide range of antibodies not representative of a single individual's serum. Nevertheless, the ability for the pooled serum to neutralize a single virus should remain constant. We were not able to explore the biological significance of individual titers. In particular, the significance of low titers for immune status remains unclear, as does the serotype-specificity of the assay. PRNTs are used to characterize infection parity, with high titers against two or more serotypes considered suggestive of secondary infection. Future research using sera of known infection status could shed light on the specificity of such classifications. Further, serum with neutralization titers outside the range used in this study may perform differently. The range of titers in this study was wide (PRNT75 range of 1∶20–1∶6000) and we observed a consistent pattern in variability across this range. Nevertheless, naturally occurring low titer antibodies (rather than the diluted high titer pools used here) may have different levels of avidity and affinity that could impact titer variability. All viruses were passages through C6/36 cells. Viruses solely passaged through mammalian cells may be differently neutralized. We only used LLC-MK2 monolayers as plaquing cells. Other laboratories use different cell lines (such as Vero cells, recommended by the WHO or BHK cells), which may behave differently. However, it is unlikely that the variability in titers would be markedly different. Laboratories that use markedly different protocols may identify different optimal PRNT evaluation points. These could be identified using the methods presented here on repeated assays on the same sera.
In conclusion, providing uncertainty estimates with both absolute and relative titer estimates would greatly aid the interpretation of individual read-outs. While the estimates provided here provide a first marker of the variance in the assay, heterogeneities in variability between laboratories will exist. By performing a small number of repeat assays (20 appears to be sufficient to obtain a precise variance estimate) on the same serum with the same virus on different plates, laboratories could generate lab-specific variability estimates without requiring excessive resources. Alternatively, where assays on identical control serum are performed as routine, the variance in titers from these assays could be calculated instead. Variance estimates could then be used to calculate confidence intervals for all reported titers and allow benchmarking of assay performance. This study demonstrates the utility of raw results. Laboratories should consider reporting plaque counts alongside titer estimates. This will allow investigators to easily compute alternative titers using different PRNT evaluation points or statistical models, facilitating comparison across laboratories. We also recommend that titers be reported on a logarithmic scale (or log differences for relative titers) to allow easy calculation and interpretation of confidence intervals.
|
10.1371/journal.pbio.2002842 | Deconstructing the principles of ductal network formation in the pancreas | The mammalian pancreas is a branched organ that does not exhibit stereotypic branching patterns, similarly to most other glands. Inside branches, it contains a network of ducts that undergo a transition from unconnected microlumen to a mesh of interconnected ducts and finally to a treelike structure. This ductal remodeling is poorly understood, both on a microscopic and macroscopic level. In this article, we quantify the network properties at different developmental stages. We find that the pancreatic network exhibits stereotypic traits at each stage and that the network properties change with time toward the most economical and optimized delivery of exocrine products into the duodenum. Using in silico modeling, we show how steps of pancreatic network development can be deconstructed into two simple rules likely to be conserved for many other glands. The early stage of the network is explained by noisy, redundant duct connection as new microlumens form. The later transition is attributed to pruning of the network based on the flux of fluid running through the pancreatic network into the duodenum.
| In the pancreas of mammals, digestive enzymes are transported from their production site in acini (clusters of cells that secrete the enzymes) to the intestine via a network of ducts. During organ development in fetuses, the ducts initially form by the coordinated polarization of cells to form small holes, which will connect and fuse, to constitute a meshwork. This hyperconnected network further develops into a treelike structure by the time of birth. In this article, we use methods originally developed to analyze road, rail, web, or river networks to quantify the network properties at different developmental stages. We find that the pancreatic network properties are similar between individuals at specific time points but eventually change to achieve the most economical and optimized structure to deliver pancreatic juice into the duodenum. Using in silico modeling, we show how the stages of pancreatic network development follow two simple rules, which are likely to be conserved for the development of other glands. The early stage of the network is explained by noisy, redundant duct connection as new small ductal holes form. Later on, the secretion of fluid that runs through the pancreatic network into the duodenum leads to the widening of ducts with the greatest flow, while nonnecessary ducts are eliminated, akin to how river beds are formed.
| Branching is a phenomenon that appears everywhere in life. Biological examples include tree leaves and branches [1], the arterial and venous systems [2], the liver [3], lung [4,5], kidney, and several glands such as the pancreas, the mammary [6], salivary [7], lacrimal [8], prostate [9], and meibomian glands [10]. Work carried out on independent organs suggests that several branched organs share principles, such as the importance of mesenchymal signals, and even molecules, such as a frequent use of fibroblast growth factor (FGF) sources [4,5]. However, important differences in morphogenesis also exist across organs. For example, branching is more stereotypic in the lung than in glands, in which the ductal tree differs between individuals. Though emphasis has been put on the outer shape of the branching epithelium, experiments in the pancreas [11,12] and salivary glands [13] suggest that the branching process may be driven from inside the gland when lumen form and connect into tubes.
The mature pancreas is a branched organ in which branches are formed of monolayers of cells assembled into tubes that are connected to form a treelike structure with exit into the duodenum. The pancreas is composed of three main components: acinar, ductal, and endocrine cells. Acinar cells at the terminal ends of the ductal tree secrete digestive enzymes into the ducts, which deliver them into the duodenum [14]. Ductal cells secrete water, the bicarbonate that neutralizes acidic gastric juices, and mucus that protects the ducts [15,16]. The basic pH of ductal secretions contributes to keeping the digestive enzymes inactive until they reach the duodenum. The endocrine cells reside in islets of Langerhans embedded near the ductal system and regulate glucose homeostasis by secretion into the bloodstream. In recent years, our understanding of pancreatic development has grown with increasing speed, driven by advances in both image acquisition and methods for tracking and categorizing cells.
Around embryonic day (E) 9.5 of mouse development, cells in the foregut endoderm change from a cuboidal into a columnar shape, forming the dorsal and mouse pancreatic buds [17,18]. The pancreatic progenitor cells then proliferate and form a stratified epithelium, with most cells losing their apical domain and connection to the duodenal lumen. Shortly after, clusters of cells begin forming microlumen. At E10.75, several of these microlumen exist, some of which are connected by polarized canals [12]. Isolated lumen continue to emerge and subsequently connect to this burgeoning network. This results in a plexus of interconnected ducts at E12.5 [11,12,17].
From E12.5, the now densely packed epithelium starts remodeling its ductal structure. This, combined with general expansion of the epithelium and associated ducts, results in fingerlike protrusions of plexus into the surrounding mesenchyme [12,17]. Running in parallel with this remodeling, the cells begin to segregate into domains with distinctive “tip” and “trunk” cell identities [19]. Tip domains contain cells that are progressively restricted to an acinar fate, while the trunk domain contains the endocrine/duct bipotent progenitors. Starting from E13.5, the tip cells become committed to the acinar fate and start a massive wave of proliferation, rapidly increasing the amount of acinar ends in the network [19]. At around E18.5, the network of the pancreas is more “arborized,” forming a ramified ductal network [20]. In the adult, the ducts are categorized in a rough hierarchical order of duct thickness, with the smallest intercalated ducts close to acini, intralobular ducts, and ending with the interlobular ducts separating the lobes of the pancreas [14].
We still know little about how the microlumen coalesce into a plexus. Kesavan and colleagues showed that epithelial polarity is needed for microlumen formation and maintenance of the ductal network [11]. However, a detailed description of how the microlumens form the ductal network is missing. Even less understood are the mechanisms for remodeling the ductal network from a plexus to a ramified treelike organ. Villasenor and colleagues provided invaluable insight into this process with a detailed anatomical description of the remodeling process [12]. They unveiled that the pancreas might be more stereotypic than previously believed and could identify trends in the patterns of branches. However, further understanding has been limited by a lack of quantitative measures.
In this work, we digitize the ductal network at 3 distinct phases in the development of its ductal tree. We chose E12.5 to represent the early stage with an almost completely interlinked plexus. E14.5 represents the intermediary network in which distinct cell types are appearing and plexus remodeling starts to be visible. Finally, E18.5 represents the (almost) fully mature network. We show that the pancreatic ductal system has stereotypic traits at each stage and that its remodeling can be quantified by standard network measures.
Using in silico modeling, we deconstruct the main steps of network development into a set of simple rules. We show that the creation of the early network from E10.5 to E12.5 can be explained by noisy, redundant duct formation as new microlumens form. We show that a little noise in both the amount of microlumen connections and where it connects is needed to reproduce a network similar to the E12.5. We subsequently show that the later transition from E14.5 to E18.5 can be reproduced by pruning the network based on a flux of fluid running through the pancreatic network into the duodenum. Taken together, we show that pancreatic ductal development can be conceptualized into simple, rule-based modeling that is likely of relevance to several other glands including secreting fluid.
To assess and quantify the development of the pancreatic ductal network, we first digitized it. The ducts were visualized by whole-mount immunostaining of mucin1 and e-cadherin, which respectively highlight the apical side of cells forming the ductal structure and the membranes of the cells lining the ducts. Although automatic segmentation succeeded in digitizing a large part of the network, it failed on fine ducts. Several of the ductal parameters rely on mapping all duct types. The networks were therefore manually skeletonized at different time points by mapping the terminal ends and intersections of the ducts. Thereby, we defined nodes and linked them by edges following the ductal system (S1 Fig). The method is labor intensive, and we therefore largely focused on the ventral pancreas, as it is almost entirely planar at its later developmental stages and is therefore easier to map manually. Some comparisons with the dorsal pancreas were done at E14.5 to test whether the network was different. The resulting networks (Fig 1) closely resemble the pancreatic structure reported in previous studies [11,12,17,20]. The network appears as a plexus at E12.5 and E14.5 and ends as a more treelike structure at E18.5.
The digitization of the pancreatic duct system enables us to derive its network properties based on nodes and links that form the network (Fig 2). A node’s degree k is the amount of connections it has to other nodes. The polygonal features formed by the nodes and their links are derived by counting network limit cycles and give a measure of how interlinked the network is [21]. The average clustering coefficient 〈C〉 (<…> denotes average) relates to the number of triangles found in the network [22]. In addition, we consider the cost of the network in terms of redundancy of paths from one point to another. This was quantified by comparing the network in question to its (euclidian) minimum spanning tree (MST) obtained by connecting the nodes in the spatial location with links that minimize the total length of all links in the network [23,24]. A similar comparison quantified network performance, defined by the transport distance along the network between all pairs of nodes, normalized by the MST.
The E12.5 and E14.5 pancreas networks contain many polygonal features (short loops), have a cost above 1, and have a performance below 1, all features of an overconnected network [24]. At these stages, there is more than one path from a point to the duodenal exit. Both the E12.5 and E14.5 networks have a dimension close to 2 (Fig 2, S2 Fig), highlighting that the planarity that is visible at E18.5 is also present in the network from early developmental stages. The dorsal pancreas, mapped at E14.5, is also an overconnected network. The dorsal pancreas has more redundant ducts than the ventral pancreas at this stage, evident from its number of polygons, average degree, and clustering coefficient. The dimension of the network remains close to 2, suggesting that the dorsal pancreas is also planar. The ventral E14.5 network has a slightly lower cost, worse performance, and slightly fewer polygons than the E12.5 after normalization to network size, which suggests that some redundant ducts may start to be eliminated. The E18.5 network has almost no polygons and a cost of 1.2. The ventral network was mapped on about half of the ventral pancreas at E18.5, always on the same area. A single E18.5 ventral pancreas has been fully mapped to see the shift in network properties from the partially mapped E18.5 (S3 Fig). Nearly all features analyzed were similar, with the exception of performance. This finding lends credibility to the use of partially mapped E18.5 ventral pancreata, with the caveat that the true performance of these networks might be even better, closer to 0.4, based on the whole ventral part analysis. This mature network contains almost no redundant connections. The network undergoes drastic changes as the pancreas matures (E12.5 → E18.5), with a loss of almost every polygonal structure and an order of magnitude shift in clustering coefficient (Figs 2 and 3).
The size of the network increases as seen in the total length Ltot of the ductal system (bearing in mind that only half of the E18.5 networks were mapped). The average distance to the network exit 〈Droot〉 increases with development, meaning that the network expands away from the exit node. The average distances between all nodes 〈D〉 increases even more than to the common exit node, reflecting that distances in the early development are reduced by redundant connections. The system loses its redundant connections and thereby has a decrease in cost and a shift in network dimension from 2 to 1.7. This transformation agrees with the visual evidence that the pancreas prunes its structure to a treelike structure [12]. The changes from E12.5 to E14.5 are less pronounced, although there is a significant loss of local loops (polygons) and a slight decrease in average degree 〈k〉. However, the absolute number of polygonal features of the E14.5 network is higher than at E12.5. This suggests that either pruning of links takes place alongside growth at the E14.5 stage or that the newly formed network forms without loops. The existence of loops at the network periphery suggests that the newly grown network also forms with redundant ducts while pruning has been initiated (S4 Fig). None of the digitized networks resembled their randomized counterparts, suggesting that there are rules behind their formation (S5 Fig).
We visualize some of these network transformations in (Fig 3). As the pancreas matures to E18.5, the network gains more nodes with degree 1 (terminal end) compared to earlier stages. Even at the mature stage, the network retains a few nodes with degree 5 and higher. Fig 3B also visualizes the normalized distribution of polygonal features in the network as the pancreas matures. Each developmental stage has a defining amount of such loops. It is therefore possible to separate each pancreatic stage based on its polygons. Another and more global network characterization is the average distance to the exit normalized by the average distance between every node 〈Droot〉/〈D〉 (Fig 3C). This property decreases substantially as the pancreas matures from stage E14.5 to E18.5. Strikingly, the average distance to the exit of the pancreas approaches the average distance between nodes. This again demonstrates the development toward a branched treelike structure, with the exit node as the common center. Taken together, our analysis shows that although the branches of the pancreas are different between individual mice [12], the pancreatic network structure is stereotypic at each stage and changes systematically as the pancreas matures.
In order to deconstruct the formation and maturation of the ductal network, we condensed our observations into a few principles and implemented these in silico (Fig 4). First, we describe the formation of the network in terms of a growth model, starting with a single lumen node (LN). LNs are subsequently added at a random position within rCM + 0.5 of the existing LNs’ center of mass, where rCM is the distance from the LNs’ center of mass to the most distant point in the network. These newly added nodes are then connected to the already generated network. Each new node forms M + 1 links with the existing network, where M is drawn from a Poisson distribution with mean λ. The assigned links are then attached to the neighboring LNs drawn from a pool of the M + 1 + Δ closest LNs, where Δ is parametrizing the spatial specificity for local links formation. Fig 4C shows the best fit together with the in vivo data for the E12.5 networks. In the supporting information, we show that an expanded parameter analysis of the in silico network approximately recapitulates the observed properties of the E12.5 pancreas network, provided that Δ is about 1. The cost measures are the limiting factor for larger Δ values (S6 Fig). Further, it seems that λ = 0.25 yields the best fit, which means that simulations best fit the biological data when new nodes only connect to slightly more than 1 node. A higher λ results in too many triangles and a too high cost measure of the network (S6 and S7 Figs), whereas a lower λ results in too few loops of any length. Therefore, in order to achieve the observed network, some noise is needed in the number of formed connections between new nodes and where they connect. We notice that the best fit does not completely pass every statistical test for all measures. These discrepancies might be explained by the lack of a growth direction in the model, which results in a roughly spherical structure of the simulated 3D network. Some secretory organs, such as the kidney, exhibit an initial phase of stereotyped duct formation followed by less stereotyped processes [25,26]. Though this has not been observed in the early pancreas [12], this may have been overlooked. We thus attempted to initialize the network from small initial binary/dichotomic patterns (L-system) of 4, 20, and 100 out of the nodes eventually generated. The decrease in polygons becomes significant only when the initial L-system makes up a large fraction of the whole network (100 out of about 320) (S8 Fig).
The ductal network of the E12.5 pancreas has an inhomogeneous duct thickness that does not indicate a hierarchy between the main duct and its terminal ends [12]. In contrast, the ductal network at E14.5 exhibits some hierarchy of duct diameter, with wider ducts close to the exit, making it possible to identify a difference between terminal ends and the main duct [12]. This hierarchy is even more distinct at the E18.5 stage, in which the main duct and the acini ends are easily identified [12]. This suggests that duct diameter relates to fluid running through the developing pancreas. We hypothesized that fluid running through the network may contribute to other time evolutions of the ductal network such as the pruning of redundant ducts. We postulated that, as the pancreas matures, redundant ducts would compete, and the smallest of competing ducts may be remodeled or merged. In order to assess this flux-based mechanism for pruning of the network, we constructed a simple in silico model (S9A Fig). The model uses the already digitized E14.5 pancreatic networks as its input. Fluid is added at the terminal ends of the network continually and is drained at the exit to the duodenum. This assumes that the acinar cells are the main contributors of fluid in the developing pancreas, but this assumption can be relaxed by simulating ductal secretion on every node, without significant change of our results (see Fig 5, S10 Fig). There is in fact evidence that ductal cells also secrete fluid during development [27]. The network is then allowed to reach steady state. This simulation provides a measure of the flux through all the ductal links (Figs 5A and S10A). We see that the high flux links lie in the inner pancreas structure, and the highest fluxes are found close to the exit of the pancreas (S11 Fig). This would impose a higher hydrostatic pressure and flow, which promotes widening of the ducts, in agreement with the observation that the largest ducts of the pancreas are closest to converging points and to the exit [12]. Similar phenomena have been reported during the transition from networks to trees in blood vessels [28–31]. According to Poiseuille’s law, a high flux of fluid is expected to result in a high internal pressure, which is lessened by increasing duct width, equilibrating at steady state to the same pressure for every duct (see Materials and methods). In agreement with this hypothesis, we show that the drained basin (approximated to the total number of total nodes or duct length upstream of a given duct) is proportional to the cube of duct diameter (S11 Fig). This is exactly what is to be expected from laminar flow in a duct system with draining (see Materials and methods).
Our assumption is further that ducts that carry low or no flow tend to be eliminated. The second part of the model therefore consists of pruning the network, removing redundant links with the lowest flux. A link is considered redundant if its removal does not fragment the network (S9A Fig). This process is repeated, removing subsequent redundant links until the network is a perfect treelike structure (Fig 5). We observe that flux-based pruning results in the network being pruned from its periphery and progressively toward the center/outlet (Fig 5B, S10B Fig). This result agrees with the experimentally observed network structure [32].
We further compare the flux-pruned network with a network that instead is pruned by randomly removing redundant links regardless of their flux (Fig 5, S10B Fig). Fig 5C quantifies that the flux-pruned network matches the mature pancreas significantly better than the random-pruned network. The random-pruned network does not correctly prioritize the exit to the duodenum (Fig 5C), and as a result, it will be less efficient at transporting fluid to the exit.
We previously established an in vitro 3-dimensional (3D) culture model in which small clusters of pancreatic progenitors proliferate, differentiate, and self-organize to form a pancreas-like structure [33]. The forming organoids branch and form ducts after 7 days of culture [33,34]. By whole-mount immunocytochemistry detection of the apical marker mucin, we observe that these organoids form a network of pancreatic ducts. This network is isotropic and lacks an exit point. Although the organoids are extracted from E10.5 pancreas and kept for 7 days in culture, they do not recapitulate the features of a network that resemble the one seen in the pancreas at E18.5. They do establish lumen and connect them into a network, but the network in the organoid exhibits a large number of loops, high cost, low performance, and high average connectivity (Figs 1 and 2). Its high cost and low performance suggest a level of connectivity that is even higher than the E12.5 pancreas. A network devoid of exit point is not exposed to a net flux, even if it is actively secreting, which is suggested by the observation that with time, the lumens widen to eventually form cysts at day 10 of culture [33]. Taken together, this suggests that in the absence of flux, pruning does not occur.
Pancreatic secretion and flux have not been studied in the embryo. The fluid secreted by the pancreas in the adult originates from ductal cells, for the most part, and from acinar cells [16,35]. Many of the channels and transporters involved in fluid secretion have been uncovered. Using transcriptional profiling, we found that many were already expressed at E10.5, E12.5, and 14.5 (S12 Fig). Several important players were up-regulated between E10.5 and E12.5 at the time of microlumen formation—notably, the apical chloride channel cystic fibrosis transmembrane conductance regulator (CFTR), which is coupled to the anion exchanger solute carrier family 26 member 6 (Slc26a6) for ductal bicarbonate secretion, as well as the calcium-activated chloride channel anoctamin-1 (Ano1)/transmembrane member 16a (Tmem16a), active also in adult acinar and ductal cells [16,35,36]. These are the main channels secreting ions and osmotically driving the water flow in the adult pancreas and in the salivary gland. The main pancreatic water channel, aquaporin 1, is also strongly up-regulated between E10.5 and E12.5, as well as the secretin receptor, which activates ductal secretion in adult. To test whether the secretory machinery is active, we stimulated CFTR channels using forskolin on E12.5 pancreatic explants (Fig 6A). This led to a dilation of ductal/acinar lumen. We also observed a dilation on spheres produced from E12.5 progenitors (Fig 6B). These spheres are mostly formed of ductal cells. These experiments reveal that the pancreatic cells secrete fluid as early as E12.5 and clarify the molecular machinery used.
Here, we show that mapping the pancreatic network unveils previously unknown features of pancreatic development. This work on a specific organ is likely to be relevant to other glands that also exhibit connecting microlumen, such as the salivary gland and more generally glands that secrete fluid. The organ-forming rules we propose here are different from the more deterministic lung branching program or the random walk proposed in the mammary gland [5,37]. We reveal that although the pancreatic branches are not stereotypic at different orders in the network and between individual mice, the pancreatic network has stereotypic properties at different developmental stages between E12.5 and E18.5. We show that the transition of the pancreatic ductal network from an interlinked mesh to a treelike structure is reflected in changing network properties. Furthermore, we uncover that this transition optimizes the distance from the acinar ends to the duodenum while reducing redundancy in the ductal network. The development of the pancreatic tubular network can be described by redundant lumen connection followed by flux-based pruning. Redundant lumen connection is simulated with a simple growth model in which new lumen appear close to the established pancreas network and immediately connect to the nearest lumen of the network. The model recapitulates the network traits of the E12.5 pancreas, but the molecular mechanisms that drive network connection—that is, the alignment of cells with coordinated apicobasal polarity in continuous tubes—are currently unknown. The network creation model predicts that if new microlumen connect to every node (λ = ∞) when they appear, then the network should form a single large lumen. This is possibly what happens in Pdx1 −/− mouse strains [38]. In this strain, the network has a single lumen at E11.5. At later developmental stages, microlumen do appear and connect to this monolumen, expanding it to a single large lumen [38].
Even a small reduction of the average connectivity λ has a profound effect on the in silico network, in particular a loss of polygonal features, a decrease in cost (fewer redundant links), and lower performance (longer distance between nodes). A reduction in λ may be the phenotype that was observed in the p120 −/− mice [39]. These mice have fewer ducts and larger ducts than their wild-type counterparts. In addition, the p120 −/− pancreas also contains fewer branches, suggesting a correlation between the inner ductal network and the outer branching morphology in the pancreas [39].
The subsequent model for flux-based pruning is a zero-parameter model that assumes that pruning of the ducts is solely based on the minimal assumption that redundant links with low flux are removed first. Thereby, we recapitulate the transition of an E14.5 pancreas network into a treelike network resembling the more mature E18.5 structure. Villasenor and colleagues qualitatively characterized the developing morphology of the pancreatic ductal network and its outer morphology [12]. Our work expands this description by providing quantitative measures of the developing ductal network. With our dataset, it is possible to compare the pancreatic network to other transportation networks such as, for example, rivers [40–43], infrastructure [24], and blood vessels [44]. This comparison suggests some similarities to other networks transporting fluids. As seen in these networks, we observe wider ducts close to the network exit. Our quantitative measurements of duct diameter reveal that the drained basin, which is expected to be proportional to flow, is proportional to the cube of duct diameter in a manner similar to blood vessels.
Though we show that network restructuring is best modeled by the elimination of ducts with the lowest flow when alternative paths exist, it raises questions as to how the ducts are eliminated. Cell death is unlikely, as there is relatively little cell death during pancreas development, and they do not form patterns of aligned cells. It is more likely that cells are rearranged and recycled to widen or elongate contiguous ducts. The recent observation that Afadin and Ras homolog gene family member A (RhoA), acting upstream of microfilament dynamics, are needed for loop elimination is in agreement with this hypothesis [45]. This raises questions as to how cells sense the lowest flow and the fact that the duct is supernumerary. It is unlikely that the cells measure an absolute amount of flow, as ducts with different flows are pruned in the model. A likely scenario is that the ducts with high flow build fluid pressure, leading to a response that widens these ducts, and that this promotes pulling forces and cell redistribution from adjacent ducts. For nonsupernumerary ducts, the initiation of closure would be expected to lead to pressure increase and reopening [46]. A flow increase may be sensed by multiple systems, including multiple ion channels, adhesion and cytoskeletal proteins, cilia, caveolae, and the membrane bilayer itself [47]. While this has been studied mainly in the blood vessels and largely in the adult, similar systems may be used in the pancreas. The elements of the flow that are sensed may be shear stress, circumferential stretch, or, though less likely, variations in the flow.
Among these sensors, only primary cilia have been shown to affect the pancreatic ductal network. Ductal cells have primary cilia on their apical side, while acinar cells do not [48,49]. Bending of cilium caused by fluid movement has been shown to cause an activation of the calcium channels in kidney tubules, affecting a number of downstream targets [50–52]. Disruption of cilia affects the structure of ducts in both the kidney and the pancreas; local ductal enlargements have been reported, but the ductal network has not been studied in whole mount, and we therefore do not know whether the resolution from a plexus to a tree is affected.
Specific pancreatic inactivation of kinesin family member 3a (Kif3a), a molecular motor used to build the cilia during development, led to cyst formation and duct enlargement. However, when cilia formation was inactivated 4 weeks after birth, no phenotype was observed [53]. This suggests that, like in the kidney, primary cilia play an important role in developing the pancreatic duct network but have a lesser role in maintaining it. Furthermore, when Kif3a inactivation was limited to the acinar and endocrine cells, no phenotype was observed [53]. Therefore, only the primary cilia of the ductal cells play an important role in pancreas tubulogenesis during development. In a similar manner, inactivation of Tg737/Polaris [54], a protein that promotes cilia assembly, leads to cyst formation. Several other mutants affect cilia and lead to cystic ducts, including hepatocyte nuclear factor 6 (HNF-6) −/− mice [55], specific pancreatic inactivation of hepatocyte nuclear factor 1 beta (HNF1B) [27], as well as bicaudal C homolog 1 (BICC1) inactivation mutants [56]. In all these mutants, the defects reported are ductal enlargements, and it would be important to know whether other ductal parameters studied here are affected.
The model also predicts that fluid must run though the pancreas from E14.5 or even earlier in order to prune correctly. Secretion in the adult pancreas originates from ductal and acinar cells. We show that the main drivers of ductal secretion—TMEM16, CFTR, and aquaporin 1—are already expressed by E12.5, and duct enlargement upon forskolin exposure suggests ductal secretion. Since spheres are primarily made of bipotent progenitors [33], we surmise that they produce a large part of the fluid. At E12.5, there are no acinar cells in the pancreas yet, but acinar cells may contribute after they differentiate. CFTR is expressed at relatively low levels in the ducts at these stages, and its importance in secretion in mice is less important than in human. It has been proposed that this is due to functional overlap with calcium-activated chloride channels (TMEM16A/ANO1) in mice [57,58]. In human and pigs, CFTR mutations are already symptomatic in the fetal pancreas as early as 19 weeks of pregnancy. Ductal obstruction and enlargement have been reported antenatally in these models, providing a natural circumstance [59] [60] suggesting that secretion starts before birth as well [61]. We did not observe dilation in explant nor spheres with secretin, cholecystokinin (CCK), or carbachol, which respectively activate secretion in adult ducts, acini, and both, even though their receptors are expressed during pregnancy. [17,56]. An interesting mutant lends support to the hypothesis that secretion test runs lead to network remodeling. Indeed, the pancreas-specific inactivation of the transcription factor nuclear receptor subfamily 5 group A member 2 (NR5A2) prevents the formation of acinar cells and intercalated ducts [62]. In this mutant, the ductal network remains hyperconnected at birth, suggesting that acinar or terminal ducts control network pruning, possibly via their secretion.
Our study suggests more elaborate tests of our flux pruning scenario. The first is to obstruct the pancreas exit into the duodenum during normal development, a difficult experiment in vivo in embryos and for which no spontaneous incidence has been reported. In the same spirit, organoids, which do not have an outlet, do not exhibit network optimization. We experimentally tried to use capillaries to create an artificial outlet, but inserting them in ducts and preventing clogging prevented success. The organoid experiment shows that the formation of the ductal meshwork occurs independently of the formation of a blood vessel network, though eventually, blood vessels are enriched along the pancreatic epithelial branches of the tree, and cross-talk is known to exist between the two tissues [63–65].
Our work provides new methods and a new framework to evaluate the formation of ductal networks in organs and sets the stage to evaluate structural defects in mutants to decipher the cellular events underlying network remodeling. Moreover, it paves the way to analyzing similarities and differences in the secretory network of other glands.
The Danish Veterinary Office approved the animal experiments performed under license number 2014-15-2934-01008.
Timed-pregnant ICR mice were killed at E10.5 (for organoids and spheres) or E12.5 (for spheres), and the cells of embryonic pancreases were recovered, dispersed, and seeded in Matrigel as described previously [33,34,66]. Organoids were cultured for 7 days and spheres for 4 days prior to imaging. For explants, E12.5 pancreases were cultured overnight on fibronectin-coated glass-bottom plates, as described previously, prior to live image acquisition [67].
Whole-mount immunostaining was performed for pancreata harvested at E12.5, E14.5 (dorsal pancreata only), and E18.5 and organoids harvested at day 7 in vitro as previously described [26]. Samples were fixed with 4% paraformaldehyde at 4°C and then washed in phosphate-buffered saline (PBS). Samples were dehydrated stepwise in 33%, 66%, and 100% methanol for 15 minutes at each step and stored at −20°C until later use. For autofluorescence quenching, they were incubated with freshly prepared Methanol:DMSO:H2O2 (2:1:3, 15% H2O2) at room temperature (RT) for 12–24 hours. Samples were washed twice in 100% methanol for 30 minutes at RT and brought to –80°C 3–5 times for at least 1 hour each time and back to RT to make antigens in the deeper parts accessible. Samples were rehydrated stepwise in Methanol/TBST (50 mM Tris-Hcl pH 7.4, 150 mM NaCl, 0.1% TritonX-100) 33%, 66%, and 100% for 15 minutes at each step at RT. After blocking overnight at 4°C in CASBLOC blocking solution, samples were incubated with primary antibodies (Ecadherin [1/200] # 610181, BD transduction laboratories and Mucin1 [1/200] #HM-1630-P0, Thermo Fisher) diluted in CAS-Block Histochemical Reagent, # 8120, Thermo Fisher for 24–48 hours. Then, samples were washed overnight in TBST and incubated with secondary (goat anti-A. hamster biotinylated [1/1,000], Invitrogen) and tertiary antibodies (donkey anti-mouse Alexa 488 [1/200] and strepatividin Alexa 647 [1/800] Invitrogen) and for 24–48 hours followed by overnight washing in TBST. Stained samples were dehydrated stepwise in Methanol as previously described and cleared in a solution of 1:2 Benzyle Alcohol and Benzyle Benzoate (BABB) for 12–24 hours prior to imaging.
Cleared samples were subsequently mounted in glass concavity slides and submerged completely in BABB to maintain refractive index matching and sample transparency. Cleared samples were imaged using a Leica SP8 confocal microscope with a 20X/0.75 oil immersion (most samples) or a 10× air (E18.5) objective at 1024 × 1024 resolution. Samples were imaged in an 8-bit format unless otherwise indicated. Images were stitched and reconstructed into 3D images using the Imaris software (Bitplane).
For live imaging of explants and spheres, we used a Zeiss LSM780 confocal microscope with a 10× air objective. Calcein 10 μM was used as background stain and incubated for 30–60 minutes and rinsed before imaging. Independent samples were imaged at least in triplicate for 40 minutes, 1 image every 2 minutes, after addition of the drug tested. We tested 10 μM Forskolin, 1 μM Ionomycin, 10 nM secretin + 10 μM Carbachol.
The inner diameter was automatically segmented and measured for each sphere at each time point using Lumen Thickness Detector script (https://github.com/gopalrk/Lumen_Detector_and_Quantification). The value of the inner diameter at t was first normalized to t0 and then normalized to control over time (n = 2 independent experiments; the total number of analyzed spheres is 68 for forskolin treatment and 59 for control). Statistically different diameter between t0 and t15 in forskolin condition and between t15 forskolin versus control; Mann–Whitney U test, p < 0.001.
Every mean and standard error of mean (SEM) shown in the paper is based on the data shown in Fig 2. All statistical tests in the paper are 2-sample, 2-tailed t tests and thus assume the data have a Gaussian distribution, which we cannot test. The annotation for p-values are as follows: *p < 0.05, **p < 0.01, ***p < 0.0001.
Pancreata were isolated at E10.5, E12.5, or E14.5 and lysed with lysis buffer RLT, and RNA was purified following the manufacturer instructions (RNeasy Plus Micro Kit, #74034, Qiagen). RNA quality was assessed using an Agilent 2100 Bioanalyzer, following the manufacturer instructions (Agilent RNA 6000 Pico Kit # 5067–1513). Extracted RNA (700 pg) was amplified using Ovation Pico SL WTA system V2 (# 3312–48, Nugen). Samples were labeled with SureTag DNA labeling kit (#5190–3391, Agilent Technologies), run on SurePrint G3 Mouse Gene Exp v2 Array (# G4852B, Agilent Technologies), and read by a SureScan Microarray Scanner (Agilent Technologies).
In the section below, we describe the calculus needed to derive the network properties shown in Fig 2, S3 Fig, and S5 Fig.
A network of N nodes and E links (edges) can be represented by its N × N adjacency matrix A
A(i,j)={1,ifnodeiisconnectedtonodej0,otherwise.
For an undirected network, A(i,j) = A(j,i). Nodes in the networks do not connect to themselves, A(i,i) = 0, and so the trace of A is always zero ∑i=1NA(i,i)=0.
A modified adjacency matrix can be constructed, called the weighted adjacency matrix Aw:
Aw(i,j)=A(i,j)d(i,j),
where d is the euclidian distance between node i and node j. The total length of the network lT is defined as the sum of all weighted links:
Ltot=12∑i=1N∑j=1NAw(i,j).
The cost of the network is the total length of all the links in the network compared to the same value of the euclidian MST composed of the same nodes [24].
The performance of the network is the average shortest path between every node compare to the same value of the MST:
P=〈l〉〈lMST〉,
where l(i,j) is the shortest path between node i and j through the network [24].
The mean distance to the root node is the mean of the shortest paths from node i to the root node.
The mean distance between every node is calculated in much the same way.
For finding polygons in the network, we follow the calculation of Alon [21]. In the following, A is the adjacency matrix, D is the degree matrix for the network, and |E| is the amount of edges in the network. nG(C3),nG(C4), and nG(C5) are the number of triangles, squares, and pentagons in the network, respectively. The following calculations assume an undirected network.
Here, D(i,i) is the degree of node i, and tr (…) is the trace of the matrix. The code calculating all the above network properties has been written in MATLAB 2014a and is in the supporting information as “NetworkProp.m”.
The expected adaptation of duct diameter in response to flux of pancreatic fluid is inspired by previous work in blood vessels [31], satisfying the following equation:
∂d(t)∂t=K[τ(t)−τdesired]d(t),
where d is the duct diameter at time t, K is an adaption constant, and τ is the shear stress experienced by the duct as a result of the fluid flow. In our setup, we consider the system only at steady state, and so
∂d(tss)∂t=0→yieldsτ(tss)=τdesired,
which means that at steady state, every duct has adapted to have the desired shear stress, whichever that value might be. Shear stress relates to flux and duct diameter though Poiseuille’s equation (assuming a likely laminar flow and a noncompressible fluid)
d3=32ηQπτ,
where η represents fluid viscosity and Q the flux of fluid running through the duct. Assuming that every part of the duct adapts to have the same shear stress (and by that extension pressure), as they are made of the same material, this shows that every duct at steady state will have a diameter that is proportional to the cubic root of the flux.
It should be noted that Hacking and colleagues [31] show that given two ducts connected in parallel with a constant flow source, one duct is bound to disappear, while the other will attain the ideal diameter for the flow. This theoretical result explains that redundant ducts can be eliminated in the pancreas, as it will have a constant flow source at steady state.
Experimentally, the duct diameter was measured using Imaris at 27 random nonredundant ducts distributed over the network of an E18.5 pancreas (S11A Fig). The code has been written in MATLAB 2017 and is in the supporting information as “DuctDiameter.m”, “DuctThickness.m”, and “DuctThicknessAnalysis,m”. The analysis shows that the cubic root is a reasonable approximation to the relation between flux, represented as nodes upstream and duct length upstream, and the corresponding duct thickness (S11B and S11C Fig).
The network diffusion model works in the following steps:
The code applying diffusion to the network has been written in MATLAB 2014a and is in the supporting information as “DiffusionOnNetwork.m”.
Given concentration ϕ at node i, which is allowed to freely diffuse, ϕ obeys the heat equation
∂ϕ(i,t)∂t=C∇2ϕ(i,t),
where ∇2 is the laplacian, and C is the diffusion coefficient.
On a network, both ϕ and ∇2 are discretized, and diffusion is constrained to the network links. The heat equation can then be rewritten
dϕ(i)dt=−C∑jL(i,j)ϕ(j),
where L(i,j) = L is the network Laplacian, and ϕ(i) is the concentration at node i.
The network laplacian for node i can be defined by the node degree d and the neighbors of node i.
An example of a network diffusion matrix is presented in S9B Fig. In order to solve diffusion numerically, the forward Euler scheme of network diffusion can be used, giving the diffusion scheme
ϕ(t+Δt)=ϕ(t)−CLϕ(t)Δt.
The network creation model works in the following steps:
The code constructing the random network has been written in MATLAB 2014a and is in the supporting information as “NetworkCreation.m”, as well as the code to initiate the system from an L-system setup instead of two nodes, “RandLsystem.m”.
The network pruning model works in the following steps:
The code applying network pruning to the network has been written in MATLAB 2014a and is in the supporting information as “PruneBasedOnFlux.m”.
A file named “ExampleScript.m” demonstrates how all the above scripts work in concert.
|
10.1371/journal.ppat.1005507 | Promotion of Expansion and Differentiation of Hematopoietic Stem Cells by Interleukin-27 into Myeloid Progenitors to Control Infection in Emergency Myelopoiesis | Emergency myelopoiesis is inflammation-induced hematopoiesis to replenish myeloid cells in the periphery, which is critical to control the infection with pathogens. Previously, pro-inflammatory cytokines such as interferon (IFN)-α and IFN-γ were demonstrated to play a critical role in the expansion of hematopoietic stem cells (HSCs) and myeloid progenitors, leading to production of mature myeloid cells, although their inhibitory effects on hematopoiesis were also reported. Therefore, the molecular mechanism of emergency myelopoiesis during infection remains incompletely understood. Here, we clarify that one of the interleukin (IL)-6/IL-12 family cytokines, IL-27, plays an important role in the emergency myelopoiesis. Among various types of hematopoietic cells in bone marrow, IL-27 predominantly and continuously promoted the expansion of only Lineage−Sca-1+c-Kit+ (LSK) cells, especially long-term repopulating HSCs and myeloid-restricted progenitor cells with long-term repopulating activity, and the differentiation into myeloid progenitors in synergy with stem cell factor. These progenitors expressed myeloid transcription factors such as Spi1, Gfi1, and Cebpa/b through activation of signal transducer and activator of transcription 1 and 3, and had enhanced potential to differentiate into migratory dendritic cells (DCs), neutrophils, and mast cells, and less so into macrophages, and basophils, but not into plasmacytoid DCs, conventional DCs, T cells, and B cells. Among various cytokines, IL-27 in synergy with the stem cell factor had the strongest ability to augment the expansion of LSK cells and their differentiation into myeloid progenitors retaining the LSK phenotype over a long period of time. The experiments using mice deficient for one of IL-27 receptor subunits, WSX-1, and IFN-γ revealed that the blood stage of malaria infection enhanced IL-27 expression through IFN-γ production, and the IL-27 then promoted the expansion of LSK cells, differentiating and mobilizing them into spleen, resulting in enhanced production of neutrophils to control the infection. Thus, IL-27 is one of the limited unique cytokines directly acting on HSCs to promote differentiation into myeloid progenitors during emergency myelopoiesis.
| Emergency myelopoiesis is inflammation-induced hematopoiesis that is critical for controlling infection with pathogens, but the molecular mechanism remains incompletely understood. Here, we clarify that one of the interleukin (IL)-6/IL-12 family cytokines, IL-27, plays an important role in emergency myelopoiesis. Among various types of hematopoietic cells in bone marrow, IL-27 predominantly and continuously promoted expansion of only Lineage−Sca-1+c-Kit+ (LSK) cells, especially long-term repopulating hematopoietic stem cells, and differentiation into myeloid progenitors in synergy with stem cell factor. These progenitors expressed myeloid transcription factors such as Spi1, Gfi1, and Cebpa/b through activation of signal transducer and activator of transcription 1 and 3, and had enhanced potential to differentiate into neutrophils, but not into plasmacytoid dendritic cells. Among various cytokines, IL-27 in synergy with stem cell factor had the strongest ability to augment the expansion of LSK cells and their differentiation into myeloid progenitors. The blood stage of malaria infection was revealed to enhance IL-27 expression through interferon-γ production, and IL-27 then promoted the expansion of LSK cells, differentiating and mobilizing them into the spleen, resulting in enhanced production of neutrophils to control the infection. Thus, IL-27 is one of the limited unique cytokines directly acting on hematopoietic stem cells to promote differentiation into myeloid progenitors during emergency myelopoiesis.
| Emergency myelopoiesis is inflammation-induced hematopoiesis, which is critical for controlling systemic infection with pathogens such as a virus, bacteria, or parasite [1,2]. In contrast to adaptive immune cells such as T cells and B cells, which can vigorously proliferate in response to their specific antigens, innate immune cells need to be replenished from hematopoietic stem cells (HSCs) and progenitors in bone marrow (BM) because of their low proliferative activity. However, the molecular mechanism of emergency myelopoiesis during infection remains incompletely understood. HSCs and hematopoietic progenitors can directly sense the presence of pathogens via pattern recognition receptors (Rs) such as Toll-like receptors (TLRs), and they can also respond to pro-inflammatory cytokines such as interferon (IFN)-α, IFN-γ, interleukin (IL)-1, tumor necrosis factor (TNF)-α, and granulocyte colony-stimulating factor (G-CSF) [1]. IFN-α and IFN-γ have pleiotropic effects on many cell types, including HSCs and hematopoietic progenitors [1]. Recently, these cytokines were demonstrated to induce an expansion of HSCs and myeloid progenitors, leading to the production of mature myeloid cells [3–6], although their inhibitory effects on hematopoiesis were previously reported [7–9]. Currently, thus, there are several conflicting positive and negative effects of IFN-α and IFN-γ in hematopoiesis [10,11]. However, these discrepancies may be explained by compensatory mechanisms, including IFN-γ-mediated secretion of other cytokines such as IL-6 [12] and fms-related tyrosine kinase 3 ligand (Flt3L) [13].
IL-27 is one of the IL-6/IL-12 family cytokines; it plays important roles in immune regulation with both pro-inflammatory and anti-inflammatory properties [14–16]. IL-27 consists of p28 and Epstein-Barr virus-induced gene 3 (EBI3), and its receptor is composed of WSX-1 and glycoprotein (gp)130, which is a common receptor subunit in many of the IL-6 family cytokines. We previously demonstrated that IL-27 plays a role in HSC regulation, and that IL-27 expands HSCs and promotes their differentiation in vitro [17]. Moreover, transgenic (Tg) mice expressing IL-27 showed enhanced myelopoiesis in BM and extramedullary hematopoiesis in the spleen [17].
In the present study, we further examined the effects of IL-27 on hematopoiesis, the molecular mechanisms, and the physiological role of IL-27 in the control of malaria infection. IL-27 acted on and expanded Lineage (Lin)−Sca-1+c-Kit+ (LSK) cells, which are highly enriched in HSCs together with very primitive hematopoietic progenitors [18,19], in BM cells in synergy with stem cell factor (SCF, c-Kit ligand) and differentiated HSCs into myeloid progenitors through activation of signal transducer and activator of transcription 1 (STAT1) and STAT3. Moreover, malaria infection induced IFN-γ production, which augmented IL-27 expression, and the IL-27 then promoted the expansion and mobilization of LSK cells into the spleen, resulting in enhanced myelopoiesis to resolve the infection. Our results revealed that IL-27 is one of the limited unique cytokines directly acting on long-term HSCs (LT-HSC), which represent the true stem cells capable of self-renewing, and promotes the expansion and differentiation of them into myeloid progenitors.
Previously, we demonstrated that stimulation of LSK cells with IL-27 and SCF induces an expansion of HSCs and hematopoietic progenitors, including short-term repopulating cells [17]. Moreover, we found that only the combination of IL-27 and SCF, but not either alone, vigorously and continuously expands BM cells to produce LSK cells and CD11b+c-Kit− cells [17]. To examine which cell populations in BM cells respond to IL-27 and SCF in more detail, BM cells were divided into two populations positive or negative for Lin markers except CD11b, and the Lin− population was further divided into four populations positive for either c-Kit or CD11b, or both positive, or both negative. Each population purified by sorting was then stimulated with IL-27 and SCF. Among these five populations, only the Lin−c-Kit+ population greatly expanded (Fig 1A). Next, the BM cells were divided into respective hematopoietic progenitors according to the expression of cell surface markers, as reported previously [20–22], and stimulated with IL-27 and SCF. Only the LSK cell population vigorously and continuously expanded over more than 6 weeks, although transient and slight expansion was seen in the cell populations of granulocyte/macrophage progenitor (GMP), common myeloid progenitor (CMP), and megakaryocyte/erythrocyte progenitor (MEP) (Fig 1B–1E). The expanding cells in the LSK cell population were further analyzed for the expression of cell surface markers. In line with the preliminary results, there seemed to be two populations, the phenotypical LSK population and the Lin+c-Kit− population (Fig 1E).
We previously demonstrated that IL-27 Tg mice, which express high amounts of IL-27 in blood, show an increased number of LSK cells in the BM and spleen [17]. To further examine the in vivo effects of IL-27 on the expansion of LSK cells, LSK cells were purified by sorting from BM cells of GFP Tg mice and transferred into wild-type (WT) and IL-27 Tg mice. The transferred GFP+ LSK cells vigorously expanded in the BM and spleen of IL-27 Tg mice, but not in those of WT mice, and approximately half of the expanding cells retained the cell surface markers for LSK phenotype (Fig 1F). These results suggest that IL-27 vigorously and continuously augments the expansion of LSK cells both in vitro and in vivo.
Because it was previously reported that IFN-α and IFN-γ induce proliferation of HSCs in vivo [3–5], we next explored the effects of various cytokines in collaboration with SCF on the expansion of LSK cells in vitro. However, IFN-α and IFN-γ augmented the expansion of LSK cells very little, and only IL-27 enhanced it vigorously over 4 weeks (Fig 1G and 1H). Moreover, although there are several cytokines, such as IL-3, IL-11, G-CSF, and TPO, that are known to transiently expand and differentiate HSCs [1], none showed an ability superior to that of IL-27 in expanding LSK cells retaining the LSK phenotype over a long period of time (S1 Fig). Thus, IL-27 has the strongest ability to augment the expansion of LSK cells.
The LSK cells expanded by IL-27 and SCF were further analyzed for the cell surface expression of various markers, and the expression levels were compared with those of primary LSK cells freshly prepared from BM of WT mice. The expression levels of macrophage colony-stimulating factor receptor (M-CSFR), CD16/32, and MHC class II in the LSK cells expanded by IL-27 and SCF were much higher than those in primary LSK cells (Fig 2A). The expression levels of CD34 and CD150 in the expanded LSK cells were slightly less than those in primary LSK cells (Fig 2A). In contrast, the expanded LSK cells were almost completely negative for Flt3 expression, whereas primary LSK cells were positive for Flt3 (Fig 2A). Thus, IL-27 and SCF expand and differentiate primary LSK cells into M-CSFR+Flt3−CD16/32+ LSK cells (myeloid progenitor cells).
Next, multipotency of the LSK cells expanded by IL-27 and SCF were examined under various differentiation conditions for migratory dendritic cells (mDCs) by granulocyte/macrophage (GM)-CSF, plasmacytoid DCs (pDCs) and conventional DCs (cDCs; lymphoid-resident DCs) were examined using Flt3L and thrombopoietin (TPO) [23], myeloid cells were examined using IL-3 and SCF, and T cells and B cells were examined by using thymic stromal cells (TSt4) with and without expressing Notch ligand Delta-like 1 (DLL1), respectively [24]. The expanded LSK cells much more rapidly proliferated and differentiated into MHC class II+CD11c+ mDCs than primary LSK cells in response to GM-CSF, although the total number of mDCs achieved seemed to be similar for both (Fig 2B and 2C). However, LSK cells stimulated with IL-27 and SCF rapidly lost the ability to differentiate into pDC and cDC (Fig 2D and 2E). These phenomena are highly consistent with the almost complete abolishment of Flt3 expression on the expanded LSK cells (Fig 2A). Under myeloid differentiation conditions, the expanded LSK cells differentiated much more greatly into neutrophils and slightly into mast cells, but less so into macrophages and basophils (Fig 2F and 2G). Similar enhanced differentiation into myeloid cells was observed using the LSK cells obtained from WT and IL-27 Tg mice (S2 Fig). In contrast, the ability to differentiate into B cells and T cells was almost completely abrogated in the expanded LSK cells (Fig 2H).
Moreover, the ability of the LSK cells expanded by IL-27 and SCF to differentiate into myeloid cells in vivo was explored by using mixed BM chimeras. The equal cell numbers of the LSK cell population expanded from CD45.1 congenic mice and BM cells from CD45.2 congenic mice were mixed and transferred into sublethally irradiated CD45.2 congenic mice. After 9 days, cell populations of neutrophils in the BM and spleen were analyzed. In agreement with the in vitro results described, the percentages of neutrophils derived from the expanded LSK cells were markedly higher than those from primary LSK cells in both BM and spleen (Fig 2I).
To explore the molecular mechanisms whereby IL-27 and SCF expand LSK cells, total RNA was prepared from the expanded LSK cells and respective hematopoietic progenitors and analyzed with real-time reverse transcriptase-polymerase chain reaction (RT-PCR). The expanded cells highly expressed the transcription factors critical for differentiation into myeloid cells such as Spi1, Gfi1, and Cebpa, but expression was much less for those important for the other types of cells such as Irf8, Tcf4, and Ikzf1 [25–27] (Fig 2J). No expression of transcription factors important for B cells and erythrocytes, Pax5 [28] and Gata1 [29], respectively, was observed. In addition, the expression of Cbepb, which is a transcription factor recently demonstrated to be regulated by cytokines and control emergency granulopoiesis [30–32], was also increased (S3 Fig).
These results suggest that LSK cells expanded by IL-27 and SCF are multipotent myeloid progenitors that have unique potential to differentiate into mDCs, neutrophils, and mast cells, and less so into macrophages, and basophils, but not into pDCs, cDCs, T cells, and B cells.
We and others previously demonstrated that IL-27 activates both STAT1 and STAT3 through WSX-1 and gp130, respectively [14,15]. Consistent with these reports, real-time RT-PCR analysis revealed that LSK populations purified from LSK cells expanded by IL-27 and SCF and primary BM cells were positive for mRNA expression of STAT1 and STAT3 (Fig 3A). Phosphorylation of STAT1 and STAT3 was also detected in primary WT LSK cells, but not WSX-1-deficient LSK cells (Fig 3B), in response to IL-27 and SCF, which were analyzed by flow cytometry. Furthermore, IL-27 alone induced phosphorylation of both STAT1 and STAT3, whereas SCF alone failed to induce phosphorylation of either one, as discussed previously [33] (S4 Fig).
To further investigate the roles of STAT1 and STAT3 in the expansion of LSK cells and the ability to differentiate into myeloid progenitors by IL-27 and SCF, we used STAT1-deficient LSK cells and conditional STAT3-knockout (STAT3 cKO) LSK cells. LSK cells from WT (129) and STAT1-deficient mice were stimulated with IL-27 and SCF. Although WT LSK cells comprised more than 90% of cells with the LSK phenotype after 7 days, STAT1-deficient LSK cells comprised half LSK phenotype cells and half Sca-1−LK (LS−K) (Fig 3C). Nevertheless, the number of expanded cells was comparable among them, probably due to the anti-proliferative effects by STAT1 signaling [34]. However, the purified STAT1-deficient LS−K cells failed to survive thereafter, even in the presence of IL-27 and SCF (Fig 3D), although the purified STAT1-deficient LSK cells expanded well (Fig 3D), as did WT 129 LSK cells (Fig 3C). Moreover, WT and STAT1-deficient LSK cells had similar abilities to differentiate into MHC class II+CD11c+ mDC cells (Fig 3E) and macrophages (Fig 3F). STAT1-deficient LSK cells showed reduced ability to differentiate into neutrophils and mast cells, but increased ability to differentiate into basophils (Fig 3F). In line with the reduced ability to differentiate into neutrophils, mRNA expression of the critical transcription factor Gfi1 was significantly reduced in STAT1-deficient LSK cells compared with that in WT LSK cells (Fig 3G). In contrast, STAT3 cKO LSK cells expanded very little in response to IL-27 and SCF (Fig 3H). Moreover, the residual surviving LSK cells almost completely lost the ability to differentiate into mDCs (Fig 3I) and myeloid cells (Fig 3J). Consistent with the abrogated abilities, these cells showed reduced expression of the critical transcription factors such as Spi1, Gfi1, and Cebpa, but not Irf8 (Fig 3K). These results suggest that both STAT1 and STAT3 are necessary for LSK cells to fully expand and differentiate into myeloid progenitor cells in response to IL-27 and SCF.
To more precisely define which cell population responds to stimulation with IL-27 and SCF, LSK cells were further divided into two populations according to CD34 expression. Although the percentage of the more primitive population of CD34− LSK cells was much less than that of CD34+ LSK cells, the CD34− LSK cells responded much better to stimulation with IL-27 and SCF and expanded more vigorously than CD34+ LSK cells (Fig 4A and 4B). Then, the LSK cells were further divided into eight populations, F1 to F8, including LT-HSCs (CD34−CD150+CD41− LSK, F1) and myeloid-restricted progenitor cells with long-term repopulating activity (MyRPs, CD34−CD150+CD41+ LSK, F4) according to the recently revised criteria [19]. Respective populations purified by sorting (Fig 4C) were stimulated with IL-27 and SCF. Only two populations, F1 and F4, vigorously expanded. F5, which corresponds to populations more differentiated toward myeloid cells such as macrophages, slightly expanded (Fig 4D and 4E). The F1 and F4 populations expanded by IL-27 and SCF had great abilities to differentiate into myeloid cells, particularly neutrophils (Fig 4E). Thus, IL-27 and SCF expand CD34−CD150+ LSK cells, including LT-HSCs and MyRPs, and differentiate them into myeloid progenitor cells, which have great potential to differentiate mainly into neutrophils.
We previously demonstrated that in the blood stage of malaria infection with the attenuated variant Plasmodium (P) berghei XAT derived from the lethal strain P. berghei NK65 IFN-γ production induced by IL-12 and phagocytic cells in the spleen are critical for controlling parasitemia [35,36]. Recently, it was reported that the blood stage of P. chabaudi infection induces mobilization of early myeloid progenitor cells out of BM, thereby transiently establishing myelopoiesis in the spleen through IFN-γ to resolve the infection [6,37]. In line with these results, WSX-1-deficient mice showed more increased parasitemia than WT mice at 7 days (Fig 5A), just prior to when the parasitemia reaches its peak after infection with P. berghei XAT (S5A Fig). In contrast, no significant difference was observed in the serum IFN-γ levels in WT and WSX-1-deficient mice (Fig 5B). The infection markedly induced the enhanced percentage and number of LSK cells in the BM and spleen of WT mice (Fig 5C and 5D). The LSK cells in the BM showed greatly augmented abilities to differentiate in vitro into neutrophils and mast cells, but had slightly reduced abilities to differentiate into macrophages and basophils after infection (Fig 5E). Moreover, the LSK cells in the spleen exhibited a much more enhanced ability to differentiate into neutrophils, macrophages, and mast cells (Fig 5E). In contrast, of note, WSX-1-deficient mice showed significantly reduced percentage and number of LSK cells in the BM and spleen compared with WT mice after infection (Fig 5C and 5D). In particular, the cell number of the neutrophils in the spleen was increased very little in WSX-1-deficient mice (Fig 5F and 5G). Consistent with this, LSK cells from BM of WSX-1-deficient mice showed reduced abilities to differentiate in vitro into neutrophils after infection compared with those of WT mice (S6 Fig). In addition, more greatly reduced abilities to differentiate in vitro into various myeloid cells were observed when LSK cells from spleen were used (S6 Fig). Moreover, mixed bone marrow chimera experiments using bone marrow cells from WT and WSX-1-deficient mice revealed that the effect of IL-27 on the expansion of LSK cells and neutrophils is actually a cell-autonomous direct effect (S7 Fig).
To further elucidate the protective role of LSK cells in malaria infection, LSK cells purified from BM cells of infected WT CD45.1 mice were injected into the infected WSX-1-deficient CD45.2 mice. Consistent with the increased percentage of neutrophils differentiated from the WT LSK cells in the BM and spleen of transferred WSX-1-deficient mice (Fig 5H), parasitemia was significantly decreased in the WSX-1-deficient mice by the transfer of WT LSK cells compared with that in non-transferred WSX-1-deficient mice (Fig 5I). Thus, the blood stage of malaria infection induces expansion, differentiation, and mobilization of LSK cells into the spleen to produce myeloid cells such as neutrophils in an IL-27-dependent manner.
Although it was previously reported that IFN-α and IFN-γ induce proliferation of HSCs in vivo [3–5], IFN-α and IFN-γ augmented the expansion of LSK cells very little in vitro, and only IL-27 enhanced it vigorously over 4 weeks (Fig 1G and 1H). However, similar to the work previously reported [37,38], IFN-γ-deficient mice showed increased parasitemia with almost no increase in the number of LSK cells in BM and spleen (Fig 6A and 6B). To clarify the molecular mechanism whereby IFN-γ induces the expansion of LSK cells, the expression of IL-27 subunits EBI3 and p28 was examined. Although the infection did not increase EBI3 mRNA expression in the BM and spleen of both WT and IFN-γ-deficient mice (S8 Fig), intriguingly, the infection greatly enhanced p28 mRNA expression in WT mice but failed to enhance it in IFN-γ-deficient mice (Fig 6C). In agreement with this, p28 protein levels in the serum were greatly increased by the infection in WT mice but not in IFN-γ-deficient mice (Fig 6D). Consistent with the in vivo role of IFN-γ, we also observed the augmentation of mRNA expression of p28, but not EBI3, and p28 protein production in the culture supernatants of WT BM cells stimulated with IFN-γ in vitro (S9 Fig).
To further clarify the role of IL-27 downstream of IFN-γ, we next performed the experiment to see the effects of forced expression of IL-27 on the susceptibility to malaria infection in IFN-γ-deficient mice. The hydrodynamic injection of IL-27 expression vector into the infected IFN-γ-deficient mice showed significantly decreased parasitemia compared with that of control vector (Fig 6E). This phenomenon was accompanied by the enhanced percentage of LSK cells and augmented numbers of LSK cells and neutrophils in both BM and spleen (Fig 6F and 6G). Thus, the blood stage of malaria infection augments the expression of IL-27 through IFN-γ, and IL-27 then promotes the expansion, differentiation, and mobilization of LSK cells into the spleen to control parasitemia.
Previously, we found that IL-27, which is in the IL-6/IL-12 family of cytokines, plays a role in the regulation of HSCs in vitro and in vivo [17]. Here, we have further elucidated that IL-27 is a unique cytokine that directly acts on LSK cells to promote their differentiation into myeloid progenitor cells called M-CSFR+Flt3−CD16/32+ LSK cells, which still retain the LSK phenotype (Fig 2A–2I). These progenitors have great potential to give rise to neutrophils, mDCs, and mast cells, but not to pDCs, cDCs, T cells, and B cells. Interestingly, among various BM progenitor cells, IL-27 and SCF vigorously and continuously expand only HSCs and primitive myeloid progenitor cells with long-term repopulating activity, such as LT-HSCs and MyRPs, respectively [19], for more than 4 weeks (Fig 4). Consistent with the ability to differentiate into myeloid progenitor cells, the LSK cells expanded by IL-27 and SCF expressed transcription factors such as Spi1, Gfi1, Cebpa, and Cebpb, which are critical for myeloid differentiation [25–27,30–32] (Fig 2J and S3 Fig). Although Cebpb was reported to be an important transcription factor for emergency granulopoiesis [30–32], STAT3 signaling was revealed to be important for its upregulation, whereas STAT1 signaling unexpectedly suppressed its expression (S10 Fig). This phenomenon seems to correlate to the expression level of the anti-apoptotic gene Bcl-2 [39,40], but not the transcription factor E2-2, which is critical for pDC differentiation [41]. Further studies are necessary to elucidate the precise roles of each STAT in the regulation of Cebpb expression. Thus, IL-27 is one of the limited unique cytokines that directly acts on the most primitive LT-HSCs; it promotes their expansion and differentiation into myeloid progenitor cells, presumably through MyRPs [19], to replenish myeloid cells such as neutrophils in the periphery during emergency myelopoiesis.
Sca-1 is an IFN-responsive molecule that is highly upregulated in many hematopoietic cells following exposure to IFNs [10,11,37]. Consequently, myeloid-restricted progenitor cells normally identified as Lin−Sca-1−c-Kit+ (LS−K) become positive for Sca-1 and can no longer be distinguished from the real multipotent LSK cells, resulting in overestimation of the latter population, and this is a problem. IL-27 was previously reported to enhance the expression of Sca-1 on T cells [42]. However, to alleviate the problem, we initially identified the cell population responsive to IL-27 and SCF among BM cells by using LSK cells and various hematopoietic progenitor cells purified by sorting. Intriguingly, it turned out that the SCF and LSK cell populations expanded vigorously and continuously in response to IL-27, and that the LS−K cell populations including GMP, CMP, and MEP only transiently and slightly responded during the first week and then disappeared thereafter (Fig 1A–1E). Moreover, in almost all in vitro experiments, we used primary LSK cells that were purified by sorting. The high responsiveness of the LSK cells to IL-27 seems to be partially due to the higher mRNA expression of WSX-1 in the LSK cells compared to that of other hematopoietic progenitor cells (S3 Fig) [17].
We previously demonstrated that during the blood stage of malaria infection with attenuated P. berghei XAT, IL-12-mediated IFN-γ production and phagocytic cells (including neutrophils) in the spleen are critical for controlling parasitemia [35,36,43]. Previous studies demonstrated that neutrophils play an important role in killing malaria parasites in mice, rats, and humans [43–45]. In marked contrast, regarding infection with lethal P. berghei NK65, IL-12-mediated IFN-γ production was shown to contribute to T-cell-dependent immunopathology [46]. However, a major role of IL-27 in infection is its suppression of excess immune responses against infection by controlling the production of pro-inflammatory cytokines [14–16]. Consistent with this, WSX-1/IL-27 was recently demonstrated to have a critical role in limiting the effector CD4+ T-cell-mediated immunopathology caused by IL-12-dependent IFN-γ production during infection with lethal P. berghei NK65 [47–49]. The present study clearly revealed that WSX-1/IL-27 contributes to clearance of parasites due to enhanced myelopoiesis during the early phase of infection with attenuated P. berghei XAT (Fig 5 and S5A Fig). However, during the late phase of infection, WSX-1/IL-27 seems to play a role in limiting the production of pro-inflammatory cytokines such as IFN-γ (S5B Fig), leading to augmented reduction of parasitemia (S5A Fig), as in the case of infection with lethal P. berghei NK65 [47–49]. Moreover, our preliminary data revealed that there were no apparent differences observed in parasitemia or expansion of LSK cells and neutrophils in the BM and spleen of WT and WSX-1-deficient mice 7 days after infection with lethal P. berghei NK65 (S11 Fig). It is conceivable that pro-inflammatory cytokines other than IL-27 were abundantly produced in the absence of IL-27 during infection with lethal P. berghei NK65 and that late-phase infection with attenuated P. berghei XAT may have redundantly compensated for the loss of IL-27 to promote myelopoiesis. However, other studies have shown that IL-27 limits migration of neutrophils from the BM to the site of inflammation by reducing production of cytokines and chemokines during influenza infection [50] and septic peritonitis [51]. IL-27 was also reported to be a negative regulator of neutrophil function [52]. Although IL-27 directly promotes myelopoiesis to produce myeloid progenitors in BM, as shown in the present study, IL-27 may indirectly regulate migration of these progenitors and neutrophils to the site of inflammation and limit neutrophil function. Thus, IL-27 has both positive and negative effects on neutrophils; therefore, the overall outcome of the effects of IL-27 is likely to be governed by the balance between these effects, depending on the disease model.
It was recently demonstrated that P. chabaudi infection induces mobilization of early myeloid progenitor cells out of BM, thereby transiently establishing myelopoiesis in the spleen through IFN-γ [37]. However, the expression of IFN-γR in the hematopoietic compartment was dispensable, whereas its expression in the irradiation-insensitive cellular compartment, including endothelial cells and stromal cells, was important [37]. Secretion of IFN-γ-induced chemokines such as CCL2 and CCL7 by non-hematopoietic cells plays a critical role in the mobilization of CCR2-expressing HSCs [37]. In this study, however, there is no experimental evidence regarding how IFN-γ regulates the activation of HSCs. In the present study, WSX-1-deficient mice showed significantly reduced numbers of LSK cells and neutrophils compared with WT mice after P. berghei XAT infection, resulting in increased parasitemia (Fig 5A–5F and S5A Fig). These results suggest that endogenous IL-27 greatly contributes to the clearance of parasitemia through augmentation of myelopoiesis. Moreover, similar to P. chabaudi infection, P. berghei XAT infection could not increase the number of LSK cells in IFN-γ-deficient mice with increased parasitemia (Fig 6A and 6B). Of note, after P. berghei XAT infection, p28 mRNA expression and its serum protein level were markedly upregulated in an IFN-γ-dependent manner (Fig 6C and 6D and S9 Fig). This is consistent with previous reports indicating that p28 gene transcription in macrophages is induced by IFN-γ and TLR ligands [53], and that IFN-γ limits Th17-mediated and Th9-mediated autoimmune inflammation through IL-27 production [54,55]. In addition, the hydrodynamic injection of the IL-27 expression vector into infected IFN-γ-deficient mice greatly recovered the number of LSK cells and neutrophils in the BM and spleen and eventually reduced parasitemia (Fig 6E–6G). Thus, during malaria infection, it is highly conceivable that the proliferative effects on LSK cells by IFN-γ are indirectly mediated by IL-27. In our study, we could not observe any direct proliferative effect of IFN-α and IFN-γ on LSK cells in vitro, as recently pointed out by others [10,11], and only IL-27 augmented the proliferative effect for more than 4 weeks (Fig 1G and 1H).
Recently, IL-27 was reported to have a polyglutamic acid domain in the p28 subunit, which is unique among cytokines, and to confer hydroxyapatite-binding and bone-binding properties and bone tropism to bone sialoprotein and the endosteal bone surface [56]. This location in the BM has been identified as a niche for HSCs [57], and these properties support the idea that IL-27 plays a critical role in the regulation of HSCs in that niche. We detected much higher expressions of IL-27 subunits (both p28 and EBI3) at mRNA levels in BM than in the spleen during the steady state, and P. berghei XAT infection greatly augmented the expression of p28 mRNA in both BM and spleen, and also its serum protein level (Fig 6C and 6D). Further studies are necessary to clarify which BM cells produce IL-27 during malaria infection; mesenchymal stromal cells might be a candidate because of their reported IL-6 production during viral infection [12], as described in the next section.
It was recently demonstrated that specific cytotoxic CD8+ T cells during an acute viral infection with lymphocytic choriomeningitis virus secrete IFN-γ, thus enhancing the production of IL-6 in BM mesenchymal stromal cells and resulting in an increased number of early multipotent progenitors and committed myeloid precursors in the BM and accumulation of myeloid cells in the periphery [12]. The IL-6Rα chain is only expressed at the stage of early multipotent progenitors and downstream myeloid precursors, and it is lacking HSCs [12]. In contrast, IL-27 most predominantly acts on only HSCs, as shown in the present study. Both IL-6 and IL-27 share gp130, which is ubiquitously expressed as a common receptor subunit. Therefore, downstream of IFN-γ, both IL-27 and IL-6 may be necessary to induce the maximum myelopoiesis to control infection. However, the mode of IL-6 action is complex and there are two major mechanisms: IL-6 classic signaling through membrane IL-6Rα and IL-6 trans-signaling through soluble IL-6Rα [58,59]. It was recently reported that IL-6Rα-deficient mice show increased resistance to P. chabaudi infection and that IL-6 trans-signaling, but not IL-6 classic signaling, contributes to a lethal outcome of infection [60]. In contrast to the viral infection, we could not detect any increased mRNA expression of IL-6 in the BM or spleen of WT and IFN-γ-deficient mice with P. berghei XAT infection (S8 Fig). A similar inability of IFN-γ to enhance IL-6 mRNA expression in WT BM cells in vitro was also observed (S9A Fig). In addition, IL-6-deficient mice showed little increased susceptibility to the P. berghei XAT infection, reduced cell numbers in the LSK cell population, and reduced neutrophils in the BM and spleen (S12 Fig). Thus, individual pathogens may utilize different mechanisms to induce emergency myelopoiesis through IL-27, IL-6, and others.
In conclusion, the present results provide a novel role and mechanism for the action of IL-27 downstream of IFN-γ in the efficient expansion of myeloid progenitor cells from LT-HSCs and MyRP cells and their mobilization into the spleen during acute malaria infection.
The animal study was approved by the Animal Care and Use Committee of Tokyo Medical University (S-230043, S-24012, S-25059, S-26003, and S-27009) and was performed in accordance with our institutional guidelines and the Fundamental Guidelines for Proper Conduct of Animal Experiment and Related Activities in Academic Research Institutions under the jurisdiction of the Ministry of Education, Culture, Sports, Science and Technology, 2006.
C57BL/6 (CD45.2) mice and C57BL/6 (CD45.1) mice were purchased from Sankyo Lab (Tokyo, Japan). The 129/Sv mice and STAT1-deficient mice (129/Sv background) were purchased from Taconic Farms (Germantown, NY, USA). IL-6-deficient mice (C57BL/6 background) were purchased from Jackson Laboratory (Bar Harbor, ME, USA). IFN-γ-deficient mice (C57BL/6 background), STAT3flox/flox mice (a mixed background of 129/Sv and C57BL/6), and GFP Tg mice (C57BL/6 background) [61] were provided by Dr. Iwakura (Tokyo University of Science), Dr. Takeda (Osaka University), Dr. Okabe (Osaka University) and Dr. Ito (Tokyo Medical University), respectively. In addition to these mice, IL-27 Tg mice (C57BL/6 background) [17], and WSX-1-deficient mice (C57BL/6 background) [62] were maintained in specific pathogen-free conditions under the care of the Laboratory Animal Center of Tokyo Medical University. STAT3 cKO cells were obtained by infecting STAT3flox/flox cells with Cre-expressing retrovirus in vitro.
Monoclonal antibodies (mAbs) for mouse c-Kit (2B8), Sca-1 (D7), CD3ε (145-2C11), CD4 (GK1.5), CD8α (53–6.7), CD19 (6D5), CD49b (DX5), Gr-1 (RB6-8C5), TER119/erythroid cell (TER-119), CD11c (N418), CD11b (M1/70), F4/80 (BM8), NK1.1 (PK136), B220 (RA3-6B2), FcεRIα (MAR-1), M-CSFR (AFS98), Flt3 (A2F10), CD16/32 (2.4G2), CD150 (TC15-12F12.2), CD41 (MWReg30), IL-7Rα (A7R34), MHC class II I-A/I-E (M5/114.15.2), Siglec H (551), Ly6G (1A8), CD45.1 (A20), and CD45.2 (104) were purchased from BioLegend (San Diego, CA). mAbs against mouse pY701-STAT1 (4a) and pY705-STAT3 (4/P-STAT3) were purchased from BD Pharmingen (San Diego, CA). mAbs against mouse CD34 (RAM34) was purchased from eBioscience (La Jolla, CA). mAbs against mouse PDCA1 (JF05-1C2.4.1) was purchased from Miltenyi Biotec (Bergisch Gladbach, Germany). APC-Cy7-conjugated streptavidin, PerCP/Cy5.5-conjugated streptavidin, and Brilliant Violet 510-conjugated streptavidin were purchased from BioLegend and used to reveal staining with biotinylated Abs. Mouse recombinant IL-27 and hyper-IL-6 were prepared as a tagged single-chain fusion protein by flexibly linking EBI3 to p28 and soluble IL-6Rα to IL-6, respectively, using HEK293-F cells (Life Technologies, Carlsbad, CA) as described previously [63,64]. Mouse recombinant SCF, IL-1β, IL-3, IL-6, IL-7, IL-11, IL-12, thymic stromal lymphopoietin (TSLP), G-CSF, M-CSF, GM-CSF, TNF-α, and human recombinant TPO were purchased from PeproTech (Rocky Hill, NJ). Human recombinant Flt3L was purchased from Miltenyi Biotec. Mouse recombinant IL-23, IL-25, and IL-33 were purchased from R&D Systems (Minneapolis, MN). Mouse IFN-α was purchased from PBL Biomedical Laboratories (Piscataway, NJ). Mouse recombinant IFN-γ was provided from Shionogi Pharmaceutical Co., Ltd. (Osaka, Japan),
Spleen and BM Lin− cells were enriched by negative selection using an autoMACS Pro (Miltenyi Biotec) with a combination of magnetic beads conjugated with mAbs against CD3ε, CD4, CD8α, Gr-1, TER119, CD11b, CD11c, NK1.1, B220, and FcεRIα. Subsequently, cells were stained with mAbs against c-Kit, and Sca-1 was used for LSK. For multiple fractions of HSC, cells were stained with CD34, c-Kit, Sca-1, CD150, and CD41 mAbs [19]. In the case of common lymphoid progenitor (CLP), macrophage-DC progenitor (MDP), and common DC progenitor (CDP), cells were stained with c-Kit, Sca-1, IL-7Rα, M-CSFR, and Flt3 mAbs [20–22]. For GMP, CMP, and MEP, cells were stained with CD34, c-Kit, Sca-1, and CD16/32 mAbs [20–22]. Sorting was performed on FACS Aria or FACS Aria III (BD Bioscience).
Cells were cultured at 37°C under 5% CO2/95% air in RPMI-1640 (SIGMA, St. Louis, MO) containing 10% fetal calf serum, 50 μM 2-mercaptoethanol (GIBCO, Grand Island, NY), and 100 μg/ml kanamycin (Meiji Seika, Tokyo, Japan). To proliferate progenitors, sorted cells were cultured with 10 ng/ml IL-27 and/or 10 ng/ml SCF. To examine the effect of various cytokines on proliferation of LSK, IL-1β, IL-3, IL-6, hyper-IL-6, IL-11, IL-12, IL-23, IL-25, IL-33, TSLP, G-CSF, TPO, TNF-α, and IFN-α were used as a final concentration of 10 ng/ml. IFN-γ was used as a final concentration of 100 U/ml. Half of the medium was changed every 3 days, with cytokines added.
For evaluation of mDC potential, sorted cells (1–5 × 103) in 96-well plates were cultured with 20 ng/ml GM-CSF for 3 to 10 days. For pDC and cDC potential, sorted cells (5 × 103) were cultured with 100 ng/ml Flt3L and 20 ng/ml TPO for 10 days. For analysis of multipotent myeloid potential, sorted cells (1 × 102–5 × 103) in 96-well plates were cultured with 10 ng/ml SCF and 10 ng/ml IL-3 for 6 days. Half of the medium was replaced every 3 days, with cytokines added. For detection of B-cell potential, sorted cells (2 × 102) were cultured with a monolayer of thymic stromal cells (TSt4) containing 2 ng/ml IL-7 for 15 days [24]. For detection of T-cell potential, sorted cells (2 × 102) were cultured with TSt4 cells expressing DLL1, TSt4/DLL1 cells, containing 2 ng/ml IL-7 and 2 ng/ml Flt3L for 18 days [24]. Half of the medium was replaced every 7 days, with cytokines added. The following cell surface markers were used to identify respective cells: mDC; MHC class II+CD11c+, pDC; Siglec H+PDCA1+CD11c+, cDC; Siglec H−PDCA1−CD11c+, neutrophil; Ly6G+CD11b+ or Gr-1+CD11b+, macrophage; F4/80+CD11b+, mast cell; c-Kit+FcεR1α+CD11b−, basophil; CD49b+FcεR1α+CD11b−, B cell; B220+CD19+, double positive T cell; and CD4+CD8+.
Flow cytometry was performed on a FACS Canto II (BD Bioscience, San Jose, CA) and data were analyzed using FlowJo Software (Tree Star, Ashland, OR). Cell number was counted using flow cytometry unless otherwise indicated. For intracellular cytokine staining, cells were fixed with Fixation Buffer (BD Bioscience) for 30 min and permeabilized with Perm Buffer II (BD Bioscience) for 30 min. Then, samples were stained with antibodies for pY701-STAT1 and pY705-STAT3.
Total RNA was prepared using RNeasy Mini Kit (QIAGEN, Hilden, Germany), and cDNA was prepared using oligo(dT) primer and SuperScript III RT (Invitrogen, Carlsbad, CA, USA). Real-time quantitative PCR was performed using SYBR Premix Ex Taq II and a Thermal cycler Dice real-time system according to the manufacturer’s instructions (TAKARA, Otsu, Shiga, Japan). Glyceraldehyde-3-phosphate (GAPDH) was used as housekeeping gene to normalize mRNA. Relative expression of real-time PCR products was determined by using the ΔΔCt method to compare target gene and GAPDH mRNA expression. Primers used in this study are listed in S1 Table.
For in vivo proliferation analysis, BM LSK cells (8 × 103) purified from GFP Tg mice were intravenously transferred into irradiated (4 Gy) WT and IL-27 Tg mice. To evaluate in vivo development, IL-27/SCF-cultured BM LSK cells (2 × 104) from CD45.1 congenic mice were transplanted into sublethally irradiated (6 Gy) CD45.2 recipient mice with the same congenic BM cells (2 × 106). As in the case of malaria infection, BM LSK cells (5 × 104) were sorted from malaria-infected CD45.1 mice and intravenously transferred into irradiated (4 Gy) WSX1-deficient mice 7 days before infection.
The retroviral vector pMX-Cre-GFP (from Dr. M. Kubo) was transfected into Platinum-E packaging cells [65] using FuGENE 6 (Promega, Madison, WI), and supernatants of these cultures were used as the source of viral particles. LSK cells sorted from BM cells of STAT3flox/flox mice were stimulated with IL-27 and SCF (10 ng/ml each) and transduced with viral particles by spin infection (2,000 rpm, 90 min, 25°C) using 8 μg/ml Polybrene at 24 hr and 48 hr later. The next day, GFP+ LSK cells were sorted and used as STAT3 cKO LSK cells.
Mice were injected intravenously with a red blood cell (RBC) suspension containing parasitized RBC (1 × 104) with the nonlethal strain P. berghei XAT [35], which is an irradiation-induced attenuated variant of the lethal strain P. berghei NK65, or the P. berghei NK65 [46]. Parasitemia was assessed by the microscopic examination of Giemsa-stained smears of tail blood after infection. The percentage of parasitemia was calculated as follows: parasitemia (%) = [(number of infected RBC) / (total number of RBC counted)] × 100.
IFN-γ-deficient mice were intravenously injected with 25 μg of p3xFLAG-CMV (Sigma Chemical Co., St. Louis, MO), or p3xFLAG-IL-27 plasmids at days 0 and 4 after malaria infection.
Amounts of IL-27 p28 in culture supernatants or serum were determined by using Quantikine kits (R&D) according to the manufacturer’s instruction.
Data are represented as mean ± SEM. Statistical analyses were performed by two-tailed Student’s t test for two groups, and by one-way ANOVA and Bonferroni’s multiple comparison tests for multiple groups. P < 0.05 was considered to indicate a statistically significant difference.
|
10.1371/journal.pgen.1004253 | Ras GTPase-Like Protein MglA, a Controller of Bacterial Social-Motility in Myxobacteria, Has Evolved to Control Bacterial Predation by Bdellovibrio | Bdellovibrio bacteriovorus invade Gram-negative bacteria in a predatory process requiring Type IV pili (T4P) at a single invasive pole, and also glide on surfaces to locate prey. Ras-like G-protein MglA, working with MglB and RomR in the deltaproteobacterium Myxococcus xanthus, regulates adventurous gliding and T4P-mediated social motility at both M. xanthus cell poles. Our bioinformatic analyses suggested that the GTPase activating protein (GAP)-encoding gene mglB was lost in Bdellovibrio, but critical residues for MglABd GTP-binding are conserved. Deletion of mglABd abolished prey-invasion, but not gliding, and reduced T4P formation. MglABd interacted with a previously uncharacterised tetratricopeptide repeat (TPR) domain protein Bd2492, which we show localises at the single invasive pole and is required for predation. Bd2492 and RomR also interacted with cyclic-di-GMP-binding receptor CdgA, required for rapid prey-invasion. Bd2492, RomRBd and CdgA localize to the invasive pole and may facilitate MglA-docking. Bd2492 was encoded from an operon encoding a TamAB-like secretion system. The TamA protein and RomR were found, by gene deletion tests, to be essential for viability in both predatory and non-predatory modes. Control proteins, which regulate bipolar T4P-mediated social motility in swarming groups of deltaproteobacteria, have adapted in evolution to regulate the anti-social process of unipolar prey-invasion in the “lone-hunter” Bdellovibrio. Thus GTP-binding proteins and cyclic-di-GMP inputs combine at a regulatory hub, turning on prey-invasion and allowing invasion and killing of bacterial pathogens and consequent predatory growth of Bdellovibrio.
| Bacterial cell polarity control is important for maintaining asymmetry of polar components such as flagella and pili. Bdellovibrio bacteriovorus is a predatory deltaproteobacterium which attaches to, and invades, other bacteria using Type IV pili (T4P) extruded from the specialised, invasive, non-flagellar pole of the cell. It was not known how that invasive pole is specified and regulated. Here we discover that a regulatory protein-hub, including Ras-GTPase-like protein MglA and cyclic-di-GMP receptor-protein CdgA, control prey-invasion. In the deltaproteobacterium, Myxococcus xanthus, MglA, with MglB and RomR, was found by others to regulate switching of T4P in social ‘swarming’ surface motility by swapping the pole at which T4P are found. In contrast, in B. bacteriovorus MglA regulates the process of prey-invasion and RomR, which is required for surface motility regulation in Myxococcus, is essential for growth and viability in Bdellovibrio. During evolution, B. bacteriovorus has lost mglB, possibly as T4P-pole-switching is not required; pili are only required at the invasive pole. A previously unidentified tetratricopeptide repeat (TPR) protein interacts with MglA and is essential for prey-invasion. This regulatory protein hub allows prey-invasion, likely integrating cyclic-di-GMP signals, pilus assembly and TamAB secretion in B. bacteriovorus.
| Bdellovibrio bacteriovorus is a small, predatory deltaproteobacterium which invades other Gram-negative bacteria wherein it replicates. Bdellovibrio can encounter their prey by fast motility, driven by rotation of a single flagellum in liquid environments [1], [2], or by slow gliding motility on solid surfaces [3], but do not show social- or S-motility a process that is shown by other deltaproteobacteria (discussed below).
In Bdellovibrio invasion into the prey cell periplasm requires T4P, thus pilus-minus cells are incapable of host/prey-dependent (HD) growth and must be cultivated on artificial media as HI - host/prey-independent - cells [4], [5]. In flagellate HD Bdellovibrio the T4P are at the non-flagellar pole and prey-invasion occurs only from that anterior pole. On surfaces a flagellum is not present and the Bdellovibrio glide bidirectionally. Both HD and HI Bdellovibrio can glide and invade prey on surfaces. Our study began by examining the genetics of surface motility control in Bdellovibrio. This work led us to find that proteins known for surface motility control in a second deltaproteobacterium, Myxococcus xanthus, have evolved to control predatory invasion of bacteria by Bdellovibrio.
Regulation of surface motility in the deltaproteobacterium M. xanthus (which is always non-flagellate), has been well characterised by pioneering work of the Søgaard-Andersen [6], Mignot [7], Zusman [8], Hartzell [9] and Kaiser [10] groups for its two types of bidirectional surface motility. These are social (S)-motility, swarming movement of streams of cells using retraction of T4P at alternate poles of the cells; and adventurous (A)-motility, characterised by the movement of individual cells on a surface. A-motility (or gliding), is thought to be powered by cell envelope-spanning motor-protein complexes, [11], [12], though the precise mechanism of movement is still being revealed [13]–[15]. In M. xanthus, T4P localize to one pole at a time. Occasionally, M. xanthus cells reverse direction; this involves a switch in the polarity of the two motility systems, including a switch in the pole at which T4P assembly occurs. Thus, M. xanthus cells can assemble T4P at both poles but at any one time, T4P are found only at one pole [16].
Recent data suggest that the four putative gliding motor-gene operons in the B. bacteriovorus HD100 genome are evolutionarily linked to those A-motility gene clusters in Myxococcus [17], with subtle distinctive absences and additions likely reflecting Bdellovibrio morphology and gliding differences.
Bdellovibrio exhibits A-motility on surfaces in a gliding process that does not use T4P [3]. In this gliding, A-motility, individual Bdellovibrio cells move bidirectionally, cells can follow each other along previous paths and reversals of individual cells and re-orientations are seen. Gliding may be a particularly important mechanism by which Bdellovibrio explores biofilms and locates bacteria to prey upon [3], [18]. It is critical for HD Bdellovibrio to be able to explore or leave solid surfaces by gliding (when its flagellum cannot operate). Unlike other non-predatory bacteria, Bdellovibrio HD cells cannot replicate outside prey without acquiring “HI mutations” to do so [19], [20], thus without surface motility they could be trapped and starve. B. bacteriovorus gliding motility is slow, with cells moving, on average, 16 µm hr−1 [3] compared to the 24–36 µm hr−1 of Myxococcus [21]. Both B. bacteriovorus and M. xanthus show reversals in gliding direction. In Myxococcus, reversals during surface motility are known, from the work of the Søgaard-Andersen and Mignot labs, to be regulated by a Ras-like GTPase, MglA, which polarises the cell during gliding [6], [7], and GTPase-activating protein (GAP) protein MglB, which activates the GTPase activity of MglA to inhibit cellular reversals [6], [7]. MglA is important for activation of both the A- and S- motility “engines” (S motility engines are T4P), at the alternating leading pole, during bidirectional movements [6], [7]. In the absence of MglA, Myxococcus is both A- and S- non-motile. This means that MglA in M. xanthus, in conjunction with interacting partner RomR, regulates the localization/pole-switching activity of both T4P and gliding engine component proteins, in this bipolar bacterium. Chemotactic signals via the Frz system control cellular reversals in M. xanthus [22] via the RomR response regulator; RomR receives signals from the chemosensory Frz system and this modulates MglA activity [23], [24]. Although romR is conserved in Bdellovibrio, the genes encoding the Frz apparatus are not.
Bdellovibrio gliding is controlled by the bacterial secondary messenger cyclic-di-GMP. A diguanylyl cyclase (dgcA) mutation abolishes gliding, rendering Bdellovibrio cells unable to glide out of a consumed prey cell bdelloplast on a surface, even 2 hours after making lytic pores in it [25].The c-di-GMP receptor CdgA (GVNEF – a degenerate GGDEF protein) was found to be present at the predatory pole of B. bacteriovorus and deletion of cdgA slowed prey-invasion significantly, showing a link between c-di-GMP signalling and predation [25].
Whilst the B. bacteriovorus HD100 genome encodes MglA (Bd3734; accession: NP_970444.1), it does not encode an MglB homologue [23]. This report caused us to ask how bipolar switching might be achieved during Bdellovibrio gliding on surfaces; and whether the non-equivalent poles of the monoflagellate Bdellovibrio in liquids might correlate with an alternative role for MglABd. Here we show that MglABd is required for predatory invasion, as well as being associated with changes in gliding reversal behaviour in B. bacteriovorus, but is not required for gliding motility per se. This activity of MglABd occurs without an MglB partner, but in a cell with a RomRBd homologue. Both of these latter proteins are important to the control of bipolar motility in Myxobacteria. However we show that RomRBd has an essential role for growth in Bdellovibrio. We also report a previously undescribed interacting protein partner of MglA, and show that MglABd and RomRBd interact with this tetratricopeptide repeat protein (TPR) which is also required for predation. TPR is expressed from an operon that encodes a TamAB transport system and again TamA was essential for growth. Implications of this for predation and the onset of predatory growth upon prey-invasion are discussed.
Whilst MglAMx is involved in regulation of T4P-mediated social motility in M. xanthus, we show that MglABd is involved in Bdellovibrio in the control of pilus extrusion for the process of T4P-mediated invasion of prey cells at the single predatory pole. We show that a complex of proteins, additional to the T4P, is required at the ‘biting’ pole to organise the prey-entry machinery.
To investigate the role of MglABd, a deletion strategy was adopted screening for possible Bdellovibrio mutants in both prey/host-dependent (HD) and host-independent (HI) growth modes. All attempts to inactivate mglA in host/prey-dependent B. bacteriovorus HD100 were unsuccessful, despite screening many more cells than required to generate other Bdellovibrio deletion strains [26] (364 revertants obtained from second crossover events, but no deletion mutants, from three separate conjugations); suggesting that MglABd is essential for an aspect of the predatory life cycle.
Three host-independent (HI) ΔmglABd strains were obtained through sucrose-suicide counter-selection from a total of 76 screened. When challenged with prey, ΔmglABd HI B. bacteriovorus strains were unable to lyse E. coli in either a soft agar prey-lawn on the surface of YPSC plates, or in liquid culture (Figure 1A). Introduction of wild type mglABd by in cis complementation method (as described previously [25]) restored predation (Figure 1B) confirming that MglABd is essential for predatory growth.
The ΔmglA HI B. bacteriovorus strain could not reduce E. coli numbers in liquid culture, though this strain could still attach to the exterior of potential prey cells (Figure 2). A parallel assay showed that 43.5% of wild-type B. bacteriovorus HI cells attached to, or had entered, E. coli prey cells after 1 hour (Figure 2A) but no ΔmglA HI strain formed prey-bdelloplasts (Bdellovibrio cause the prey to round-up into ‘bdelloplast’ structures after invasion), even after 22 hours. Figure 2 also shows that both the ΔmglA HI and ΔpilA HI (Δbd1290, which is known to lack pili and is obligately host-independent [4]) could still attach to E. coli prey cells, albeit at a lower frequency. This suggests that pili are not a prerequisite for attachment, (although they are required for prey-invasion [4], [5]), and suggests that the ΔmglA HI predatory defect is not due to the inability of the Bdellovibrio cell to attach to prey cells.
The nature of the predatory defect of the ΔmglA HI strain was analysed further by microscopy, using a fluorescent E.coli S17-1::pMAL_p2-mCherry prey strain [27]. Addition of the ΔmglA HI strain to E.coli S17-1::pMAL_p2-mCherry and incubation for 22 hours demonstrated that although ΔmglA HI cells could attach to the outside of a prey cell, they could not invade to form bdelloplasts (Figure 2B). A wild-type HI B. bacteriovorus strain (HID50) successfully invaded E. coli cells and killed them (as shown in Figure 1) and at the 22 hour stage was shown to have formed bdelloplasts from 26.3% of the remaining E. coli, compared to zero bdelloplasts for the ΔmglA HI strain. Thus the deletion of mglABd abolished a process required for prey-invasion.
The ΔmglA HI strain showed a similar phenotype to that observed in a pilus-minus (ΔpilA) strain, which was known to be unable to invade prey cells [4]. We hypothesised that B. bacteriovorus ΔmglA might be defective in the synthesis or extrusion of pili, preventing prey cell invasion. This seemed plausible given that MglA regulates both the pole-switching of the A-motility and Type IV pilus-mediated S-motility systems in M. xanthus. Transmission electron microscopy of HI Bdellovibrio cultures grown to an OD600 of 0.2–0.3 showed that a wild type HI control had pili in 14.3% of cells, whilst ΔmglA HI had pili in only 2.3% of cells analysed (p = 0.02). These data suggested that MglABd regulates formation of pili; loss of mglA reduces the number of piliated cells. But, in contrast to the ΔpilA strain which completely lacks pilus fibres, the total inability of ΔmglA cells to invade, despite the presence of a low (but significant) frequency of piliated cells, suggests that these few pili present in the ΔmglA cells are not competent to facilitate invasion. This could be due to a defect in pilus retraction upon attachment to prey surfaces, or a requirement for another MglA-controlled factor to mediate invasion. Candidate MglABd-interacting proteins for invasive processes are discussed later.
Knowing that MglABd controls pilus-mediated bacterial invasion in B. bacteriovorus, but that in M. xanthus both pilus-mediated S-motility and gliding A-motility are MglA controlled, we used time-lapse microscopy to observe ΔmglA and wild-type B. bacteriovorus strains for gliding motility on 1% agarose/CaHEPES. Surface motility in B. bacteriovorus begins after a period of incubation on an agarose surface and allows exploration of environments for potential prey.
In contrast to recent studies in Myxococcus xanthus which showed that a ΔmglAMx strain is non-motile on surfaces [7], and a mglAG21V strain displays hyper-reversals during A-motility [6], we found that Bdellovibrio ΔmglA cells showed sustained gliding runs on surfaces (Figure 3), indicating that MglABd is not absolutely required for Bdellovibrio cells to glide.
A Bdellovibrio strain with C-terminally His8-tagged MglABd, expressed from the endogenous bd3734 promoter in cis, with a plasmid promoter-driven wild-type copy of mglABd, could be grown predatorily, in contrast to the ΔmglA strain which was non-predatory. In a previous study in M. xanthus, the presence of tagged MglAMx protein in conjunction with wild-type MglAMx allowed gliding to remain fully functional [7]. In contrast to the sustained gliding motility of the ΔmglABd strain (Figure 3A), the predatory B. bacteriovorus HD100 MglA-His8 showed increased reversals during gliding: on average 9.0 reversals hr−1 (n = 28), significantly more than wild-type HD100 cells with an average of 3.2 reversals hr−1 (n = 21) (p<0.001) (Figure 3B,C). The same hyper-reversal phenotype was also observed in B. bacteriovorus HD100 MglA-mCherry cells (data not shown).
MglABd (Bd3734) shares significant sequence similarity (Figure 4) with MglAMx (MXAN1925 accession: YP_630169.1), with 64% protein identity and 82% similarity (NEEDLE global alignment). The majority of residues shown to be important for MglAMx function [28], [29] are conserved in MglABd (Figure 4A–D).
The P-loop region (19GXXXXGKT26) of MglAMx was shown by Søgaard Andersen and co-workers to be important for GTP hydrolysis, and for MglA function [28], and substitutions in this region, such as G21V, were reported to decrease hydrolysis [6]. The P-loop region of MglABd contains a natural serine at residue 21; the corresponding G12S substitution in eukaryotic G protein Ras activates Ras protein [30], essentially locking the protein in a GTP-bound state, in the same way as a Ras G12V substitution. This suggests that MglABd exists in a permanently GTP-bound state. The G21-equivalent residue is a conserved glycine across 7 deltaproteobacterial genera (Figure S1A) which all also have a conserved mglB gene, though in Bdellovibrio the equivalent residue is a serine.
The difference at residue 21 in the MglABd sequence suggested to us a reason why we did not observe conservation of the gene encoding MglB in Bdellovibrio, as the GAP activity of an MglB would likely be ineffective on a permanently GTP-bound MglA protein such as that suggested by the MglABd sequence with S at position 21. We thought that it might also explain the lack of a Bdellovibrio Frz system [23], which stimulates motility reversals in M. xanthus, as a mutation, causing MglAMx G21V, bypasses the requirement for Frz for reversals in that deltaproteobacterium [6]. Thus we turned to examine the presence of mglB in the deltaproteobacterial relatives of Bdellovibrio. We also tested the conserved RomRBd protein, while also looking for other proteins, specific to Bdellovibrio, with which MglABd might interact. In M. xanthus, RomR is found at both poles of the cell and interacts with both MglA and MglB to link the Frz system to regulate polarity control [23], [24].
The majority of sequenced deltaproteobacteria genomes contain both mglA and mglB, and these are often co-transcribed at the same locus, including in M. xanthus where the MglBMX protein has an important role in motility [6], [7], [31]. Although the mglA gene product in B. bacteriovorus HD100 shares extensive sequence similarity with other MglA proteins, there is no mglB homologue in the HD100 genome, despite neighbouring genes (dnaX, recR, mglA and a DUF149-encoding gene) showing conserved synteny to other deltaproteobacteria that do have an mglB. The closely related B. bacteriovorus Tiberius [32] also lacks an mglB homologue. The predatory, invasive, marine bacterium Bacteriovorax marinus is also closely related to B. bacteriovorus, although the Bdellovibrio and Bacteriovorax genera have diverged separately from Myxobacteria. A 16S rRNA phylogenetic tree of the deltaproteobacteria shows the ancestral lineage leading to Bdellovibrio and Bacteriovorax diverged from the ancestral lineage leading to the clade including Myxococcus xanthus [33] and in that divergent Bdellovibrio branch we detect mglB loss (Figure S1A). We found that in B. marinus, which also has an mglA gene (BMS_0054), there is an adjacent putative mglB homologue (BMS_0053), both genes lying downstream of recR (Figure S1B).
BMS_0053 shares only limited sequence similarity with other MglB Roadblock domain proteins (BMS_0053, 168 residues, shares 22% identity and 43% similarity (NEEDLE global alignment) with M. xanthus MglB protein, 159 residues, Figure S1C). This highly divergent MglB homologue in Bacteriovorax is likely still functional, since no frameshift or nonsense mutations have arisen in the B. marinus lineage, and protein sequence length is conserved; however, its function is unclear. We are unable to test whether mglB is under positive selection (dN/dS>1) in Bacteriovorax because synonymous substitution rates are saturated for available sequence comparisons (dS>2). The Bacteriovorax MglA homologue is much more conserved (66% identity and 83% similarity to MglABd) and may function in an analogous predatory role to that of B. bacteriovorus.
As MglABd had both similarities and differences to MglAMx, we sought to identify proteins that interact with MglA homologue Bd3734 in B. bacteriovorus as we reasoned that these proteins might have a predatory role. We used a pull-down co-purification assay with proteins from the predatory B. bacteriovorus strain producing MglABd with a C-terminal His8 tag from the endogenous mglABd promoter, mentioned above. For the co-purification assay, a host-independent isolate of the MglABd-His8 strain was used, as previous array data showed that mglABd transcription is up-regulated in wild type HI cells, (which remain predatory but are longer than attack phase Bdellovibrio). Whole cell lysates of this HI strain were used in the assay, in which the bait His-tagged protein MglABd binding to TALON-NX cobalt-charged resin allowed interacting proteins to be identified (Figure S2) that were not present in the control without the His-tag.
MglABd co-purified with Bd2492 (accession: NP_969302.1) (Figure S2) - a B. bacteriovorus protein with a hypothetical annotation, with predicted tetratricopeptide repeat (TPR) domains typically involved in protein-protein interactions. Bands were excised from the gel and analysed by LC-MS/MS. Corresponding regions of the wild-type HID13 control lane were also analysed, and neither MglABd nor Bd2492 were found in these regions, suggesting that MglABd and Bd2492 (TPRBd) interact in vivo.
The mglA ORF and bd2492 ORF were cloned into pUT18C and pKT25 vectors containing T18 and T25 fragments of adenylate cyclase, respectively [34]. The bacterial two-hybrid assay for MglA and Bd2492 showed a strong signal (Figure S3A–B) suggesting that the two B. bacteriovorus proteins interact. This interaction was supported by the observation that MglA co-purifies with His6-tagged Bd2492 in nickel-affinity chromatography of E. coli lysates heterologously expressing these two proteins from plasmid pD2492N/3734 (Figure S4). Gel filtration and SDS-PAGE of purified MglA and Bd2492-His6 indicated that the MglA-Bd2492 complex has an Mw of approximately 63 kDa and exists predominantly as a heterodimeric complex of 1∶1 stoichiometry (data not shown).
As the B. bacteriovorus mglA mutant was non-predatory, we tested whether bd2492 (encoding TPRBd) was essential for predatory growth. All attempts to inactivate bd2492TPR in host-dependent B. bacteriovorus HD100 were unsuccessful (68 revertants obtained from second crossover events, but no deletion mutants). Two host-independent (HI) Δbd2492 strains were obtained through sucrose-suicide counter-selection from a total of 10 screened. When challenged with prey, Δbd2492TPR HI strains were unable to lyse E. coli in liquid culture (Figure S5). As with the ΔmglA HI strains, the Δbd2492TPR HI strains could still attach to E. coli prey cells (attachment assay; 26.6% of E. coli cells had attached Bdellovibrio), but could not invade to form bdelloplasts (invasion assay; 0/389 E. coli cells). The B. bacteriovorus Δbd2492TPR HI strain was still able to glide on 1% agarose CaHEPES (data not shown).
TPR gene bd2492 is co-transcribed with bd2494 and bd2495 (Figure S6). The same gene synteny is also found in M. xanthus (MXAN_5763-5766) and B. marinus SJ (BMS_0137-140) (Figure 5) where the gene encoding a TPR domain protein is followed by genes encoding homologues of Bd2494 and Bd2495. In M. xanthus, the genes encoding homologues of Bd2492 and Bd2494 (MXAN_5766 and MXAN_5764) are interrupted by a gene encoding a putative Sec system ATPase, MXAN_5765.
B. bacteriovorus gene bd2492 encodes a hypothetical 353 amino acid tetratricopeptide repeat (TPR) protein; TPRpred (http://tprpred.tuebingen.mpg.de/tprpred) was used to predict TPR domains [35]. TPRpred confirmed that both BMS_0137 (524 residues; accession: YP_005034048.1) and MXAN_5766 (1031 residues; accession: YP_633903.1) are also predicted to contain TPR domains. All three TPR domain proteins do not have predicted signal sequences, as predicted by SignalP [36].
Bd2494 is a predicted transmembrane protein with a DUF490 domain. Both BMS_0139 (accession: YP_005034049.1) and MXAN_5764 (accession: YP_633901.1) also contain predicted DUF490 domains. Bd2495 is a surface antigen variable number repeat domain protein of the (outer membrane protein) Omp85 (TamA/BamA/YaeT) superfamily, hereafter termed TamABd; homologues of which are conserved in both B. marinus (BMS_0140; accession: YP_005034050.1) and M. xanthus (MXAN_5763; accession: YP_633900.1).
As mentioned in the introduction, RomRMx interacts with the MglAMX signalling system to regulate surface motility in response to Frz system signals [23], [24], but the Frz system is not conserved in Bdellovibrio. We assessed the interaction of the RomRBd (Bd2761; accession: NP_969553.1) with the MglA-interacting protein TPRBd (Bd2492) by bacterial two-hybrid (Figure S3A). RomRBd shares homology with the REC domain and C-terminal region of RomRMx, whilst the remainder of the protein is less well conserved (Figure S7). RomRBd and TPRBd interact in the BTH assay (Figure S3A, C). We found that RomRBd and MglABd interacted weakly, but not significantly (p = 0.18) (Figure S3B–C).
Fluorescent tagging of RomRBd and TPRBd with C-terminal mCherry revealed that both proteins are localised at only one pole of the cell. Co-incubation with E. coli prey cells confirmed that both RomRBd-mCherry and TPRBd-mCherry are found at the anterior, prey-interaction pole of B. bacteriovorus cells (Figure 6). Fluorescent tagging of MglABd with mCherry typically showed cells with diffuse fluorescence localization in cells directly after applying to 1% agarose/CaHEPES (i.e. not gliding) (Figure 6); 63% of HD100 MglA-mCherry Bdellovibrio had diffuse fluorescence, the remainder showing a unipolar focus (28.4%) or bipolar foci (8.6%).
We found earlier that the Bdellovibrio ΔmglA strain does not show a hyper-reversal or non-motility phenotype (Figure 3). Thus, the regulation of MglABd localization in the control of gliding reversals (in the absence of MglB and Frz) is likely to employ an alternative signalling system to that of M. xanthus. Previous work suggested that this could be c-di-GMP as we have shown [25] that lack of GGDEF protein Bd0367 DgcA abolished gliding exit from bdelloplasts.
We had also had previously noted a link between a c-di-GMP binding protein and prey-invasion in Bdellovibrio [25]. Degenerate GGDEF (GVNEF) protein CdgA, Bd3125 (accession: NP_969891.1), is located at the prey invading pole of B. bacteriovorus and lack of this polar protein causes a very significant slowing of prey-invasion with bdelloplast formation taking 40–90 minutes compared to 30–40 minutes for wild type [25]. We concluded in that paper that “CdgA organises processes at the Bdellovibrio “nose” that are crucial to rapid prey-invasion”. In our current study, we found that both RomRBd and TPRBd (though not MglA) interacted with CdgA in the bacterial two-hybrid assay (Figure S3), supporting this idea. Whether RomRBd has a role in the regulation of gliding motility will be the subject of a subsequent study, but our interaction data suggested a link between RomRBd and predatory growth (as ΔcdgA was affected in predation [25]), so we tested for a romRBd deletion strain.
Given the CdgA and TPRBd interactions found at the B. bacteriovorus invasive pole, we speculated that RomRBd would be required for prey-invasion. Attempts to delete romRBd, both predatorily (HD) and host-independently (HI), were unsuccessful (HD 104; HI 120 revertants screened), suggesting that RomRBd is required for both predatory and host-independent Bdellovibrio growth.
As RomRBd interacted, by BTH, with TPRBd, encoded in an operon with the tamAB genes, we speculated that the TamAB complex would also be required for predatory growth. Attempts to delete tamABd also proved unsuccessful (HD 140; HI 97 revertants screened), suggesting that TamABd is also essential for both phases of Bdellovibrio growth.
Here we report that B. bacteriovorus use homologues of adventurous/social motility-control proteins for the process of predatory invasion of other bacteria. Whilst non-invasive M. xanthus utilise the proteins MglA and MglB to control bipolar, bidirectional surface motility [6], [7] in Bdellovibrio MglABd has evolved to function without an MglB homologue (the mglB gene is absent) to regulate prey entry at a single pole.
There are three lines of evidence to suggest this: (1) The deletion of mglABd caused a non-prey-invasive phenotype (Figure 1) and severely reduced pilus formation on the cell surface; (2) the natural substitution in MglABd of serine for glycine (Figure 4) at the position equivalent to residue 21 in MglAMx suggests that MglABd exists in a permanently GTP-bound state, and is not involved in the GTPase cycle which is key to the alternate bi-polar switching of motility proteins in M. xanthus [6], [7]; (3) RomR-mCherry is unipolar in B. bacteriovorus (Figure 6), in contrast to its asymmetric bipolar localization in Myxococcus, controlling MglAMx positioning.
We had hypothesised that RomRBd might be involved in regulating pole activity to control gliding motility. As RomRBd was found at the predatory pole only, this suggested an alternative role. We could not detect a significant interaction between RomRBd and MglABd by BTH, but we did detect a significant interaction with Bd2492 TPR protein (Figure S3), which is also at the anterior pole (discussed later).
The RomRBd location at the anterior pole of B. bacteriovorus puts it where prey-invading T4P are located. Lotte Søgaard-Andersen's group showed that an mglAMx deletion mutant resulted in unipolar RomRMx, with RomRMx and T4P, (used in that bacterium for bipolar social motility), found at the same pole [37]. Sequence- and localization- differences between unipolar RomRBd and MglABd (in the absence of an MglB) in B. bacteriovorus, versus those in M. xanthus (which has MglB), might explain why T4P are only found at the anterior Bdellovibrio pole where they control prey-invasion.
Deletion of romRBd abolished Bdellovibrio growth in both HI and predatory conditions, but in M. xanthus romR is viable with abolition of gliding motility and reduction of T4P-dependent social motility [23], [24]. Thus RomRBd, which does show some sequence divergence from RomRMx (Figure S7), could be reporting T4P activity and prey-invasion, at the anterior pole, back to initiate Bdellovibrio growth. It should be recalled that predatory “attack phase” Bdellovibrio do not replicate outside prey, but initiate replication when prey are entered [20].
Our BTH interaction data were too weak to prove a significant interaction between RomRBd and MglABd. This could be interpreted to mean that RomRBd transiently docks with MglABd when RomRBd is complexed at the pole, that other partner proteins are required to contribute to this interaction, or that they do not interact, in contrast to published data for MglAMx [23], [24]. Our finding that RomRBd is unipolar fits with evidence in M. xanthus that MglBMX is required for bipolar localization of RomRMx [23] and the apparent loss of MglB from the prey-invasive Bdellovibrio lineage in evolution. The Bdellovibrio-like invasive B. marinus has a putative mglB gene, the product of which shows only limited sequence similarity to other MglB Roadblock domain proteins (Figure S1). This mglBBm gene is highly divergent from mglbMx but likely still functional. It may be undergoing selection to evolve an alternative function, while the B. marinus mglA gene is maintained for a predatory role analogous to that in B. bacteriovorus.
MglA and MglB were shown to be conserved by Keilberg and co-workers in many deltaproteobacteria but also occur in some evolutionarily distant bacteria such as the green non sulphur bacteria, Acidobacteria and Deinococcus-Thermus group, Figure S3 in Ref [23]. The authors calculated the following: out of a total of 70 species with (at least one) predicted MglA homologue 87% = 61/70 species have MglB and an MglA. Of the 9 without MglB, 4 bacteria had MglA G21 with no MglB; 5 had MglA A/S21 with no MglB. Of these 9 with no MglB, only B. bacteriovorus and one other species, (a soil Acidobacterium named Candidatus koribacteria versatilis), have predicted RomR homologues. Thus bacteria with RomR and MglA and B may have interacting protein complexes that move between poles; but our study on B. bacteriovorus is the first to examine the situation in a bacterium where MglA and RomR are present but MglB is not.
As mentioned above, we detected an interaction with an additional protein that could contribute to the localization of MglABd and RomRBd at the single prey-invasion pole of Bdellovibrio. This was with the unipolar tetratricopeptide repeat TPR protein, Bd2492 (TPRBd) shown using both His-tag pull-downs and BTH for MglA and BTH for RomR. TPRBd could sequester either MglABd or RomRBd at the prey-invasive pole, regulating their freedom to interact with each other, or promoting an interaction on the TPRBd surface. Deletion of bd2492TPRBd abolished prey-invasion in the same manner as ΔmglABd (Figure S5, Figure 1).
It was not possible to monitor localization of fluorescently tagged proteins informatively in the HI derivative strains of the non-predatory ΔmglABd and Δbd2492 mutants. This is because HI derivatives have pleomorphic cell morphotypes (HI cells naturally differ greatly in length and shape) [38], and indeed some long HI cells are predatory at both poles [25].
The bd2492 gene is located upstream of, and is co-expressed (Figure 5, Figure S6) in an operon with, gene bd2494, which encodes a transmembrane protein with a C-terminal DUF490 domain, homologous to the TamB component of the TamAB autotransporter-secretion system [39]. Bd2494 might dock with TPRBd at the prey-invasive nose. The last gene in the operon (bd2495) encodes a 7-POTRA (polypeptide-transport-associated)-domain, outer membrane protein (OMP) member of the Omp85 superfamily. The Omp85 protein family includes the BamA component of the BAM complex, known to receive and assemble beta barrel proteins during outer membrane growth [40]. The family also includes the TamA component of the TamAB complex, which aids autotransporter secretion [39]; and two-protein secretion (TPS) proteins [41].
The TamA and TamB genes are typically adjacent in proteobacteria [39], suggesting that the adjacent B. bacteriovorus bd2494-2495 genes encode a TamAB-like transporter.
Thus our finding that MglABd and RomRBd interact with a TPR protein (Figure S3), encoded from the 5′ gene of a tamAB-like operon, suggests that the Bd2494-2495 TamAB-like transport activity might be required for OMP/autotransporter proteins involved in predation. This may account for our observation that some pili are present on the ΔmglA mutant but that despite this, it does not invade due to an effect on TamAB-dependent predatory protein transport. Similarly, the Δbd2492 mutant was also non-predatory (Figure S5), but attached to prey. This suggests that either: TPRBd and MglABd are important in the positioning of proteins (probably Bd2494-5 TamABBd) at the predatory pole of the B. bacteriovorus cell to facilitate prey entry; or that binding of RomRBd and MglABd to TPRBd affects its activity, and that of the TamABBd complex, regulating predatory protein secretion.
Reinforcing our observation (mentioned above) that RomRBd is essential, we found, by attempting to delete bd2495, that TamABd was also essential for both HD and HI growth of Bdellovibrio. This suggests that the activity of the TamAB complex (possibly involving a TPR-mediated interaction with RomRBd) is required for secretion of proteins required for prey-invasion and both predatory and HI growth. Potential candidates for TamAB export are proteins involved in the synthesis/secretion/maturation of extracellular polysaccharide (EPS) or polyelectrolytes; an earlier study proposed that RomR was responsible for stimulating polyelectrolyte secretion in M. xanthus [37]. We cannot yet define whether RomRBd activates a TamAB dependent process that is essential for predatory and HI growth, or whether it reports on the activity of a TamAB complex, via its interaction with TPRBd, to regulate Bdellovibrio growth. This will be the subject of a further extensive genetic study.
Although a TPR protein interaction with MglAMx or RomRMx has not been previously reported, the MXAN_5766 gene encoding a TPR domain protein, from a gene cluster with similar synteny to the B. bacteriovorus bd2492-2495 genes (Figure 5), has previously been implicated in M. xanthus S-motility by transposon studies carried out by the Hartzell group [42]. The low percentages of TPR ORF similarity/identity between MXAN_5766 and Bd2492 could reflect the greatly different protein sizes and may indicate interactions with additional protein partners in M. xanthus. However, in M. xanthus, similar TPR interactions with RomR, MglA and TamAB-like proteins could play a role in bipolar motility control. Whether or not this is the case in M. xanthus, it is clear that TPR, and likely TamABBd, proteins play an important role in defining the single, active, predatory pole of Bdellovibrio.
We propose a predatory regulatory ‘hub’ of proteins at the B. bacteriovorus prey-invasive pole (Figure 7), with the TamAB-associated Bd2492-5 TPRBd protein complex involved in the organisation/assembly of OMPs or autotransporters at the predatory pole. This is reflective of TamAB protein functions in other bacteria (discussed in [39]). Such protein secretion could facilitate predation directly or produce other extra-cellular compounds such as EPS or polyelectrolytes, as mentioned above, which contribute to predatory invasion. Predatory proteins could be secreted in outer membrane vesicles (OMVs); Sar and Arf GTPases (homologous to MglA) have functions in vesicle transport [43] and M. xanthus vesicles likely have an extra-cellular predatory role in the “wolf-pack” [44] hunting of M. xanthus [45]. Our studies show that the directed prey-invasion of Bdellovibrio requires a protein encoded by a tamAB operon, suggesting synergies in TamAB-mediated predation and cell interaction processes of B. bacteriovorus and M. xanthus which is worthy of further investigation. Regulatory protein hubs are reported to control pili and flagella in other bacteria [46].
Considering evolutionary differences that led Bdellovibrio to prey-invasion via a single pole, we also suggest that the absence of mglB in B. bacteriovorus (and the high degree of divergence of this gene in B. marinus), is because MglB is no longer required for pole switching of pili: B. bacteriovorus pili are found at only one - the non-flagellar, prey-invasive pole [4]. This is also concordant with B. bacteriovorus cells being incapable of S-motility (which would require pole-switching of T4P) and instead using T4P at a single pole for prey-invasion. However, the absence of an MglB homologue does suggest that an alternative mechanism for regulating reversals during gliding motility is likely to exist.
The mechanism by which reduced incidence of pili or a change in their retraction state is caused, in the B. bacteriovorus ΔmglABd strain, remains to be determined. Capeness and coworkers have recently shown that regulation of Bdellovibrio pilus retraction status does correlate with prey-invasion [26]. Pilus retraction occurs through secretin PilQ [47], which is required for predation in B. bacteriovorus [18]. The OM-assembly of a pilus-biogenesis protein such as PilQ could be affected by the Bd2492-5 TamAB complex activity. Alternatively, OMPs required for secretion of EPS might be perturbed at the Bdellovibrio pole, preventing pilus retraction; EPS is required for pilus retraction in M. xanthus [48]. These considerations will be the subject of a subsequent study.
The MglA/RomR-TPR interactions reported in this paper may have evolved from ancient interactions common to ancestors of M. xanthus and Bdellovibrio, and are now used in B. bacteriovorus for prey-invasion control. They may also underlie the motility and “wolf-pack” predation of Myxobacteria, but the function of the M. xanthus TPR protein homologue remains to be explored. Pioneering work by Mignot/Theodoly has shown that adhesion during gliding motility is mediated by slime deposition [14], [15] on a solid surface and that gliding directionality is controlled by MglAMx [6], [7] and other interacting proteins. In nature gliding of M. xanthus may occur on top of prey bacterial biofilms and we hypothesise that the Bd2492-5 TamAB associated system may have a role in producing vesicles, not only for gliding, but to damage prey cells as part of the M. xanthus wolf-pack lytic process.
In M. xanthus, chemotactic phospho-transfer signalling, involving Frz proteins, governs the localization of soluble RomRMx, MglAMx and MglBMx proteins to alternately activate or deactivate each cell pole for surface-motility directionality [23], [24]. In B. bacteriovorus, we detected an interaction between RomRBd and the CdgA GVNEF domain c-di-GMP binding protein (Figure S3) which has been shown to affect prey entry [25]. There is no Frz system in Bdellovibrio [23] but our finding that CdgA binds RomRBd (Figure S3) suggests that this c-di-GMP signalling pathway could contribute to RomRBd localization in the control of the prey-invasive pole. Further work is underway to define any signalling-link to RomRBd and CdgA from our previous observations that c-di-GMP synthases control gliding motility, predation and the switch from predatory to host-independent growth [25].
The data we present here show how the “phenotype space” and function of B. bacteriovorus MglA has diverged from that in M. xanthus. MglABd functions in the control of unipolar prey-invasion: a critical process in the predatory lifecycle of B. bacteriovorus. Our present observations indicate (Figure 7) that MglABd, RomRBd and the interacting TPR-domain protein TPRBd and TamABBd complex act at a single pole in B. bacteriovorus to facilitate prey-invasion via a mechanism that has diverged from that which controls M. xanthus S-motility.
Bacterial strains and plasmids used are listed in Table S1. Primers used for gene manipulation or PCR amplification are listed in Table S2.
Markerless deletion strains of mglABd and bd2492 (encoding TPRBd) were generated using a modified technique of that of the Pineiro lab [49], and as described previously [25]. Construction of each mutant is described in full in Text S1.
Fluorescent protein tags were generated as described previously [25] by cloning of a whole gene fused to mCherry at the 3′ end. Construction of each tag is described fully in Text S1.
To observe the fluorescence of B. bacteriovorus mCherry-tagged strains during attachment to E. coli prey cells, 1 ml of a B. bacteriovorus predatory culture (containing 2.5×108 pfu ml−1) was concentrated 20-fold and added to a microcentrifuge tube containing 30 µl CaHEPES and 40 µl E. coli S17-1 pZMR100 (from a culture grown for 16 hours at 37°C 200 rpm in YT broth supplemented with Km50) diluted to OD600 2.0 in CaHEPES, before incubating at 29°C for 5 minutes to allow attachment to occur. Cells were immobilised on a 1% agarose/CaHEPES pad and images were taken on using a Nikon Eclipse E600 epifluorescence microscope with a 100× objective lens and an hcRED filter (excitation 550 to 600 nm; emission 610 to 665 nm) with a Hamamatsu Orca ER camera. Images were analysed using Simple PCI software (version 5.3.1 Hamamatsu).
Procedures for attachment, invasion and predation assays of HI Bdellovibrio cells on E. coli prey are described in Text S1. 3 biological replicates were performed.
B. bacteriovorus gliding motility was observed on 1% agarose/CaHEPES by timelapse microscopy as previously described [3]. Briefly, 1 ml of an predatory culture (containing 2.5×108 pfu ml−1) was concentrated 10-fold (HI cultures were not concentrated) and 8 µl was spotted onto the agarose pad. Measurements of gliding reversals were calculated after cells had been gliding for >1 hr.
To analyse percentages of piliated cells, each HI strain was back-diluted and grown to OD600 0.1–0.5 in PY broth at 29°C 200 rpm. Cells were then stained with 2.0% phosphotungstic acid (PTA) on carbon formvar copper grids (Agar Scientific) and analysed for the presence/absence of a pilus structure, as described previously [26].
Procedures for bacterial two-hybrid and protein co-purification are described in Text S1.
|
10.1371/journal.pntd.0002032 | African Programme for Onchocerciasis Control 1995–2015: Model-Estimated Health Impact and Cost | Onchocerciasis causes a considerable disease burden in Africa, mainly through skin and eye disease. Since 1995, the African Programme for Onchocerciasis Control (APOC) has coordinated annual mass treatment with ivermectin in 16 countries. In this study, we estimate the health impact of APOC and the associated costs from a program perspective up to 2010 and provide expected trends up to 2015.
With data on pre-control prevalence of infection and population coverage of mass treatment, we simulated trends in infection, blindness, visual impairment, and severe itch using the micro-simulation model ONCHOSIM, and estimated disability-adjusted life years (DALYs) lost due to onchocerciasis. We assessed financial costs for APOC, beneficiary governments, and non-governmental development organizations, excluding cost of donated drugs. We estimated that between 1995 and 2010, mass treatment with ivermectin averted 8.2 million DALYs due to onchocerciasis in APOC areas, at a nominal cost of about US$257 million. We expect that APOC will avert another 9.2 million DALYs between 2011 and 2015, at a nominal cost of US$221 million.
Our simulations suggest that APOC has had a remarkable impact on population health in Africa between 1995 and 2010. This health impact is predicted to double during the subsequent five years of the program, through to 2015. APOC is a highly cost-effective public health program. Given the anticipated elimination of onchocerciasis from some APOC areas, we expect even more health gains and a more favorable cost-effectiveness of mass treatment with ivermectin in the near future.
| In 1995, the World Health Organization launched the African Programme for Onchocerciasis Control (APOC) with the aim to control morbidity due to the parasitic infectious disease onchocerciasis (river blindness). APOC aims to set up sustainable national control programs against onchocerciasis in 16 countries in sub-Saharan Africa, covering over 100 million people who are at risk for infection. The main control strategy is mass treatment with the drug ivermectin, which is donated by the pharmaceutical company Merck. Coordination of the mass treatment programs is made possible by financial contributions from donor and beneficiary countries. We estimated that between 1995 and 2010, APOC has had a huge impact on population health in sub-Saharan Africa, preventing 8.2 million years worth of healthy life from being lost due to disease and mortality, at a cost of about US$257 million. We predicted that this health impact will double during the subsequent five years, at a cost of about US$221 million. This makes APOC one of the most cost-efficient large-scale public health programs in the world. We may expect even greater health gains in the future, given the anticipated extension of the APOC mandate with the aim to eliminate infection where possible.
| Onchocerciasis is caused by Onchocerca volvulus, a filarial nematode restricted to human hosts. The adult female worms reside in subcutaneous nodules where they produce millions of microfilariae during their on-average ten-year life span [1]. The microfilariae are found predominantly migrating through the skin and eyes and are transmitted by biting flies of the genus Simulium (the vector), an obligatory part of the parasite's life cycle. Onchocerciasis is responsible for a considerable burden of disease, mainly because of visual impairment, blindness, disfiguring skin lesions, and severe itching, which are the results of continuous exposure to microfilariae. Most of the global burden of onchocerciasis (>99%) is found in sub-Saharan Africa. In the West African savanna, where onchocerciasis is of a severely blinding form (savanna type), fear of blindness previously led to abandonment of fertile river basins. However, by now, onchocerciasis has been largely eliminated from West Africa by the Onchocerciasis Control Programme (1974–2002), which relied on intense vector control and mass treatment with the drug ivermectin [2].
In the more central and eastern parts of Africa, where onchocerciasis is usually of the less blinding form (forest type), there was no control or control only at a limited scale until the inception of the African Programme for Onchocerciasis Control (APOC) in 1995. APOC is a morbidity control program scheduled to be active until 2015, requiring that by that year, participating countries support and coordinate control measures independently. Since 1995, APOC has mapped infection with O. volvulus in 20 countries [3] and has coordinated interventions in 16 countries where onchocerciasis is considered a public health problem (Angola, Burundi, Cameroon, Central African Republic, Chad, Congo, Democratic Republic of Congo, Equatorial Guinea, Ethiopia, Liberia, Malawi, Nigeria, South Sudan, Sudan, Tanzania, and Uganda), covering endemic areas inhabited by about 71.5 million people in 1995. APOC's main strategy is to implement annual mass treatment with ivermectin.
Ivermectin kills microfilariae and permanently reduces the production of microfilariae by adult female worms, slowing down transmission and preventing morbidity [4], [5]. Annual mass treatment with ivermectin is implemented through a community-directed treatment approach, empowering communities to take responsibility for ivermectin delivery and to decide how, when, and by whom ivermectin treatment is administered. Mass treatment with ivermectin is enabled by donation of the drug by the pharmaceutical company Merck through the Mectizan Donation Program. Furthermore, coordination of the program is funded by donor countries (through the World Bank) and national onchocerciasis task forces (including beneficiary governments and non-governmental development organizations). To demonstrate APOC's importance, validate the efforts of endemic communities and national task forces, and maintain commitment of all stakeholders, it is essential to establish the health impact and cost of APOC.
Here, we present the estimated impact of APOC on population health and the costs involved up to 2010, with extrapolated trends up to 2015. An impact assessment would ideally be based on observed trends of infection and morbidity, but such longitudinal data are of limited availability in APOC areas. We therefore estimated trends of infection and morbidity based on APOC data of pre-control levels of infection and history of mass treatment, and literature-derived associations between infection and morbidity and the effect of treatment on infection and morbidity. For our calculations, we used ONCHOSIM, an established micro-simulation model for transmission and control of onchocerciasis [6], [7].
The impact of APOC was estimated at project level (a project being an implementation unit for mass treatment with ivermectin), while taking account of the prevailing type of onchocerciasis (i.e., savanna versus forest or mixed forest/savanna, with different patterns of morbidity) and the project-specific history of control. Project populations were further stratified by endemicity groups, to take account of differences in the pre-control prevalence of morbidity (which is non-linearly associated with infection) and the potential impact of mass treatment (e.g., the impact is relatively lower in highly endemic areas due to more residual transmission between treatment rounds). We considered four endemicity levels: non-endemic (prevalence of onchocercal nodules in adult males <1%), hypoendemic (nodule prevalence ≥1% and <20%), mesoendemic (nodule prevalence ≥20% and <40%), and hyperendemic (nodule prevalence ≥40%).
We estimated the size of the population at risk for infection in the 107 geographical project areas covered by APOC, for the years 1995–2010 (see File S1). These estimates were based on records kept by community-appointed drug distributors, aggregated to the project level. From the same data, we took the reported number of individuals who were treated with ivermectin during mass treatment (File S1) and calculated the average therapeutic coverage of mass treatment in each project per calendar year (i.e., the fraction of the population at risk that was treated). Based on data from extensive pre-control mapping studies, we estimated the fraction of the population in the different endemicity categories and the mean pre-control infection level in each endemicity category (File S1).
For the years 2011–2015, we assumed that population size will increase according to the latest known national growth rate (as reported by the United Nations World Population Prospects, published 11 May 2010, accessed 24 October 2011). If therapeutic coverage in 2010 was already at or above 75%, we assumed that coverage in the years 2011–2015 will remain equal to that in 2010. For those few project in which this was not yet the case, we assumed that between 2011 and 2015, therapeutic coverage will be scaled up by 10 percentage points per year (conservative compared to reported coverage patterns in projects that started mass treatment between 1995 and 2010), to a maximum of 75% (conservative compared to the longest-running projects that reported stable coverage levels around 80% in 2008–2010).
For each unit of analysis (project, onchocerciasis type, endemicity), we simulated trends in infection, morbidity, and mortality in the ONCHOSIM model [6]–[8], considering the project-specific history of mass treatment (File S1). For each endemicity stratum, ONCHOSIM was calibrated so that it could reproduce the average pre-control level of infection (File S1). Furthermore, ONCHOSIM was calibrated to reproduce the association between the prevalence of infection and morbidity (visual impairment, blindness, and itch) as estimated by analysis of literature data (File S1). Based on previous studies with ONCHOSIM, we assumed that ivermectin instantly kills all microfilariae and permanently reduces the capacity of adult female worms to release microfilariae by 35% in treated individuals (with cumulative effects for repeated treatments) [4], [7]. Individual participation in mass treatment was assumed to depend on age, sex (pregnant women and children under the age of five were assumed to be excluded from treatment), random non-compliance (i.e., temporal factors), and systematic non-compliance (i.e., fixed individual factors other than age and sex e.g. inclination towards participation). Systematic non-compliance was assumed to play a larger role when overall treatment coverage was lower (i.e. when there is lower inclination to participate in general), and vice versa [6], [8]. No simulations were performed for hypoendemic areas, as ONCHOSIM predicts that transmission of infection is unsustainable without migration of infected flies and/or human, and information on migration was lacking. Instead, we assumed that the prevalence of infection and morbidity in hypoendemic areas was 1/3 of that in mesoendemic areas, both pre-control and during control. For non-endemic areas, we assumed that prevalence of infection and morbidity was always zero.
We combined the predicted trends in prevalence of infection, morbidity, and mortality with information on the number of people at risk, yielding an estimate of the absolute number of cases of infection, morbidity, and deaths in each stratum. After aggregation of these results over all APOC projects, we calculated the burden of disease in terms of disability-adjusted life years (DALYs), which in our case is the sum of years lived in disability due to troublesome itch, visual impairment, and blindness, weighted by the loss of quality of life due to each symptom: 0.068, 0.282, and 0.594, respectively [9]; and years of life lost due to excess mortality from blindness (File S1). Every incident case of blindness was attributed 8 years of life lost, based on the average age of onset of blindness in ONCHOSIM, the associated life-expectancy (16 years) of a healthy person of the same age, and an estimated 50% reduction in remaining life-expectancy due to blindness (File S1). The estimated annual burden of disease was compared to the burden in a counterfactual scenario in which the pre-control prevalence of infection and morbidity did not change (i.e., as if there were no mass treatment), yielding an estimate of the averted disease burden. All DALY estimates in the present study are undiscounted.
We assessed the influence of uncertain model assumptions on the estimated health impact, by means of univariate and multivariate sensitivity analyses (File S1). In a univariate sensitivity analyses, we assumed extreme, though plausible parameter values for each of the selected parameters. In a multivariate sensitivity analysis, the analysis was repeated, based on 200 sets of random parameter values. Parameter values were randomly drawn from triangular distributions with modes equal to the values used in the main analysis, and minimum and maximum values equal to those used in the univariate sensitivity analyses. To arrive at a crude estimate of the uncertainty in the estimated health impact, the results of the multivariate sensitivity analysis were expressed as the 2.5 and 97.5 percentiles of results from 200 repeated analyses.
We estimated the financial costs for coordination of ivermectin mass treatment taken on by APOC and national onchocerciasis task forces (beneficiary governments and non-governmental development organizations), based on APOC financial reports for The World Bank, which acts as fiscal agent for APOC. Because governments of beneficiary countries will eventually have to finance and coordinate ivermectin mass treatment, costs were estimated from a program perspective, not accounting for community costs and costs of donated drugs. For the years 1995–2003 and 2010, cost data for national onchocerciasis task forces were not available and were assumed to be proportional to APOC expenditures by a factor based on data available for other years. Expenditures for 2011–2015 were estimated based on the expected number of treatments in that period multiplied by the estimated cost per treatment in 2010. All costs are reported in nominal values, by which we mean that the presented costs are the amounts that were actually spent (i.e. uncorrected for inflation, and undiscounted).
In 1995, the total population size in the APOC target area was 71.5 million (Figure 1), with 30% of the APOC target population living in hyperendemic communities, 31% in mesoendemic communities, 38% in hypoendemic communities surrounded by mesoendemic or hyperendemic areas, and 1% living in non-endemic communities. About 30% of the APOC population lived in savanna areas and 70% in forest or forest–savanna mosaic areas (Table 1). Before the inception of APOC in 1995, about 32 million people (45%) in APOC areas were infected with onchocerciasis, with 404,000 people (0.6%) blind because of onchocerciasis, another 889,000 (1.2%) suffering from visual impairment, and 10 million people (14%) suffering from troublesome itch. In the same year, a total of 1.6 million DALYs (22.8 DALYs per 1,000 persons) were lost due to onchocerciasis: 694,000 because of troublesome itch, 684,000 from blindness, and 251,000 due to visual impairment.
Mass treatment effectively started in 1997 (80,000 treatments) and was scaled up over the years, reaching an overall therapeutic coverage of about 73% in 2010 (75.8 million treatments; Figure 1). We estimated that the therapeutic coverage will increase to 78% by 2015 (92.5 million treatments). By 2010, about 65% of the population lived in areas subjected to 10–13 rounds of mass treatment, 17% in areas subjected to 6–9 rounds of mass treatment, 18% in areas subjected to 3–5 rounds of mass treatment, and less than 1% in areas subjected to only 1–2 rounds of mass treatment (Table 1). Cumulatively, about 500 million treatments with ivermectin were given between 1995 and 2010, with another 430 million expected to follow in the period 2011–2015. Considering the differences between projects in start year and patterns of scaling up of mass treatment, the prevalence of infection for APOC as a whole declined gradually and non-linearly over time, from 45% in 1995 to 31% in 2010, and to 18% in 2015 (Figure 2). Similarly, the prevalence of troublesome itch was reduced from 14% to 6% to 2%, and prevalence of visual impairment was reduced from 1.2% to 0.8% to 0.6%. Because of excess mortality among the blind and the fact that ivermectin prevented blindness in individuals who were already visually impaired, the prevalence of blindness declined more rapidly than that of visual impairment: from 0.6% to 0.3% to 0.2%.
In the counterfactual scenario without mass treatment, in which levels of infection and morbidity were stable, the absolute number of DALYs lost due to onchocerciasis would have increased over the years with population growth. In contrast, in the scenario that considers mass treatment with ivermectin, the absolute number of DALYs lost was predicted to decrease over the years. Due to these divergent trends, the number of DALYs averted by mass treatment with ivermectin was predicted to increase year by year (Figure 3). Overall, mass treatment with ivermectin averted 8.2 million DALYs between 1995 and 2010 (3.2 million due to itch, 4.4 million due to blindness, 0.6 million due to visual impairment). Moreover, we expect that APOC will avert another 9.2 million DALYs in the period 2011–2015, adding up to an expected total of 17.4 million averted DALYs by 2015 (Table 2). In relative terms, the disease burden of onchocerciasis was reduced from 22.8 DALYs per 1,000 persons in 1995 to 9.6 DALYs per 1,000 persons in 2010, and is expected to be further reduced to 5.0 DALYs per 1,000 persons by 2015.
Univariate sensitivity analyses identified the following parameters as having the most influence on the estimated health impact: the population at risk, pre-control levels of infection, and the associations between infection and itch and eye disease (Figure 4). The multivariate sensitivity analysis showed that the estimated cumulative number of DALYs averted could be up to 25% higher or lower, when we considered the separate sources of uncertainty simultaneously (6.0–9.8 million DALYS cumulatively averted by 2010, and 13.1–21.3 million DALYs cumulatively averted by 2015; Figure 4).
Between 1995 and 2010, coordination of mass treatment cost roughly US$257 million (Table 2), of which US$175 million was disbursed by APOC and US$82 million by national onchocerciasis task forces (cost of donated drugs and government salaries not included). Assuming that costs will rise proportionally with the number of treatments, mass treatment was expected to cost another US$221 million between 2011 and 2015, adding up to a total cost of US$478 million by 2015.
We estimated the health impact and cost of mass treatment with ivermectin for the 20-year period that APOC is scheduled to run as a morbidity control program (1995–2015). Our simulations suggest that mass treatment with ivermectin has markedly reduced the prevalence of infection with O. volvulus, troublesome itch, visual impairment, and blindness in APOC areas, averting an estimated 8.2 million DALYs due to onchocerciasis by 2010 at a nominal financial cost of about US$257 million (excluding cost of donated drugs). We expect that APOC will avert another 9.2 million DALYs between 2011 and 2015, at a nominal financial cost of US$221 million.
Our estimate of APOC's health impact only considered eye disease and troublesome itch, and would be even higher if other clinical manifestations of onchocerciasis would have been taken into account. For instance, disfiguring skin disease also contributes to the disease burden of onchocerciasis and is known to be reduced by ivermectin [10]–[13]. Further, epilepsy may be associated with onchocerciasis, as suggested by a growing but still uncertain base of evidence [14]. However, we chose to include only the most important disease manifestations for which data were available for model calibration (i.e., eye disease and troublesome itch). Furthermore, we did not include the effect of ivermectin on diseases that are co-endemic with onchocerciasis, such as soil transmitted helminthiases, ectoparasitic infections, and lymphatic filariasis [15]. Other minor factors leading to an underestimation of the health impact are that we only considered the effect of ivermectin on the capacity of adult female worms to release microfilariae and its microfilaricidal effect, whereas ivermectin may additionally have a modest effect on adult worm viability [16], [17]. Furthermore, we ignored between-village variation in coverage, which is perhaps most extreme in the phase of scaling up: in some projects, treatment started in a subpopulation with high coverage, while the other part of the population did not yet receive mass treatment (which is more efficient than treating the entire project population at an equivalent average coverage). We may have somewhat overestimated the number of life years lost due to excess mortality from blindness during and after mass treatment, causing a small underestimation in the number of DALYs averted. This is because we appointed a fixed number of life years lost to every new case of blindness, while regular ivermectin treatment is expected to postpone the onset of blindness to a higher age, reducing the number of life years lost due to blindness. Furthermore, we did not consider a possible association between excess mortality and (high) microfilarial load [18], [19].
There are several factors that may (partly) counterweigh the underestimation of the health impact of APOC described above. Therapeutic coverage may have been over-reported by community members responsible for the distribution of ivermectin, either because of incomplete estimates of the community population or to inflate their own performance. Yet, the estimated health impact of APOC by 2015 would decrease by only 0.8 million averted DALYs if we assume that coverage were to be systematically 10% lower than reported. Also, we ignored any mass treatment prior to the inception of APOC, whereas in reality, ivermectin distribution had already started in a limited number of foci (here morbidity levels had already been reduced somewhat, but not on account of APOC). Taking all above sources of under- and over-estimation into account, we believe that the true health impact of APOC is still slightly higher than our calculations.
The validity of our results, as in any simulation study, depends on the quality of the model and its assumptions. ONCHOSIM was first developed in the early nineties and has earned trust over the years from the large scale control programs. ONCHOSIM has been used to successfully mimic observed epidemiological data from various locations [4], [20]–[22], and has been used for policy making in the West-African Onchocerciasis Control Programme [7]. Efforts to validate the model continue. We have recently compared ONCHOSIM predictions to longitudinal data from Senegal and Gambia [23] and found that model-predicted trends in mf prevalence during 14 to 16 years of mass treatment were broadly consistent with the observed trends, although the mf prevalence sometimes seemed to decline slightly faster than predicted (unpublished data). Furthermore, our model predictions for trends in itch were comparable to the reported average trend in APOC sentinel areas [13]; after five to six years of mass treatment at 70–80% coverage, itch prevalence was reported to decline from 16% to 7%, and we predicted a decline from 14% to 6.5% for areas with similar pre-control levels of infection and history of mass treatment. Likewise, our model adequately reproduced trends in onchocercal blindness during vector control in West Africa (File S1). Although the above suggests that our model predictions are realistic, our estimates remain subject to uncertainty and it would be good to have them confirmed by more field data, especially regarding trends in morbidity during mass treatment.
Even though the model seems to be reliable, we should consider potentially important sources of uncertainty in our analysis. An often debated factor concerns the effect of ivermectin on adult worms. The univariate sensitivity analysis showed that the assumed treatment effects of ivermectin on the capacity of adult worms to release microfilariae influenced the estimated health impact only marginally. We did not study the effects of assuming no cumulative effects of ivermectin on worm fecundity, whereas it has been suggested that the latter may be the case [24]. However, if we had, ivermectin efficacy parameters would have been calibrated such that the model-predicted trends in mf prevalence and density were still in agreement with observed trends [4], [22], and therefore predicted trends in infection levels and morbidity should not have differed much from the current model's predictions. The sensitivity analysis showed that alternative assumptions for the effect of ivermectin on itch (the only reversible symptom under consideration) also influenced the estimated health impact only marginally. The most influential assumptions in our analysis were related to the estimated size of the population at risk, pre-control levels of infection, and the assumed associations between infection and morbidity, which were all based on data. Even though the multivariate sensitivity analysis suggested considerable overall uncertainty in our estimate of the health impact (±25%), the magnitude of the predicted impact was always large.
With an estimated 8.2 million DALYs averted in a 15-year period and a predicted doubling in the subsequent 5 years, the predicted health impact of APOC is impressive. According to our calculations, mass treatment against onchocerciasis cost about a nominal US$31 per undiscounted DALY averted between 1995 and 2010. According to World Health Organization guidelines [25], this is highly cost-effective, as it is below the per capita gross domestic product of most countries covered by APOC (27–1,545 international dollar per capita; Global Health Observatory Data Repository, accessed 2 August 2012). Furthermore, this cost-effectiveness is comparable to or even better than those for several other public health interventions. For example, the life-time cost-effectiveness of prophylaxis against mother-to-child transmission of HIV in a resource-limited setting has been estimated at US$52 per undiscounted DALY (incremental cost-effectiveness ratio of World Health Organization guidelines versus minimal standard of care) [26]. The cost-effectiveness of large-scale, long-term (30-year period) public health interventions targeting other neglected tropical diseases has been estimated at US$4–US$29 per DALY (mass drug administration against lymphatic filariasis), US$38 per DALY (case detection and treatment for leprosy), US$260 per DALY (vector control against Chagas disease), and US$48–US$303 (vector control against lymphatic filariasis) [27]. Mass treatment against onchocerciasis is of even better value (US$27 per DALY) if expected health gains and costs for the period 2011–2015 are included. In view of the anticipated elimination of infection so that mass treatment can be stopped altogether, the cost-effectiveness will be even better than our calculations suggest [23].
The objective of APOC is to establish country-led systems for onchocerciasis control by 2015, which means that countries and their partners will have to carry full financial responsibility by that year. Our results indicate that cost per treatment with ivermectin in APOC areas is affordable (US$0.51 per treatment, excluding cost of donated drugs) and comparable to the costs of existing national mass treatment programs for the elimination of lymphatic filariasis (US$0.06–US$2.23 per treatment) [28]. Mass treatment with ivermectin, however, also involves costs for society not covered by the program. From published data for two Nigerian communities, we derived that these costs are about US$0.23 per treatment (excluding start-up costs) [29]. Based on this estimate, the sum of program and community costs for mass treatment with ivermectin was approximately US$370 million from 1995 to 2010 and will be another US$320 million for 2011–2015. In addition to costs, there are significant benefits for society that countries need to take into account, such as prevented productivity losses resulting from blindness and itch. Blindness in rural Africa has previously been assumed to result in an annual productivity loss of US$150 per case [30]. Likewise, the productivity loss due to itch among coffee plantation workers in an Ethiopian site has been estimated at around US$5.32 per month per case [12]. Combined with our predictions of health impact, these figures suggest that by 2015, APOC will have averted a staggering US$2.2 billion due to productivity losses from blindness (US$517 million) and itch (US$1.7 billion, assuming productivity losses in 25% of people with itch). In other words, beneficiary countries should expect economic benefit from mass treatment that outweighs any costs.
Clearly, all of the above calculations apply only under the condition that countries do not themselves pay for the drug ivermectin. The amount of ivermectin donated up to 2010 represents a value of US$2.1 billion, assuming 2.8 tablets per treatment and a commercial price per tablet of US$1.50 plus US$0.005 shipping costs (personal communication with Dr. A. Hopkins, director of the Mectizan Donation Program). This amount is eight times the program costs for coordinating mass treatment. Likewise, for the period 2011–2015, the value of donated ivermectin will be an additional US$1.8 billion. Therefore, mass treatment with ivermectin can be sustained only with donation of ivermectin, which Merck has pledged to continue for as long as necessary.
We expect that levels of infection in the APOC target area will have fallen drastically by 2015 (overall prevalence of adult female worms 18%). The implication is that by that time, transmission of infection may be almost interrupted in areas with favorable conditions for elimination, such as high coverage of mass treatment, sufficient treatment rounds, and/or low to medium pre-control levels of infection [31]. Until recently, elimination of onchocerciasis from Africa was thought to be impossible by means of mass treatment alone, considering the large size of the transmission zones, mobility of the vectors and human populations, and poor compliance with mass treatment [32]. Following reports of elimination of onchocerciasis from foci in Mali and Senegal by mass treatment alone [23], however, interest has renewed in elimination of onchocerciasis from Africa [33]. Following this, WHO has recently been advised to extend APOC mandate by ten years to 2025 with the new aim of eliminating infection with O. volvulus, where possible. With this new motivation, we may indeed expect focal elimination of infection, resulting in even more health gains from mass treatment with ivermectin in the future and the possibility of being able to end mass treatment altogether.
According to our simulations, APOC has had a remarkable impact on population health in Africa between 1995 and 2010. This health impact is expected to double during the subsequent five years. Further, APOC is a highly cost-effective public health programs, and given the anticipated elimination of onchocerciasis from APOC areas, we expect even more health gains and a more profitable cost-effectiveness of mass treatment with ivermectin in the near future. Our study fully supports the advice to continue APOC activities for another ten years.
|
10.1371/journal.ppat.1002896 | CPSF6 Defines a Conserved Capsid Interface that Modulates HIV-1 Replication | The HIV-1 genome enters cells inside a shell comprised of capsid (CA) protein. Variation in CA sequence alters HIV-1 infectivity and escape from host restriction factors. However, apart from the Cyclophilin A-binding loop, CA has no known interfaces with which to interact with cellular cofactors. Here we describe a novel protein-protein interface in the N-terminal domain of HIV-1 CA, determined by X-ray crystallography, which mediates both viral restriction and host cofactor dependence. The interface is highly conserved across lentiviruses and is accessible in the context of a hexameric lattice. Mutation of the interface prevents binding to and restriction by CPSF6-358, a truncated cytosolic form of the RNA processing factor, cleavage and polyadenylation specific factor 6 (CPSF6). Furthermore, mutations that prevent CPSF6 binding also relieve dependence on nuclear entry cofactors TNPO3 and RanBP2. These results suggest that the HIV-1 capsid mediates direct host cofactor interactions to facilitate viral infection.
| In order to infect a host cell, HIV-1 must interact with and exploit cellular cofactors. Mutations within capsid, the protein shell that surrounds the virus, have been shown to affect virus usage of these cofactors and susceptibility to host antiviral proteins. However, with the exception of the Cyclophilin A-binding loop, there is no defined protein interface on the capsid that mediates interactions with cellular proteins. Here, we describe the identification of a conserved interface on HIV-1 capsid that binds the host protein CPSF6 and determines viral dependence on nuclear transport cofactors. This data illustrates how host-virus interactions allow HIV-1 to hitch a ride into the nucleus and reveals a potential new target for antiviral drugs.
| The HIV-1 genome enters target cells encapsulated in a fullerene capsid cone composed of capsid (CA) protein. The role of the capsid in early events of the HIV-1 replication cycle is not known, nor is it clear exactly how long the capsid remains associated with the infectious particle, or where the capsid disassembles. Early biochemical studies led to the view that the capsid is an inert shell, required during particle assembly and target cell entry, whereupon it rapidly falls apart (‘uncoats’) to permit reverse transcription [1], [2]. However, recent data suggest that uncoating may occur later than previously thought, either during transport of the reverse transcribing virus to the nucleus, or once the reverse transcribing virus has docked at the nuclear pore [3], [4], [5], [6], [7]. This accommodates the possibility that the capsid may facilitate transit of the core towards the nucleus by interacting with the cell's cytoskeletal transport system [5]. An intact capsid would also be expected to maintain a high stoichiometry of reverse transcriptase enzyme to viral template, which is necessary for overcoming the rate limiting steps in reverse transcription [4], [8].
Like all lentiviruses, HIV-1 is able to infect non-dividing cells, which requires the exploitation of active host cell nuclear import processes [9]. CA mutations have been identified that specifically affect nuclear entry in non-dividing cells [10], [11], [12], suggesting a link between CA and nuclear import. However, apart from the Cyclophilin (Cyp)-binding loop, there are no known interfaces through which the CA can interact with host cell cofactors. CA mutations outside of the Cyp-binding loop have been suggested to exert their effect by altering capsid stability or particle assembly. For example, it has been proposed that CA mutations that decrease the ability of HIV-1 to enter the nucleus (Q63A/Q67A) or infect non-dividing cells (A92E, G94D or T54A/N57A) do so by causing the capsid to uncoat faster or slower than wild type [10], [11], [12], [13]. However, many CA mutations that give clear infectivity defects are located on an exposed surface in the CA hexamer structure (Figure 1, [14], [15]) and therefore seem unlikely to purely affect capsid stability.
Recent genome-wide screens have implicated a number of nuclear import components as HIV-1 cofactors, including the nuclear pore proteins RanBP2 (also called NUP358) and NUP153 and the karyopherin TNPO3 [16], [17], [18]. In each case, there is evidence that the requirement for these import cofactors map to CA [6], [19], [20], [21], [22], [23], [24], [25]. Of particular interest, a single CA mutation (N74D) has been shown to affect the sensitivity of HIV-1 to depletion of RanBP2, Nup153 and TNPO3 [6], [19]. Mutation N74D arose spontaneously during passage of HIV-1 in cells expressing CPSF6-358, an artificially truncated version of cleavage and polyadenylation specific factor 6 (CPSF6, also known as CF Im) that perturbs HIV-1 nuclear entry [19]. CPSF6 is a pre-mRNA processing protein that dynamically shuttles between the nucleus and the cytoplasm [26] and contains a C-terminal nuclear-targeting arginine/serine-rich (RS-) domain [27], [28] of the type bound by TNPO3 [29], [30]. CPSF6 lacking this RS-domain is no longer confined to the nucleus but is also found in the cytoplasm [28]. CPSF6-358 (which is truncated at position 358 and therefore lacks the C-terminal RS-domain) was also found to be cytoplasmic and restricted HIV-1 before nuclear entry [19]. It is therefore significant that the HIV-1 CA N74D mutation not only allows escape from CPSF6-358 restriction but also results in the loss of viral dependence on several cofactors involved in nuclear entry (RanBP2, Nup153 and TNPO3).
Here we show that CPSF6 binds specifically to a novel protein-protein interface on the N-terminal domain of HIV-1 CA. We show that this interface is structurally and functionally conserved across lentiviruses from different genera and is accessible in the context of an intact CA hexamer. We propose that CPSF6 interacts with incoming capsid during the post-entry stages before uncoating. Structure-guided mutagenesis of this interface reveals that CA residues that mediate CPSF6 binding also mediate dependence on TNPO3 and RanBP2. Finally, addition of an ectopic nuclear localization signal (NLS) to CPSF6-358 recovers its nuclear localization and restores HIV-1 infectivity, suggesting that the functional outcome of CPSF6 interaction with HIV-1 depends on whether CPSF6 is trafficking into the nucleus or remaining in the cytosol. Together, our data reveals that HIV-1 CA possesses a previously undescribed, conserved protein-protein interface that dictates cofactor dependence. Furthermore, it suggests that HIV-1 uses CA to interact directly with host cofactors and exploit cellular nuclear import pathways.
To test our hypothesis that CPSF6 interacts directly with the HIV-1 CA we used a combination of biophysical, structural and cellular infection approaches. Expression of soluble and correctly folded full-length CPSF6 protein was not possible for technical reasons. However, it has recently been shown that CPSF6 residues 301–358 are sufficient for restriction in a TRIM-fusion assay and that conservation of residues 313–327 is necessary for full activity [31]. We therefore synthesised a peptide corresponding to this putative HIV-1-interacting region (CPSF6313–327) and tested binding to recombinant HIV-1 CA N-terminal domain (CAN) by isothermal titration calorimetry (ITC). We found that CPSF6 residues 313–327 were sufficient for direct binding to HIV-1 CAN (Figure 2A) with low affinity (362 µM). Next, we tested binding of CPSF6313–327 to HIV-1 CA mutant N74D, which escapes CPSF6-358 restriction. This single mutation all but abolished binding to CPSF6313–327 (Kd>5 mM) (Figure 2B). This suggests that CPSF6 binding to HIV-1 CA is specific and that N74D allows escape from CPSF6-358 restriction by preventing CA interaction with CPSF6-358. To determine whether CPSF6 binding is conserved across diverse lentiviruses, we measured the interaction between CPSF6313–327 and CANs from HIV-2, SIVmac and FIV. All three lentiviral CANs bound to CPSF6313–327, with an affinity of 219–350 µM (Figure 2C). Binding of CPSF6313–327 to HIV-2 and SIVmac agreed with published data showing that these viruses are restricted by CPSF6-358 [19]. Binding of CPSF6313–327 to FIV CAN was unexpected, given that FIV, like HIV-1 N74D, is insensitive to CPSF6-358 restriction [19]. The reason for this discrepancy is unclear, however it has previously been shown that N74D replicates with wild-type efficiency in HeLa cells whereas it does not replicate appreciably in macrophages, suggesting that CPSF6 may be required for HIV-1 replication in primary cells [6]. It is therefore possible that FIV similarly only requires CPSF6 to replicate in primary cells. This would agree with growing data that there are multiple nuclear entry pathways that may be redundant in certain cell lines [6], [19].
To understand how CPSF6313–327 binds directly to HIV-1 CAN, we solved the crystal structure of the complex between HIV-1 CAN and CPSF6313–327 at 1.8 Å resolution (Figure 3). In the complexed structure, CPSF6313–327 lies in a binding site comprising a narrow channel formed on one side by helix 4 and on the other by helices 3 and 5 and the helix 5/6 turn (Figure 3A). Three discrete pockets in the centre of the channel are filled by CPSF6 residues V314, L315 and F321 (Figure 3B). The channel is closed at one end around residue Q63 and extends the length of helix 4, until the beginning of the CypA-binding loop at V86 where it opens into solvent. The interface, as defined by CPSF6, is bordered by CA residues 53, 56–57, 66–67, 70, 73–74, 105, 107, 109 and 130 (Figure 3C). CPSF6313–327 itself does not possess any secondary structure but forms a relatively compact loop due to intramolecular interactions centering on the Q319 side chain, which hydrogen bonds to the amide nitrogen of F316 and the carbonyl oxygens of V314 and Q323, pinning the two halves of the peptide together (Figure 3D). Additional constraining intramolecular interactions are made between the peptide oxygen of F316 and the amide nitrogen of Q319, and between the peptide oxygen of P320 and the amide nitrogen of Q323. Formation of these interactions is facilitated by proline residues P317 and P320, which introduce kinks into the backbone, and by glycine residues G318 and G322, which confer backbone flexibility. The N and C termini of CPSF6313–327 project directly out of the binding channel (Figure 3E), suggesting that CPSF6313–327 is a protruding structure within the full-length CPSF6 protein. This supports a model in which CPSF6 residues 313–327 can access the CA interface in the context of intact, full-length CPSF6.
CPSF6313–327 is highly hydrophobic, containing only two polar residues (Q319 and Q323). Therefore, it makes a number of hydrophobic interactions with CA, including via V314, L315 and F321, which project into the channel at the centre of the binding interface (Figure 3B). In addition to hydrophobic burial of CPSF6 side chains, CPSF6 is also held in place by a number of hydrogen bonds between side chains in HIV-1 CAN and the backbone amide and carbonyl groups of CPSF6313–327, some of which are water-mediated (Figure 4). Significantly, the side chain of CA residue N74 makes a bifurcated hydrogen bond with the main chain of L315 in CPSF6 (Figure 4B), which explains why the N74D mutation resulted in loss of binding to CPSF6313–327 and escape from restriction by CPSF6-358 (Figure 2B and [19]). Two water-mediated interactions are also made between the backbone amide of V314 in CPSF6 and the main chain carbonyl of N74, and between the backbone carbonyl of V314 and the side chain of T107 (Figure 4B). CPSF6 makes two further interactions with side chains from helix 4 in HIV-1 CAN: one between the backbone nitrogen of G318 in CPSF6 and the CA Q67 side chain; and the other between the peptide oxygen of Q319 in CPSF6 and the CA K70 side chain (Figure 4C). CA N57 is another key interaction residue for CPSF6 binding. Similar to CA N74, the side-chain of N57 mediates a bifurcated hydrogen bond with the backbone of F321 in CPSF6. This positions the benzyl side chain of F321 for hydrophobic burial beneath the aliphatic side chain of K70. The identification of CA N57 as an important residue for CPSF6 binding is of interest, as mutation N57A impairs HIV-1 infection of nondividing cells [6]. Finally, several water-mediated interactions are made between the backbone of G322 and the side chains of N53 and Y130 and the main chain carbonyls of A105 and S109 (Figure 4D).
It is also worth noting the similarity between the location of the CPSF6-binding interface and exposed CA mutations that have been shown to affect infectivity (Figure 1). Many of the side chains that are directly involved in CPSF6 binding (N57, Q67, K70 and N74) were found to reduce HIV-1 infectivity in a comprehensive alanine-scan of CAN [14] (Figure 4 and Figure 5B and C). Furthermore, alanine-scan mutants that map to the CPSF6 binding site are distinct in that their reduced infectivity is not fully explained by structural or assembly defects. Mutants Q63A/Q67A and N74A have 5–35 fold decreased infectivity but normal levels of particle production and no assembly defects [14]. Similarly, while mutants T54A/N57A and K70A have fewer conical capsids, this was a minor defect (∼4-fold with respect to wild type HIV-1) compared to their effect on infectivity (which was reduced by 20–80 fold) [14], [32]. This lack of correspondence between magnitude of structural defect and loss of infection supports the conclusion that CPSF6 defines an interface in which residues have a role in mediating protein interaction necessary for optimal infection.
To address the accessibility of the CA:CPSF6 interface in the context of the hexameric CA, we superposed our HIV-1 CAN:CPSF6313–327 complex structure onto the recently solved structure of the HIV-1 CA hexamer (pdb: 3H47 [15]) (Figure 5). The monomers of the hexamer are arranged radially from a centre comprised of the packed N-terminal CA domains. The CPSF6-binding interface is found on the outside edge of the hexamer, where it is exposed to solvent and highly accessible for protein-protein interaction (Figure 5A). The CPSF6-binding interface is not involved in intra-hexamer or inter-hexamer interactions, the latter of which occur exclusively between C-terminal CA domains and build up the capsid lattice found in assembled virions (Figure 5D). This suggests that CPSF6 binding does not require the dissociation of subunits from the assembled capsid lattice. This suggests that binding of CPSF6-358 to the CPSF6 interface on capsid does not in and of itself restrict virus replication, for example by directly affecting capsid stability and uncoating, but rather competitively inhibits recruitment of endogenous CPSF6 and/or other cofactors necessary for productive nuclear import and integration.
To confirm that disruption of the CPSF6:CA interaction impacts on virus replication, we mutated residue F321 in CPSF6-358. The aromatic side chain of F321 forms extensive hydrophobic interactions with CA, suggesting that it may be essential for CPSF6:CA binding. We found that CPSF6-358 F321N was unable to restrict HIV-1, demonstrating that F321 is a key residue for restriction of virus (Figure 6A and B). Next, we investigated how loss of CPSF6:CA interaction alters known post-entry cofactor dependencies. Mutation N74D is located at the centre of the CPSF6-binding interface and abolishes binding of CPSF6313–327 to CA. N74D also results in loss of dependence on TNPO3, RanBP2 and Nup153, suggesting that the CPSF6-binding interface may be involved in HIV-1 nuclear entry [6], [19], [31]. To test this, we investigated whether there is a correlation between mutation of CPSF6-binding interface residues, binding to CPSF6313–327, and viral dependence on nuclear entry cofactors TNPO3 and RanBP2. Using our structure, we designed CA mutations with the aim of specifically knocking out CPSF6 binding. Five residues were selected for mutation (N57, Q67, K70, N74 and T107), on the basis that (1) they bind CPSF6 via their side chain and not their main chain and (2) are not obviously involved in maintaining CA structure. With the exception of N74D [19], all residues were mutated to alanine in accordance with previously published mutations [14]. We also made an additional M66F mutation in order to occlude the hydrophobic pocket filled by CPSF6 residue F321. Modelling of this mutant on to our structure suggested that the only F66 rotamer that would permit normal HIV-1 folding would be one that resulted in a steric clash with F321. We tested the effect of these mutations on in vitro affinity to CPSF6313–327 (Figure 6C) and the sensitivity of VSV-G pseudotyped HIV-1 to CPSF6-358 restriction and TNPO3 and RanBP2 depletion (Figure 6D). All of the mutants showed reduced CPSF6313–327 affinity and CPSF6-358 restriction, confirming that the mutations had acted to impair CPSF6 binding. Strikingly, all of the mutations also resulted in either the loss (N57A and N74D) or reduction (Q67A, K70A and T107A) of dependence on TNPO3 and RanBP2, a phenotype previously only shown for N74D and N57A [6], [19].
Although all CA mutations tested were found to reduce both the affinity of CA for CPSF6313–327 and the ability of CPSF6-358 to restrict, a direct correlation between the magnitude of the two was not observed. For instance, M66F reduced the affinity of CA to CPSF6313–327 by 7-fold but only recovered infection in the presence of CPSF6-358 by 3-fold. One possibility for this difference may be that the mutation has a reduced effect on binding in the context of hexameric or intact capsid. Indeed, in the hexamer model, helix 4 (where M66 is located) packs against helix 8 leading to differences in the orientation of side-chains around M66, such as Q63, with respect to the N-terminal capsid domain structure (Figure S1). Thus, although we predicted a rotamer conformation for M66F that would occlude binding in the N-terminal domain, rotamer occupancy may differ in the virion. Mutations Q67A and K70A abolished sensitivity to CPSF6-358 whilst only showing a minimal effect on binding to CPSF6313–327 (Figure 6C and D). This may be illustrative of the fact that CA mutations can affect both the stability of the capsid and the affinity of interactions at the protein-protein interface. Therefore, the effect of a capsid mutation on one process, such as uncoating, may mask the effect of the same mutation on another process, such as CPSF6-358 restriction, if the mutation leads to uncoating before binding of CPSF6-358 occurs. In support of this, Q67A is a mutant that is known to give rise to an unstable core [13], [14]. Other examples of this phenomenon exist; for instance, the unstable capsid mutant P38A is poorly restricted by TRIM5α inside cells [33]. Similarly, PF-3450074 shows significantly diminished inhibition of the unstable capsid mutant E45A, even though this mutation has no effect on affinity of the drug for HIV-1 particles [34]. In this respect, N74D is a particularly useful mutant as it has near wild type infectivity levels, but a dramatic reduction in affinity to CPSF6313–327 and restriction by CPSF6-358 and the best correlation between the two. Importantly, the direct correlation observed between escape from CPSF6-358 restriction and the lack of sensitivity to TNPO3 and RanBP2 depletion support a link between the CPSF6-binding interface in CA and the utilization of specific nuclear entry pathway components by HIV-1.
To provide further evidence that the CPSF6 interface is important we investigated its conservation in HIV-1 and other primate lentiviruses. Physiologically relevant protein-protein interfaces are more conserved than non-interacting surfaces [35]. Sequence mapping of ∼100 unique CAN sequences onto the complexed structure showed that the CPSF6-binding interface is highly conserved within HIV-1 CA (Figure 7A) suggesting that it is a functionally important interface required for efficient HIV-1 infection. Alignment of the CAN sequences of other primate lentiviruses reveals that the CA residues in HIV-1 that interact with CPSF6 are also highly conserved in both HIV-2 and SIVmac (Figure 7B). The two HIV-1 mutations with the greatest effect on both CPSF6 binding and restriction are N57A and N74D. To determine if these residues are functionally conserved, we introduced the equivalent mutations (N56A and N73D) into HIV-2 and SIVmac. As can be seen, N56A and N73D potently reverse CPSF6-358 restriction of both viruses (Figure 7C). This data suggests that the CPSF6 binding interface is highly conserved in HIV-1, HIV-2 and SIVmac.
CPSF6 is known to shuttle in and out of the nucleus [26] and contains a C-terminal nuclear-targeting RS-domain [27], [28] of the type bound by TNPO3 [29], [30]. We therefore investigated whether the link between dependence on CPSF6, TNPO3 and nuclear pore proteins is because binding of CPSF6 to HIV-1 facilitates nuclear entry. This hypothesis is suggested by the fact that deletion of the nuclear-targeting domain of CPSF6 results in a truncated cytosolic form (CPSF6-358) that reduces viral titre [19], [28]. Over-expressed cytosolic CPSF6-358 might act as a dominant negative, preventing the use of endogenous CPSF6 by HIV-1. We hypothesised that retargeting truncated CPSF6 to the nucleus by attaching a different NLS motif might prevent restriction and loss of titre. To test this, we determined the infectivity of HIV-1 in HeLa cells exogenously expressing full length CPSF6 (CPSF6-FL), the truncated form of CPSF6 (CPSF6-358) and HeLa cells expressing CPSF6-358 with the SV40 NLS sequence ‘PKKKRKVG’ at the C-terminus (CPSF6-358-NLS), and compared the subcellular localization of CPSF6 inside these cells. Whilst CPSF6-358 localized to both the cytosol and the nucleus, CPSF6-FL and CPSF6-358-NLS were entirely nuclear (Figure S2A). Furthermore, we observed that HIV-1 titre was reduced 6.3-fold in cells expressing CPSF6-358, whereas efficient infection was observed in cells expressing CPSF6-FL or CPSF6-358-NLS (Figure S2B and C). The recovery of efficient infection upon restoration of CPSF6 nuclear transport is consistent with a model in which CPSF6 is a cofactor for HIV-1 nuclear import. This result by itself does not rule out the possibility that CPSF6, when expressed in the cytosol, serendipitously binds to and inhibits HIV-1. CPSF6 may only inhibit HIV-1 if aberrantly localized in the cytosol, for instance in cells depleted of TNPO3. However, knocking out interaction with CPSF6 through a single point-mutation (N74D) results in a virus that loses dependence on TNPO3 and RanBP2 and that cannot replicate in macrophages, consistent with the hypothesis that CPSF6 is a cofactor for primate lentiviruses [6].
Several drugs have been identified that directly bind to HIV-1 CA [36], [37], [38]. In each case they are thought to inhibit viral replication by altering the stability of the capsid. Intriguingly, we observe that the most recently described drug, PF-3450074, binds within the CPSF6-binding interface [38]. Even more remarkably, one of the phenyl rings of the drug superposes almost exactly with the phenyl ring of CPSF6 residue F321, a critical residue for CPSF6-CA interaction (Figure 8A and B). Based on these data, we hypothesised that PF-3450074 may be a competitive inhibitor of a cellular cofactor, most likely CPSF6. To test this, we investigated whether PF-3450074 competes with CPSF6313–327 for CA binding and whether the drug occupies the same interface as CPSF6313–327, as defined by the CAN mutants used in this study. The synthesis for PF-3450074 has not been published nor was it possible to obtain the compound from Pfizer, therefore we performed a 4-step synthesis (as described in Methods) from which we obtained >400 mg of material at >95% purity. Similar to published data, we found that the drug bound to wild type HIV-1 CAN with an affinity of 5 µM (Figure 8C). PF-3450074 was also able to completely inhibit binding of CPSF6313–327 to CA (Figure 8D). The antiviral activity of PF-3450074 is therefore consistent with the hypothesis that CPSF6 is an important HIV-1 cofactor.
Binding experiments with different capsid mutants revealed that mutations N57A or K70A were sufficient to abolish PF-3450074 binding completely (Figure 8C). This is in agreement with the PF-3450074:HIV-1 CAN crystal structure, which shows that residues N57 and K70 form direct hydrogen bonds with the drug (Figure 8B). However, mutation of other key CPSF6 interface residues had little effect on drug binding; N74D and M66F reduced the affinity by 2- and 3-fold respectively, while Q67A had no effect despite Q67 forming a (weak) water-mediated hydrogen bond with the drug (Figure 8B and C). The reduced effect of the M66F mutant on the drug with respect to CPSF6313–327 (3-fold versus 7-fold) may be because the peptide places greater constraints on the flexibility of the binding pocket than the drug. Thus, M66F may be better at accommodating a small drug than a large peptide. One mutation, T107A, resulted in an increased affinity to the drug (Kd = 1 µM), possibly due to the removal of a slight steric repulsion between one of the aromatic moieties in PF-3450074 and the T107 side chain. These data show that PF-3450074 occupies only one pocket within a larger protein interface bound by CPSF6. Consequently, it may be possible to develop more effective high-affinity drugs by addressing this entire interface as a drug target, either by compound or fragment screening or rational drug design.
Although there is extensive experimental evidence that the HIV-1 capsid is more than a packaging device to carry viral protein and nucleic acid into the cell, no interaction interfaces other than the Cyp-binding loop have been identified on its surface. For instance, despite considerable effort the structural interface between capsid and the restriction factor TRIM5α remains incompletely characterized. Here we have described the identification of a conserved interface within the N-terminal CA domain and shown that interface mutations alter HIV-1 interaction with CPSF6 and cofactors RanBP2 and TNPO3. Previously, CA mutation N74D has been shown to escape CPSF6-358 restriction whilst simultaneously relieving dependence on nuclear transport factors such as TNPO3 and a functional nuclear pore [6], [19]. We have shown that CA residue N74 makes an essential interaction with CPSF6, both by solving the crystal structure of a CPSF6:CA complex and by showing that mutation N74D abolishes binding to CPSF6. Furthermore, mutation N74D results in a virus that has no defect in infectious titre on immortalized cells but that cannot replicate in macrophages [6]. This suggests that CPSF6 may be an important cofactor in HIV infection. Further evidence in support of CPSF6 as a cofactor is that addition of an ectopic NLS to CPSF6-358 restores both nuclear localization of CPSF6-358 and HIV-1 infectivity. However, we cannot rule out that this might be due to a reduction in the concentration of cytosolic CPSF6-358 that would otherwise prevent functional interaction of endogenous CPSF6 with CA.
CPSF6 is transported into the nucleus and contains an RS-domain, which is known to interact with karyopherins like TNPO3 [29], [30]. A compelling model for the role of the CPSF6 interface in HIV-1 replication is therefore that binding to CPSF6 facilitates active nuclear transport. Such a model would explain why mutation N74D results in concomitant loss of both CPSF6 interaction and dependence on TNPO3 and RanBP2; if the virus cannot bind CPSF6 then it cannot recruit TNPO3 to pass through the nuclear pore. We have further substantiated the connection between CPSF6 binding and TNPO3 and RanBP2 dependence through structure-guided mutagenesis of the CPSF6-binding interface. This has identified five CA mutations that have the same pleiotropic effects as N74D (N57A, M66F, Q67A, K70A and T107A). If CPSF6 is an HIV-1 cofactor then it would allow seemingly conflicting data that report different viral targets for TNPO3 requirement and interaction to be resolved: the requirement for TNPO3 maps to CA [23], but TNPO3 has been found to bind integrase (IN) and not CA [17]. Recruitment of TNPO3 to CA-bound CPSF6 could explain why CA determines TNPO3 requirement, while also accommodating a role for IN as the direct viral binding partner of TNPO3. In this context, it is important to note that the interaction of CPSF6313–327 with CA, whilst specific, occurs with weak affinity. The affinity of the full-length protein for assembled virions is presumably significantly higher. The low affinity we have measured may be augmented by avidity, as there are many potential CPSF6 binding sites per virion, and CPSF6 is known to be part of a heterotetrameric protein complex together with CFIm25 [39]. Alternatively, the addition of other proteins such as TNPO3 may stabilize the CPSF6:CA complex. For instance, it is possible that a larger complex comprising CPSF6, CA, TNPO3 and/or IN exists inside the cell and future structural investigation of this possibility is likely to be highly informative.
Recent findings suggest that HIV-1 may utilize flexible nuclear import pathways [6], [19], [20]. Redundancy in HIV-1 infection is conceptually appealing as it provides a mechanism for viral escape from host immunity and effective zoonosis. For instance, the RS-domain of CPSF6 may be recognised by more than one karyopherin. Likewise, other RS-domain containing proteins may use the CPSF6 interface. This may be helpful for the virus but it adds to the complexity of investigating cofactor dependence. In this respect, interface mutations may be particularly useful to unpick which factors operate in a shared pathway and which are redundant. The seeming interdependency of multiple host factors on a single capsid mutant, N74, suggests that they operate in a single pathway, which the virus utilizes for efficient infection. Given the host factors involved, this single pathway most likely involves nuclear import of the virus.
Irrespective of the role of CPSF6 itself, the conservation and location of the CPSF6 interface, together with the effects of mutations that disrupt it, suggest that it plays an important role in HIV-1 infection. The importance of the CPSF6 interface in HIV-1 infection is supported by the fact that random drug screening recently resulted in the discovery of a drug, PF-3450074, that inhibits infection and mimics very closely the core F321 residue of CPSF6 (Figure 7A) [38]. The CPSF6-binding interface on HIV-1 CA possesses several important attributes that make it an ideal antiviral drug target, in that it is highly conserved, functionally important and druggable. Since PF-3450074 occupies only a subset of the entire CPSF6-binding site, it is unlikely to be as effective a drug as one that inhibits the entire interface. Our complexed HIV-1 CAN:CPSF6313–327 crystal structure provides a molecular delineation of the CPSF6 interface that may be useful in the development of antiviral therapeutics.
HIV-1 and HIV-2 CAN were expressed in BL21 (DE3) E. coli cells and purified as described (price et al). SIVmac and FIV CAN were expressed with an N-terminal His tag in BL21 (DE3) E. coli cells and purified by capture on Ni-NTA resin (Qiagen) followed by gel filtration. All HIV-1 CAN mutants were purified as per the wild type protein.
Proteins were prepared by dialysis against a buffer containing 50 mM potassium phosphate (pH 7.4), 100 mM NaCl and 1 mM DTT. The chemically synthesized CPSF6313–327 peptide (Designer Bioscience) was dissolved in the same buffer. ITC experiments were conducted on a MicroCal ITC-200 as described [38], with CPSF6313–327 (10 mM) in the syringe and CAN (600 µM) in the cell, unless otherwise indicated. Drug PF-3450074 was synthesized in-house and binding to CAN proteins carried out with protein (200 µM) in the syringe and drug (30 µM) in the cell. Data were analyzed using Origin data analysis software (MicroCal).
Crystals of HIV-1 CAN:CPSF6313–327 grew at 17°C in sitting drops. Protein/peptide solution (0.37 mM HIV-1 CAN and 4 mM CPSF6313–327 in 20 mM HEPES pH 7, 50 mM NaCl, 1 mM DTT) was mixed with reservoir solution (20% w/v PEG 3350, 0.2 M potassium phosphate dibasic) in a 1∶1 mix, producing 0.55 mm×0.15 mm×0.05 mm crystals within one week. Crystals were flash-frozen in liquid nitrogen and data collected on an in-house Mar-345 detector to a resolution of 1.8 Å. Crystal data collection and refinement statistics are provided in Table S1. The dataset was processed using the CCP4 program suite [40]. Data were indexed and scaled in MOSFLM and SCALA, respectively. The structure was determined by molecular replacement in PHASER using HIV-1 CAN (pdb: 2GON) as a model. Structural figures were prepared using PyMOL (MacPyMOL Molecular Graphics System, 2009, DeLano Scientific LLC).
HeLa cells were transfected with EXN-based expression plasmids containing HA-tagged CPSF6 constructs and transduced cells were selected with 1 mg/ml G418 (Gibco). Gene expression was confirmed by western blot using α-HA monoclonal antibody 16B12 (Covance). HeLa cells stably depleted for TNPO3 or NUP358 were made using short hairpin sequences expressed from MLV vector pSIREN RetroQ (Clontech) as described [6] and depletion confirmed using mouse TNPO3 antibody ab54353 (Abcam) and a NUP358 antibody kindly given by Frauke Melchior. VSV-G pseudotyped GFP-encoding lentiviral vectors based on HIV-1 NL4.3 were prepared in HEK 293T cells, as described [41].
Cells were seeded in 6-well plates at 1×105 cells/well and inoculated with GFP-reporter virus in the presence of 5 µg/ml polybrene. The virus dose was selected so as to infect ∼30% of unmodified cells and the percentage of GFP-positive cells enumerated 48 h later by flow cytometry. Unless otherwise indicated, experiments were performed in triplicate and one representative experiment is shown in each case. Titers are plotted as infectious units per ng of reverse transcriptase activity ± standard deviation.
Cells were plated on glass coverslips, washed with PBS and fixed with 4% PFA in PBS before being permeabilized with 0.5% Triton in PBS for 10 min at room temperature, washed with PBS and then blocked with 5% BSA in PBS containing 0.1% Tween (PBST) for 1 h at room temperature. Cells were incubated for 1 h with the first antibody (α-HA 16B12) at 1∶250 dilution, washed three times with PBST and then incubated for 1 h with the secondary antibody (Alexa-488 conjugated anti-mouse IgG (Invitrogen)) at 1∶400 dilution. Coverslips were mounted onto glass slides using Vectashield mounting medium with DAPI (Vector Labs) and imaged using a Zeiss 780 confocal microscope equipped with a 63×/1.4 NA Plan-Apochromat oil-immersion objective. Images were taken under identical conditions to aid comparison. Images were prepared using ImageJ (NIH).
PF-3450074 was obtained in a 4-step synthesis as described in the Supplementary Methods (Text S1). Small molecule LC-MS was carried out using the Agilent system using a Phenomenex Jupiter 150×2 mm, C18, 5 µm column. Variable wavelengths were used and MS acquisitions were carried out in positive and negative ion modes. Kieselgel 60 F-254 commercial plates were used for analytical TLC, UV light and/or potassium permanganate stain was used to follow the course of the reaction. Flash chromatography (FC) was performed with silica gel grade 9385 pore size 60 Å, 230–400 mesh. The structure of each compound was confirmed by 1H & 13C NMR (400 & 100 MHz, Bruker spectrometer). Mass spectra were obtained on an Agilent 1200 series LC-MS system.
Protein Data Bank: Coordinates for HIV-1 CAN:CPSF6313–327 have been deposited (PDB ID code 4b4n).
|
10.1371/journal.pgen.1006040 | Testing Rare-Variant Association without Calling Genotypes Allows for Systematic Differences in Sequencing between Cases and Controls | Next-generation sequencing of DNA provides an unprecedented opportunity to discover rare genetic variants associated with complex diseases and traits. However, the common practice of first calling underlying genotypes and then treating the called values as known is prone to false positive findings, especially when genotyping errors are systematically different between cases and controls. This happens whenever cases and controls are sequenced at different depths, on different platforms, or in different batches. In this article, we provide a likelihood-based approach to testing rare variant associations that directly models sequencing reads without calling genotypes. We consider the (weighted) burden test statistic, which is the (weighted) sum of the score statistic for assessing effects of individual variants on the trait of interest. Because variant locations are unknown, we develop a simple, computationally efficient screening algorithm to estimate the loci that are variants. Because our burden statistic may not have mean zero after screening, we develop a novel bootstrap procedure for assessing the significance of the burden statistic. We demonstrate through extensive simulation studies that the proposed tests are robust to a wide range of differential sequencing qualities between cases and controls, and are at least as powerful as the standard genotype calling approach when the latter controls type I error. An application to the UK10K data reveals novel rare variants in gene BTBD18 associated with childhood onset obesity. The relevant software is freely available.
| In next-generation sequencing studies, there are typically systematic differences in sequencing qualities (e.g., depth) between cases and controls, because the entire studies are rarely sequenced in exactly the same way. It has long been appreciated that, in the presence of such differences, the standard genotype calling approach to detecting rare variant associations generally leads to excessive false positive findings. To deal with this, the current “state of the art” is to impose stringent quality control procedures that much of the data is eliminated. We present a method that allows analyzing data with a wide range of differential sequencing qualities between cases and controls. Our method is more powerful than the current practice and can accelerate the search for disease-causing mutations.
| Recent technological advances in next-generation sequencing (NGS) have made it possible to conduct association studies on rare variants, which hold great potential to explain the missing heritability of complex traits and diseases [1]. However, it is prohibitively expensive to conduct high-depth, whole-genome sequencing (WGS) for large-scale association studies [2]. Therefore, many WGS studies have reduced the overall average depth to as low as 4–10× [3, 4]. Other studies have adopted whole-exome sequencing (WES), in which only the protein coding regions were sequenced but at high depth (e.g., ≥ 30×) [5, 6]; nevertheless, even though the average depth may be high, the large variability in capture efficiency may cause some genes or some regions within a gene to have much lower depth than the average [7].
The case-control design remains the most commonly used approach to studying rare variant associations. Due to the high cost of sequencing, many studies have focused sequencing effort on cases. Some studies sequenced cases at higher depth than controls by design, when the cases are unique and there is interest in identifying novel mutations [4]. Some studies even sampled only cases for sequencing and intended to compare them with publicly available NGS data on general populations such as the 1000 Genomes [3]. In both cases, the controls typically have systematically different sequencing qualities (e.g., depth and base-calling error rate) from the cases. Even when their average depths are similar, the actual depth could vary in individual regions across platforms, resulting in regions with differential depths in cases and controls by chance. This can easily occur when using different exome capture kits for cases and controls; if one kit can capture a certain exonic region better than the other, then there will be a systematic difference in read depth between cases and controls in this region.
The prevailing practice of analyzing NGS data for association with rare single-nucleotide variants (SNVs) is to first call underlying genotypes (e.g., using SAMtools [8] or GATK [9]), and then treat the called values as known in gene- or region-based tests such as the burden test [10, 11]. Genotype calling is difficult when read depth is low because minor allele reads are indistinguishable from sequencing errors. Genotype calling is especially challenging for rare SNVs, first because their locations cannot be easily inferred [12], and second because little information can be borrowed from other variants through linkage disequilibrium (LD) [3]. In case-control studies with differential sequencing qualities, the genotype calling process can introduce confounding that causes inflated type I error in downstream association tests [13]. Recall that confounding occurs when a variable is correlated with both the case-control status and the genotype. When read depths are different in cases and controls, the dependence of genotyping quality on the depth establishes the depth as a confounder. Likewise, the base-calling error rate has the same confounding effect as the depth. Even when read depths and error rates are comparable between cases and controls, differences in genotype calling algorithms or quality control (QC) filters (e.g., phred score cutoffs) can lead to differential genotyping errors that could also act as a confounder. For these reasons, publicly available NGS data have generally been under-utilized as controls for association studies. To reduce genotyping errors, one typically applies QC procedures to filter out SNVs at which many samples are covered by low depth of reads or called with low quality scores [5, 6]. The use of any reasonable QC procedure will remove a large number of variants, especially rare ones, and results in loss of important information.
An example is the UK10K Project [4], which sequenced cases at ∼ 60× and controls at ∼ 6×. In analysis of called genotypes, we obtained severely inflated type I error without QC (see Results). The UK10K Statistics Group adopted a series of QC procedures and controlled the type I error, but their QC removed 76.9% variants. Another example is the study of amyotrophic lateral sclerosis [6], which employed several sequencing platforms with unequal case-control ratios. Even when the average depth was as high as 144.6×, there were still at least 7.66% bases excluded from analysis due to depth less than 10×.
To avoid the confounding effect induced by calling genotypes, Derkach et al. [14] proposed to replace the genotypes in the standard score statistic by their expected values given observed read data, and developed a robust variance for the score statistic to account for differential variances of the expected genotypes in high- and low-depth samples. However, they still used called genotypes to determine SNV locations, which approach tends to yield more false positive SNVs among the low-depth group than the high-depth group and again cause confounding. To ensure accuracy of the called SNV locations, they resorted to stringent QC procedures, which would result in substantial information loss.
In this article, we provide a likelihood-based approach to testing rare variant associations that directly models sequencing reads without calling genotypes. We consider the (weighted) burden test statistic, which is the (weighted) sum of the score statistic for assessing effects of individual variants on the trait of interest. Our read-centric approach enables us to exploit genomic loci covered by low depth of reads and explicitly account for sequencing differences (i.e., read depth and error rate) between cases and controls.
Full implementation of a read-centric approach requires solutions to a number of problems. Because SNV locations are unknown, we first develop a simple, computationally efficient screening algorithm to estimate their locations using read data alone. Because an imbalance in putative SNVs can arise due to differences in read depths and error rates between cases and controls, the burden statistic may not have mean zero even in the absence of association. Thus, we develop a novel bootstrap procedure for assessing the significance of the burden statistic. Specifically, in each bootstrap iteration, we propose to first generate a dataset with the same coverage patterns as the original data, but where the loci are all monomorphic. By comparing the false-positive SNVs found in the monomorphic dataset to the SNVs detected in the original data, we show how to estimate the number of true SNVs and the allele frequencies of the true SNVs in the original data. With this information, we can then generate a final bootstrap dataset in which the allele frequencies at true SNVs match those in the original data, but are identical in cases and controls. The entire procedure is repeated to generate multiple bootstrap datasets. Finally, we compare the burden statistic from the original data to those from the bootstrap datasets to assess significance. The complete flowchart is depicted in Fig 1. Our method can encompass all informative loci including singletons and doubletons if desired; additionally, we can down-weight or mask loci that are unlikely to be deleterious.
We showed through extensive simulation studies that our bootstrap tests are robust to a wide range of differential sequencing qualities between cases and controls, and are at least as powerful as the standard genotype calling approach when the latter controls type I error. We further applied the new methodology to a case-control data from the UK10K Project comparing children with severe early onset obesity to population-based controls. We identified a gene, BTBD18, that passes the exome-wide significance threshold and that is also a plausible candidate for childhood onset obesity.
We first consider a single (bi-allelic) SNV. Let G be the genotype (coded as the number of minor alleles) at the variant site and let D be the disease status. We denote the genotype distribution under Hardy-Weinberg equilibrium (HWE) by Pπ(G), where π is the minor allele frequency (MAF). Note that the HWE assumption has a minimal effect for rare variants, as homozygotes of minor alleles are not expected. Instead of observing G, we observe the total number of reads mapped to the SNV and the number of reads carrying the minor allele, denoted by T and R, respectively. Similar to SAMtools, GATK, and seqEM [15], we assume that R given T and G follows a binomial distribution
P ϵ ( R | T , G ) = Binomial ( T , ϵ ) if G = 0 Binomial ( T , 0 . 5 ) if G = 1 Binomial ( T , 1 - ϵ ) if G = 2 , (1)
where ϵ is the probability that a read allele is different from the true allele and is referred to as the error rate. The “errors” here comprise both base-calling and alignment errors. We treat ϵ as a free parameter that is locus-specific and will be estimated from the read data [15].
To account for case-control sampling, we adopt the retrospective likelihood with individual contribution
Pr ( R i | T i , D i ) = ∑ g = 0 , 1 , 2 Pr ( R i | T i , g , D i ) Pr ( g | T i , D i ) = ∑ g = 0 , 1 , 2 Pr ( R i | T i , g ) Pr ( g | D i ) ,
where the second equation follows from two assumptions: first, the binomial distribution for read count data depends only on the underlying genotype, not on the disease status; second, the genotype distribution depends only on the disease status, not on the read depth. Thus, the likelihood based on n subjects takes the form
L CC ( π 1 , π 0 , ϵ 1 , ϵ 0 ) = ∏ i ∈ D 1 ∑ g = 0 , 1 , 2 P ϵ 1 ( R i | T i , g ) P π 1 ( g ) ∏ i ∈ D 0 ∑ g = 0 , 1 , 2 P ϵ 0 ( R i | T i , g ) P π 0 ( g ) , (2)
where D 1 and D 0 denote the sets of cases and controls, respectively, πd denotes the allele frequency for D = d, and (π1, ϵ1) and (π0, ϵ0) are separate parameters for cases and controls. Note that in writing Eq (2) we assume that the depth T is independent of the genotype G. Also note that this formulation obviates the need to model other covariates (e.g., age and environmental exposures) as long as they are not confounders. The null hypothesis of the association test is H0: π1 = π0. We re-parameterize (π1, π0) in terms of (α, β) such that π0 = eα/(1 + eα) and π1 = eα+β/(1 + eα+β); then the null hypothesis is H0: β = 0. The score function for β under H0, as derived in S1 Text, can be written as
S = ∑ i = 1 n ( D i - n 1 n ) G ˜ i , (3)
where
G ˜ i = ∑ g = 0 , 1 , 2 g P ϵ ˜ D i ( R i | T i , g ) P π ˜ 0 ( g ) ∑ g = 0 , 1 , 2 P ϵ ˜ D i ( R i | T i , g ) P π ˜ 0 ( g ) , n1 is the number of cases, and ( π ˜ 0 , ϵ ˜ 1 , ϵ ˜ 0 ) are restricted maximum likelihood estimates (MLEs) under the null; these restricted MLEs can be obtained via the expectation-maximization (EM) algorithm described in S2 Text. G ˜ i can be interpreted as the posterior dosage of the minor allele (estimated under the null hypothesis); as the read depth increases, G ˜ i converges to the underlying genotype Gi and S reduces to the standard score statistic ∑ i = 1 n ( D i - n 1 / n ) G i. Finally, we construct the burden statistic W as a (weighted) sum of the score statistics at a set of SNVs in the gene of interest. The variance estimator V for W is calculated as the empirical variance of the efficient score functions [16]. When true SNVs are used, the test statistic Z = W / V is asymptotically normal with mean 0 and variance 1.
The score statistic of the Derkach test [14] has the same form as Eq (3), as it also uses the posterior dosage G ˜ i. The only difference is that the Derkach test substitutes the genotype likelihood P ϵ ˜ D i ( R i | T i , g ) that is provided in the output of standard genotype calling packages [8, 9], which calculate error rates based on phred scores.
In reality, the locations of rare SNVs are not available without calling genotypes. In order to include the maximum set of variants in the burden test without calling genotypes, we develop a screening algorithm to screen every locus (i.e., base pair) in the genome and filter out only loci that are “uninformative” in the sense that they yield S = 0 and thus do not contribute to the test statistic. Specifically, we consider the likelihood L S ( π , ϵ ) = ∏ i = 1 n ′ ∑ g = 0 , 1 , 2 P ϵ ( R i | T i , g ) P π ( g ) which is based on a homogenous group (i.e., cases or controls only) of n′ subjects. Let π ˜ be the MLE based on LS(π, ϵ)under the constraint that π ∈ [0, 1] and note that π ˜ = 0 indicates no mutation in this group at this locus. Fortunately, we can easily determine whether π ˜ = 0 without iteratively solving for π ˜. By definition, π ˜ also maximizes the profile likelihood pl(π) = maxϵ log LS(π, ϵ). Because we have shown in S3 Text that pl(π) is a concave function of π, a negative derivative of pl(π) at π = 0 leads to π ˜ = 0. At π = 0, the ϵ maximizing log LS(π, ϵ) can be easily determined because, in the absence of any minor alleles, all reads carrying the minor allele must be errors. Therefore, we check the sign of the derivative of pl(π) at π = 0 for cases and controls separately and screen out the loci at which both signs are negative. If π ˜ = 0 in both cases and controls, then π ˜ 0 = 0 in the combined sample, where π ˜ 0 was defined in the text following expression Eq (3). From π ˜ 0 = 0, we have G ˜ i = 0 for all individuals and thus S = 0. This screening algorithm only involves evaluating simple (derivative) functions twice at each locus without any iteration, and is thus computationally extremely efficient.
Although most monomorphic loci are “uninformative” and will be screened out, there are exceptions. It is possible that a truly monomorphic locus has π ˜ > 0 in one disease group or both, if by chance some individuals have more errors than expected. If a truly monomorphic locus has π ˜ > 0 in the control group but π ˜ = 0 in the case group, the score statistic S of this locus will have a negative mean. Such loci will accumulate over the gene when controls have systematically lower depth (or higher error rate) than cases, and then the expected value of the burden statistic W will be substantially biased below zero, even when allele frequencies are identical among cases and controls at true SNVs. Consequently, screening for SNVs in the presence of differential sequencing qualities between cases and controls will invalidate the asymptotic version of our test.
We thus propose a bootstrap procedure for assessing the significance of the observed test statistic Z. The idea is to generate bootstrap datasets that mimic the original data in terms of read depth and error rate, have the same number of truly monomorphic loci and true SNVs, but have no difference in allele frequencies among cases and controls. To this end, we condition on the observed depth T and simulate the minor-allele read count R using the estimated error rates ϵ ˜ 1 and ϵ ˜ 0 once the underlying genotype G is simulated. However, it is nontrivial to simulate G, because we do not know how many loci in the gene are true SNVs and what are allele frequencies at these SNVs. To obtain this information, we first form a “monomorphic” dataset by simulating R at every locus in the gene assuming that all Gs are zero; thus, each read for the minor allele is an error that occurs with rate ϵ ˜ 1 or ϵ ˜ 0, depending on the disease status. This dataset should provide a good approximation to the truly monomorphic loci in the original data, as the proportion of true SNVs in the original data should be small. Let Ms be the number of loci that are screened in from the original data and let Fs(π) be the cumulative distribution function (CDF) of estimated MAFs at the Ms loci. Let Mm and Fm(π) be their counterparts in the monomorphic dataset. The CDF of allele frequencies at true SNVs, denoted by Fp(π), is related to Fs(π) and Fm(π) through the equation
F s ( π ) = ϕ F m ( π ) + ( 1 - ϕ ) F p ( π ) ,
where ϕ is the proportion of monomorphic loci among loci that are screened in. This equation expresses the fact that the distribution of observed (non-zero) allele frequencies Fs(π) in the original data is a mixture of the distributions for allele frequencies of true SNVs Fp(π) and artifactual SNVs Fm(π) that actually correspond to monomorphic loci. We estimate ϕ by ϕ ^ = M m / M s and Fp by F ^ p ( π ) = ( 1 − ϕ ^ ) − 1 { F ^ s ( π ) − ϕ ^ F ^ m ( π ) }, where F ^ s and F ^ m are empirical CDF estimators of Fs(π) and Fm(π) respectively. To ensure that F ^ p ( π ) is monotonically increasing, we refine F ^ p ( π ) by fitting an isotonic regression to data points of ( 1 - ϕ ^ ) - 1 { F ^ s ( π ) - ϕ ^ F ^ m ( π ) } evaluated at the pooled (Ms + Mm) MAFs by the pooled-adjacent-violator algorithm (PAVA) [17]. After the largest value of MAF, we set F ^ p ( π ) = 1. Finally, starting from the monomorphic dataset, we select M ^ p = M s - M m loci to be SNVs, sample π from F ^ p, and re-generate G and R at these SNVs to form a final bootstrap dataset. Note that, for a small π, we may need to resample G repeatedly until each truly polymorphic locus screens in. The bootstrap statistic is then calculated based on all the loci that were screened in from the final bootstrap dataset. The entire procedure is repeated to generate multiple bootstrap replicates.
Although bootstrap tests are computationally intensive in general, we can save considerable time by adopting a sequential stopping rule [18]. We stop after generating Lmin bootstrap replicates, if these early replicates suggest a large p-value. When Lmin = 5, the number of replicates at termination has a median of only 10 for a gene having no SNVs that affect the trait. We also use a closed sampling scheme, in which we restrict the total number of bootstrap replicates to be at most Kmax. If we stop when Lmin bootstrap statistics exceed the observed Z and Kobs (≤Kmax) replicates have been collected, we set the p-value to Lmin/Kobs. If we stop when Kmax replicates are reached and only Lobs (<Lmin) values exceed Z, we set the p-value to (Lobs + 1)/(Kmax + 1).
The MLEs of error rates may not recover the true distribution of error rates, which is essential for generating valid bootstrap replicates. In particular, when the true error rates are very small (e.g., ∼ 0.02%), the MLEs tend to be over-dispersed. Therefore, we propose the following “adjusted” empirical Bayes (EB) estimator of the error rate to be used in bootstrap (instead of the MLE), which is calculated separately among cases and controls. We assume a prior beta distribution for error rates, i.e., ϵj ∼ Beta(a, b), where j = 1, …, M, M is the total number of loci in the gene, and a and b are hyperparameters that can be consistently estimated by the method of moments (see S4 Text). While the EB estimator is easily obtained (S4 Text), it is known that the distribution of EB estimators is over-shrunk [19]. Louis and Shen [19] proposed estimators that have good distribution, rank and expected value, but these are cumbersome to compute. We use a simplified version of the Louis and Shen estimator in which we first calculate the EB estimators but then replace the (ordered) EB estimators by (ordered) quantiles of the prior beta distribution evaluated using the method-of-moments estimators of a and b. Because the sample size M is typically on the order of a few hundred, a and b are accurately estimated, ensuring that the distribution of the adjusted EB estimates will closely resemble the prior (true) distribution of error rates.
We have observed that a small proportion of read data (R, T) do not fit the binomial model (1). This may be due to genotype mosaicism (i.e., the presence of two or more populations of cells with different genotypes in one individual), experimental artifacts, sample contamination, or copy number variants. To detect data that do not fit the binomial model, for each individual at each locus that screens in, we calculated a likelihood-ratio-type statistic for the goodness of fit to the binomial model
Q = 2 log R / T R 1 - R / T T - R / max g = 0 , 1 , 2 e g ( ϵ ) R 1 - e g ( ϵ ) T - R ,
where eg(ϵ) = ϵ, 0.5, and 1 − ϵ for g = 0, 1, and 2, respectively. Then, we mask an individual at a variant (by setting T and R to zero) if Q is greater than 10 and remove a variant altogether if more than 5 individuals are masked at that locus. We can also identify individuals with problematic data by checking for the presence of an excessive number of Qs greater than 10.
The proposed methods are implemented in the C/C++ program TASER, which is publicly available at http://web1.sph.emory.edu/users/yhu30/software.html.
We carried out extensive simulation studies to evaluate the performance of our proposed methods in realistic settings. We used the coalescent simulator cosi [20] to generate a base population of 100,000 European haplotypes with length 10 kb. We assumed that the 10 kb region corresponds to a gene with 3 exons that are separated by 2 introns, with introns being 3 times the length of exons. This setup gave us a total of 2,730 loci in exons, among which there are 44 SNVs with MAFs ≤ 0.05 in the base population. To generate individual genotypes, we sampled from the 100,000 haplotypes allowing recombination in introns (but not in exons). To generate disease outcomes, we considered a risk model that assumed equal attributable risk (AR) for each SNV: log { P ( D = 1 ) / P ( D = 0 ) } = α + ∑ j = 1 m G j log ( 1 + AR / 2 π j ), where m is the total number of SNVs, Gj and πj are the genotype and MAF of the jth SNV, and α was set to −3 to achieve a disease rate of ∼ 5%. This risk model implies that a more rare SNV has a stronger effect than a less rare SNV. The process was repeated until 500 cases and 500 controls were collected.
The sequencing reads T and R were generated to mimic real NGS data. We considered average read depths of 6×, 10×, and 30×, and average error rates of 0.02% and 0.016% (as observed in the UK10K cases and controls, respectively). While these very low error rates are characteristic of the newest Illumina platforms, we also considered average error rates of 1% and 0.5% that exist in historical NGS data [21]. We sampled the locus-specific error rate ϵ from a beta distribution that yields the pre-specified average rate. We sampled the individual depth T by a two-step strategy which first simulates the locus-specific mean depth c from a beta distribution (re-scaled to achieve the pre-specified average depth) and then simulates individual T’s from a negative-binomial distribution with mean c. The first step permits the accessibility of sequencing to depend on local nucleotides, and the second step allows for dispersion in the individual count data. For specific parameter values in these distributions, refer to S5 Text. Note that at each locus we sampled ϵ and c independently for cases and controls, mimicking the scenario in which the two groups have been sequenced as part of different studies (e.g., on different platforms), even when the average values are the same between the two groups. Finally, we sampled R given (T, G, ϵ) according to Eq (1).
We considered eight methods. First, we assumed that the 44 SNV locations were known and applied the asymptotic version of our method, the method using called genotypes that extends the multi-sample, single-locus genotyper seqEM [15] to allow for different error rates in cases and controls, the Derkach method using genotype dosages, and the method using true genotypes as a gold standard; we refer to them as New, CG, Dose, and True. Note that, to ensure fair comparisons, we used the error rates from our method in the implementation of the Derkach test, whose score statistic is then the same as our S in Eq (3). Thus, although Derkach et al. used a slightly different variance estimator for the score statistic, New and Dose are asymptotically equivalent. Next, we considered the more realistic case that the SNV locations are unknown. We applied our method including the screening and bootstrap procedures and refer to it as New-SB. While this method aims to maximize the set of true SNVs, it may also include a sizable number of monomorphic loci that can adversely affect the power of association testing. We thus explored a modification of New-SB, which adds a thresholding step that excludes loci with estimated MAFs <(2n)−1 and is referred to as New-STB. The threshold of (2n)−1 corresponds to the MAF of a singleton variant and can effectively remove the majority of monomorphic loci that accidentally pass the screening algorithm, although at a cost of potentially losing some true singletons. In addition, we applied the method of called genotypes and the Derkach method based on loci that were screened in and refer to them as CG-S and Dose-S.
We focused on the weighted burden test of SNVs with MAFs ≤ 5%, in which each SNV is inversely weighted by π j ( 1 - π j )[11]; results of the unweighted test are provided in S1 and S2 Tables. We first evaluated type I error of the burden test using the aforementioned methods and summarized the results in Table 1. All of the new methods (New, New-SB, New-STB) have correct type I error, regardless of how different the sequencing depths and error rates are between cases and controls. The genotype calling methods (CG, CG-S) generally have inflated type I error when the average depths are different between cases and controls. Their type I error tends to be inflated even when the average depths and error rates are the same but there are random differences in individual regions between cases and controls; the inflation in such a case is more noticeable for the unweighted test (S1 Table) than for the weighted test (Table 1), because the SNVs with higher MAFs contribute more to the inflation and they are down-weighted in the weighted test. Only when cases and controls have exactly the same sequencing feature at every locus, which can be achieved by sequencing cases and controls together, should the genotype calling methods have correct type I error. The Derkach approach worked well when the SNV locations are known, but its type I error rate can be as much as 88 times the nominal level when the locations are unknown. In Table 2, we give additional results on the behavior of our test statistics under the null hypothesis. We see that the test statistic in the presence of screening is negatively biased from zero when controls have lower average depth than cases, which confirms the need for our bootstrap test. We also see in Table 2 that, when the average error rate is high, the screening procedure screened in a large number of monomorphic loci, and that the thresholding procedure effectively removed many such loci. Finally, we see that the bootstrap procedure accurately estimated the number of truly polymorphic loci. S1 Fig shows that the MLEs of error rates are more dispersed than the true error rates (especially contain too many zeros when the average is 0.02%), the EB estimator imposed a strong shrinkage effect, and that our adjusted EB estimator accurately recovered the true distribution. S2 Fig shows that, when the average error rate is 1%, the monomorphic loci that were screened in are typically associated with small π ˜’s, the majority of which are smaller than the threshold of (2n)−1.
Fig 2 contrasts the power of different methods. The thresholding strategy implemented in New-STB significantly improved the power of New-SB at error rate of ∼ 1% and performed as well as New-SB at ∼ 0.02%. In the presence of differential depths between cases and controls, the power of CG-S and Dose-S can even decrease as the effect size starts to increase from zero and both are substantially lower than the power of New-SB and New-STB at median and high effect sizes. In the presence of equal average depths and error rates, the power of CG-S and Dose-S are comparable to that of New-SB and New-STB at error rate of ∼ 0.02% and noticeably lower at ∼ 1% (even at high depth of ∼ 30×). Power curves pertaining to unweighted burden tests are displayed in S3 Fig, which shows similar patterns to Fig 2 but lower power due to the weighted nature of our risk model for simulating the disease status. While the results described up to now pertain to simulation settings where the locus-specific ϵ and c are sampled independently for cases and controls (even when the average values are the same between the two groups), we also considered the setting in which ϵ and c are the same between cases and controls at each locus. This would occur when the two groups have been sequenced together through the exact same pipeline. As shown in S4 Fig, the power of New-SB and New-STB are always greater than or equal to the power of CG-S.
The UK10K project [4] was funded by the Wellcome Trust Sanger Institute in 2010 to help investigators better understand the link between low-frequency and rare genetic changes and complex human diseases by applying NGS on 10,000 people in the United Kingdom (UK). We focused on the samples collected by the Severe Childhood Onset Obesity Project (SCOOP), all of whom have severe, early onset obesity (i.e., body mass index Standard Deviation Scores [22] > 3 and obesity onset before the age of 10 years). For controls, we utilized the population-based cohort collected in the TwinsUK study (randomly excluding one twin from each twinship) from the Department of Twin Research and Genetic Epidemiology at King’s College London. Both cases and controls are UK-based populations and part of the UK10K project. While the cases were whole-exome sequenced at average depth of 60×, the controls were whole-genome sequenced at average depth of 6×.
We used SAMtools to generate the pileup files from the BAM files and extracted read count data, filtering out reads that are PCR duplicates, that have mapping score < 30, that have improperly mapped mates, or that have phred base-quality scores < 30. We restricted our analysis to the consensus coding sequence gene sets [23] and further masked repeat regions, regions covered by monomorphic read alleles, and regions not covered by any reads, resulting in a total of ∼ 14 million loci exome wide. We recorded read count data for these loci such that, for example, a locus covered by 10 reads of allele A and 1 read of C was coded as A10C1. Read count datasets in this format are much more manageable than the BAM files; our formatted, zipped files required only 126 GB of disk space, compared to ∼ 14 TB for the BAM files. We obtained data in this format for 784 cases and 1,669 controls. We found that 87 cases had excessive read data that do not fit the binomial model (i.e., Q > 10) and we excluded these subjects (plus 1 additional case which is possibly in the same batch as the 87 cases) from further analysis; see S6 Text for more details. Thus the analysis described here was based on 696 cases and 1,669 controls.
We considered two versions for the weighted burden test, one including all variants and one including only variants that are annotated as “probably damaging” or “possibly damaging” by PolyPhen [24]. We applied our methods, New-SB and New-STB, to scan all genes for association with severe childhood onset obesity. We set Kmax = 10,000,000, which is sufficient for detecting p-values that pass the exome-wide threshold that is on the order of 10−6. The analysis of damaging variants took a total of 1,713 hours on an IBM HS22 machine or equivalently 8.6 hours on 200 such machines in a computing cluster. We also applied the genotype calling method (CG-S) and the Derkach method (Dose-S) as described in Simulation Studies. Further, we analyzed the genotypes in the VCF files downloaded from the UK10K website. These genotypes were called by SAMtools, filtered by GATK VQSR, and imputed by Beagle [25], by the UK10K investigators with cases and controls being processed separately. We refer to this approach as CG-VCF.
We screened in a total of 474,508 loci, among which 465,967 (98.2%) loci passed our read-based QC procedure. The 465,967 loci span over 16,318 genes; 431,311 passed the threshold of (2n)−1 and 288,535 were estimated to be polymorphic. Considering damaging variants only, 238,753 loci were screened in and passed QC; 219,540 passed the threshold and 143,822 were estimated to be polymorphic. Note that the CG-VCF analysis was based on the same set of 465,967 loci, although some of them had been called monomorphic and were thus not included in the VCF files. As a result, the CG-VCF analysis included 167,980 loci, of which 79,271 were predicted as damaging.
The quantile-quantile plots are displayed in Fig 3. The observed p-values for New-STB and New-SB agree very well with the global null hypothesis of no association (genomic control λ = 1), except at the extreme right tails. By contrast, the observed p-values for Dose-S, CG-S, and CG-VCF show very early departures from the global null distribution, reflecting severe inflation of type I error. Fig 4 shows that the test statistics are negatively biased from zero, which explained the poor performance of Dose-S.
Among all p-values generated by our methods, the smallest one, 2.0 × 10−7, was obtained for gene BTBD18 by New-STB using damaging variants only, and this p-value passed the exome-wide significance threshold of 3.1 × 10−6 (0.05/16,318) after Bonferroni correction. Looking into the raw read data on this gene, we found that among cases the WES resulted in extremely low depth (∼ 0.34×). (This kind of regions is not uncommon; indeed, 1.9% of all loci that were screened in have depth ≤1× in cases.) We found that at each of four loci (57512143, 57512745, 57513287, and 57513568 when mapped to the hg19 reference genome), there is a case individual covered by two reads and both are minor allele reads. These four suggestive minor allele homozygotes made large contributions to the score statistic and drove the gene-level association signal. As gene BTBD18 has also been found to over-express in obese children elsewhere (NCBI GEO Profile ID: 64932244), it makes a plausible candidate for childhood onset obesity. Table 3 lists BTBD18 and other top ten genes ranked by New-STB using damaging variants. Note that the standard genotype calling approach (CG-VCF) would have precluded BTBD18 from association analysis due to the low depth data in cases. Using all SNVs, BTBD18 was also ranked highest by New-STB, with the same four loci driving the association signal, but the p-value did not pass the exome-wide significance threshold because of the inclusion of other neutral variants.
We have presented a robust and efficient approach to association testing of rare variants that is based on analyzing raw sequencing reads directly, without calling genotypes. Our bootstrap procedure guarantees that the corresponding association tests have correct type I error under a wide range of sequencing differences between cases and controls. Our simulation studies showed that the proposed methods perform better than or as well as the genotype calling method in terms of power, when the latter shows no significant increase in type I error (e.g., when the average read depths and error rates are the same between cases and controls). These results can be understood by noting that converting reads into genotype data is a coarsening of the read data, which can result in information loss even when there is no differential error between cases and controls. These results suggest that, if the main goal is burden-based association testing (which is, in most cases, the goal of sequencing studies), then our proposed methods may be an attractive alternative to analyses based on called genotypes, even in studies where cases and controls have been “well-matched” for average depths or, further, have been sequenced together.
When applied to real data, our read-based procedure allows use of far more loci than methods based on calling genotypes, because we do not filter out variants covered by low depth of reads or called with low quality scores. For example, in analysis of the UK10K data, we only filtered out 1.8% of loci that were screened in; our final analysis included data from 465,967 loci. By contrast, the UK10K Statistics Group had to pare down to only 132,984 loci in order to achieve accurate type I error in the standard genotype calling approach, even though their analysis included almost 2,000 additional control participants from the Avon Longitudinal Study of Parents and Children (ALSPAC).
We have presented our methods in the context where all cases are from a single source and all controls are from another source. In practice, it is also common to use cases or controls from multiple sources, all from different platforms. The methods we have presented here can readily be extended to such scenarios by estimating a separate error rate for each data source, and then generating bootstrap datasets with the same source characteristics as the original data. We plan to implement this in future work.
When developing our methods, we made some simplifying assumptions. First, we assumed independence (i.e., no LD) across rare variants when generating bootstrap replicates. This is reasonable because rare variants typically do not exhibit strong LD with each other [26]. However, if strong LD occurs, it is possible to generate SNVs that have the same amount of LD as the original data by sampling haplotypes instead of single SNVs. The SNVs in the bootstrap sample can be placed in the same order (by allele frequency) as the original data.
Second, we assumed that base-calling errors are independent across loci. In reality, the base-calling errors might be correlated due to factors such as library preparation and sequence context. However, this assumption only affects the efficiency of our method, not its validity. We also assumed that the errors are symmetric, i.e., the probability of a read for the major allele being mis-called as the minor allele is the same as the probability of the minor allele being mis-called as the major allele. For analyzing rare variant data, this assumption has a negligible effect as rare variant homozygotes are extremely rare. Further, our methods estimate error rates directly from the read data, and thus ignored phred scores that characterize the base-calling quality and alignment scores that calibrate alignment quality. In our analysis of the UK10K data, we filtered out reads with alignment scores < 30 and phred scores < 30. We have shown in other work [27] that phred scores and low-score reads can provide additional information. It would be possible to include a model of the variability in error rates that is explained by base-calling and alignment quality scores in our current approach.
Finally, we do not account for confounders such as principal components for ancestry. In the UK10K data, all samples are UK-based Caucasians and are therefore not expected to have strong population stratification. It is also possible to extend our methods to allow confounders, by generating bootstrap replicates that have the same amount of confounding as the original data. We plan to describe such approaches in a subsequent report.
Our bootstrap procedure is parametric in the sense that its validity depends on correctly modeling the error and allele frequency distributions required to generate the bootstrap replicates. In addition, any added power that could be realized by relaxing assumptions like no LD across variants and independent error rates across loci would require special modification of our procedure. Further, we have assumed there are no confounding covariates; we plan to extend our approach to account for confounding covariates in future work. Finally, even with a sequential stopping rule, our bootstrap procedure may still be computationally intensive when the p-value to be estimated is very small. It may be possible to adopt a dynamic scheduling system so that nodes that are calculating a region having a large p-value would then shift their resources to regions where early bootstrap replicates suggest a small p-value.
We have focused on the burden test in this article. Because our score statistic may not have mean zero after screening, it is nontrivial to construct the sequence kernel association test (SKAT) [28]. A valid SKAT statistic requires the score statistic be properly centered; we are currently developing methods to center the score statistic within our bootstrap approach.
|
10.1371/journal.pntd.0006071 | A temporal comparison of sex-aggregation pheromone gland content and dynamics of release in three members of the Lutzomyia longipalpis (Diptera: Psychodidae) species complex | Lutzomyia longipalpis is the South American vector of Leishmania infantum, the etiologic agent of visceral leishmaniasis (VL). Male L. longipalpis produce a sex-aggregation pheromone that is critical in mating, yet very little is known about its accumulation over time or factors involved in release. This laboratory study aimed to compare accumulation of pheromone over time and determine factors that might influence release in three members of the L. longipalpis species complex.
We investigated male sex-aggregation pheromone gland content at different ages and the release rate of pheromone in the presence or absence of females under different light conditions by gas chromatography-mass spectrometry (GC-MS). Pheromone gland content was determined by extraction of whole males and pheromone release rate was determined by collection of headspace volatiles. Pheromone gland content appeared age-related and pheromone began to accumulate between 6 to 12 h post eclosion and gradually increased until males were 7–9 days old. The greatest amount was detected in 9-day old Campo Grande males ((S)-9-methylgermacrene-B; X ± SE: 203.5 ± 57.4 ng/male) followed by Sobral 2S males (diterpene; 199.9 ± 34.3) and Jacobina males ((1S,3S,7R)-3-methyl-α-himachalene; 128.8 ± 30.3) at 7 days old. Pheromone release was not continuous over time. During a 4-hour period, the greatest quantities of pheromone were released during the first hour, when wing beating activity was most intense. It was then substantially diminished for the remainder of the time. During a 24 h period, 4–5 day old male sand flies released approximately 63 ± 11% of the pheromone content of their glands, depending on the chemotype. The presence of females significantly increased pheromone release rate. The light regime under which the sand flies were held had little influence on pheromone release except on Sobral 2S chemotype.
Accumulation of pheromone appears to occur at different rates in the different chemotypes examined and results in differing amounts being present in glands over time. Release of accumulated pheromone is not passive, but depends on biotic (presence of females) and abiotic (light) circumstances. There are marked differences in content and release between the members of the complex suggesting important behavioural, biosynthetic and ecological differences between them.
| The Dipteran subfamily Phlebotominae includes the genera Lutzomyia and Phlebotomus among which several species are important vectors of parasitic and bacterial pathogens. The sand fly Lutzomyia longipalpis is considered the main vector of visceral leishmaniasis (VL) in the New World. Based on the main component of the male sex-aggregation pheromone gland, different sex pheromone-producing populations (chemotypes) of L. longipalpis are recognized in Brazil. Given the importance of the sex-aggregation pheromones in the biology of this species complex, we present here the first attempt to study how pheromone accumulates in the glands over time and factors that might influence its release in the three most common chemotypes from Brazil. Our results demonstrated that pheromone first starts to accumulate a few hours post-eclosion (6–12 h) and this continues over 15 days. Pheromone release is a dynamic process which varies between the 3 chemotypes depending on biotic factors, such as light regime and presence/absence of conspecific females. This work provides valuable information, critical to our understanding of the behaviour and ecology of L. longipalpis sand flies and which will contribute to investigations to improve field-based pheromone control and monitoring of L. longipalpis sand flies.
| There are over 800 known phlebotomine sand fly species, but only approximately 56 Lutzomyia and Phlebotomus species are proven or suspected vectors of human leishmaniasis [1]. Among them, L. longipalpis is the primary vector of Leishmania infantum, the etiological agent of visceral leishmaniasis (VL) in the Americas [2]. The presence of L. longipalpis has been recorded in 12 countries, including Argentina, Bolivia, Colombia, Costa Rica, El Salvador, Guatemala, Honduras, Mexico, Nicaragua, Paraguay, and Venezuela. It is also widely distributed throughout Brazil [1]. During the last two decades it has colonised urban environments and expanded its geographical range, which has resulted in an increase in the number of cases of canine and human VL [3, 4].
The taxonomic status of L. longipalpis has been uncertain since it was first described by Mangabeira in 1969 [5] and recent analysis of molecular and genetic markers [6], morphological features [7], copulation songs [8], and chemical communication [9] all indicate that L. longipalpis is a complex of recently evolved cryptic species [10]. However, there is no consensus on the number of species in the complex or their geographic distributions [7, 11, 12].
In many Dipteran species, sex pheromones, together with visual, tactile and acoustic signals, play an important role in courtship behaviour [13]. Sex-aggregation pheromones occur in male L. longipalpis [9, 14] and may be widespread in the genus Lutzomyia. There is chemical evidence that they also occur in L. cruzi [15], L. pseudolongipalpis [16], L. pessoai [17], L. lichyi [18], L. lenti, L. carmelinoi [19] and L. cruciata [20]. They have also been found in the closely related genus Sergentomyia, e.g. S. minuta and S. fallax [21]. These species produce volatile terpenoid compounds that are structurally similar to the sex-aggregation pheromones of the L. longipalpis species complex. Based on behavioural experiments, there is some evidence that pheromones may also play a role in the mating of Phlebotomus papatasi [22], the main vector of the Old World cutaneous leishmaniasis [1].
In the L. longipalpis species complex the sex-aggregation pheromones have been studied for both their taxonomic value and to exploit their vector control potential [23]. Analysis of the main terpene component of the sex-aggregation pheromone gland extract has shown that there are at least four distinct chemotypes of L. longipalpis [9, 24]: i) (1S,3S,7R)-3-methyl-α-himachalene (3MαH), a novel bicyclic methylsesquiterpene with a 16 carbon skeleton (C16, molecular weight (mw): 218). In Brazil, this chemotype has been found only in Bahia State; ii) (S)-9-methylgermacrene-B (9MGB), a novel monocyclic methylsesquiterpene (C16, mw: 218). It is also the most widespread chemotype in the Americas, and is found in Argentina, Colombia, Paraguay, Venezuela, Honduras, Costa Rica and Brazil. This chemotype is common in Brazil, particularly in the centre, south and to a lesser extent, in the northeast of the country; iii) a partially characterised diterpene (C20, mw: 272) is the second most widely distributed chemotype in Brazil [25]. It is mainly found in the northeastern states although recent reports indicate that it is present in the southeast [4]; iv) a related partially characterised diterpene has also been found only in specimens from Jaíba, Minas Gerais State, Brazil [25].
A racemic version (containing both the R and S isomers) of the 9-methylgermacrene-B sex pheromone has been synthesised in bulk, shown to be active in the field and formulated for long-term controlled release [23]. It is currently being evaluated for its potential to reduce the risk of canine VL infection in a lure-and-kill vector control tool and for enhanced monitoring. Optimising the controlled release formulation of the synthetic pheromone for both of these functions relies on knowledge of how pheromone is produced and released by individuals and groups of males under different conditions.
The sex-aggregation pheromone is synthesised and stored in glandular tissue underlying the abdominal tergites and is transported via cuticular ducts to modified structures, “papules”, on the cuticle surface [19, 26]. Pheromone glands in young, 0–6 h old, males are undifferentiated but appear to be fully differentiated in 4-day old males [27]. Pheromone production is currently believed to start after 12 h and increase continuously for 3 days when it reaches a plateau [28].
No studies have been undertaken to determine if differences in pheromone gland content occur between different members of the species complex nor to determine which factors, if any, might contribute to subsequent pheromone release. It has been suggested that wing-fanning, which occurs in males during courtship, may help to distribute pheromone [29–32] and it has also been suggested that frequent mating attempts during courtship might increase pheromone release but deplete glandular pheromone reserves [29]. Males with depleted gland contents are less successful at obtaining mating attempts than males with fuller glands [29].
In this study, we have investigated the accumulation of pheromone and the dynamics of release for each of three members of the L. longipalpis species complex. Specifically, we addressed the following questions: a) do males of different chemotypes have the same total amount of pheromone (gland content) over time? b) how much pheromone is released from the glands into the atmosphere? c) is pheromone release affected by the presence of conspecific females? and d) do light conditions have an effect on pheromone release?
All three colonies (chemotypes) of L. longipalpis used in this study were originally established from females collected using miniature CDC light traps in chicken shelters. The Jacobina (3MαH), and Campo Grande (9MGB) colonies were established from groups of females collected in Jacobina, Bahia State, (11° 11' S, 40° 31' W) and Campo Grande, Mato Grosso do Sul State (20° 28’ S, 54° 37’ W). The Sobral 2S (diterpene) colony was established from Sobral, Ceará State (3° 41’ S, 40° 20’ W).
The Jacobina colony was originally established in 1974 [33] (estimated 380th generation). The Campo Grande colony was established in 2009 (estimated 63rd generation) and the Sobral 2S colony was established in 2013 (estimated 27th generation).
Previously, chromatographic analysis of the pheromone gland contents of individual males from both Jacobina and Campo Grande colonies established that both were allopatric and contained representatives of only one chemotype [17, 23]. In Sobral the diterpene and 9MGB chemotypes are sympatric. The diterpene chemotype males can be distinguished from the 9MGB males by the number of pale patches on the abdomen. The diterpene chemotype males have two patches (2S), whereas the 9MGB chemotype males have only 1 (1S). Using careful iso-female rearing [34] we were able to establish a diterpene producing colony. Evidence collected from the Sobral field site indicates that the 2 chemotypes do not cross-mate [9, 24].
Sobral 2S males produce a Burst-type copulatory song and the Jacobina males produce a Pulse-type copulatory song, categorised as P1 because of trains of pulses with usually two or three cycles per pulse [12]. The Campo Grande chemotype copulatory song has not yet been categorised.
The sand flies were maintained at 28 ± 2°C, 80 ± 5% relative humidity (RH) and a 12:12 light:dark (L:D) photoperiod in an insectary at Lancaster University (United Kingdom). Immature stages were maintained in rearing pots in which the bottom was filled with a layer (2 cm) of dampened Plaster of Paris to maintain humidity. Females and males were pooled together 3–4 days after eclosion in Barraud cages (18 x 18 x 18 cm) and the females were routinely blood-fed on anaesthetized mice to maintain the colony.
Sand fly blood feeding for colony maintenance was performed according to the guidelines and regulations of the Animals in Science Regulation Unit (ASRU) and in accordance with the terms of a regulated licence (PPL 40/3279) in compliance with the UK Home Office, Animals (Scientific Procedures) Act (ASPA) regulations. All procedures involving animals were reviewed and approved by the Animal Welfare and Ethical Review Board (AWERB) at Lancaster University.
Analysis of male sex-aggregation pheromone gland extracts and headspace entrainment extracts was performed on an Agilent 7890A/5975C GC-MS (Agilent Technologies UK Ltd, Cheshire, UK) operating in electron impact mode. Chromatographic analysis was conducted on a non-polar HP-5MS capillary column, 30 m x 0.25 mm i.d., 0.25 μm film thickness (Agilent, UK), using H2 as carrier gas at 1 ml min-1. Samples were introduced via an on-column injector set at 40°C. The temperature program was an initial temperature (40 oC) held for 2 min and then increased at 10°C min-1 to a final isothermal temperature (250 oC) held for 10 min.
To remove potential contaminants, all glassware was carefully cleaned prior to use by washing in a 10% detergent solution. It was then rinsed with distilled water, dried with acetone and finally heated in an oven at 180ºC for 12 h.
The components of the the entrainment apparatus were connected with fluorinated ethylene propylene (FEP) tubing that was cleaned internally by rinsing with hexane (BDH HiPerSolv, 97%, VWR, Lutterworth, UK) before and after each entrainment.
A new clean 50 ml r/b flask was used for every experimental replicate to ensure that the potential presence of residual pheromone did not interfere with the outcome of successive entrainments.
To obtain male sand flies of known age, larval rearing pots were inspected every 2 h, and newly emerged males were transferred to nylon netting cages (18 x 18 x 18 cm) inside plastic bags. These males were kept under the same temperature, RH and L:D conditions as the colony. Humidity was maintained by placing dampened laboratory filter paper inside the plastic bags and sugar (50% fructose solution) was freely available on a small piece of cotton wool inside the holding cage. Males were selected for experimentation at known ages; 0–2, 6–8, 12–14, 18–20, 24–26 hours old and 2, 3, 5, 7, 9, 12 and 15 days old.
To determine the pheromone gland content, six individuals (replicates) of each age category were analysed by GC-MS. Males were individually placed in Pasteur pipette ampoules and covered with a drop (ca. 10 ul) of analytical grade hexane (≥ 99% purity, SupraSolv, Merck, Germany). The ampoules were flame-sealed and left for 24 h at room temperature prior to analysis. Subsequently, the ampoules were opened and the solvent was gently evaporated under N2 to a volume of ca. 1 ul. The entire sample was then injected into the GC-MS system.
Groups of 10 males, which appeared to be healthy and vigorous, were immobilised by cooling in a freezer for 30 s, and transferred from the holding cage into a clean 50 ml r/b glass flask using a battery powered aspirator. Pheromone released from the males was collected from the headspace volatiles using a portable entrainment apparatus (Barry Pye, Kings Walden, Herts, UK). Clean air was pushed through the 50 ml r/b glass flask containing the male sand flies into a glass tube filled with Tenax, an adsorbent polymer (ORBO 402, Sigma-Aldrich Ltd., Dorset, UK). All joints in the apparatus were sealed with Teflon tape and the airflow at the outlet of the Orbo 402 tube was measured accurately with a bubble flow meter and adjusted with a rotameter (GPE Ltd., Leighton Buzzard, UK) to 400 ml min -1. Adsorbent tubes were used only once. The entrainment was done in an insectary at 26–28°C and 60–80% RH under controlled light conditions.
Volatiles adsorbed on the Orbo 402 tubes were eluted in 2 ml of pure analytical grade hexane. These extracts were collected in small, clean glass Pasteur pipette vials and concentrated under a gentle stream of N2 to 10 ul, and then 1 ul (1 male equivalent) was injected into the GC-MS system. Peak areas of the main terpene component present in the extracts were compared to those of known amounts of both caryophyllene (40 ng/ul) and 9MGB (40 ng/ul) as external standards. The major terpene peak usually represents more than 90% of the total terpenes present, and behavioural biossays have demonstrated that it is the active component of the extract [35, 36]. Analytical standards (n-alkane series C8-C20) of known concentration (10 ng ul -1) were also injected at the begining and end of each analytical session to provide comparative retention time data, an accurate average for the peak areas, and as a check on the GC-MS system performance.
To determine the effect of the presence of females on pheromone release, three 4-day old virgin females were placed in the 50 ml r/b flask along with the 10 males (males + females). The females remained with the males throughout the experiment and the Orbo 402 tubes were replaced each hour as previously described.
To determine if light or dark had an effect on pheromone release we entrained pheromone from males and males + conspecific females in two lighting conditions light (L) or dark (D). Entrainments were carried out on groups of males and males + females. For each light regime 4–6 experimental replicates were carried out.
The pheromone content of 5-day old males (n = 1 male, 3 replicates) was compared with the amount of pheromone released during 24 h by 5-day old males (12:12 LD) (n = 10 males, 3 replicates) for each of the three chemotypes studied. The mortality within the entrainment glass r/b flask varied between 10% to 20%.
A Kruskal-Wallis non-parametric ANOVA was used to compare the amount of pheromone: 1) extracted from the glands of individual males of the three chemotypes at different ages, 2) extracted from 5-day old males with the amount of pheromone released during 24 h by 4–5 day old males, 3) released during each hour of the 4 h period by 4–5 day old males and males + conspecific females, under the two light regimes for each of the three chemotypes. Results are expressed as mean ± standard error (X ± SE) ng male -1 hour -1. All statistical analyses were done using SPSS (v15.0, SPSS Inc.) software. Alpha was set at P < 0.05.
GC-MS analysis confirmed that the retention times (Rt) of the major peaks of each of the chemotypes were 14.58 min (3MαH, Jacobina), 15.82 min (9MGB, Campo Grande) and 20.17 min (diterpene, Sobral 2S) (Fig 1). Retention time and mass spectral data allowed us to accurately identify the relevant pheromone peak in both pheromone gland and headspace extract chromatograms even when they were present in trace amounts.
The GC-MS analysis of the Jacobina male extract showed that up to 12 other terpene compounds were present in minor quantities (four were present in trace amonts only and were unquantifiable), which eluted both before and after the main peak (Fig 1A). In the Campo Grande male extracts, we noticed two small terpene peaks that eluted before the major 9MGB peak (Fig 1B). In the extract of Sobral 2S males seven other minor diterpene components were apparent (Fig 1C). These data are consistent with previous observations from field collected L. longipalpis and confirmed that the terpene components of the glands had not changed over the time they were kept in a laboratory colony, Jacobina [37], Sobral 2S [24, 35] and Campo Grande [23]. None of the pheromone extracts from any of the three populations contained significant quantities of the predominant pheromone molecule(s) characteristic of any of the other populations.
Analysis of the pheromone gland content revealed that males of the three chemotypes produced and stored pheromone throughout the 15 days of the experiment, although the amount of pheromone stored varied with age (Fig 2). Traces of pheromone were detected in extracts from individual Campo Grande and Sobral 2S males at 6–8 h post-emergence but were not detected in individual Jacobina males until 12–14 h after emergence (Fig 2).
Each chemotype displayed a distinct pattern of pheromone accumulation over time; generally, pheromone gland content increased for 7–9 days, after which a small reduction in quantity of pheromone was observed in most of the oldest specimens studied. Overall, there was no significant difference in the amount of pheromone present in the glands of the three chemotypes despite large differences in quantities on specific days.
Jacobina gland content (Fig 2A) peaked at 7 days but the amount of pheromone (128.8 ± 30.3 ng) was not significantly different to the other two chemotypes (χ2 = 5.81, df = 2, P = ns). Campo Grande gland content (Fig 2B) peaked at 9 days (203.5 ± 57.4 ng) and Sobral 2S male’s pheromone gland content (Fig 2C) peaked at 7 days (199.9 ± 34.3 ng).
The biggest increase in pheromone gland content was between 1st and 2nd day old Jacobina males when the average amount of pheromone stored in the glands increased from 10.7 ± 3.0 ng to 102.3 ± 18.9 ng, a 10-fold increase. The Jacobina gland content remained relatively constant for the remainder of the time with a slight dip in pheromone content in older males. A similar increase was observed in Campo Grande males, where the amount of pheromone stored in 2-day old males showed a 10-fold increase (from 11.1 ± 3.4 ng to 117.7 ± 28.7 ng) compared to 1-day old males. Pheromone continued to be produced and content reached a peak at 9 days, after which time it also declined. Pheromone accumulation in the Sobral 2S males was markedly different to the other two chemotypes. There was no increase in the gland content between 1 and 2 days, but the amount of pheromone increased gradually and reached a peak at 7 days, after this time gland content was greatly reduced.
When males were first placed in the 50 ml r/b entrainment flask, they were very active, and appeared to compete with each other to establish space around themselves. This activity occurred both when females were present and absent. Within the first hour of the entrainment, males spent 15–45 min fanning their wings, a behaviour that is associated with pheromone release. Afterwards, they remained mostly motionless for several hours, periodically repositioning themselves within the flask.
The general pattern of pheromone release for all three chemotypes was that it was greatest during the 1st h of the entrainment. In Jacobina and Campo Grande chemotypes release was significantly reduced during the following 3 h, with release rate values approaching zero in some samples after 2–4 h of continuous entrainment (Figs 3 and 4).
The amount of pheromone released by the Sobral 2S chemotype males was much lower than for the other chemotypes and the decrease in release over time was more gradual (Fig 5). Nevertheless, this hourly pattern of pheromone release was observed in all three chemotypes in all experimental regimes, both L and D regimes, and in male and male + female entrainments.
Males always released more pheromone when females were present during the 1st h of the entrainment for all three chemotypes. However, after the 1st h, the presence of females made no significant difference to the release of pheromone for either the Jacobina or Campo Grande chemotypes (Figs 3 and 4). By contrast, Sobral 2S males in the presence of females continued to release more pheromone per hour than males alone although in general without significant differences (Fig 5).
For the Jacobina chemotype, 1.8 times more pheromone was released when females were present under light conditions (χ2 = 5.4, df = 1, P ≤ 0.05) and 2.3 times when females were present in dark conditions (χ2 = 3.0, df = 1, P ≤ 0.05) (Fig 3). For the Campo Grande chemotype, 2.1 times more pheromone was released when females were present in light (χ2 = 5.4, df = 1, P ≤ 0.05), and 1.2 times more when females were present in dark (χ2 = 0.3, df = 1, P = ns) (Fig 4). For Sobral 2S, 2.3 times more pheromone was released when females were present in light (χ2 = 5.7, df = 1, P ≤ 0.05) and 3.2 times more when females were present in dark (χ2 = 3.8, df = 1, P ≤ 0.05) (Fig 5).
Light conditions appeared to have little effect on pheromone release by the Jacobina and Campo Grande chemotypes (Table 1). However, they were important for pheromone release by the Sobral 2S chemotype where males and males + females released more pheromone under light compared to dark conditions (Table 1).
For males and males + females of both the Jacobina and Campo Grande chemotypes there was no significant difference in the amount of pheromone released for any of the 4 h periods of both light regimes (Table 1). For the Sobral 2S chemotype, males released almost twice as much pheromone during the 1 st h in light compared to the dark (2.6 ± 0.6 ng vs 1.0 ± 0.3 ng; χ2 = 5.0, df = 1, P ≤ 0.05). A higher increase was found during the 2 nd h (1.9 ± 0.8 ng vs 0.06 ± 0.1 ng; χ2 = 5.2, df = 1, P ≤ 0.05). When females were present there was a significant difference in the pheromone released in light vs dark conditions only during the 3rd h of entrainment (2.3 ± 1.1 ng vs 0.06 ± 0.1 ng; χ2 = 6.2, df = 1, P ≤ 0.05) (Table 1).
During the 1st h of entrainment, a significant difference in the average amount of pheromone released by the males and males + females was observed between the three chemotypes. Campo Grande males released the greatest quantities of pheromone, followed by Jacobina males and then Sobral 2S males.
The average pheromone release of individual, 4–5 day old males was significantly different for each of the three chemotypes (χ2 = 22.6, df = 2, P ≤ 0.001). Campo Grande males released 40.3 ± 12.6 ng, Jacobina chemotype males released 16.0 ± 4.8 ng and Sobral 2S chemotype males 2.0 ± 0.9 ng.
The average amount of pheromone released by 4–5 day old males kept with females was also significantly different for each of the three chemotypes (χ2 = 15.1, df = 2, P ≤ 0.001). As before, the Campo Grande chemotype males released more pheromone (66.6 ± 17.7 ng), followed by the Jacobina chemotype males (31.9 ± 8.4 ng) and the Sobral 2S chemotype (4.8 ± 1.3 ng).
After 24 h of entrainment, 5-day old Campo Grande males had released 92.6 ± 32.1 ng of pheromone; the Jacobina males 61.3 ± 37.2 ng and the Sobral 2S males 90.9 ± 38.9 ng. Five days old, Campo Grande males had 125.5 ± 37.6 ng stored in their pheromone glands, Jacobina males 97.1 ± 7.8 ng and Sobral 2S 131.0 ± 16.9 ng. Thus, overall, males had released 73.7%, 63.3% and 67.4% of pheromone relative to their gland contents, respectively.
This study shows that in L. longipalpis, the content of the pheromone gland, which is likely to be closely related to biosynthesis and release, is influenced by both internal and external factors. The age of commencement of pheromone production varies between members of the complex. The amount of pheromone present in the gland depends on the chemotype, and the age of the male sand fly. Pheromone gland content is not linearly related to age and in both of the methylsesquiterpene-producing chemotypes (Jacobina and Campo Grande) there is a period of significant increase in gland content, which may reflect either a period of increased pheromone production or improved ability to store the pheromone. In the Sobral 2S chemotype, the gland content increased over a much longer period, from 2 to 7 days. This may reflect changes in the rate of production and/or accumulation of pheromone within the gland; in any case, the pattern is markedly different from the other two chemotypes.
In this study we have also demonstrated that the amount of pheromone released by males depends on several factors, which include whether or not females are present and the activity of the males. The light conditions in which they were held only affected pheromone release in the Sobral 2S chemotype.
Previous studies on the ultrastructure of pheromone glands of L. longipalpis have shown that male sand flies have structures that could be involved in the storage and subsequent release of sex-aggregation pheromone [38, 39]. These structures appear to develop in synchrony with pheromone gland cell maturation and pheromone content [28]. As we detected the presence of traces of pheromone 6–8 h after eclosion in Campo Grande and Sobral 2S males, and after 12–14 h in the Sobral 2S chemotype, we confirm previous work in which pheromone synthesis was shown to commence 10–14 h post emergence [28, 38, 39]. As technical developments in GC/MS detectors lead to improved sensitivity, it is likely that pheromone production will be seen to start at an even earlier age.
Previous studies have shown that males up to 4-days old [28] or 9-days old [38] have pheromone present in their glands but we have shown in this study that pheromone is present in the glands of 15-day old males. The biological significance of this remains to be determined but clearly suggests that production, storage and potentially release of pheromone is lifelong for males despite younger males having greater success at obtaining matings [39].
The pheromone gland content in the Campo Grande and Jacobina chemotypes increased by 1000% during a 24 h period between 1 and 2 days after emergence. By comparison, the gland content of Sobral 2S chemotype males increased more slowly (300% between day 2 and 3) and started when the males were older. The amounts of pheromone continued to increase in males of all three chemotypes up to 7–9 days old, and thereafter remained constant or gradually declined. This drop, which was particularly noticeable in Sobral 2S males, may partly account for the lack of mating success in older male sand flies [39]. A drop in pheromone gland content with advancing age is common in many groups of insects, such as flies, e.g. Isoceras sibirica [40], and moths, e.g. Ctenopseustis spp., Planotortrix octo and Epiphyas postvittana [41]. The differences in pheromone gland content between the three chemotypes and in particular between the methylsesquiterpene producing populations (Jacobina and Campo Grande) and the diterpene population (Sobral 2S) suggest important behavioural or reproductive differences.
A detailed analysis of the terpene composition of some wild populations of L. longipalpis from Brazil was previously described [24]. The quantity of pheromone found in our Sobral 2S laboratory colony males (132.0 ± 37.9 ng male-1 and 199.8 ± 90.9 ng male-1 at 5 and 7 days old, respectively) was similar to the amounts found in wild-type mixed-age males (167.9 ± 52.4 ng male-1).
The amounts of 9MGB from wild specimens collected from different parts of Brazil varies considerably, e.g. males from Lapinha (Minas Gerais State) produced 116.5 ± 13.5 ng male-1 of pheromone which was significantly more than wild type Sobral 1S males (47.8 ± 10.6 ng male-1). Our laboratory colonised Campo Grande males produced 125.5 ± 9.6 ng male-1 at 5 days and 184.4 ± 10.0 ng male-1 at 7 days. This may indicate that wild sand flies contain less sex-aggregation pheromone than laboratory colonised sand flies as a consequence of environmental factors, e.g. temperature, relative humidity, diet, habitat, season and/or physiological stage [42–44]. However, these differences could also occur because the 2 Sobral populations are sympatric whereas the others are allopatric [24, 45, 46]. In addition, these differences may reflect the population substructuring seen across Brazil [12].
Male behaviour in the entrainment flasks was typical of “lekking” males and involved parading, wing flapping, wing fanning and/or wing vibrating, walking forward in short bursts and changing directions, and fighting with other males [39, 47, 48]. Wing flapping and/or fanning has been suggested by some authors to be a way of distributing pheromone [30, 33] and males that fan their wings more than competitor males are more likely to be successful in obtaining a mate [29]. Our results suggest that during this intense activity the males partly deplete their pheromone glands, as there is a notable decline in the amount of pheromone released after the 1st h of entrainment compared to the next 3 h.
After this initial period of activity, males remained largely motionless for several hours. This stage called “quieting” was reported as a possible indication of pheromone communication because the males that were distributed with regular spacing around the flask, may have established a dominance hierarchy [47]. Although the males may use this “quieting” time to replenish pheromone reservoirs after periods of heavy demands, our results suggest that even after 24 h the pheromone in the glands are not completely depleted. It would be useful to measure the amount of pheromone in the glands during and after the resting period to see whether gland content is restored and how long the recovery period is.
More pheromone was released by males of all three chemotypes when conspecific females were present. This was most noticeable during the 1st h when increases of between 15 and 300% were observed. When both sexes were held together, an unquantified increase in male activity including wing fanning and other behaviours were observed, which may account for the greater release of pheromone.
Overall, our study showed that light had little or no effect on pheromone release from males of the Jacobina and Campo Grande chemotypes. However, Sobral 2S males released more pheromone either when alone or with females in light conditions compared to dark. Whether or not the light regime also influences pheromone production by the Sobral 2S chemotype remains to be determined.
In the future it will be interesting to determine the role of other factors on pheromone gland content and release. Studies on other insects have shown that the density and numbers of males and male diet cause changes in the sexual signalling and mating behaviour [49, 50]. In L. longipalpis other factors such as body size [39] and gland or tergite width [29] have been reported to have no effect. From these studies it is clear that pheromone gland content and release are dynamic and responsive processes.
The work presented here is the first attempt to provide a comparative analysis of the sex-aggregation pheromone gland content and factors that might influence the release of the pheromone of the three most widespread chemotypes of L. longipalpis from Brazil. Although this study was conducted under laboratory conditions on only one population of each of the chemotypes, it is clear that a number of factors can influence pheromone release and that it is a dynamic and not a passive process. We have shown that the presence of females and light conditions influence pheromone release. The observations suggest that significant behavioural, biosynthetic and ecological differences between the three chemotypes and in particular between the methylsesquiterpene chemotypes (Jacobina and Campo Grande) and the diterpene chemotype (Sobral 2S) occurs. It is likely that there are other factors that may also be important but which we have not investigated, for example numbers of males and the presence of host odour on the pheromone gland content. Further work is also needed to identify the circumstances under which males produce and store pheromone and how other factors could positively or negatively impact both of these. In this study we did not investigate the effect of any of these factors on pheromone gland content and it will be interesting in the future to examine these factors also.
An exploration of other aspects of the sex-aggregation pheromones, such as their volatility and the subsequent dispersion of these chemicals, is critical to understanding their behavioural and ecological importance.
We have looked at three different chemotypes of L. longipalpis and each of these represent a member of the species complex. However, the precise nature of the L. longipalpis species complex is unclear and there are several contradictory views on the numbers of sibling species and their associations that have recently been reviewed [10].
A better understanding of chemical communication will also be useful for developing synthetic pheromone for control and monitoring applications, e.g. an understanding of differences in rates and patterns of pheromone release by the methylsesquitepene and diterpene producing members of the L. longipalpis complex may require different approaches to the practical application of these chemicals.
Finally, we believe that sex-aggregation pheromones are a valuable tool for defining members of the L. longipalpis species complex. The different pheromones represent important differences between members of the complex because they act as a premating isolating barriers. These variations are underpinned by significantly different biosynthesis in males and receptor biology in females. The differences between the chemotypes that this study highlighted are therefore valuable additional supporting evidence to ultimately help define the members of the complex.
|
10.1371/journal.pntd.0002119 | Activity of Oxantel Pamoate Monotherapy and Combination Chemotherapy against Trichuris muris and Hookworms: Revival of an Old Drug | It is widely recognized that only a handful of drugs are available against soil-transmitted helminthiasis, all of which are characterized by a low efficacy against Trichuris trichiura, when administered as single doses. The re-evaluation of old, forgotten drugs is a promising strategy to identify alternative anthelminthic drug candidates or drug combinations.
We studied the activity of the veterinary drug oxantel pamoate against Trichuris muris, Ancylostoma ceylanicum and Necator americanus in vitro and in vivo. In addition, the dose-effect of oxantel pamoate combined with albendazole, mebendazole, levamisole, pyrantel pamoate and ivermectin was studied against T. muris in vitro and additive or synergistic combinations were followed up in vivo.
We calculated an ED50 of 4.7 mg/kg for oxantel pamoate against T. muris in mice. Combinations of oxantel pamoate with pyrantel pamoate behaved antagonistically in vitro (combination index (CI) = 2.53). Oxantel pamoate combined with levamisole, albendazole or ivermectin using ratios based on their ED50s revealed antagonistic effects in vivo (CI = 1.27, 1.90 and 1.27, respectively). A highly synergistic effect (CI = 0.15) was observed when oxantel pamoate-mebendazole was administered to T. muris-infected mice. Oxantel pamoate (10 mg/kg) lacked activity against Ancylostoma ceylanicum and Necator americanus in vivo.
Our study confirms the excellent trichuricidal properties of oxantel pamoate. Since the drug lacks activity against hookworms it is necessary to combine oxantel pamoate with a partner drug with anti-hookworm properties. Synergistic effects were observed for oxantel pamoate-mebendazole, hence this combination should be studied in more detail. Since, of the standard drugs, albendazole has the highest efficacy against hookworms, additional investigations on the combination effect of oxantel pamoate-albendazole should be launched.
| The roundworm Ascaris lumbricoides, the whipworm Trichuris trichiura and the two hookworm species Ancylostoma duodenale and Necator americanus are responsible for the most common infections worldwide and place more than 5 billion people at risk. To control these infections, at risk populations are treated regularly with anthelminthic drugs, mostly albendazole and mebendazole. Since both drugs have a low therapeutic effect against T. trichiura, alternative drugs should be discovered and developed. Possible strategies are to re-evaluate forgotten compounds and to thoroughly study drug combinations. We evaluated the activity of the “old”, veterinary drug oxantel pamoate against T. muris, Ancylostoma ceylanicum and Necator americanus in vitro and in vivo. In addition, we studied the activity of oxantel pamoate combinations with the four standard treatments for soil-transmitted helminthiasis. Our results confirm that oxantel pamoate has excellent trichuricidal properties. We show that the drug lacks activity against hookworms. It is therefore necessary to combine oxantel pamoate with an anti-hookworm drug. Synergistic effects were observed with oxantel pamoate-mebendazole in our study. Additional preclinical studies should be launched with oxantel pamoate-mebendazole as well as oxantel pamoate-albendazole, since albendazole is the most widely used and efficacious anti-hookworm drug.
| Infections with the three major soil-transmitted helminth (STH) species, Ascaris lumbricoides, Trichuris trichiura and the hookworms Necator americanus and Ancylostoma duodenale are among the most common parasitic diseases in areas of rural poverty in developing countries [1]. In regions where soil-transmitted helminthiasis is endemic, preventive chemotherapy, i.e. regular anthelminthic drug administration to all people at risk of morbidity, is one of the key strategies [2]. In 2009 it was estimated that 204 million school-aged children were treated for soil-transmitted helminthiasis [3]. The benzimidazoles, albendazole and mebendazole are the most widely used drugs in preventive chemotherapy programs. At present, two alternative drugs, pyrantel pamoate and levamisole are available but currently have a less prominent role since they require weight-based dosing [4]. Despite their excellent safety profile, these drugs have serious limitations with regard to their efficacy. When delivered as a single dose, as in preventive chemotherapy programs, all four compounds have a limited effect against infections with T. trichiura as shown in a recent meta-analysis [5]. In addition, drug resistance is a concern [4]; [6]. Efforts are therefore ongoing to discover and develop the next generation of anthelminthic drugs [7]. Promising strategies to identify potential anthelminthic drug candidates are to assess compounds derived from animal health, to re-evaluate forgotten compounds and to thoroughly study drug combinations [7], [8].
Oxantel is the meta-oxyphenol analog of pyrantel. It was discovered in the early 1970s by Pfizer and showed high activity in T. muris-infected mice and T. vulpis-infected dogs [9], [10]. Subsequent exploratory clinical trials demonstrated that the drug was safe and effective in the treatment of trichuriasis [11]–[14]. For example, complete cure was observed in 10 T. trichiura-infected patients treated with 20 mg/kg oxantel pamoate [11]. In veterinary medicine oxantel pamoate was later combined with pyrantel pamoate, which has, with the exception of activity against Trichuris spp., a broad spectrum of activity against different nematodes [15]. Today oxantel-pyrantel is widely available as a dewormer for dogs and cats. The combination of oxantel-pyrantel was also evaluated in a few clinical trials against human STH infections [13], . For example, a decade ago oxantel-pyrantel (10 mg/kg) was tested in school-aged children on Pemba island. The combination achieved cure rates of 38.2% and 12.7% against infections with T. trichiura and hookworms, respectively [19]. To our knowledge, despite the interesting trichuricidal properties of oxantel, combinations of this drug with other recommended anthelminthic drugs have not been evaluated to date.
The aim of the present study was to investigate the trichuricidal potential of oxantel pamoate combined with the four WHO recommended anthelminthic drugs for the treatment of hookworm, T. trichiura and A. lumbricoides infections (albendazole, mebendazole, levamisole or pyrantel pamoate) as well as combinations of ivermectin and oxantel pamoate. Ivermectin, the first line drug for strongyloidiasis, is known to have trichuricidal properties and combinations of albendazole-ivermectin and mebendazole-ivermectin have been tested clinically [20]. In a first step the EC50 (ED50) values of oxantel pamoate against T. muris were determined in vitro and in vivo. Next to oral administration we also tested the activity of intraperitoneal oxantel pamoate in mice. We then elucidated whether oxantel pamoate combined with albendazole, mebendazole, ivermectin, levamisole or pyrantel pamoate interacts in an additive, antagonistic or synergistic manner in vitro using the combination index equation [21]. Additive and synergistic combinations were followed up in vivo. In addition, the activity of oxantel pamoate was studied against A. ceylanicum and N. americanus in vitro and in vivo.
Albendazole and levamisole were purchased from Fluka (Buchs, Switzerland), oxantel pamoate, mebendazole, ivermectin and pyrantel pamoate were obtained from Sigma-Aldrich (Buchs, Switzerland). Note that, the pamoate salts of oxantel and pyrantel contain only 35.8% and 34.7% of the active ingredients, oxantel and pyrantel base, respectively.
For in vitro studies, drug stocks (5–10 mg/ml) were prepared in 100% DMSO (Sigma-Aldrich, Buchs, Switzerland) and stored at 4°C pending usage. For in vivo studies, the drugs were suspended in 10% Tween 80 [80% EtOH (70∶30 v/v)] (Buchs, Switzerland) and 90% dH2O shortly before treatment.
Four week-old female C57BL/10 mice and 3 week-old male Syrian golden hamsters were purchased from Charles River (Blackthorn, UK and Sulzfeld, Germany, respectively). Before infection, animals were allowed to acclimatize for one week in our animal facility. They were kept in groups of maximum ten (mice) or three (hamsters) in macrolon cages with free access to water and rodent food pellets (Rodent Blox from Eberle NAFAG, Gossau, Switzerland).
Experiments were performed in an attempt to comply with the 3R rules for animal experiments. The current study was approved by the cantonal veterinary office Basel-Stadt (Switzerland) based on Swiss cantonal and national regulations (permission no. 2070).
All the data obtained were analyzed by Excel (Microsoft Office, 2007). In vitro data obtained from the individual motility assays were averaged and normalized to the controls. IC50s (median-effect dose), defined as the concentration of a drug required to decrease the mean worm's motility to 50% at the 72 hour time point, were calculated with the CompuSyn software (CompuSyn, version 3.0.1). The combination index (CI) was calculated for the combination chemotherapy data with CompuSyn. To test the significance of the WBRs in vivo, the Kruskal-Wallis (several treatment doses vs. controls) or the Mann-Whitney U test (one treatment dose vs. control) was applied, using StatsDirect (version 2.4.5; StatsDirect Ltd; Cheshire, UK).
Since the introduction of albendazole, mebendazole, levamisole, and pyrantel pamoate in the human armamentarium to treat STH infections 3–4 decades ago, successes in the discovery and development of a novel nematocidal drug have been limited. The danger of resistance development therefore raises concern for the availability of effective therapies in the future. Furthermore, all four above-mentioned drugs have a limited activity against Trichuris spp when administered as single oral doses. To accelerate the discovery of novel anthelminthic treatments potential drug candidates have recently been examined in vitro, in vivo and in clinical trials. Disappointingly, nitazoxanide, a potential drug candidate identified through systematic literature searches [7] as well as a combination of albendazole and nitazoxanide revealed low trichuricidal activity in a randomized placebo controlled trial on Pemba [31]. Furthermore, monepantel, a safe nematocidal drug recently marketed for veterinary use showed a very poor activity against Ascaris suum and T. muris in vitro and in vivo [25]. Hence, neither nitazoxanide nor monepantel can be recommended for the treatment of infections with STH.
In the present work, another potential candidate, oxantel pamoate, widely used in veterinary medicine was evaluated against T. muris and hookworms in vitro and in vivo. Note that one limitation of our study (and helminth drug discovery in general), is that in vitro testing relied on motility scoring using microscopy, which is a subjective examination procedure [32].
Oxantel pamoate revealed an excellent trichuricidal activity in mice. We calculated an ED50 of 4.7 mg/kg in T. muris-infected mice. A similarly low ED50 of 1.7 mg/kg was reported previously in this model [33]. For comparison, the WHO recommended drugs for the treatment of STH infections are characterized by much higher ED50 values against T. muris in vivo, namely 345 mg/kg for albendazole, 79 mg/kg for mebendazole, 46 mg/kg for levamisole and >300 mg/kg for pyrantel pamoate [25], [30]. Ivermectin, used in the treatment of strongyloidiasis and filarial infections, displayed a comparable ED50 value of 4 mg/kg in our T. muris model [29]. A dose of 10 mg/kg oxantel pamoate administered intraperitoneally lacked activity in T. muris-infected mice. For comparison, the same i.p. dose of ivermectin resulted in a high reduction of the worm load (>93%). This demonstrates that in contrast to ivermectin oxantel pamoate does not kill the worm via the blood stream.
Oxantel pamoate lacked in vivo activity against both hookworm species A. ceylanicum and N. americanus. This finding is in line with a previous study in A. caninum-infected mice [34]. Interestingly, N. americanus adults were affected by the drug in vitro while no activity was observed on A. ceylanicum. To our knowledge, the activity of oxantel pamoate against hookworms has not been studied in humans.
Oxantel pamoate showed also no effect against the third major soil-transmitted helminth species, A. lumbricoides in humans (all 53 patients treated with oxantel revealed Ascaris eggs in the stools collected posttreatment regardless of the dose administered) [11]. It is therefore necessary to combine oxantel pamoate with a partner drug with a therapeutic profile that covers roundworms and hookworms. In the present work we have, for the first time, thoroughly evaluated the potential of oxantel pamoate in drug combinations. This work builds on a series of laboratory investigations on the potential of combination chemotherapy for the treatment of STH infections. We have for example recently examined combinations of marketed drugs in in vitro and in vivo studies against T. muris [8].
Interestingly, antagonistic effects were observed in the present work with oxantel pamoate-pyrantel pamoate against T. muris in vitro, hence this combination was not pursued further. However, we cannot exclude a better trichuricidal effect in vivo for this combination, in particular as a pharmacodynamic interference at the target is unlikely. Oxantel is classified as an N-subtype AChR agonist, while pyrantel is considered an L-subtype suggesting differences in drug action [35]. The combination of oxantel pamoate-pyrantel pamoate is widely used in veterinary medicine and has also been studied in several human clinical trials. For example, in Korea oxantel pamoate-pyrantel pamoate at 20 mg/kg achieved a cure rate of 75% and egg reduction rate of 97% against T. trichiura infections and cleared A. lumbricoides infections [13]. A high egg reduction rate against T. trichiura following oxantel pamoate-pyrantel pamoate at 20 mg/kg was also reported in a Malaysian study [18]. A lower effect of this combination administered at 10 mg/kg was observed on Pemba with cure rates of 96.3, 38.2 and 12.7% against A. lumbricoides, T. trichiura and hookworm, respectively [19]. In two Korean trials both oxantel monotherapy as well as an oxantel-pyrantel combination were used, however since different formulations were used (syrup versus tablets), different dosages applied and sample sizes were small no conclusion can be drawn whether the combination was superior to oxantel monotherapy [13], [14].
Antagonistic effects were observed in vivo using combinations of oxantel pamoate-albendazole, oxantel pamoate-levamisole and oxantel pamoate-ivermectin. Since the molecular basis for the actions of these drugs is not yet fully elucidated it is impossible to explain the antagonistic interaction profile observed for these combinations. On the other hand, the oxantel pamoate-mebendazole combination revealed highly synergistic effects against T. muris in vivo. It is striking that the two benzimidazole derivates behave so differently when administered as partner drugs in oxantel pamoate combinations to T. muris infected mice given that both drugs, despite their differences in pharmacokinetics [36], have identical targets. However, our results should be interpreted with caution. First of all, drug scheduling and drug vehicle, solubility, host behavior, environmental factors and genetic variations might influence the level of activity [37]. In addition, though the median effect method used in the present work is the most commonly used, our data are based on a single method only and one could have considered applying another method, such as the isobologram method to re-analyze the data [38]. Finally, note that these findings are based on a single ratio of the combined agents (ED50 values) and it might be worthwhile to assess other ratios of the drug dosages.
In conclusion, our study confirms that oxantel pamoate has excellent trichuricidal properties. In the T. muris mouse model oxantel pamoate showed a higher activity than the standard drugs albendazole, mebendazole, levamisole and pyrantel pamoate. Since the drug has no activity against hookworms it is necessary to combine oxantel pamoate with a partner drug revealing anti-hookworm properties. Synergistic effects were observed for oxantel pamoate-mebendazole. Despite of our results pointing to an antagonistic behavior of oxantel pamoate-albendazole additional investigations on the effect of this combination might be considered (e.g. evaluation of a different dosing ratio or schedule) since of the standard drugs albendazole has the highest efficacy against hookworms [5]. Systemic drug interactions between oxantel pamoate and partner drugs are unlikely given that the absorption of oxantel pamoate is very poor [11]. Nonetheless, preclinical studies should carefully elucidate metabolic and pharmacokinetic interactions of oxantel pamoate and the benzimidazoles.
|
10.1371/journal.pntd.0007369 | Tracing the environmental footprint of the Burkholderia pseudomallei lipopolysaccharide genotypes in the tropical “Top End” of the Northern Territory, Australia | The Tier 1 select agent Burkholderia pseudomallei is an environmental bacterium that causes melioidosis, a high mortality disease. Variably present genetic markers used to elucidate strain origin, relatedness and virulence in B. pseudomallei include the Burkholderia intracellular motility factor A (bimA) and filamentous hemagglutinin 3 (fhaB3) gene variants. Three lipopolysaccharide (LPS) O-antigen types in B. pseudomallei have been described, which vary in proportion between Australian and Asian isolates. However, it remains unknown if these LPS types can be used as genetic markers for geospatial analysis within a contiguous melioidosis-endemic region. Using a combination of whole-genome sequencing (WGS), statistical analysis and geographical mapping, we examined if the LPS types can be used as geographical markers in the Northern Territory, Australia. The clinical isolates revealed that LPS A prevalence was highest in the Darwin and surrounds (n = 660; 96% being LPS A and 4% LPS B) and LPS B in the Katherine and Katherine remote and East Arnhem regions (n = 79; 60% being LPS A and 40% LPS B). Bivariate logistics regression of 999 clinical B. pseudomallei isolates revealed that the odds of getting a clinical isolate with LPS B was highest in East Arnhem in comparison to Darwin and surrounds (OR 19.5, 95% CI 9.1–42.0; p<0.001). This geospatial correlation was subsequently confirmed by geographically mapping the LPS type from 340 environmental Top End strains. We also found that in the Top End, the minority bimA genotype bimABm has a similar remote region geographical footprint to that of LPS B. In addition, correlation of LPS type with multi-locus sequence typing (MLST) was strong, and where multiple LPS types were identified within a single sequence type, WGS confirmed homoplasy of the MLST loci. The clinical, sero-diagnostic and vaccine implications of geographically-based B. pseudomallei LPS types, and their relationships to regional and global dispersal of melioidosis, require global collaborations with further analysis of larger clinically and geospatially-linked datasets.
| Burkholderia pseudomallei is a pathogenic soil bacterium that causes the disease melioidosis, which occurs in many tropical regions globally and in recent years has emerged in non-tropical regions. Melioidosis has been predicted to affect 165,000 people every year resulting in an estimated 89,000 deaths. Person to person transmission is rare with most cases linked to exposure to the bacterium from the environment. The genetic background of B. pseudomallei has been well studied and variably present genes have been linked to distinct melioidosis disease states and geographic regions, however we still need a stronger understanding of the association of genes with geography. Three lipopolysaccharide types exist in B. pseudomallei and the prevalence of the lipopolysaccharide genes vary between melioidosis endemic regions, but it is unknown if the lipopolysaccharide genes can be used as geographical markers in a single melioidosis-endemic region. In this study, we used a combination of whole-genome sequencing, statistics and geographical mapping to elucidate if the three lipopolysaccharide genes can be used as geographical markers within the Northern Territory, Australia. We show that the three LPS types have distinct but overlapping geographical footprints within a single melioidosis region and can be used as geographic markers alongside a number of other important variably present B. pseudomallei genes.
| Burkholderia pseudomallei is the causative agent of melioidosis, an often fatal disease that is endemic in tropical regions globally, including the “Top End” of the Northern Territory, Australia [1]. Clinical presentations of melioidosis are highly varied [2–5] with pneumonia being the most common presentation [6]. Melioidosis mortality rates range from <10% in the NT to >40% in regions of Asia and Southeast Asia where prompt diagnosis and accessibility to drugs and hospital care can be limited [2, 6, 7]. Global melioidosis mortality has been predicted to be considerably higher than mortality from dengue and leptospirosis combined and similar to that from measles [8]. In addition, the bio-threat status of B. pseudomallei and the lack of a current vaccine makes this organism of high importance to public health in many regions globally.
Variable genetic markers in B. pseudomallei with known geographical associations include the mutually exclusive Burkholderia intracellular motility factor A (BimA) variants bimABm and bimABp, the filamentous hemagglutinin 3 (fhaB3) gene and the Burkholderia thailandensis-like flagellum and chemotaxis (BTFC/YLF) gene clusters [9, 10]. These genetic markers have been used to elucidate strain relatedness, origin and virulence potential, with bimABm being strongly linked to neurological melioidosis in Australian studies [9]. The lipopolysaccharide (LPS) of B. pseudomallei, a type II O-polysaccharide [11], consists of A, B and B2 variants [12]. As with bimABm/bimABp, YLF/BTFC, and fhaB3, the distribution of LPS genotypes varies between Australia and Southeast Asia [9, 12] and it is unknown what LPS types predominant in other endemic regions. In Thailand and Australia, LPS type A is the most abundant, followed by LPS B, whereas LPS B2 has not been detected in Thailand but has been detected in Australia [12]. Despite regional differences in LPS genotype prevalence, the distribution of these types across different geographical regions within a single melioidosis-endemic region has not been thoroughly investigated. Elucidating the geospatial patterns of LPS genotypes is potentially important for B. pseudomallei virulence studies, the development of LPS-based sero-diagnostics, and for consideration of vaccines that have an LPS component.
In this study, we used a combination of WGS, bivariate logistic regression (with the cluster feature) and geographical mapping to examine the geographical distribution of LPS types in the Top End region. We first investigated LPS geographic correlations using a large number of clinical isolates (n = 1005) from our 28-year Darwin Prospective Melioidosis study (DPMS). We then investigated LPS type distributions for 340 environmental strains collected from known Top End locations to determine whether LPS genotypes for clinical isolates, which have usually only presumptive infection locale data, matched that of the environmental isolates, where location is definite. Lastly, we determined bimABm/bimABp diversity in the 340 environmental strains to determine whether the bimA variants and LPS types share similar geographical footprints in the NT.
This study was approved by the Human Research Ethics Committee of the NT Department of Health and the Menzies School of Health Research (HREC 02/38).
A total of 1,345 B. pseudomallei isolates from the Top End region were used in this study for bivariate analysis and geographical mapping. 1,005 clinical isolates were included in this study and 999 (excluding the 6 B2 strains) were used for bivariate statistical analysis, and environmental isolates (n = 340) were used for geographical mapping. The 1,005 clinical isolates were primary B. pseudomallei isolates corresponding to 1,005 patients enrolled in the 28-year DPMS. The clinical isolates were assigned to one of four Top End geographic regions based on the presumptive location of infection, which is assigned based on the individual patient’s history of illness and activities at their residential location or elsewhere if considered relevant. These regions have been segregated and assigned as outlined by the Australian Bureau of Statistics (https://www.abs.gov.au/; statistical area level 3): Darwin and surrounds; Darwin remote (includes West Arnhem), Katherine and Katherine remote (representing one region, Fig 1), and East Arnhem (Fig 1). Melioidosis is very uncommon in the southern half of the NT and this region was not included in our study.
The 340 Top End environmental isolates were processed and subsequently B. pseudomallei was confirmed as previously described [13–17]. The location for each of the 340 environmental isolates was determined using Global Positioning System coordinates, with mapping of each isolate coordinate using ArcGIS (https://www.arcgis.com/index.html) and maps were exported as JPEG and manipulated using INKSCAPE 0.92 (https://inkscape.org/release/inkscape-0.92/). For comparison, we included an additional 61 isolates of Queensland (n = 38), Western Australian (n = 12) and Papua New Guinean (n = 11) origin.
The assignment of LPS types for the 1,005 clinical genomes has previously been determined [18] using the Basic Local Alignment Search Tool (BLAST). The clinical genomes were not assigned a bimA variant as this has previously been performed on our clinical strains [10]. LPS and bimA genotypes were determined for each of the 340 Top End environmental genomes, with an LPS type being assigned to those additional 61 isolates from Western Australia, Papua New Guinea and Queensland, Australia. In brief LPS A (wbil to apaH in K96243 [GenBank ref: NC_006350]), LPS B (BUC_3392 to apaH in B. pseudomallei 579 [GenBank ref: NZ_ACCE01000003]), LPS B2 (BURP840_LPSb01 to BURP840_LPSb21 in B. pseudomallei MSHR840 [GenBank ref: GU574442]), bimABm (BURPS668_A2118 in B. pseudomallei MSHR668 [GenBank ref: NZ_CP009545]) and bimABp (BPSS1492 in B. pseudomallei K96243) were BLAST-searched against an in-house B. pseudomallei database containing all genomes in this study using the nucleotide BLAST (BLASTn) parameter.
For each clinical strain, the presumptive geographical location of infection and its sequence type (ST) (https://pubmlst.org/bpseudomallei/) were analysed using Stata version 14.2 (StataCorp LP, College Station, TX, USA). Bivariate logistics regression using the cluster feature was performed for LPS type with locale using the Darwin and surrounding region as the reference level, and bivariate Pearson’s χ2 (frequency >5 in all cells) was performed for LPS type with ST or BimA variants: A P < 0.05 was considered significant.
We performed comparative analyses of 175 B. pseudomallei genomes that represent both the global diversity of this bacterium and within-Northern Territory population diversity (S1 Table). Of these 175 genomes, 144 were publicly available and 31 were sequenced as part of this study. Genomic DNA for 31 isolates lacking WGS data was extracted as previously described [19]. These isolates were sequenced at Macrogen, Inc. (Gasan-dong, Seoul, Republic of Korea) or Australian Genome Research Facility Ltd. (Melbourne, Australia) using the Illumina HiSeq2000 and Illumina HiSeq2500 platforms (Illumina, Inc., San Diego, CA) and these data are available on NCBI (S1 Table)
Comparative analysis of the 175 genomes was carried out to investigate geographical clustering of the LPS types, to determine the genetic relatedness of suspected ST homoplasy cases [20, 21], and to ascertain if suspected ST homoplasy cases were intercontinental or intracontinental. Orthologous biallelic single-nucleotide polymorphisms (SNPs) were identified from the WGS data using the default settings of SPANDx v3.2 [22]. The closed Australian B. pseudomallei genome MSHR1153 [23] was used as the reference for read mapping. A maximum-parsimony (MP) phylogenetic tree was reconstructed based on 219,075 SNPs identified among the 175 genomes using PAUP* 4.0.b5 [24]. Recombinogenic SNPs were identified using Gubbins v2.2.0 (default parameters) [25]. STs were determined using the BIGSdb tool [26].
BLAST analysis demonstrated that overall in the Top End, LPS A was dominant in both environmental B. pseudomallei strains (A = 89%; LPS B = 10% and LPS B2 = 1%) and in clinical strains (LPS A = 87%; LPS B = 12% and LPS B2 = 1%) (Table 1). Using the clinical B. pseudomallei dataset, we performed bivariate logistics regression (with the cluster feature) of LPS type A and B correlations and presumptive infecting locale, using Darwin urban as the reference level. Low LPS B2 numbers (n = 6) precluded statistical analysis of this genotype, leaving 999 DPMS isolates for the analysis. Of these six B2 isolates, five were from presumptive infections occurring in Katherine and Katherine remote communities, with the remaining LPS B2 strain linked to infection in the Darwin and surrounding region. Bivariate logistics regression adjusted with the cluster feature demonstrated that the odds of getting a clinical isolate with LPS B in the Darwin remote, Katherine and Katherine remote and East Arnhem regions, respectively is 4.7 (95% CI 2.4–9.6; p<0.001), 14.6 (95% CI 6.1–34.7; p<0.001) and 19.5 (95% CI 9.1–42.0; p<0.001) times higher compared to the odds in the Darwin and surrounding region. LPS B prevalence was highest in patients infected in the Katherine and Katherine remote and East Arnhem regions (n = 79; 64% of LPS B isolates), whereas LPS A prevalence was highest in patients infected in the Darwin and surrounds and Darwin remote regions (n = 756; 86% of LPS A isolates).
Geographical mapping of LPS types for the environmental isolates also supported the link between LPS and geography. LPS type A was predominant in Darwin and surrounds (Table 1 and Fig 2B), whilst LPS B was found in all four geographical regions but with numbers highest in the Katherine and Katherine remote region (Fig 2A and 2B, number of LPS B strains at each site is indicated). LPS B2 from environmental strains was only observed in the Katherine remote region (n = 1). In non-Top End comparator strains, LPS A was also the dominant genotype (range = 63–91%), followed by B (range = 9–24%) and B2 (range = 0–18%). Numbers of the LPS type A, B, and B2, respectively, were as follows: Western Australia (10, 2, 0), Queensland (24, 9, 5) and Papua New Guinea (8, 1, 2).
Bivariate analysis of clinical isolates encoding LPS types A and B revealed statistically significant associations between LPS types and STs. LPS A was significantly associated with STs only found in the Darwin and surrounding region, including ST-109 (n = 126; 14% of LPS A’s; P<0.001), ST-36 (n = 81; 9% of LPS A’s; P<0.001) and ST-132 (n = 67; 8% of LPS A’s; P<0.001), whilst LPS B was associated with 88 rare STs that are uncommon in the Darwin regions (S2 Table). These associations therefore reflected the LPS geographical associations observed in clinical strains. Although LPS B2 was not included in the statistical analysis it was found in strains belonging to rare STs (ST-331 n = 1; ST-737 n = 2; ST-770 n = 2; ST-118 n = 1) with four of the STs unique to the Katherine and Katherine remote region. Of the STs identified in the 1,005 DPMS clinical strains, four contained multiple LPS types amongst 18 strains. Strains belonging to ST-468, ST-807 and ST-734 contained LPS A or LPS B, whilst strains belonging to ST-118 contained LPS A or LPS B2. The LPS and ST discrepancy was also observed in the Top End environmental strain cohort, with three strains belonging to ST-1485 containing LPS A or B.
BLAST analysis demonstrated that bimABp was the dominant BimA variant in Top End environmental samples (94.8%), similar to the bimABp prevalence in clinical isolates from a previous study, whereby 95.7% of clinical isolates from Darwin and surrounds possessed bimABp [10]. Bivariate statistical analysis demonstrated that the proportion of LPS B strains carrying bimABm (n = 8, 21%) is higher compared to LPS A (n = 8, 2.6%) and this was significant (P<0.001). We next determined if bimABp or bimABm has a similar geographical footprint to LPS A or LPS B in the Top End. [10]. Geographical mapping of the bimA variants in the environmental strains demonstrated that bimABp was linked with Darwin and surrounds (Fig 3B), having a similar geographical footprint as LPS A. In contrast, bimABm was associated with the East Arnhem and the Katherine and Katherine remote region (13 of 17 (76%) environmental bimA strains), similar to LPS B. We also noted that the majority of the LPS B strains in the Katherine and East Arnhem region also carried bimABm (n = 8/13; 62%), and these strains were concentrated in two geographical regions (Fig 3, star [n = 6/9] and triangle [n = 2/2]).
A phylogenomic analysis was performed to investigate the distribution of LPS types and strain relatedness on the whole-genome level (Fig 4). Overall, good clade structure based on geographical origin was observed, with strains belonging to the same ST clustering together and within the same LPS type, but with LPS A and B being dispersed throughout the phylogeny (Fig 4). However, this analysis revealed that 5/6 LPS B2 genomes of Top End origin clustered together, suggesting that LPS B2 is geographically restricted in the Northern Territory (Fig 4). Similarly, the LPS B2 strains with a Queensland, Australian origin grouped closely together.
The strong link between LPS type, geography and ST suggested that homoplasy may be present in the 18 clinical strains and three environmental strains that belonged to the five STs with more than one LPS type (Table 2). Based on comparative analysis of 175 genomes, isolates with the same ST but mixed LPS type did not group together (Fig 4: “ST homoplasy cases”) and were separated by a large number of SNPs. This is comparable to previously observed MLST homoplasy in B. pseudomallei [20, 21]. To assess the effect of recombination on phylogenetic inference, SNP density filtering was applied to remove recombinogenic SNPs with Gubbins (v.2.3.1), which did not considerably alter the topology of the phylogenetic tree. We also noted that the isolates belonging to ST-807 and ST-468 were from three Top End geographical regions and varied by a large number of SNPs with the largest SNP difference being ~36,000 and ~29,000 respectively (Table 2). This phenomenon represents further examples of intracontinental homoplasy but includes homoplasy occurring in more closely related geographical regions than previously described [20].
Zoonotic and person-to-person transmission of B. pseudomallei is extremely rare [28], with almost every case of melioidosis therefore reflecting a single infecting exposure event from the environment in a specific geographical location. Variably present genetic markers have been documented in B. pseudomallei and include the mutually exclusive BimA variants [29], the BTFC/YLF gene clusters [9] and the variably present FhaB3 locus [30]. The prevalence of these genetic markers varies between Australia and Asia and BTFC/YLF prevalence varies between clinical and environmental strains, and in the case of bimA and fhab3 they have also been linked to clinical variation in disease phenotype [9, 10]. The geographical footprint of fhaB3 and bimA variants has been previously determined in the Northern Territory, with bimABm associated with melioidosis in remote communities while bimABp and absence of fhaB3 is associated with melioidosis in the Darwin urban region [10].
The LPS O-antigen moiety of B. pseudomallei is highly varied with three serotypes described [12]. Nevertheless we recently found that previously published in vitro differences in virulence between LPS A and LPS B did not translate to differences in mortality in DPMS patients, supporting the dominant role in human melioidosis of host risk factors in determining disease severity and outcomes [18]. Regional differences in LPS serotype prevalence have been noted in Southeast Asia and Australia with LPS A being the most prevalent, followed by LPS B and with LPS B2 being the rarest in Australian strains. As yet LPS B2 has only been detected in B. pseudomallei strains from Australia and Papua New Guinea [12]. Despite the reporting of these regional differences in LPS prevalence, no studies have investigated the geographical footprint of the LPS types in a contiguous melioidosis-endemic region. We used a combination of statistical analysis, WGS and geographical mapping to examine the geographical footprint of the LPS genotypes in the Top End of the Northern Territory.
We first examined the prevalence of the LPS types across the environmental and clinical samples, and our data confirm that the prevalence of the LPS types is consistent between the Northern Territory environmental (LPS A = 89%; LPS B = 10% and LPS B2 = 1%) and clinical isolate collection (LPS A = 87%; LPS B = 12% and LPS B2 = 1%) [18], with LPS A being the most prevalent. High prevalence of LPS A is reflected in it also being present in many near-neighbour species [31], with the possibility of horizontal gene transfer of LPS A between these related species. LPS A was also distributed throughout the B. pseudomallei phylogeny, also supporting horizontal transfer of the entire LPS A loci. Based on our large clinical dataset (n = 999) we found strong geographical clustering of LPS types in the Northern Territory, Australia, with the highest odds of having a clinical isolate with LPS B being in the East Arnhem region in comparison to the Darwin region (OR 19.5, 95% CI 9.1–42.0; p<0.001). A similar trend was observed for strains isolated from the Northern Territory environment (n = 340), and in the environmental strains LPS B was not detected in urban Darwin. The successful use of LPS genotypes as geographical markers of B. pseudomallei within the Northern Territory can now be replicated in other melioidosis regions globally to ascertain the global LPS footprint alongside other variable markers such as bimA and fhab3.
LPS provides protective immunity to melioidosis in in vitro models of disease making it a potential vaccine candidate. It has also been demonstrated that antibodies from the two dominant LPS types, LPS A and B are not cross-reactive due to structural differences of the O-antigen, with consequent implications for an LPS based vaccine [12]. The sole presence of LPS A in the Darwin city region where the majority of melioidosis cases occur in the Northern Territory [6], suggests that an LPS A based vaccine strategy would be appropriate in this region. However, in regions where LPS A and LPS B strains are both present such as the Darwin remote region, an LPS A alone vaccine strategy is unlikely to provide adequate protection. For other melioidosis endemic regions globally, determination of the LPS genotype footprints will be critical if LPS vaccines are to be used.
Despite the strong link noted between ST and LPS type, we detected an LPS discrepancy in 21 isolates belonging to five rare STs: ST-118 (n = 3), ST-734 (n = 10), ST-468 (n = 5), ST-807 (n = 3) and ST-1485 (n = 3). Phylogenomic analysis of 175 genomes confirmed that the isolates belonging to the five STs represented five new ST homoplasy occurrences, explaining the LPS discrepancy noted in the five STs. Isolates of each ST were separated by a large number of SNPs (~29,000-~39,000), which is characteristic of Australian isolates that belong to different STs [20, 21]. Despite the large number of SNPs separating isolates within each ST, they all still resided within the Australian clade, representing five novel intracontinental ST homoplasy cases. These findings further highlight the low resolution of MLST and the need for high resolution techniques such as WGS to resolve unexpected genotyping results, particularly for B. pseudomallei which is a highly recombinogenic pathogen [32].
In addition to the phylogeographical role of the different B. pseudomallei LPS genotypes, the potential for differential virulence amongst LPS genotypes and potential interactions with other variable genotypes requires consideration. In vitro models of LPS types have been associated with differential virulence [12, 33, 34]. However, as noted above this did not translate to clinical significance in our large clinical cohort (>1000 melioidosis patients) where mortality, rates of bacteraemia and septic shock were the same for patients with LPS A and LPS B [18]. Nevertheless, it is known that virulence mechanisms work in concert for B. pseudomallei to invade and cause disease and further genomic studies are required to assess whether there are regional combinations of virulence genes including LPS variants that will have predictive capability for variable disease presentations and severity. For example, an association between bimABm and neurological melioidosis has been previously described [10], and a new finding from this study was that eight environmental strains from the Katherine and Katherine remote and East Arnhem region had a combination of bimABm and LPS B. Whether this specific bimABm−LPS B combination has implications for neurotropism of B. pseudomallei now requires elucidation.
The present study encompassed a number of limitations. One limitation of this study is that the assigned presumptive location of infection for cases of melioidosis is dependent on good epidemiological history obtained from patients and will sometimes not be the true location of infection. Nevertheless, the close correlation of geographical LPS footprints between the clinical case isolates and the environmental isolates (which are 100% location accurate) is reassuring that the prospective nature of the Darwin melioidosis study is providing accurate epidemiological data. A second limitation is that the four regions (Darwin and surrounding [persons, 148,564; geographic size, 316,391hectares], Darwin remote [persons, 17,902; geographic size, 11,229,485hectares], Katherine and Katherine remote [persons, 20,839; geographic size, 32,625,027hectares] and East Arnhem [persons, 14,519; geographic size, 3,360,659hectares] https://www.abs.gov.au/) vary substantially in both geographic size and population density, which results in a greater number of isolates obtained from the Darwin and surrounding region and ultimately clustered data. We accounted for the clustered nature of our data by using bivariate logistics regression with the cluster feature.
In conclusion, by combining genomic data with corresponding strain geographical information we have found that in the tropical north of Australia the LPS types have distinct geographical footprints as well as ST associations, adding to the already known variably present genetic markers fhaB3 and the bimA variants. A novel and interesting finding was that the geographical footprint of LPS B and bimABm are similar and in remote Top End locations. The clinical, sero-diagnostic and vaccine implications of geographically-based B. pseudomallei LPS and other gene differentials and their relationships to regional and global dispersal of melioidosis require global collaborations with further analysis of larger clinically and geospatially-linked datasets.
|
10.1371/journal.ppat.1003761 | The Type-Specific Neutralizing Antibody Response Elicited by a Dengue Vaccine Candidate Is Focused on Two Amino Acids of the Envelope Protein | Dengue viruses are mosquito-borne flaviviruses that circulate in nature as four distinct serotypes (DENV1-4). These emerging pathogens are responsible for more than 100 million human infections annually. Severe clinical manifestations of disease are predominantly associated with a secondary infection by a heterotypic DENV serotype. The increased risk of severe disease in DENV-sensitized populations significantly complicates vaccine development, as a vaccine must simultaneously confer protection against all four DENV serotypes. Eliciting a protective tetravalent neutralizing antibody response is a major goal of ongoing vaccine development efforts. However, a recent large clinical trial of a candidate live-attenuated DENV vaccine revealed low protective efficacy despite eliciting a neutralizing antibody response, highlighting the need for a better understanding of the humoral immune response against dengue infection. In this study, we sought to identify epitopes recognized by serotype-specific neutralizing antibodies elicited by monovalent DENV1 vaccination. We constructed a panel of over 50 DENV1 structural gene variants containing substitutions at surface-accessible residues of the envelope (E) protein to match the corresponding DENV2 sequence. Amino acids that contribute to recognition by serotype-specific neutralizing antibodies were identified as DENV mutants with reduced sensitivity to neutralization by DENV1 immune sera, but not cross-reactive neutralizing antibodies elicited by DENV2 vaccination. We identified two mutations (E126K and E157K) that contribute significantly to type-specific recognition by polyclonal DENV1 immune sera. Longitudinal and cross-sectional analysis of sera from 24 participants of a phase I clinical study revealed a markedly reduced capacity to neutralize a E126K/E157K DENV1 variant. Sera from 77% of subjects recognized the E126K/E157K DENV1 variant and DENV2 equivalently (<3-fold difference). These data indicate the type-specific component of the DENV1 neutralizing antibody response to vaccination is strikingly focused on just two amino acids of the E protein. This study provides an important step towards deconvoluting the functional complexity of DENV serology following vaccination.
| Despite decades of research, there remains a critical need for a dengue virus (DENV) vaccine. Vaccine development efforts are complicated by a requirement to protect against four DENV serotypes (DENV1-4), and incomplete immunity as a risk factor for severe disease. Antibodies play a major protective role against DENV. However, they also have been implicated in severe clinical manifestations of DENV infection. The antibody response to DENV is composed of antibodies that neutralize only the infecting DENV serotype (type-specific), as well as those that are cross-reactive. Cross-reactive antibodies are hypothesized to contribute to severe dengue following heterologous infections. Identifying DENV epitopes that are targets of type-specific neutralizing antibodies may facilitate vaccine development and the identification of correlates of protection. In this study, we identified amino acids on DENV1 recognized by type-specific neutralizing antibodies elicited by DENV1 vaccination. Our results indicate that the type-specific DENV1 response is remarkably focused on just two regions of the DENV1 envelope protein. Furthermore, a significant contribution of antibodies with this specificity was a common feature among vaccine recipients. This study identifies targets of neutralizing antibodies elicited by DENV1 vaccination and provides an important first step toward identifying epitopes recognized by each component of a tetravalent vaccine.
| Dengue virus (DENV) is a mosquito-transmitted flavivirus responsible for 390 million human infections each year [1]. Four related serotypes (DENV1-4) circulate in virtually all tropical and sub-tropical regions of the world [2]. While DENV infection is often subclinical, clinical symptoms of dengue fever (DF) include a self-limiting febrile illness, myalgia, rash, and retro-orbital pain [3]. A more severe clinical illness (dengue shock syndrome/dengue hemorrhagic fever) involving capillary leakage, thrombocytopenia, and hemorrhage has been associated with secondary infections by a heterologous DENV serotype and higher viral loads in vivo [4], [5]. The incidence of severe DENV disease is rising globally due to increasing co-circulation of multiple DENV serotypes in endemic areas [2], [6]. Currently, there are no specific treatments or approved vaccines for DENV infection.
Flaviviruses encapsidate a single-stranded RNA genome of positive-sense polarity. This ∼11 kb genomic RNA is translated as a single open reading frame that is cleaved in infected cells by cellular and viral proteases into at least ten proteins [7]. The virus encodes three structural proteins (envelope (E), premembrane (prM), and capsid (C)) that associate with a lipid envelope and the viral genome to form the virion [8]. Flavivirus assembly occurs on virus-induced membranes derived from the endoplasmic reticulum (ER) [9]–[13], resulting in the budding of non-infectious immature virus particles into the lumen. The E protein of immature virions exists as heterotrimeric spikes in complex with the prM protein; sixty of these spikes are organized on the virion with icosahedral symmetry [14]–[16]. During egress through the secretory pathway, prM is cleaved by a cellular furin-like protease to generate the mature infectious virus particle [17]–[19]. Mature virions are characterized by a dense array of antiparallel E protein dimers orientated roughly parallel to the surface of the virion [20]–[22]. In many cases, this arrangement of E proteins imposes steric constraints for epitope recognition by antibodies [23], [24]. The virion maturation process is inefficient for many mosquito-borne flaviviruses, including DENV. Partially mature viruses with structural features of both mature and immature particles may be infectious and differentially interact with antibodies as a function of their prM content (reviewed in [25]).
The humoral response plays an important role in protection against flaviviruses (reviewed in [26]). Development of a neutralizing antibody response is an established correlate of protection following vaccination against yellow fever virus (YFV), Japanese encephalitis virus (JEV), and tick-born encephalitis virus (TBEV) [27]–[30]. Passive transfer of monoclonal antibodies (mAbs) is protective in several animal models of flavivirus infection, including DENV [31]–[36]. The flavivirus E protein (Figure 1a) is the principal target of neutralizing antibodies [37]. Studies with murine and human mAbs have identified neutralizing epitopes on all three structural domains of E (DI–III) [23], [24], [35], [36], [38]–[47]. Recent studies with human mAbs suggest the antibody repertoire may differ from that observed in mice [39], [48]–[52], and have identified a quaternary epitope composed of surfaces on two adjacent E proteins [46], [48], [52]. Antibodies that bind prM also have been identified frequently in studies of human mAbs and typically possess limited neutralization potential in vitro [39], [49], [51], [53]. Beyond a capacity to directly neutralize the infectivity of virions, antibodies may protect the host via effector functions orchestrated by the constant region of the antibody molecule [54]–[57].
Development of a DENV vaccine has been a focus of considerable effort for decades. While the dramatic success of other flavivirus vaccines [28], [58], [59] and the pioneering work of Sabin and colleagues [60] suggest effective DENV vaccination is possible, several unique challenges exist. A DENV vaccine must simultaneously protect against four different viruses. In addition, vaccination must not sensitize the recipient to more severe manifestations of disease in the event of breakthrough. The antibody response to DENV results in the production of both type-specific (TS) and cross-reactive (CR) antibodies that may vary significantly with respect to their capacity to neutralize virus infection [35], [36], [38], [61]–[63]. CR antibodies are hypothesized to contribute to severe clinical outcomes of DENV infection via a process called antibody-dependent enhancement of infection [64]. The extensive cross-reactivity of the DENV antibody response complicates serological studies of these viruses and the identification of immune correlates of protection [65], [66]. For example, a recent large phase IIb trial of a live-attenuated tetravalent DENV vaccine revealed modest protective efficacy (∼30% overall) with no protection at all observed for DENV2 [67]. Importantly, despite the absence of protection, vaccine-induced neutralizing antibody was observed for DENV2. Whether this increase in neutralizing titer was associated with a TS- or CR-response is unknown. This trial underscores the importance of understanding the functional complexity of the DENV antibody response.
In this study we sought to identify the immunodominant epitopes recognized by TS-neutralizing antibodies elicited by DENV vaccination of humans. We constructed libraries of DENV1 variants containing substitutions in the E protein at surface exposed residues that differ between the DENV1 and DENV2 components of the NIAID tetravalent vaccine candidate [68]. We then screened this library for a reduction in sensitivity to neutralization by sera from DENV1 vaccine recipients, but not DENV2 immune sera from vaccinated subjects. Remarkably, these studies identified two amino acids (E126 and E157) that when mutated significantly reduced the DENV1 TS-neutralizing response of more than 77% of recipients of a monovalent DENV1 vaccine. Our studies, for the first time, identify functionally significant epitopes that comprise a TS-neutralizing response and provide insight into the complexity of the DENV humoral response.
The goal of this study was to identify epitopes recognized by DENV1 TS-neutralizing antibodies elicited by vaccination. Our analyses employed immune sera collected during the clinical evaluation of individual components of the NIAID tetravalent vaccine candidate (reviewed by [68]). The attenuated DENV1 in this vaccine (rDEN1Δ30) is derived from the South Pacific genotype 4 Western Pacific (WP) strain [69]. The DENV2 component of the tetravalent formulation (rDEN2/4Δ30(ME)) is the Southeast Asian genotype 2 New Guinea C strain (NGC) [70]. The envelope proteins of these viruses differ by 158 amino acids (Figure 1b); 68 amino acids that differ between these two strains are predicted to be accessible to solvent on the surface of the mature DENV virion (Figure 1c, see Materials and Methods) [71]. Both rDEN1Δ30 and rDEN2/4Δ30(ME) vaccines were tested as monovalent formulations in phase I clinical studies in humans [72], [73]. To identify amino acids recognized by TS-neutralizing antibodies, we created a panel of DENV1 variants to replace surface-accessible residues on the mature DENV1 virion with those found on DENV2 NGC (detailed below). This panel of mutants was then screened using pooled immune sera collected during monovalent DENV1 and DENV2 vaccine studies for those exhibiting reduced sensitivity to neutralization by DENV1, but not DENV2, immune sera.
Construction and functional characterization of libraries of DENV1 E protein variants were performed using DENV reporter virus particles (RVPs). RVPs are produced by complementation of a self-replicating sub-genomic flavivirus replicon with a plasmid encoding viral structural genes. These virus particles are capable of only a single round of infection and allow virus entry to be scored as a function of reporter gene expression. Flavivirus RVPs have been used extensively to study the functional properties of anti-flavivirus antibodies [33], [74]–[78]. This technology allows for the rapid generation of structural gene variants by exchanging the plasmids used in complementation studies, simplifying the construction and characterization of a large library of DENV1 mutants. Furthermore, because the RVPs are not passaged in cell culture, the genetic stability of a particular mutation does not limit its utility in neutralization studies.
Immune sera used to characterize and screen DENV1 mutant libraries were pooled from recipients of DENV1 or DENV2 vaccine candidates (two and three vaccine recipients, respectively) [72], [73]. The neutralization titer (NT50) of antibodies in these pooled serum samples was determined using wild type (WT) DENV1 WP and DENV2 NGC RVPs. As expected, antibodies present in the DENV1 immune sera efficiently neutralized DENV1 RVPs (Log NT50 2.67±0.04, n = 11) (Figure 2a and b). Cross-reactive neutralizing antibodies in the sera, measured using DENV2 RVPs (Log NT50 1.32±0.05, n = 11), were significantly less potent when compared to DENV1 neutralization (23-fold; p<0.0001). A similar pattern of TS- and CR- neutralization was observed in reciprocal studies with DENV2 immune sera (Figure 2c and d), which neutralized DENV2 RVPs (Log NT50 3.06±0.03, n = 11) at significantly greater dilutions than DENV1 RVPs (Log NT50 1.80±0.04, n = 9) (20-fold, p<0.0001).
To disrupt TS epitopes recognized by DENV1 immune sera, we created a library of RVPs in which the 68 surface-accessible residues that vary between the DENV1 and DENV2 components of the vaccine (Figure 1) were used to guide construction of a panel of 54 DENV1 variants containing only one, two, or three substitutions each; in cases where adjacent residues were selected for mutagenesis they were introduced into the same RVP. All 54 DENV1 variants were infectious, albeit to differing degrees (n = 2–5, Figure 3a). The sensitivity of each variant in this panel to neutralization by DENV1 immune sera was tested, and three patterns emerged. The majority of DENV1 variants (81%) were neutralized to the same extent as WT DENV1 tested in pairwise experiments (Figure 3b and e). Eight variants in this panel (15%) revealed a small (1.5- to 2-fold) but reproducible increase in sensitivity to neutralization (Figure 3c and e). These variants were also typically more sensitive to neutralization by the DII-fusion loop reactive mAb E60 (Figure S1). Mechanisms with the potential to increase neutralization sensitivity in this context are discussed below. Finally, two mutants, E126K and E157K, exhibited a statistically significant decrease in sensitivity to neutralization by pooled DENV1 sera (3.5-fold (n = 10, p<0.0001) and 2.4-fold (n = 10, p<0.0001), respectively) (Figure 3d and e, Figure 4c and d). Neither of these two mutations conferred an altered sensitivity to neutralization by mAb E60 (Figure S1).
To investigate whether a single DENV1 variant encoding both E126K and E157K mutations had an even further reduction in neutralization sensitivity, we combined them into a single construct (Figure 4a). RVPs produced using this DENV1 E126K/E157K variant were able to infect cells, albeit with greatly reduced titer as compared to WT DENV1 RVPs (>1,000 fold, p<0.001 for all time points, n = 2–4) (Figure 4b). Neutralization studies using single and double mutants revealed the DENV1 E126K/E157K variant was more resistant to neutralization by DENV1 immune sera than either single mutant (Log NT50 1.79±0.06, n = 11) (p<0.0001 for both comparisons) (Figure 4c and d).
Our screening data with pooled immune sera suggested antibodies that bound epitopes containing residues E126 and E157 were a functionally significant component of the polyclonal antibody response to DENV1 vaccination. To validate this interpretation, we performed control experiments to address aspects of flavivirus biology with the potential to reduce the apparent sensitivity of flaviviruses to antibody-mediated neutralization.
First, we conducted experiments to demonstrate that the reduced sensitivity of the E126K/E157K variant to neutralization was not an artifact of its low titer and/or an overall disruption of the antigenic surface of the virion. An assumption of quantitative neutralization studies is that the observed neutralizing activity is dependent solely on the concentration of antibody and its affinity/avidity for viral antigens; the results of neutralization studies should be independent of the amount of antigen/virus in the experiment [79], [80]. Excess antigen in virus preparations (due to a reduced specific infectivity) may shift antibody dose-response curves towards higher concentrations of antibody if present in sufficient amounts to reduce the concentration of free antibody in solution [77]. To rule out the confounding effects of excess antigen in preparations of the E126K/E157K variant, we performed control studies using TS-neutralizing mAbs that have been mapped to the DIII lateral ridge [35]. Neutralization by these mAbs should not be affected by the DI/DII mutations. Both mAb DENV1-E103 (Figure 5a, n = 8) and mAb DENV1-E105 (Figure 5b, n = 5) neutralized WT DENV1 and DENV1 E126K/E157K RVPs to an equivalent extent (p = 0.14 and 0.57, respectively). Confirmatory studies with six additional DENV1 mAbs yielded similar results; in each instance we observed less than a two-fold difference in the EC50 of WT and the E126K/E157K variant (Figure S2, n = 2). Altogether, these results demonstrate that studies of the neutralization sensitivity of the DENV1 E126K/E157K variant are not confounded by excess antigen arising from a reduction in the specific infectivity of the virus.
We have shown previously that changes in the efficiency of virion maturation can markedly impact the sensitivity of flaviviruses to neutralization by antibodies [74]. Because mutations in the E protein have the potential to impact the efficiency of prM cleavage, we investigated the maturation state of the DENV1 E126K/E157K variant using biochemical and functional approaches. The prM content of the DENV1 E126K/E157K variant was found to be similar to WT by Western blot (Figure 5c). The DII-FL-reactive mAb E60 neutralizes flaviviruses in a maturation state-dependent fashion and serves as a sensitive functional probe for changes in virion structure that expose this otherwise cryptic epitope on the mature virion [74], [81]. Manipulating the efficiency of prM cleavage modulates the potency of E60 against DENV1 WP; increasing the efficiency of virion maturation reduces sensitivity of the virion to neutralization (Figure 5d). That no significant difference in the potency of E60 was observed when studying WT DENV1 (EC50 = 1.3×10−9±2×10−10 M, n = 8) and DENV1 E126K/E157K (EC50 = 1.9×10−9±5×10−10 M, n = 8) (p = 0.68) (Figure 5e) suggests the efficiency of virion maturation of WT and the E126K/E157K variant are similar on infectious viruses. Finally, using RVP populations in which the efficiency of prM cleavage was greatly enhanced (furin RVPs) or reduced via treatment of cells with furin inhibitor (FI RVPs) and DENV1 immune sera from eight vaccine recipients, we established that the TS-neutralizing activity in these sera was not markedly sensitive to the maturation state of the virus particle (Figure 5f). In each instance, furin- and FI-RVPs were neutralized equivalently. Altogether our data strongly suggest that the reduction in sensitivity of the DENV1 E126K/E157K variant to neutralization by TS antibodies present in vaccine immune sera is not an artifact of a change in the efficiency of virion maturation.
The accessibility of antibody epitopes on the virion may also be modulated by changes in the structure of the virion as it samples multiple conformations at equilibrium (known as viral “breathing”) [81]–[85]. The impact of viral dynamics on antibody-mediated neutralization is most readily observed with antibodies that bind epitopes predicted to be poorly exposed on the surface of the mature virion. The neutralization potency of E60 is also sensitive to viral structural dynamics [81]. That E60 neutralized both WT and the DENV1 E126K/E157K variant with equivalent efficiency (Figure 5e) indicates that the reduced sensitivity of this variant to neutralization by DENV1 immune sera is not a result of a change in the extent of viral “breathing”.
Finally, studies with pooled DENV2 immune sera revealed these mutations only altered sensitivity to TS antibodies and not CR antibodies. The sensitivity of DENV1 E126K/E157K RVPs to neutralization by cross-reactive DENV2 immune sera was similar to WT DENV1 RVPs (1.3-fold, n = 8, p = 0.08) (Figure 5g). Together, these experiments support the conclusion that mutation at residues E126K and E157K specifically reduces neutralization by TS antibodies elicited by DENV1 vaccination. Furthermore, the reduced sensitivity of variant E126K/E157K to immune sera from DENV1, but not DENV2, vaccinated individuals was not an artifact of changes in the structural dynamics of the virion, the efficiency of virion maturation, gross changes in the antigenic structure of the virion, or antigen excess in the neutralization assay.
To determine the prevalence of neutralizing antibodies that target epitopes incorporating residues E126 and E157, we next performed longitudinal and cross-sectional studies of immune sera obtained from individual recipients of a DENV1 live-attenuated vaccine candidate [86]. Participants in this vaccine study were administered vaccine on days 0 and 180, as described previously [86]. Neutralization studies with sera from three subjects collected at days 28, 42, 180, 208, and 222 post-vaccination were performed using WT DENV1, WT DENV2, and DENV1 E126K/E157K RVPs. These three subjects were selected for study because they represented both high and low neutralizing antibody responses among participants in the phase I study. Dose response curves and summary NT50 values of sera from longitudinally tested subjects are shown in Figure 6. As anticipated, sera from post-vaccination samples neutralized DENV1 more efficiently than DENV2. In each instance, neutralization activity declined over time and was not boosted significantly by a second dose of vaccine, consistent with previous reports (Figures 6 and 7) [86].
Longitudinal analysis revealed that the sensitivity of the E126K/E157K variant to neutralization by DENV1-immune sera varied with time and among the three subjects tested. For Subject 43, the contribution of antibodies sensitive to mutations at positions E126 and E157 increased with time post-vaccination. By day 222, DENV2 and DENV1 E126K/E157K RVPs were neutralized to an equivalent degree (Figure 6a and b). The reduced sensitivity of DENV1 E126K/E157K RVPs to neutralization was observed at earlier time points in the sera of Subject 57, suggesting the TS-antibody response in this individual was focused early on epitopes impacted by mutating these two residues (Figure 6b). In contrast, while the sensitivity of DENV1 E126K/E157K RVPs to neutralization was reduced in comparison to WT at all time points sampled from Subject 39 (from 2–6 fold), mutation of these two residues alone did not ablate the entire TS response (Figure 6b).
To better understand the contribution of epitopes defined by the E126K and E157K mutations in the TS response to the DENV1Δ30 vaccine in a larger sample set, we assayed sera collected from an additional 21 DENV1 vaccine recipients at two time points post-vaccination (day 42 and 222; day 222 was collected 42 days after the second vaccine dose was administered). Antibody dose-response curves were generated for each sample using WT DENV1, WT DENV2, and DENV1 E126K/E157K RVPs as described above. The NT50 values calculated from this cross-sectional study are presented in Figure 7a and b. As expected, vaccine sera most efficiently neutralized DENV1; DENV1 E126K/E157K and DENV2 RVPs were significantly less sensitive to neutralization at both time points studied (p<0.0001 for both comparisons) (Figure 7a). While comparisons of the neutralization titer against DENV1 E126K/E157K and DENV2 revealed differences at day 42 post-vaccination (p<0.0001), by day 222 no significant difference was observed (p = 0.15).
On average, we observed a 3.5-fold decrease in neutralization titers measured with DENV1 RVPs between day 42 and day 222 post-vaccination. This decay was more rapid than the change in DENV2-reactive titers measured at these two time points (mean 2.4-fold decrease) (p<0.01). Of interest, the ability of DENV1 sera to neutralize the E126K/E157K variant reactive antibody declined most dramatically between days 42 and 222 (a mean 4.8-fold decrease). This decline in sensitivity to neutralization was more rapid that that observed with either DENV1 (p<0.05) or DENV2 (p<0.0001) RVPs. This data is in agreement with our interpretation that over time, antibodies in DENV1-immune sera become more focused on epitopes containing residues E126 and E157. On average, RVPs containing mutations at these positions become more difficult to neutralize with time post-vaccination.
Analysis of individual neutralization titers revealed the same pattern (Figure 7b). On day 42, the serum of roughly half the recipients neutralized DENV1 E126K/E157K RVPs less efficiently than WT DENV1 RVPs. The contribution of antibodies binding epitopes impacted by mutation of E126 and E157 was monitored as the difference in neutralization sensitivity of the DENV1 E126K/E157K mutant and DENV2 RVPs (highlighted in red, yellow, and green to indicate 0–3, 3–6, and >6 fold differences in NT50). In 46% of the day 42 samples tested, the neutralization titer of the E126K/E157K variant was found to be within 3-fold of the cross-reactive DENV2 neutralization titer (Figure 7b). By day 222 post-vaccination, 77% of the NT50 values obtained with the E126K/E157K variant were within 3-fold of the DENV2 neutralization titer, confirming the trend observed in the longitudinal analysis described above. Together, these data identify a significant functional contribution of epitopes containing residues E126 and E157 in the TS DENV1 response that increases with time.
Interpreting the serology of flaviviruses is hindered by extensive cross-reactivity and significant differences in the functional properties of antibodies that bind different epitopes on the virion [23], [35], [36], [39]–[42]. Several approaches have been employed to deconstruct the complexity of the polyclonal antibody response to DENV. Biochemical studies using recombinant proteins or subviral particles have identified mutations in the E protein that reduce recognition by antibodies in DENV immune sera in an ELISA or Western blot assay format [23], [87], [88]. A limitation of this type of approach is that it does not account for the marked difference in neutralization potential among antibodies of varying specificity. A neutralization response driven by the contribution of a low concentration of potent neutralizing antibodies may be difficult to detect using biochemical studies.
The antibody repertoire elicited by DENV infection or vaccination has also been investigated using approaches that allow study of the functional properties of antibodies in immune sera. Exciting advances have been made towards understanding the DENV antibody repertoire through the analysis of the specificity of human mAbs [39], [49], [51]. A strength of this approach is that it enables functional analysis of antibody specificities that make up the polyclonal immune response. While this approach reveals the specificity of DENV-reactive memory B-lymphocytes, it is unknown how faithfully these methods capture the breadth of the polyclonal humoral immune response in vivo, and how screening bias impacts the mAbs selected for study. A recent study of 26 DENV-reactive human mAbs obtained from subjects immunized with the rDEN1Δ30 vaccine candidate studied herein identified only three mAbs that neutralize DENV [50]. The functional specificity of the polyclonal DENV immune response has also been investigated by depletion of immune sera using recombinant proteins. These studies suggested antibodies that bind E-DIII are not a significant component of the neutralizing antibody response and play only a modest role in protection in the AG129 mouse model of DENV infection/disease [31], [89]. This finding was confirmed using infectious clones encoding mutations in E-DIII [90].
The number of epitopes that contribute to the neutralizing activity of the polyclonal antibody response to DENV remains unknown. To date, insight into where neutralizing antibodies present in human sera bind DENV has been limited principally to negative data. The goal of the present study was to identify the epitopes recognized by neutralizing antibodies of the type-specific DENV1 response following vaccination. We undertook a mutagenesis approach to identify amino acid substitutions that conferred a reduction in sensitivity to neutralization by DENV1 immune sera, but not cross-reactive antibodies present in DENV2 sera. We created a library of DENV1 RVPs encoding mutations at residues predicted to be surface-accessible on the mature virion that differ between DENV1 and DENV2. We identified single amino acids in E-DI (E126) and E-DII (E157) that when mutated together abrogated a large portion of the TS-neutralizing antibody response in 77% of vaccinated subjects.
That two amino acid substitutions were sufficient to significantly reduce the TS-antibody response in the majority of vaccinees was a surprise. These data suggest the TS-antibody response is remarkably focused. Similar conclusions were reached in studies of the complexity of the polyclonal antibody response to HIV-1 and influenza [91]–[94]. Residues E126 and E157 are very conserved among DENV1 viruses; the mutations identified in our study were found only once (E126K) or not at all (E157K) among 1,398 DENV1 sequences available for in silico analysis [95]. Thus, the ability of DENV1 to escape from neutralization by mutation may be limited by the functional pressure of cross-reactive antibody and a substantial fitness cost. A more detailed analysis of the functional consequences of mutations at these residues is underway.
While the TS-immune response of a majority of volunteers in our study was focused significantly on epitopes affected by mutations at E126 and E157, these changes had a somewhat reduced impact on the potency of immune sera from five volunteers. This suggests that additional residues are involved in the fine specificity of the response and will require further study. Limited secondary screening with a subset of the panel of DENV1 variants identified residue 203 as a significant contributor to TS-neutralization patterns of Subject 38 (Figure S3), but not the others. As this residue is located within 13 Å from residue 126, it is possible that mutations at this position impact recognition of the same or overlapping epitopes in E-DII.
While, to our knowledge, residues 126 and 157 on DENV1 E protein have not yet been identified in neutralization escape studies with DENV mAbs, recent studies have identified a complex epitope in proximity to E157. Studies of the human anti-WNV mAb CR4354 identified an epitope at the junction of E-DI and E-DII that exists only on the intact virion, and not on soluble E protein [96]. Four human mAbs have since been characterized that bind in the same region. The footprint of the TS DENV1 mAb DENV HM14c10 has been solved using cryo-electron microscopy reconstruction and has confirmed the discontinuous structure of the epitope and its similarity to CR4354 [47], [52]. Neutralization escape mutants have been identified for HM14c10 as well as three additional human mAbs thought to recognize quaternary epitopes on DENV; each escape variant was mutated around the DI–II hinge region and within the CR4354 footprint [47], [48]. The E157K mutation identified in our study is close to the hinge region. However, it remains unclear if antibodies that bind DENV1 at this location engage a quaternary epitope similar to CR4354, as this residue falls just outside the DI–DII hinge epitopes defined by the footprints of WNV CR4354 or DENV HM14c10. Furthermore, the 12 surface-exposed residues that differ between DENV1 and DENV2 present in the published DI–DII hinge epitope footprints did not themselves markedly reduce sensitivity to neutralization by DENV1 sera in this screen.
The dense arrangement of E proteins on the virion complicates our understanding of antibody recognition. Many well-characterized epitopes are not predicted to be accessible for antibody binding using existing static models of the mature DENV structure [23], [76], [97]. For example, antibodies that bind the E-DII fusion loop are a significant component of the CR repertoire produced by mouse and human following flavivirus infection, yet bind an epitope not predicted to be accessible for antibody recognition [23]. The accessibility of cryptic epitopes like the DII-FL has been shown to be governed by several factors [74], [81]. Viruses exist as an ensemble of structures at equilibrium (reviewed by [84], [98]). “Breathing” of envelope proteins incorporated into the flavivirus virion has been shown to alter epitope accessibility and sensitivity to neutralization by monoclonal and polyclonal antibody [81], [82], [99]. In addition, many antibodies are sensitive to the maturation state of the virus particle [74]. Engineered mutations or naturally occurring variation have the potential to alter the heterogeneity or dynamics of the virion, and therefore may impact epitope exposure. While we have identified amino acids that contribute to the type-specific recognition of the DENV1 Western Pacific strain, epitopes containing these amino acids may play a reduced or enhanced role in the context of other viral E protein sequences. Of interest, ∼30% of the vaccine recipients studied within neutralize a related DENV1 strain 16007 more efficiently than the WP strain used in the DENV1 vaccine candidate (data not shown). This enhanced recognition may reflect differences in the “breathing” of these two viruses that alter the accessibility of epitopes on the virus particle, as suggested in a recent study [100]. As our analysis was focused on surface accessible differences between the DENV1 and DENV2 components of the vaccine, it is possible that amino acid variation among DENV1 strains and differences in the extent of “viral breathing” (and therefore epitope exposure) will impact the pattern of type-specific recognition of different DENV1 strains. It remains to be determined whether the two resides characterized within will be useful as a generalizable signature of a type-specific humoral response for all DENV1 viruses. Evaluating the contribution of this structural complexity towards neutralization sensitivity is a critical component of understanding the antigenic surface of flaviviruses and how this varies among strains within and between serotypes.
The identification of major targets of the TS-neutralizing antibody response to DENV1 vaccination represents an important step toward a more complete understanding of the humoral immune response to DENV. Whether epitopes including residues E126 and E157 contribute significantly to the neutralizing antibody response to natural infection or vaccination with all four DENV serotypes remains to be determined. The four serotypes of DENV share a 63% or greater amino acid identity and presumably the same overall virion structure. The epitopes engaged by TS-neutralizing antibodies elicited by infection of other DENV serotypes are unknown. Of note, the introduction of the reciprocal mutations at positions K126 and K157 of DENV2 NGC did not markedly shift the neutralization curve of DENV2 immune sera obtained from vaccine recipients (Figure S4). These data suggest that antibodies that contribute to the TS-neutralizing antibody response may bind distinct epitopes on all four DENV serotypes. Beyond the identification of significant epitopes in DENV, the data and experimental approaches described within provide insight into mutagenesis approaches for deconstructing the functional components of the polyclonal DENV antibody response and will be used to guide future studies on the functionally important epitopes involved in the neutralization of the other three DENV serotypes.
All cell lines were maintained at 37°C in the presence of 7% CO2. HEK-293T cells were passaged in complete Dulbecco's modified Eagle medium (DMEM) containing Glutamax (Invitrogen, Carlsbad, CA), supplemented with 7.5% fetal bovine serum (FBS) (HyClone, Logan, UT) and 100 U/ml penicillin-streptomycin (PS) (Invitrogen, Carlsbad, CA). Raji-DCSIGNR cells were passaged in RPMI-1640 medium containing Glutamax (Invitrogen, Carlsbad, CA), supplemented with 7.5% FBS and 100 U/ml PS.
Sera from recipients of phase I studies of candidate monovalent DENV1 or DENV2 vaccines were obtained for study. Initial screening studies were performed using sera pooled from two or three recipients of the DENV1 and DENV2 vaccines, respectively, collected 2–3 years post-vaccination. Neutralizing antibody responses from 24 participants of a DENV1 vaccine study were studied individually [86]; sera were collected for study on the indicated day post-vaccination.
Clinical studies were conducted at the Center for Immunization at the Johns Hopkins Bloomberg School of Public Health under an investigational new drug application reviewed by the United States Food and Drug Administration. The clinical protocol and consent form were reviewed and approved by the NIAID Regulatory Compliance and Human Subjects Protection Branch, the NIAID Data Safety Monitoring Board, the Western Institutional Review Board, and the Johns Hopkins University Institutional Biosafety Committee (ClinicalTrials.gov identifiers; NCT00473135, NCT00920517). Written informed consent was obtained from each participant in accordance with the Code of Federal Regulations (21 CFR 50) and International Conference on Harmonisation guidelines for Good Clinical Practice (ICH E6).
Plasmids encoding a WNV sub-genomic replicon and DENV1 WP CprME structural genes have been described previously [77], [81], [101]. An expression construct of the CprME gene of the DENV2 NGC strain was constructed using similar methods and will be described elsewhere (VanBlargan and Pierson, unpublished data). Plasmids encoding structural gene variants with up to three amino acid substitutions were produced by site-directed mutagenesis using the Quikchange Mutagenesis kit (Stratagene, La Jolla, CA) according to the manufacturer's instructions. All plasmids used in this study were propagated in Stbl2 bacteria grown at 30°C (Invitrogen, Carlsbad, CA).
To identify amino acids recognized by TS-neutralizing antibodies in vaccine sera, we targeted residues for mutagenesis that differed between the DENV1 and DENV2 strains used in the NIAID tetravalent vaccine candidate. The envelope proteins of the DENV1 WP strain and the DENV2 NGC strain differ by 158 amino acids. As a metric to narrow our mutagenesis efforts, we focused on residues predicted to be exposed on the surface of the virion. Surface accessibility was estimated using solvent accessible surface areas of the residues determined from the crystal structure of the E protein dimer (PDB ID: 10AN), with a cut-off value of 30 Å2 (UCSF Chimera package) [102], [103]. Residues were then further restricted by modeling the dimer onto the mature virion (PDB ID: 1THD) [71]. Residues exposed on the surface of the virion were then selected, narrowing the list of candidates to 68 [71]. Admittedly, because our selection scheme was based on a static model of the mature virion, this minimalist approach to selecting a core panel of residues had the potential to be complicated by the structural heterogeneity and dynamics of the virus particle. Virion structure is influenced by the maturation state and structural dynamics of the virion [15], [74], [81]. Both have the potential to increase the number of residues that may contribute to TS-antibody recognition. Additionally, structural studies of DENV2 at 37°C revealed the virus not only becomes considerably more heterogeneous at this temperature but also appears to adopt a distinct structure(s) at this temperature [104], [105].
DENV RVPs were produced by complementation of a WNV replicon with plasmids encoding the structural genes of DENV as described previously [81], [101]. Briefly, pre-plated HEK-293T cells were transfected with plasmids encoding the WNIIrepG/Z replicon and DENV CprME in a 1∶3 ratio by mass, using Lipofectamine LTX (Invitrogen, Carlsbad, CA) in accordance with the manufacturer's instructions. Four hours post-transfection, culture media were replaced with a low-glucose formulation of DMEM containing 25 mM HEPES (Invitrogen, Carlsbad, CA), 7% FBS, and 100 U/ml PS, and incubated at 30°C. RVP-containing supernatant was collected at multiple time points starting at 72 hours post-transfection, filtered using a 0.22 µm syringe filter, and stored at −80°C. The low-glucose DMEM was replenished after each harvest.
To generate more homogenous populations of RVPs with low to undetectable uncleaved prM, complementation experiments were modified to include a plasmid expressing human furin (furin DENV) in a 1∶3∶1 ratio of replicon, CprME, and furin by mass. In order to generate RVP populations that retained significant levels of uncleaved prM (FI-DENV), medium from transfected cells was exchanged with medium supplemented with 50 µM furin inhibitor (FI) Dec-RVKR-CMK at four hours post-transfection (Enzo Life Sciences, Farmingdale, NY).
The efficiency of prM cleavage in RVP preparations was determined by Western blotting as previously described [74], [106]. Briefly, DENV RVPs were partially purified over a 20% sucrose cushion by ultra-centrifugation. Pelleted RVPs were lysed with in buffer containing 1% Trition, 100 mM Tris, 2 M NaCl, and 100 mM EDTA. The protein content of lysates was analyzed using E- and prM-reactive mAbs (4G2 and GTX128093 (Genetex), respectively) at 1 µg/ml. The efficiency of prM cleavage was evaluated on blots normalized by loading equivalent E protein.
The infectious titer of each RVP stock was determined using Raji cells that express the flavivirus attachment factor DCSIGNR as described previously [78], [101], [107]. Cells were infected with serial two-fold dilutions of transfection supernatant, incubated at 37°C for two days, and then scored for GFP expression by flow cytometry. Only data from the linear portion of the virus dose-infectivity curves were used to compare RVP titers. Infectious titer was calculated using the formula: IU/sample volume = (percent GFP positive cells)×(number of cells)×(dilution factor).
DENV RVP stocks were diluted and incubated with serial dilutions of mAb or serum for one hour at room temperature prior to the addition of Raji-DCSIGNR cells, unless specified otherwise. Infections were carried out at 37°C for 48 hours and infectivity was scored as the fraction of GFP-expressing cells determined using flow cytometry. Antibody-dose response curves were analyzed using non-linear regression analysis (with a variable slope) (Graphpad Software, La Jolla, CA). Data are expressed as the concentration of antibody (EC50) or serum dilution (NT50) required to reduce infection by half.
Statistical analyses were performed using Prism software (GraphPad). Log EC50 or Log NT50 values were compared using Student's t-test when comparing two samples. For comparisons of more than two samples, Log NT50 values were compared by one-way ANOVA followed by Tukey's multiple comparisons test or, where indicated, Šidák correction for multiple comparisons.
|
10.1371/journal.pntd.0003404 | Genome and Phylogenetic Analyses of Trypanosoma evansi Reveal Extensive Similarity to T. brucei and Multiple Independent Origins for Dyskinetoplasty | Two key biological features distinguish Trypanosoma evansi from the T. brucei group: independence from the tsetse fly as obligatory vector, and independence from the need for functional mitochondrial DNA (kinetoplast or kDNA). In an effort to better understand the molecular causes and consequences of these differences, we sequenced the genome of an akinetoplastic T. evansi strain from China and compared it to the T. b. brucei reference strain. The annotated T. evansi genome shows extensive similarity to the reference, with 94.9% of the predicted T. b. brucei coding sequences (CDS) having an ortholog in T. evansi, and 94.6% of the non-repetitive orthologs having a nucleotide identity of 95% or greater. Interestingly, several procyclin-associated genes (PAGs) were disrupted or not found in this T. evansi strain, suggesting a selective loss of function in the absence of the insect life-cycle stage. Surprisingly, orthologous sequences were found in T. evansi for all 978 nuclear CDS predicted to represent the mitochondrial proteome in T. brucei, although a small number of these may have lost functionality. Consistent with previous results, the F1FO-ATP synthase γ subunit was found to have an A281 deletion, which is involved in generation of a mitochondrial membrane potential in the absence of kDNA. Candidates for CDS that are absent from the reference genome were identified in supplementary de novo assemblies of T. evansi reads. Phylogenetic analyses show that the sequenced strain belongs to a dominant group of clonal T. evansi strains with worldwide distribution that also includes isolates classified as T. equiperdum. At least three other types of T. evansi or T. equiperdum have emerged independently. Overall, the elucidation of the T. evansi genome sequence reveals extensive similarity of T. brucei and supports the contention that T. evansi should be classified as a subspecies of T. brucei.
| The single-cell parasite Trypanosoma evansi is the disease-causing trypanosome with the widest geographical distribution. The disease, called surra, has significant economic impact primarily due to infections of cattle, horses, and camels. Morphologically the parasite is indistinguishable from bloodstream stage T. brucei, a parasite causing sleeping sickness in humans and the disease nagana in animals. T. brucei, however, is strictly bound to sub-Saharan Africa where its obligate vector, the tsetse fly, resides. The lack of a complete mitochondrial genome in T. evansi further distinguishes this parasite from T. brucei. Important questions regarding the biology of T. evansi include how it escaped from Africa, whether this has happened more than once, and how exactly it is related to T. brucei. To help answer these questions we have sequenced the T. evansi nuclear genome. Our phylogenetic analysis demonstrates that T. evansi, and the closely related horse parasite T. equiperdum, evolved more than once from T. brucei. We also demonstrate extensive similarity to T. brucei, including the maintenance of numerous genes that T. evansi no longer requires. Therefore, despite the significant functional and pathological differences between T. evansi and T. brucei, our analysis supports the notion that T. evansi is not an independent species.
| Trypanosomatid parasites Trypanosoma evansi and T. equiperdum are responsible for animal diseases with extensive pathological and economic impact and closely related to the T. brucei group [1], [2]. The latter includes three subspecies: the human parasite T. b. rhodesiense, the zoonotic parasite T. b. gambiense, and the animal parasite T. b. brucei. Together T. brucei, T. evansi, and T. equiperdum comprise the subgenus Trypanozoon. The exact nature of the phylogenetic relationship between these three species has been the subject of ongoing debate, with some evidence suggesting that T. evansi and T. equiperdum are monophyletic and other evidence suggesting that they are polyphyletic and have emerged multiple times from T. b. brucei [3]–[6]. Trypanosomatids are a family within the protist group Kinetoplastida, the eponymous feature of which is a large and complex network of circular DNAs (kinetoplast or kDNA) inside their single mitochondrion. Two key biological features distinguish T. evansi and T. equiperdum from the T. brucei group. Firstly, their transmission is independent from the tsetse fly as obligatory vector. T. evansi is predominantly transmitted by biting flies and causes surra in a wide variety of mammalian species (the name of the disease varies with geographical area), while T. equiperdum causes a sexually transmitted disease called dourine in horses [1], [3], [7]. The altered mode of transmission has enabled both parasites to escape from the sub-Saharan tsetse belt and become the pathogenic trypanosomes with the widest geographical distribution. Secondly, all strains of T. evansi and T. equiperdum investigated so far are dyskinetoplastic, i.e., lacking all (akinetoplastic) or critical parts of their kDNA [8]. The loss of kDNA is thought to lock T. evansi and T. equiperdum in the bloodstream life cycle stage, presumably because the absence of kDNA-encoded components of the oxidative phosphorylation system prevents ATP generation in the tsetse midgut [9]. Nonetheless, whether dyskinetoplasty preceded the switch to tsetse-independent transmission or vice versa is unresolved [8], [10], [11].
The kDNA that comprises the mitochondrial genome in T. brucei consists of numerous concatenated circles of two kinds: maxicircles that encode genes primarily involved in oxidative phosphorylation and minicircles that encode guide RNAs (gRNAs) [12]. The majority of maxicircle mRNAs undergo RNA editing to insert or delete uridine residues as specified by template gRNAs in a process catalyzed by multiprotein complexes called editosomes [13]–[15]. One kDNA-encoded transcript that requires editing is the F1FO-ATPase subunit 6, which is essential in both bloodstream and insect stage T. brucei [16]–[19]. However, it has recently been shown that mutations found in the nuclear-encoded ATPase subunit γ of some T. evansi and T. equiperdum strains can compensate for the loss of kDNA, explaining their viability [20].
In an effort to better understand the causes and consequences of tsetse-independent transmission and kDNA-independent viability, we sequenced the genome of T. evansi strain STIB805. This strain was isolated in 1985 from an infected water buffalo in the Jiangsu province of China, shown to completely lack kDNA (i.e. to be akinetoplastic), and suggested to belong to a possibly clonal group of T. evansi with worldwide distribution [4], [21], which is why it was chosen for this study. The comparative genome analysis between this strain and the T. b. brucei TREU 927/4 strain reference genome [22] revealed extensive similarities. While the sizes of the chromosomes differ between T. evansi and T. brucei, the gene content within their respective genomes are largely similar, as 92.7% of T. evansi CDS have an identifiable ortholog in T. brucei. Analysis of T. evansi variant surface glycoprotein (VSG) sequences shows broad conservation of N-terminal sub-types, with extensive phylogenetic similarity and no evidence of any species-specific expansion of clades. An analysis of T. evansi CDS corresponding to the identified T. brucei mitochondrial proteome revealed that virtually all are retained, despite the lack of requirement in an akinetoplastic trypanosome for respiratory complexes I-IV or any proteins involved in maintenance or expression of the mitochondrial genome. Phylogenetic analyses with several genetic markers conclusively show that extant strains of T. evansi or T. equiperdum are not monophyletic and evolved on at least four independent occasions. Together, the results presented here show few critical differences between T. evansi and T brucei, indicating that dyskinetoplasty and concomitant tsetse-independent transmission are significant phenotypic changes underpinned by relatively subtle genomic alterations.
The rearing of animals was regulated by Czech legislation (Act No 246/1992 Coll.). All housing, feeding and experimental procedures were conducted under protocol 90/2013 approved by Biology Centre, Czech Academy of Sciences and Central Commission for Animal Welfare of the Czech Republic.
Trypanosomes (T. evansi strain STIB805) were purified from mice by DEAE (DE52) cellulose [23]. Total DNA was extracted as described elsewhere [24]. Briefly, the cells were lysed using SDS, and incubated with proteinase K and RNase. DNA was harvested after phenol extraction and ethanol precipitation. Four runs of single-end 454 sequencing plus 2 runs of paired-end 454 sequencing were obtained using GS FLX(+) System following the manufacturer's instruction (Roche) and generated 1,904,327 reads (225,826 paired end, 1,678,501 single end) [25]. Approximately 10 µg of genomic DNA was sheared by nebulization into desired fragments sizes (∼400 bp for single-end 454, ∼3 kb for paired-end) and adaptor oligos ligated to create the library for sequencing. Additional sequence data was obtained by shearing genomic DNA to ∼200–300 bp fragments sizes for sequencing on an Illumina GAIIx producing 19,701,740 tags with an ordered read length of 76mers. Illumina reads for T. b. brucei strains TREU 927/4 and Lister 427 (provided by the Wellcome Trust Sanger Institute, Hinxton, UK) were downloaded from the European Nucleotide Archive (accession nos. ERX009953 and ERX008998).
Genome assemblies and identification of sequence polymorphisms (SNPs and indels) were carried out with CLC Genomics Workbench (CLC bio). Reads for T. evansi STIB805, T. b. brucei TREU 927/4 and T. b. brucei Lister 427 were mapped against the T. brucei TREU 927/4 version 4 reference (Tb927) using the following mapping parameters: global alignment, similarity fraction = 0.9, length fraction = 0.5, insertion cost = 3, deletion cost = 3, mismatch cost = 2. De novo assemblies for each strain were created using either all reads or those binned during the reference-based assemblies, using the following parameters: automatic word size = yes, bubble size = 50, similarity fraction = 0.9, length fraction = 0.5, deletion cost = 3, insertion cost = 3, mismatch cost = 2. De novo contigs were aligned to Tb927 using standalone NCBI BLAST version 2.2.25 and Artemis Comparison Tool release 12.0 [26]. For RPKM (reads per kilo base per million) analysis, aligned STIB805 and TREU 927/4 reads were assigned to sequential 1-kb bins along the length of the Tb927 reference. For each chromosome, the log2 ratio of binned reads from the two read-sets was calculated for each bin, and normalized to a median log2 ratio of 0 by offsetting all values by the median log2 ratio. SNPs and indels were called with CLC bio using the following parameters: minimum coverage = 10, maximum coverage = 100, minimum variant frequency = 30%, minimum central quality = 20, minimum average quality = 15.
Coding sequence prediction of T. evansi genome was done using a combination of de novo gene prediction approach and reference based gene transfers. The Rapid Annotation Transfer Tool (RATT) was used to transfer the gene boundaries and functional annotations from Tb927 onto T. evansi chromosomes (target hereafter) [27]. To prevent the transfer of paralogs to the same target region, thus resulting in multiple overlapping/duplicate gene calls, the annotation transfer was performed pairwise between one chromosome from Tb927 and the corresponding chromosome from the target. This resulted in the transfer of more than 96.4% (9427/9776) of genes from Tb927_v4 onto TevSTIB805ra. Though we observed extensive conservation of synteny between Tb927 and target genomes, this approach would likely have missed genes in targets that were shuffled by chromosomal fission and fusion that occurred during evolution. We performed another RATT transfer considering all Tb927 chromosomes and one target chromosome at a time. We then parsed out genes that were predicted uniquely by this approach (i.e. from regions where the initial RATT transfer failed to predict any genes). Combining these two approaches allowed us to predict genes that were shuffled across the chromosomes while avoiding predictions that overlap each other. We then used an in-house consensual de novo gene prediction suite called 'AutoMagi' to predict protein coding genes from T. evansi [28]. AutoMagi internally uses three gene prediction algorithms (genescan, testcode, codonusage) and predicts a consensus gene model out of individual gene predictions. The codon usage table required by AutoMagi was generated by ‘cusp’ using the RATT output of T. evansi STIB805 assembly [29]. We then used an in-house built prolog system to combine the de novo gene predictions and reference-based gene predictions. The genes that are unique to RATT predictions were automatically included in the final set. Genes that are unique to AutoMagi were compared with the NCBI non-redundant (NR) database (accessed 23 February 2011) using 'BLASTp' (default criteria: EXPECT = −10, WORD SIZE = −3, MATRIX_NAME = BLOSUM62, GAP COST = Existence:11 Extension:1) algorithm. All of the 'BLASTp' results were reviewed manually, and genes meeting the following two criteria were retained: (a) an E-value of 5e-6 or lower to the matched NR sequence; (b) coding strand identical to nearest neighbors on either side (i.e. on same poly-cistronic unit). This manual curation removed 165 putative CDS from making into the finalized TevSTIB8805ra.
The gene calls that overlapped between RATT and AutoMagi were subdivided into the following 4 groups. 1) Identical: overlapping exactly, 2) AM_subsetof_RATT: AutoMagi prediction is entirely contained within RATT prediction, 3) RATT_subsetof_AM: RATT prediction is entirely contained within AutoMagi prediction 4) StaggeredOverlap: both predictions overlap in a staggered fashion. In the first two cases (Identical & AM_subsetof_RATT), AutoMagi gene calls were ignored and RATT models were retained. In the third case (RATT_subsetof_AM) coordinates from AutoMagi were combined with annotation information from the RATT model. Genes in the fourth category (Staggered) were subjected to a thorough manual review process. The review process involved (i) 'BLASTp' search against NCBI's NR database to identify coordinates for queries and subject; (ii) a ClustalW multiple sequence alignment of both candidate genes with their potential homolog(s) from T. brucei. Manual review of BLAST and ClustalW was performed in each case to decide either to split/merge/choose one of the AutoMagi/RATT predictions. These newly derived coordinates were then combined with the annotation information from the RATT model. GeneIDs of the final set of protein coding genes were unified and ordered from left to right end of the chromosome.
Entire genome and all the predicted genes are publicly available at TriTrypDB (http://tritrypdb.org/tritrypdb/). The fastq files containing the Illumina read data for T. evansi STIB805 are available at: http://www.ebi.ac.uk/ena/data/view/ERA000101.
Bloodstream parasites at 2×108 cells/ml were purified using DEAE cellulose (DE52) chromatography, and subsequently used to prepare chromosome blocks as previous described [30]. DNA from T. b. evansi and T. b. brucei cells was embedded in low-melt agarose blocks (final concentration of 5×107 cells/ml) according to [31], and was resolved using a 1% Megabase agarose (Bio-Rad) gel with 0.5X TBE buffer in the CHEF-DRIII system (Bio-Rad). S. cerevisiae DNA was used as a size marker. Pulse field gel electrophoresis (PFGE) was run at 14°C under the following conditions: switch time A increased from 28.6 s to 228 s for 24 hrs, followed by switch time B, with increase from 28.6 s to 1,000 s for another 24 hrs. The angle was set to 120° and voltage gradient to 3 V/cm. The PFGE gel was stained with ethidium bromide after the run.
VSG sequences were extracted from the genome assembly using hidden markov models (HMM) and HMMER 3.0 [32]. HMMs were constructed for a- and b-type VSG respectively using multiple sequence alignments of T. brucei TREU 927/4 and T. b. gambiense DAL972 sequences [33]. All open reading frames>100 bp were marked up and the predicted amino acid sequences were searched for matches to either HMM using HMMER 3.0. Significant matches were checked manually to ensure that each VSG was complete and gene boundaries were correct. Intact VSG were extracted and aligned approximately using Clustal X [34]. The aligned sequences were combined with existing alignments of T. brucei TREU 927/4 a- and b-type VSG [35] and modified by eye.
C-terminal domains were trimmed (due to recombination these present an inconsistent phylogenetic signal), resulting in a- and b-VSG alignments of 470 and 492 characters, respectively. Neighbour-joining trees were estimated for each amino acid sequence alignment using PHYLIP v3.6, with a JTT rate matrix and 100 non-parametric bootstrap replicates. A frequency distribution of species-specific clade sizes, (i.e. three T. evansi VSG clustered together with a T. brucei TREU 927/4 sequence as the sister lineage has a clade size of 3) was calculated to express the degree of intercalation of sequences from the two strains.
Putative VSG orthologs were extracted from the phylogenies. In situations where single VSG from T. evansi STIB805 and T. brucei TREU 927/4 were sister taxa, supported by a bootstrap value>95, these genes were interpreted as orthologs. The rates of synonymous and non-synonymous nucleotide substitutions per site were estimated for these orthologous pairs. The ratio of these rates (ω) was estimated using Ka/Ks Calculator v1.2 [36] using GY and MS methods. For comparison, this was repeated for 151 pairs of non-VSG orthologs chosen at random.
The dihydrolipoamide dehydrogenase (LipDH) CDS (Tb927.11.16730) was PCR-amplified from total parasite DNA using primers 5′-ATA AAG CTT ATG TTC CGT CGC TGC-3′ (forward) and 5′-ATA AGA TCT TTA GAA GTT GAT TGT TTT GG-3′ (reverse) and Phusion polymerase (New England Biolabs, NEB). In cases where direct sequencing of the amplicon revealed heterozygosity, sequence information for individual alleles was obtained after cloning. After removal of the primer sequences, LipDH CDSs were aligned using ClustalX [37]. Phylogenies were reconstructed using the program MrBayes [38] via the Topali platform [39], implementing the GTR substitution model and a discrete gamma rate distribution model with four rate categories (to account for rate heterogeneity among sites) as the most appropriate nucleotide substitution model. Four independent MCMC chains of 1×106 generations were sampled every 100th generation. A 50% majority rule consensus tree was derived after the first 25% of trees were discarded as burn-in. The ATP synthase γ subunit sequence (Tb927.10.180) was PCR-amplified from total parasite DNA with primers 5′-GCG GAA TTC GAA GCA GAT GAC ACC TAA-3′ (forward) and 5′-GCG GAA GAC CTT GCT GCG GAG CCA CTC T-3′ or 5′-GGC GAC ATT CAA CTT CAT-3′ (reverse) and Phusion polymerase (NEB). The sequence was determined by direct amplicon sequencing or, in cases of heterozygosity, after cloning. A partial 812-bp sequence of cytochrome oxidase 1 (COX1) was obtained by PCR and used for phylogenetic analysis as previously described [40]. Briefly, we assessed phylogenetic relationships among T. equiperdum and T. brucei isolates using a haplotype network constructed using the statistical parsimony approach implemented in TCS v. 1.21 [41]. Subnetworks were created with 95% confidence limit and then unconnected subnetworks>10 mutations apart were connected by relaxing the confidence limit. To verify that the haplotypes containing T. equiperdum isolates were on phylogenetically distinct branches, we estimated a phylogenetic tree using the Bayesian approach implemented in MrBayes v. 3.2 [42]. PartitionFinder v. 1.0.1 [43] determined the Hasegawa, Kishino and Yano nucleotide substitution [44] with invariant sites (HKY+I) without partitioning by codon positions as the most appropriate model for the MrBayes analysis. Microsatellite genotyping was carried out exactly as described previously [40]. Briefly, isolates were typed for eight microsatellite markers [45]. Principal component analysis (PCA) was performed in R using the package adegenet, as described [40].
PFGE was performed to visualize the pattern of chromosomes in the akinetoplastic T. evansi STIB805 strain. Multiple bands corresponding to megabase chromosomes ranging in size from ∼1 to ∼5 Mb could be visualized (Fig. 1). The most noticeable differences in the chromosomal pattern in comparison to T. brucei are the five intermediate chromosome bands that range from ∼300 kb to ∼800 kb. Additionally, two size groups of minichromosomes (∼100 kb and ∼200 kb) were observed in T. evansi STIB805. Such a degree of variability is within the range observed among strains of T. b. brucei [46].
Genomic DNA isolated from T. evansi STIB805 was subjected to both 454 and Illumina sequencing, generating a combined total of 21,445,221 reads. Using the T. brucei TREU 927/4 genome sequence [22] (version 4; from here onwards abbreviated Tb927 for convenience) as a scaffold, a reference-based assembly of these T. evansi reads was generated. This assembly, called TevSTIB805ra, incorporates 17,856,165 reads and has an average coverage of 57.2 reads per nucleotide across the entire assembly and of 48.4 reads per nucleotide for the ‘core regions’ [i.e. regions excluding telomeres, subtelomeres, and internal regions that consist of repetitive coding sequences such as VSGs, expression site-associated genes (ESAGs) or retroposon hot spot genes (RHS)].
An assembly approach based on the genome of a closely related reference strain allows the reliable identification of homologous gene pairs, of any differences that might exist between these sequences, and of genes that are missing or highly diverged in the genome of interest. This approach has limited power in identifying structural differences between genomes and, by definition, cannot identify sequences that are present in the genome of interest but not in the reference genome. For that reason we have supplemented the reference-based approach with the analysis of contigs that were assemble de novo. As detailed below, this allowed the confirmation of Tb927 genes that are absent in in T. evansi STIB805 and the identification of candidate genes that might be present in the latter but absent in the former. However, genome rearrangements that are not associated with differences in gene content will have been missed and the T. evansi STIB805 genome as published on TriTrypDB may indicate gene synteny where it does not exist.
Annotation of the TevSTIB805ra T. evansi genome was performed using a combination of RATT and AutoMagi, followed by manual curation. RATT (Rapid Annotation Transfer Tool; [27]) identified likely orthologs in the T. brucei reference genome and transferred their functional annotations to the T. evansi genome, while AutoMagi predicted genes de novo using the consensus of three gene prediction algorithms (genescan, testcode, codonusage) [28]. A total of 10,110 CDS were identified, with 9368 CDS annotated as T. brucei orthologs by RATT and subsequent manual inspection (Table 1, columns C, E and F; S1 Table), and 742 CDS uniquely predicted by AutoMagi (Table 1, column D; S2 Table). Thus, 92.7% of the identified T. evansi CDS have an identified ortholog in T. brucei, while 7.3% were uniquely detected by de novo gene prediction. Analysis of the 742 AutoMagi CDS by BLAST searching of the GenBank non-redundant database revealed at least one hit (E-value ≤5e-6) for each of these CDS to a gene from a Trypanosoma species (T. b. gambiense 568 CDS; T. b. brucei 168 CDS; T. equiperdum, T. vivax, or T. congolense 6 CDS). The most common annotations among these BLAST hits were for ‘hypothetical unlikely’ (415 CDS) or other hypothetical (150 CDS) sequences (S2 Table). Orthologous sequences had not been annotated in the Tb927 reference, presumably due to more stringent criteria for gene calling. Thus, most if not all of the de novo predicted genes in TevSTIB805ra are also present in T. brucei.
Of the 9368 T. evansi CDS identified as T. brucei orthologs, 8421 CDS were classified as non-repetitive genes; the remaining 947 CDS were VSG, ESAG, RHS, or duplicate sequences. A comparison of the 8421 non-repetitive CDS between T. brucei and T. evansi revealed 7970 (94.6%) had a nucleotide identity of 95% or greater, 320 (3.8%) had a nucleotide identity between 70–95%, and 131 (1.6%) had a nucleotide identity less than 70% (Fig. 2; S1 Fig.). After RATT annotation transfer, a total of 503 (5.1%) Tb927 GeneIDs did not have identified orthologs in the TevSTIB805ra annotated genome (Table 1, column G; S2 Table). The majority of these (406) represent repetitive genes (e.g. VSG, ESAG, and RHS) or ‘hypothetical unlikely’ genes, which were not analyzed further. Five CDS correspond to predicted pseudogenes. T. evansi STIB805 orthologs for 49 Tb927 CDS were shown to be wholly or partially missing in TevSTIB805 reads that mapped to these gene loci (see below). T. evansi homologs for the remaining 43 Tb927 GeneIDs were ‘missed’ by RATT and AutoMagi, but were subsequently identified by manual examination of the T. evansi sequence. Thus, the majority of T. brucei CDS have extremely similar T. evansi orthologs, and very few T. brucei CDS were not found in T. evansi. This result is consistent with a very close phylogenetic relationship between these parasites.
The initial RATT approach only identified T. evansi CDS with an annotated Tb927 homolog in the syntenic position. To identify potential CDS in TevSTIB805ra that are syntenic to unannotated sequences in Tb927, and are homologous to annotated CDS in other (non-syntenic) locations, RATT was performed a second time by comparing each TevSTIB805ra chromosome to the entire Tb927 genome. This approach complemented de novo gene calling by AutoMAGI and identified 112 CDS in TevSTIB805ra (Table 1, columns E and F; S2 Table). The majority of these CDS were annotated as hypothetical proteins (64 CDS) or repetitive sequences such as VSG, ESAG, and RHS (39 CDS). For almost all loci, manual inspection revealed that either (i) a syntenic CDS existed in Tb927 that was unannotated or (ii) a syntenic, annotated CDS did exist in Tb927 but the presence of a highly similar sequence elsewhere caused a RATT artefact. An exception was TevSTIB805.6.770, where the syntenic sequence in Tb927 was found to be disrupted by frame-shifts, but in T. b. gambiense DAL972 to be annotated as Tbg972.6.420 (hypothetical protein, unlikely)
Two approaches were used to find genes in the T. evansi genome that may not be present in the T. brucei genome. Firstly, reads from T. evansi that did not match Tb927 were de novo assembled to make TevSTIB805dn and putative CDS called with AutoMagi (Table 2 and S2 Table). The 22 CDS with a homolog in Tb927 that was not a repetitive or hypothetical gene included multiple copies of pseudogenes for putative UDP-Gal/UDP-GlcNAc-dependent glycosyltransferase and lytic factor resistance-like protein, an adenosine transporter, calmodulin, a DNA topoisomerase, a leucine-rich repeat protein, ribulose-phosphate 3-epimerase, major surface protease B, and a galactokinase pseudogene. 36 of the 202 CDS that did not get a hit against Tb927 had a BLASTx hit in the NCBI non-redundant database: 23 CDS matched hypothetical Trypanosoma spp. and 9 appeared to be contaminants from other - mostly bacterial - organisms. The remaining 4 BLAST hits were for T. vivax RNA-dependent DNA polymerase and T. evansi VSGs (2 hits) and diminazene-resistance-associated protein.
Because Tb927 was created using a traditional sequencing approach and only covers the 11 large chromosomes [22], sequences found in the de novo assembly of T. evansi deep sequencing reads might not be unique to this strain or species, but rather reflect a difference in sequencing methodology or stem from intermediate-sized or mini-chromosomes. To address this, Illumina reads from sequencing T. brucei TREU 927/4 that also did not match Tb927 were de novo assembled into Tb927dn. Comparison of Tb927dn to TevSTIB805dn showed that 45.6% of the binned T. evansi reads matched to Tb927dn. The remaining T. evansi reads were then assembled into TevSTIB805dn_sub and putative CDS called with AutoMagi (Table 2 and S2 Table). Of the 84 BLASTp matches to non-repetitive genes, 71 are annotated as hypothetical proteins, 2 are lytic factor resistance-like proteins, 1 is a putative transporter, and 10 are annotated as putative (or pseudogene) UDP-Gal or UDP-GlcNAc-dependent glycosyltransferase. Glycosyltransferases in T. brucei are frequently found in subtelomeric regions that are difficult to assemble due to repetitive sequences, and are known to be highly variable among trypanosome strains [33], [47]. Of the 117 TevSTIB805dn_sub CDS for which no match was identified in Tb927, 18 had a BLASTx hit in the NCBI non-redundant database: 9 CDS matched hypothetical genes from various Trypanosoma species and 1 matched a diminazene resistance-associated protein that was previously suggested to convey diminazene aceturate (Berenil) resistance to certain strains of T. evansi [48]. Thus, gene prediction found very few CDS in de novo T. evansi assemblies that are not present in T. brucei, and analysis of these CDS revealed differences reminiscent of those observed among strains of the same species.
When T. evansi STIB805 de novo contigs were filtered for minimum coverage of at least 5x, the number of genes without obvious homologs in Tb927 was reduced considerably (Table 2). To test the potential absence of these genes in T. b. brucei more rigorously, short read data sets for strains TREU 927/4 and Lister 427 were searched for matches to the 22 candidates from TevSTIB805dn and the 6 candidates from TevSTIB805dn_sub. Only seven CDS candidates remained where the sequence was either entirely absent from both T. b. brucei datasets, or did not contain an undisrupted ORF (S2 Table, highlighted). Whether these candidates are indeed functional CDS, and whether they are generally absent from T. brucei ssp. and present in T. evansi, and therefore are of potentially diagnostic value, requires further investigation. A non-redundant list of the ORFs that did not have a BLASTx hit in NCBI (200 ORFs in total) is provided as S1 Data File.
A total of 49 CDS found in Tb927 were not identified in TevSTIB805ra. These loci were analyzed individually by (i) specifically searching for matching T. evansi STIB805 reads (S2 Table); (ii) comparing reads per kilobase per million (RPKM) coverage plots of these regions for T. brucei and T. evansi; (iii) aligning contigs of a full T. evansi STIB805 de novo assembly to the respective regions in Tb927. These analyses confirmed the absence or disruption of several of these CDS in T. evansi STIB805 (including the iron/ascorbate oxidoreductase loci, the Tb927.9.7950/Tb927.9.7960 repeat, the Tb927.4.3200-.3270 region, both Tb927.8.490 and Tb927.8.500, and the Tb927.8.7300-.7330 region), as illustrated in S2–S9 Fig., with the procyclin loci described in detail below. These cases have in common that the loci in question show considerable variation among Trypanozoon strains and CDS appear to be absent in T. b. brucei Lister 427 and/or T. b. gambiense DAL972 as well [33]. The differences observed in T. evansi STIB805 compared to Tb927 therefore are unlikely to be relevant for kDNA loss or tsetse-independent transmission.
In T. evansi STIB805, the procyclin loci either lack or have disrupted versions of several CDS found in T. brucei. GPEET and EP procyclins and associated genes are encoded in loci on chromosomes 6 and 10, respectively, in various T. brucei strains [49]. Procyclin proteins are GPI-anchored coat glycoproteins that are expressed exclusively in the procyclic insect form of T. brucei, and they have been hypothesized to be involved in protection against tsetse fly midgut hydrolases [50]. Experiments in T. brucei have shown that knocking out all of the procyclin genes (Null mutants of GPEET and EP3 on chromosome 6; EP1 and EP2 on chromosome 10) causes no growth defect in vitro and permits completion of the entire life cycle, but causes a selective disadvantage during co-infection with wild type cells in the tsetse fly midgut [51]. These procyclin loci also contain procyclin-associated genes and a gene related to expression site associated gene 2 (PAG3 and GRESAG2 on chromosome 6; PAG1, 2, 2*, 4, and 5 on chromosome 10) [49]. The functions of the PAG proteins and GRESAG2 are unknown; although transcripts of PAG1–3 have been shown to increase during differentiation to procyclic forms, published experiments using cell lines with all PAG genes knocked out reported no obvious abnormal phenotypes in vitro or in vivo [49]. In multiple T. brucei strains, the chromosome 10 procyclin loci are heterozygous, with one chromosome containing EP1/EP2/PAG1/PAG5/PAG2*/PAG4 and the other chromosome containing EP1/EP2/PAG2/PAG4 (PAG2 being a fusion of the 5′ part of PAG1 and the 3′ part of PAG2*) [49], [52]–[56]. In T. evansi STIB805, chromosome 10 appears to be homozygous, with only the EP1/EP2/PAG2/PAG4 locus present, and the associated absence of the segment containing the 3′ part of PAG1, PAG5 and the 5′ part of PAG2* (Fig. 3). The full STIB805 de novo assembly contained a single 14.9 kb contig corresponding to the EP1/EP2/PAG2/PAG4 locus (S2 Fig.). Also on chromosome 10, the EP2 in T. evansi contains a stretch of 12 divergent amino acids in the domain N-terminal to the EP repeat; this region is highly conserved in T. brucei [57], [58]. Although the function of this domain is unknown, these 12 amino acids are found in all sequenced T. brucei genomes with very few variations. The chromosome 6 procyclin locus contains a triplication of three genes (EP3/PAG3/GRESAG2) in T. brucei TREU 927/4, with GPEET present only in front of the last unit (Fig. 3). Copy numbers of these genes may vary among T. brucei strains, as Southern analysis showed that these are single copy genes in the AnTat1.1 strain [49]. In TevSTIB805ra, coverage of this locus is much reduced (Fig. 3, S3 Fig.), indicating either divergence, reduced copy number, or both. The locus did not assemble into a single contig in the de novo approach and because of the repeat nature of this locus in Tb927, assigning sequence differences to particular gene copies was not possible. Nonetheless, several non-synonymous mutations in EP3 are evident, and the PAG3 and GRESAG2 genes have frameshifts and deletions. In contrast, analysis of a subset of other T. evansi orthologs to CDS shown to be upregulated in PF relative to BF T. brucei (Tb927.1.2310, Tb927.1.2350, Tb927.1.2560, Tb927.1.580, Tb927.10.10260, Tb927.10.10950, Tb927.10.4570, Tb927.11.16130, Tb927.11.8200, Tb927.4.1800, Tb927.4.1860, Tb927.4.4730, Tb927.5.1710, Tb927.5.2260, Tb927.6.510, Tb927.8.5260, Tb927.4.4730, Tb927.8.8300, Tb927.9.15110, and Tb927.9.8420) [59] detected no notable changes.
Any Tb927 gene encoding a product that is exclusive involved in (i) maintenance or expression of kDNA, (ii) life cycle progression, or (iii) insect-stage specific energy metabolism, e.g. oxidative phosphorylation, is redundant in T. evansi. Although many of these genes encode mitochondrial proteins, numerous other mitochondrial activities are expected to remain essential even in an akinetoplastic bloodstream trypanosome (see Discussion). Extensive proteomic analyses have identified 978 mitochondrial proteins in T. brucei [60]–[63]. A search of TevSTIB805ra and TevSTIB805dn identified orthologous gene sequences for all of these proteins (S6 Table). A detailed examination of a subset of genes representing categories i-iii above (with the exception of the F1 subunits of respiratory complex V, which remain essential in T. evansi [19]) suggests that the vast majority, if not all of these genes remain functional (see S6 Table for details). To obtain further evidence for or against mutational decay in these genes, we calculated Ka, Ks, and Ka/Ks for the following orthologous gene pairs in Tb927 and T. evansi: 208 CDS associated with kDNA expression or function; 942 unambiguous orthologous gene pairs for mitochondrial proteins; and 6331 unambiguous orthologous gene pairs corresponding to non-VSG and non-mitoproteome annotations (Table 3). These orthologous gene pairs were selected using greater stringency (reciprocal best matches in BLASTp searches with a minimum e-value threshold of 1×10−4) than the analysis shown in Fig. 2, to ensure that paralogous gene pairs were excluded (these would possibly increase Ka and Ks estimates and skew the distribution; S6 Table). We found no evidence of relaxed selection for T. evansi CDS associated with kDNA expression or function or the mitoproteome as a whole. Both Ka and Ks were lower for CDS associated with mitochondrial expression, and Ka/Ks was also lower relative to the genomic background, indicating that purifying selection is generally stronger among mitoproteome CDS, perhaps suggesting a higher proportion of essential genes compared to the control set. Furthermore, we made the same comparison between Tb927 and T. b. gambiense DAL972 (both of which retain functional kDNA) and the values for Ka, Ks and Ka/Ks show no significant differences (by t-test) with the values for T. evansi.
The predicted protein structures of VSGs in T. evansi STIB805 conform to the canonical structures in T. b. brucei and T. b. gambiense, and include all five recognized N-terminal sub-types (N1-5). In T. evansi STIB805 we identified 525 a-type VSGs (i.e. N-1-3 and 5; S2 Data File) and 505 b-type VSGs (i.e. N4; S3 Data File); of these 453 (86%) and 451 (89%), respectively, are full-length. Given that the assembly of subtelomeric and mini-chromosomal regions is fragmentary, these numbers may underestimate the real number of VSGs, although the total is comparable with both T. b. brucei TREU 927/4 [22] and T. b. gambiense DAL972 sequences [33]. The VSG repertoire is largely conserved between the T. b. brucei TREU 927/4 reference and T. evansi STIB805, as the VSGs are interspersed among each other in neighbour-joining molecular cladograms (S10 Fig.). Another indication of the similarity of the VSG repertoire is that very few clades of strain-specific VSG are larger than 2–3 (Fig. 4a and b). Large clades of VSG from a single genome would suggest divergence of VSG repertoire through gene duplication. Hence, 376 and 384 ( = 85.1%) T. evansi STIB805 and T. b. brucei TREU 927/4 a-VSGs, respectively, are most closely related to an ortholog in the other strain, or are paraphyletic to a clade of such sequences (i.e. strain-specific clade size = 1). Only 40 T. evansi STIB805 VSGs and 51 T. b. brucei TREU 927/4 VSGs form a clade with a paralog from the same genome (i.e. strain-specific clade size = 2), suggesting a single gene duplication since the strains separated. The same patterns occur with b-VSGs. Orthology between VSGs does not mean that they are unaffected by recombination, only that enough sequence homology persists for two orthologs to cluster together. Indeed, among 151 putative a-VSG orthologs 33 (21.8%) have dissimilar C-terminal types, while among 112 b-VSG orthologs 32 (28.6%) showed similar evidence for recombination having occurred since these two strains split from their common ancestor. Thus, analysis of the VSG repertoire reveals no evidence for the evolution of specific VSG gene clusters or subfamilies in T. evansi, similar to what had been observed for T. b. gambiense [33].
This comparison of orthologous VSGs from T. evansi STIB805 and T. b. brucei TREU 927/4 provides an overview of the molecular evolutionary forces incident on the VSG archive. Fig. 4c and d show that substitution rates among orthologous VSGs are typically higher than the background defined by comparison to the 6331 non-VSG orthologous gene pairs previously used in the mitoproteome analysis above. The mean average rate for VSG is significantly greater than other genes (p<0.05) and comparison of Ka/Ks (ω) explains this. The distribution of ω is left-skewed for non-VSG genes, which reflects the strong purifying selection on non-synonymous substitutions that typifies most genes that perform essential functions. The distribution for VSG is normal, indicating that purifying selection is much weaker on average, although still in effect. This difference in the distribution of ω is significant (p<0.005) and shows that VSGs evolve in a more neutral fashion than other genes, resulting in higher substitution rates. However, there is no more evidence for positive selection among VSGs than for 'normal' genes, which reflects our current understanding that VSG sequence evolution in T. brucei is driven predominantly by the diversifying effects of recombination [35], rather than directional selection driven by host immune responses.
Consistent with earlier reports [64] that have suggested RoTat 1.2 (NCBI accession AF317914) as a predominant and diagnostic T. evansi VSG, we identified a closely related ORF in STIB805. The first 1250 bp of a 1431-bp ORF on a 3.9-kb de novo contig are nearly identical to the published RoTat1.2 sequence, while the 200 bp at the 3′ end are nearly identical to Tb927.8.240, annotated as ‘VSG, degenerate’ in TriTrypDB (S11 Fig.). Thus, the gene in STIB805 had most likely obtained a different C-terminal region through recombination, an important mechanism in VSG evolution [35]. We did not identify any close homologs to the 5′ region of RoTat1.2 in the publicly available T. b. brucei TREU 927/4, T. b. brucei Lister 427 or T. b. gambiense DAL972 datasets.
A comparison of the TevSTIB805ra to the Tb927 reference genome revealed 354,809 single nucleotide polymorphisms, 68,938 of which were non-synonymous in identified CDS. 32,892 of these non-synonymous SNPs were found in 904 CDS classified as repetitive (e.g. RHS, VSG, ESAG), while 36,046 were found in 6630 CDS from non-repetitive genes (S3 Table). This number is slightly lower than the ∼45,000 non-synonymous SNPs reported for T. b. gambiense orthologs of non-repetitive T. b. brucei TREU 927/4 genes [33]. Additionally, there were 45,269 short indels (insertions/deletions of up to eight nucleotides), of which 5544 were in CDS regions. Of these, 4963 (89.5%) were not an indel of mod 3, and would therefore be expected to cause a frameshift. 3247 (58.6%) of the indels were found in 656 CDS classified as repetitive, while 2297 (41.4%) were found in 1091 CDS from non-repetitive genes (S4 Table). The latter consisted of 1154 insertions, 1000 deletions, and 143 that were complex (a combination of insertion, deletion and/or SNP). The Tb927 reference contains information for one allele only for any heterozygous locus. We therefore also identified SNPs and indels for TREU 927/4 using publicly available Illumina reads and the same criteria that were applied for STIB805. We identified 26,522 non-synonymous SNPs and 1630 indels. Of these, 814 and 58, respectively, were homozygous, thus reflecting discrepancies between the original Sanger sequencing data and the Illumina reads. The cause of these discrepancies, which could arise from a number of potential sources, including DNA isolation from different cultures of the same strain, is unknown. Only 1267 non-synonymous allele variations (SNP or indel) in CDSs (repetitive or non-repetitive) were shared by STIB805 and TREU 927/4, 8710 variations affected the same position, but were different, and 89,043 affected a different position. All non-synonymous SNPs and indels identified in STIB805 and TREU 927/4 are compared in S5 Table, listed by chromosome and gene. In summary, while numerous small sequence variations exist between T. evansi STIB805 and the T. b. brucei reference genome, these are comparable in number to differences between the T. b. brucei and T. b. gambiense subspecies.
The phylogenetic relationship of various T. evansi and T. equiperdum isolates to each other and to isolates of T. brucei subspecies is controversial. We compared T. evansi STIB805 CDSs with their T. b. brucei TREU 927/4 orthologs as retrieved from Tb927 in order to identify a genetic marker that would likely be informative for the elucidation of the phylogenetic origins of T. evansi and T. equiperdum, and for determining their evolutionary relationships with established T. brucei sub-species. Selection criteria included (i) sufficient SNPs over a length of 1–3 kb (to minimize number of internal primers necessary for amplification and sequencing), (ii) lack of paralogs, and (iii) minimal alignment gaps. One promising candidate, the LipDH gene (Tb927.11.16730), a shared component of four mitochondrial multi-enzyme complexes [65], was selected for sequence analysis. Both LipDH CDS alleles were amplified and sequenced from a total of 40 isolates from various geographical locations (13 T. b. brucei, 3 T. b. gambiense group 1, 4 T. b. rhodesiense, 5 T. equiperdum, 15 T. evansi; S7 Table). Since mutations in subunit γ of the mitochondrial ATP synthase complex were recently identified as important factors in the viability of T. evansi and T. equiperdum, the sequence for both γ alleles was also determined for all isolates.
32 unique LipDH haplotypes were identified. The majority of strains (36/41) had two different alleles and four strains appeared to be homozygous at this locus, all of which were T. b. brucei (although it cannot be ruled out that the primers used were selective for one allele in these cases). Phylogenetic analysis (Fig. 5) showed all but three of the haplotype sequences fell into one of five major clusters with strong support (posterior probabilities ≥0.9), which we refer to as clades V, W, X, Y and Z. Importantly, 14 different haplotypes are derived from the T. evansi/T. equiperdum isolates, which are found in three different clades as well as outside of the clades. This feature is incompatible with monophyly of T. evansi or T. equiperdum. Some subspecies did have relatively restricted phylogenetic diversity, for example T. b. gambiense type 1 had only five closely related haplotypes and certain clades contained sequences from a restricted number of the subspecies; e.g. clade X only contained T. b. rhodesiense and T. b. brucei; clade Y included only T. b. gambiense type 1 and T. b. brucei. This may reflect relatively more recent origins of certain populations, bottleneck events and/or sampling bias, and is in line with most previous work ([40] and references therein).
Eight unique T. evansi/T. equiperdum genotypes (haplotype pairs) were found which, by disregarding minor sequence differences, were reduced to four major genotypes: W/Z, V/V, V/Z, or atypical (Fig. 5; S7 Table). These LipDH genotypes correlated very well with the four ATP synthase γ genotypes found in these strains (Fig. 5B; see also S7 Table).
The largest group (Group 1), consisting of 13 T. evansi isolates (including STIB805) and 2 T. equiperdum isolates, is characterized by LipDH genotype W/Z and γ genotype A281Δ/WT. The single exception, Tev48, had two Z haplotypes (Hap24+Hap29) distinguished by two SNPs, which may be a result of loss of heterozygosity (LOH) followed by two mutations, or it may represent a cloning artefact. Non-T. evansi/T. equiperdum strains that had sequences within these clades originated from both East and West Africa in the case of clade Z, and the T. b. gambiense samples from Ivory Coast in the case of clade W. The remaining T. evansi/T. equiperdum strains were split into three further groups according to the LipDH/ATP synthase γ genotypes as follows. Group 2 is characterized by LipDH V/Z and γ genotype A273P/A273P. The two T. equiperdum strains carrying this genotype (Teq21 and Teq22) were heterozygous for two alleles derived from well-separated clades: one ‘V’ allele (Hap17 or Hap18, 1 SNP difference) and one ‘Z’ allele (Hap14 or Hap15, 1 SNP difference). The divergence between the V and Z alleles (11 or 12 discriminating SNPs), suggests that these strains have evolved out of a T. b. brucei background that had relatively old and distinct LipDH alleles, possibly via a recombination event. Interestingly, the most closely related non-T. evansi/T. equiperdum alleles are found in current East African populations as well as in three T. b. brucei from Ivory Coast. Group 3 is characterized by LipDH V/V and γ genotype WT/WT. The one T. equiperdum strain with a WT γ genotype (Teq23 = STIB841; probably synonymous with the OVI strain [5]) was heterozygous with two similar ‘V’ clade alleles that are closely related to the V alleles in Teq21, Teq22, Tbr02 (Kenya) and Tbb08 (Zambia) (S7 Table). Group 4 is characterized by atypical LipDH and γ genotype M282L/WT. This genotype was found in a single T. evansi strain (Tev42 = KETRI 2479) with Hap27+Hap28 and could not be reliably grouped with any other strains.
An interesting aspect of the T. evansi/T. equiperdum genotypes is that most strains possess two divergent alleles that are more similar to those found in other, non-Tev/Teq strains than they are to each other. These non-Tev/Teq strains are themselves heterozygous, with second alleles that are clearly distinct from either of the Tev/Teq alleles. A clear example is the most common Teq/Tev LipDH genotype, W/Z, that is carried by 13 out of 15 T. evansi analyzed and 2 out of 5 T. equiperdum analyzed. This finding is incompatible with allelic divergence due to long-term clonal evolution, rather it shows that recombination between ancestral parasites occurred either at or immediately prior to the origin of this genotype which has subsequently undergone clonal expansion in the absence of the possibility of sexual exchange in the tsetse vector. Three of the four T. b. rhodesiense strains (Tbr01, 03, 04) also exhibit alleles split across different clades as does Tbb06. Interestingly, all Tbb except Tbb06 have alleles from the same clade, possibly indicative of inbreeding.
The COX1 gene, encoded in the mitochondrial maxicircle DNA, was recently used as a highly informative marker to investigate the phylogeography of T. brucei subspecies [40]. Although most, if not all, T. evansi isolates lack a maxicircle, the presence of at least one T. equiperdum isolate with at least a partial maxicircle in three of the above groups prompted us to determine the COX1 haplotypes for these strains and to investigate their relationship with the Trypanozoon isolates from the earlier study (S7 Table). A maximum parsimony network analysis strongly supported the notion of three independent evolutionary origins for these three groups (Fig. 6). T. equiperdum STIB818 (Teq24; COX1 haplotype 23) links Group 1 with COX1 clade A, which is composed of isolates of all three T. brucei subspecies found across all of sub-Saharan Africa [40]. Although Groups 2 and 3, which share a related ‘V’ LipDH genotype, are both linked to COX1 clade C (composed of T. b. brucei and T. b. rhodesiense isolates from eastern and southern Africa), they are well separated within this clade, indicating independent evolutionary origins. T. equiperdum STIB841 (Teq23) shares both its COX1 haplotype 14 and its LipDH haplotype 19 with Tbr02 from Zambia and Tbb08 from Kenya, both members of the Kiboko group, suggesting relatively recent common ancestry. Note that Tev42 (KETRI2479), the lone representative of the ‘atypical’ LipDH genotype, is not represented in this network since this isolate lacks a maxicircle and therefore the COX1 gene.
Incorporation of microsatellite data for a subset of T. evansi/T. equiperdum isolates into an established PCA network [40] also gave results that are inconsistent with monophyly of either species (Fig. 7). Most T. evansi isolates, together with Teq24 (STIB818), formed a cluster (grey circle) related to, but somewhat distinct from, non-Kiboko T. b. brucei (light blue circle) and T. b. rhodesiense (red circle). The single exception among T. evansi was again Tev42 (KETRI2479), which localized near the centre of the non-Kiboko cluster. Teq21 (BoTat1.1) was also more related to non-Kiboko T. b. brucei, but relatively distant from Tev42. The PCA analysis, consistent with the COX1 and LipDH data, suggested a relatively close evolutionary relationship of T. equiperdum STIB841/OVI (Teq23) with the Kiboko group of T. b. brucei (dark blue circle).
The results of the phylogenetic analyses are summarized in Table 4, along with the ATPase subunit γ genotypes and, where known, dominant minicircle class and RoTat 1.2 VSG genotype. Combined, these markers suggest that the T. evansi/T.equiperdum isolates investigated in our study can be arranged into four distinct groups of independent evolutionary origin (see Discussion)
We sequenced and analyzed the genome of T. evansi strain STIB805, a representative of a large, clonal group of close relatives of T. brucei that are transmitted independently of tsetse flies. Our examination revealed a number of insights into the biology and phylogenetic origin of this akinetoplastic trypanosome and provides further support for the proposed classification of T. evansi and T. equiperdum as subspecies of T. brucei [4]. A comparison with the T. b. brucei TREU 927/4 reference shows broad similarity, with 94.6% of non-repetitive CDS having a nucleotide identity of 95% or greater and 78.9% having ≥99% identity (Fig. 2). This high degree of identity mirrors similarities previously observed between T. b. brucei TREU 927/4 and T. b. gambiense DAL972 (86.4% of CDS had ≥99% identity) [33]. Phylogenetic analyses of T. evansi STIB805 VSG sequences also show extensive similarity to T. b. brucei. A cladistic representation of VSG sequences from T. evansi and T. b. brucei shows these sequences broadly interspersed with little evidence for subspecies-specific grouping, which underscores the close evolutionary relationship of the two strains (S10 Fig.). Despite the lack of kDNA and the absence of the life cycle stages in the insect vector, where the T. b. brucei mitochondrion is exclusively active in oxidative phosphorylation, T. evansi has maintained the coding capacity for nearly all of the mitochondrial proteome found in T. b. brucei. To some extent this reflects the fact that many mitochondrial activities are expected to remain essential even in an akinetoplastic bloodstream trypanosome. Such activities include the glycine cleavage complex [65], the alternative oxidase [66] and ubiquinone biosynthesis [67], fatty acid metabolism [68], [69], the F1-ATP synthase and ADP/ATP carrier [20], iron sulfur cluster biosynthesis [70], and all activities required to maintain and duplicate the mitochondrion itself. Nonetheless, a focused analysis of those genes known to be involved in kDNA maintenance or expression, or in oxidative phosphorylation, identified not a single case of gene loss and very few cases of mutations with predicted consequences for protein function. Of the 35 kDNA replication and transcription CDS examined, only the mitochondrial DNA polymerase beta-PAK appeared to be functionally compromised (S6 Table). Similarly, all 133 of the CDS identified as the mitochondrial ribosome in T. b. brucei [62], [71] are maintained in T. evansi, with the largest difference observed being a relatively small region of TevSTIB805.11_01.4800 (Tb927.11.4650) that alters 24 amino acids in the middle of the coding sequence. The paucity of disruptions among such a large protein complex strongly suggests that the mitochondrial ribosomes in T. evansi STIB805 are fully functional despite the lack of known substrates. Editosome function also appears essentially intact in T. evansi. Previous experiments have shown that dyskinetoplastic T. b. brucei and another strain of T. evansi contain functional editing complexes, which is consistent with retention of proteins no longer required [72].
Components of the mitochondrial electron transport chain are of particular interest in T. evansi, as the key compensatory change that permits survival in the absence of kDNA resides in complex V (mitochondrial FOF1-ATP synthase) [20]. The electron transport chain is composed of five complexes in the inner mitochondrial membrane, and in T. brucei these complexes are differentially expressed between life cycle stages [9]. In bloodstream forms, ATP is thought to be produced entirely by glycolysis. The classical respiratory chain is not functionally active, and the ATP synthase operates “in reverse” by hydrolyzing ATP to pump protons into the mitochondrial matrix in order to maintain the essential mitochondrial membrane potential, Δψm [19], [73], [74]. This activity requires the kDNA-encoded FO subunit 6. Maintenance of Δψm in the absence of a functional FO in T. evansi STIB805 involves an A281 deletion in the gamma subunit of the F1 subcomplex, but whether this mutation alone is sufficient to fully compensate for kDNA loss is presently unclear [20]. Our study has identified mutations in the α and β subunits of F1 that could function in concert with the A281 deletion to permit survival of T. evansi STIB805, and this can be tested experimentally. Nuclear genes encoding components of complexes I-IV of the oxidative phosphorylation system appear to be intact in T. evansi. Although complex I is expressed in bloodstream T. brucei, a recent study demonstrated that it is non-essential for bloodstream form metabolism, and no growth defects were observed in cell lines in which complex I components were eliminated [75]. Genes for subunits of complexes II-IV are not required in bloodstream form parasites, and eleven kDNA-encoded subunits are absent in T. evansi, yet the integrity of the nuclear-encoded genes for complexes I-IV has been maintained in T. evansi STIB805.
Ka/Ks analysis further highlights the conservation of the T. evansi mitoproteome in general and of a subset of proteins involved in kDNA maintenance, expression, and function in particular: we detected no significant difference between Ka, Ks, or Ka/Ks values for T. evansi STIB805 and T. b. gambiense DAL972 relative to T. b. brucei. A potential explanation is that selective pressure remains for a large number of the mitoproteome CDS, including those involved in kDNA maintenance and function, perhaps due to additional functions outside of the mitochondrion. For example, the mitochondrial topoisomerase II was recently shown to be important for normal growth even in the absence of kDNA [20]. Alternatively, retention of functional sequences could reflect a relatively short evolutionary separation between T. brucei and T. evansi STIB805, high stability of the gene maintenance and expression pathways in these organisms, or some unknown aspect of their biology. For example, robust mechanisms to maintain gene functionality would make sense in an organism with a complex life cycle like T. brucei that spends many generations in an environment where only a subset of genes are required.
The depth of similarity between T. evansi STIB805 and T. b. brucei TREU 927/4 creates a background that highlights some of the observed differences. The differences in chromosome pattern between the two T. evansi and T. b. brucei strains are comparable to strain variations observed within the T. b. brucei subspecies (Fig. 1) [46]. Previous studies investigating T. brucei chromosomes have indicated that gene content is more conserved than chromosome size [30], [76], similar to other kinetoplastid species [77], [78]. This size polymorphism is probably caused by genome-wide rearrangements over evolutionary time. The majority of the megabase chromosomes appeared largely similar between T. evansi STIB805 and T. b. brucei TREU 927/4, which is consistent with similarities subsequently observed in sequence. It remains to be shown whether the chromosomal differences observed among trypanosomatids reflect any functional consequences, which are more obvious at the gene level.
One clear example of such CDS differences that are found in the T. evansi STIB805 genome is the loss of a significant number of procyclin associated genes (PAGs). In multiple T. brucei strains, the chromosome 10 procyclin locus is heterozygous, with one chromosome containing EP1/EP2/PAG1/PAG5/PAG2*/PAG4 and the other chromosome containing EP1/EP2/PAG2/PAG4 [49], [52]–[56]. In T. evansi STIB805, chromosome 10 is homozygous for the EP1/EP2/PAG2/PAG4 locus, resulting in the associated absence of PAG5 and PAG2*. Also on chromosome 10, the EP2 gene in T. evansi contains a stretch of 12 divergent amino acids in the domain N-terminal to the EP repeat; this region is highly conserved among T. b. brucei strains [57], [58]. The PAG3 and GRESAG2 genes on chromosome 6 have frame-shifts and deletions. Although the function of procyclins and PAGs are not fully clear, the hypothesized role in protection within tsetse fly midgut is a logical extrapolation [50]. Because T. evansi has jettisoned this life-cycle stage, a life persisting as bloodstream parasites provides no selection to maintain these genes. Nonetheless, the fact that several procyclin genes and PAGs have been disrupted in T. evansi STIB805 stands in stark contrast to the integrity of other genes dispensable in an akinetoplastic bloodstream trypanosome, such as the genes specifically required for kDNA maintenance or expression, or for insect stage-specific energy metabolism (see above). Procyclin genes - in contrast to most other protein coding genes - are transcribed by RNA polymerase I, which might be associated with idiosyncratic features at these loci. Thus, either the loss of the genes gives a selective advantage to parasites locked in the bloodstream stage, or their genomic loci are intrinsically less stable.
Several other genes appear to be missing in T. evansi STIB805, but all cases we identified represent members of gene families that show considerable variation among T. brucei subspecies and strains. For example, presence of the complete IAO/AT2 array on T. b. brucei TREU 927/4 chromosome 2 and of the single IAO gene on chromosome 9 (Tb927.9.14600) is highly variable among T. b. brucei and T. b. gambiense strains [33]. Similarly, tandem duplications on T. b. brucei TREU 927/4 chromosomes 3 (Tb927.3.5690-.5730) and 6 (Tb927.6.1310-.1390) are absent from both STIB805 and T. b. gambiense DAL972 [33]. Hence, with the notable exception of procyclins and PAGs, most cases of genes that are present in Tb927 but absent from T. evansi STIB805 are probably more a reflection of frequent differences in gene content within the subgenus Trypanozoon, rather than events specific to T. evansi. It is unknown if this variability is of any biological significance.
A search for ORFs on contigs assembled de novo from T. evansi STIB805 Illumina reads identified 200 candidates for T. evansi-specific genes (S1 Data File). These candidates lacked close homologs in other trypanosome species based on the absence of BLASTp hits (E-value cut-off 0.001) in Tb927 or in the NCBI database (Table 2 and S2 Table). Seven of these candidates came from contigs with at least 5x minimal coverage and mapping of T. b. brucei TREU 927/4 or Lister 427 Illumina reads to these ORFs resulted in no or only incomplete coverage. The size of these ORFs is generally small (306–501 bp) and the surrounding contig regions either lack homology to Tb927 chromosomes entirely or match to repetitive regions such as VSG arrays. Further studies are needed to elucidate whether any of these candidates are functional CDS, and whether they are generally absent from T. brucei ssp. and present in T. evansi strains, and therefore are of potential diagnostic value. We confirmed the presence of the RoTat 1.2 gene in the STIB805 genome, which encodes a VSG used for serodiagnosis of surra [79]. Although the N-terminal domain is almost identical to the canonical RoTat 1.2 sequence (NCBI accession no. AF317914), the STIB805 protein has a different C-terminal domain. Based on our phylogenetic analysis we would caution that the RoTat 1.2 gene may not reliably distinguish surra from dourine (see below).
The relationship of T. evansi strains to each other as well as to strains classified as T. equiperdum and to the T. brucei group is controversial [5], [11], [80]. Hoare had suggested that T. evansi evolved from a T. brucei infection in camels that had temporarily entered the sub-Saharan tsetse belt. This strain subsequently became adapted to mechanical transmission by biting flies and disseminated by caravans to areas outside of sub-Saharan Africa. Hoare further speculated that the sexually transmitted T. equiperdum may have evolved from an equine strain of T. evansi by developing tropism for genital tissues [81]. Subsequent molecular analyses demonstrated that most strains of T. equiperdum, in contrast to T. evansi, possess at least a partial maxicircle (reviewed in [8]), which is inconsistent with a scenario in which the former evolved from the latter. Brun et al. proposed a reversed scenario in which T. evansi arose from a clone of T. equiperdum that had lost its maxicircle [3]. More extensive phylogenetic studies suggested that the T. evansi/T. equiperdum group is not monophyletic, leading to the hypothesis that some T. equiperdum isolates are misclassified T. evansi strains whereas others are a subspecies of T. brucei [5], [82]. A more recent phylogenetic analysis based on SL RNA repeats from various isolates of T. evansi, T. equiperdum, T. b. brucei and T. b. gambiense found no species-specific clusters, and the authors suggested that the various T. evansi and T. equiperdum strains may have evolved on numerous independent occasions and should be classified as subspecies of T. brucei [1], [4].
The phylogenetic analysis presented in the present study, which was based on four different genetic markers, strongly suggests that T. evansi/T. equiperdum evolved from within the T. brucei group on at least four independent occasions and from genetically distinct T. brucei strains (Table 4 and Fig. 5–7). Isolates within each group share not only the genotypes assessed in the present study, but also the dominant class of minicircle (if present) and the RoTat 1.2 genotype. The grouping presented in Table 4 is also consistent with a phylogeny based on random amplified polymorphic DNA and a multiple endonuclease typing approach [5], [82]. Not all markers listed in Table 4 have been assessed for all strains, but all the data available are consistent with this grouping, which provides a solid framework for further studies. The isolates we sampled cover a time period of 110 years and a vast geographical area (S7 Table), but it seems likely that other independent lineages of T. evansi/T. equiperdum have emerged earlier, and that independent extant lineages exist that were not sampled in the present study.
Another important observation from the phylogenetic data summarized in Table 4 is that Group 1 contains isolates representing both T. evansi and T. equiperdum. A similar intermingling of isolates classified as T. evansi and T. equiperdum was observed in other phylogenetic studies and led the authors to hypothesize that these T. equiperdum strains had been misclassified and are actually T. evansi [5]. Another possibility is that T. evansi can under certain circumstances evolve from T. equiperdum by converting from a tissue parasite back to a blood parasite, and from sexual transmission to mechanical transmission by biting flies, as suggested by Brun et al. [3]. It remains mysterious why sexual transmission in horses by and large appears to be correlated with the presence of a (sometimes partial) maxicircle, and mechanical transmission by biting flies with a lack thereof. Based on our current knowledge of kinetoplast function in trypanosomes the maxicircle becomes vestigial in the absence of minicircle heterogeneity, and apparent minicircle homogeneity is a hallmark of both T. evansi and T. equiperdum (reviewed in [8]). Finding answers to these important questions will require a more complete understanding of kinetoplast function, and a thorough molecular and epidemiological analysis of new T. evansi and T. equiperdum isolates.
Regardless of which of the above explanations for the simultaneous presence of T. evansi and T. equiperdum isolates in Group 1 is correct, it is evident that they both evolved independently on at least two distinct occasions, and neither is therefore monophyletic. This scenario is reminiscent of the situation with T. b. gambiense, which evolved from T. b. brucei on two independent occasions [40], [83]. An important difference is that T. b. gambiense, at least theoretically, can mate with T. b. brucei to form hybrids, and there is strong evidence that this occurs for group 2 T. b. gambiense [83]. In contrast, T. evansi and T. equiperdum can no longer develop in the tsetse vector, considered a requirement for mating, and are therefore now genetically isolated. A consistent solution to the species problem in trypanosomes, as in other organisms, has not been reached, and classification of species vs. (sub)species has often been subjective and utilitarian rather than logical [84],[85]. However, maintaining the rank of species for T. evansi and T. equiperdum, despite clear evidence for lack of monophyly in both cases, seems inconsistent even within the Trypanozoon group itself. We therefore recommend re-classification as subspecies, i. e. T. b. evansi and T. b. equiperdum.
The present study did not identify any obvious candidates for genetic changes that might underpin the switch from tsetse-dependent to mechanical transmission by biting flies in T. evansi. A mathematical model for mechanical transmission suggested a high level of parasitaemia in the blood as an important condition for its success [86]. T. brucei regulates its parasitaemia through a quorum sensing mechanism that results in density dependent differentiation into a non-proliferative ‘stumpy’ form that is pre-adapted to survival in the tsetse vector [87]. Adaptations in the signaling pathway responsible for this sensing could provide a mechanism for increased parasitaemia, and numerous components of this pathway have recently been identified [88]. Indeed, while stumpy forms were occasionally observed for T. evansi, they are usually absent [81]. However, an entirely uncontrolled proliferation in the blood would probably result in untimely death of the host, again detrimental to transmission. On the other hand, a multi-species parasite like T. evansi could have (i) an uncontrolled proliferation in some of its hosts, and thus be fatal, as in horses, and (ii) a more controlled proliferation in others, such as certain bovines, which may act as reservoir. It therefore seems plausible that successful adaptation to mechanical transmission requires striking a new balance between survival of parasite and host, resulting in a new equilibrium that involves a very large range of receptive hosts exhibiting variable susceptibility.
Taken together, the genome analysis and accompanying phylogenetic studies presented in this work revealed important insights into the biology and evolution of T. evansi STIB805 and dyskinetoplastic trypanosomes in general. Their re-classification as subspecies of T. brucei, i. e. T. b. evansi and T. b. equiperdum, seems clearly justified considering the vast similarities observed at the genome level and the lack of monophyly confirmed by phylogenetic analyses. Important questions that remain include the molecular basis of tsetse-independent transmission and the evolutionary timescale of the appearance of the T. evansi/T. equiperdum groups. The genome data presented here provide an important tool for future studies aimed at resolving these questions.
|
10.1371/journal.pgen.1007212 | Genetic compensation triggered by actin mutation prevents the muscle damage caused by loss of actin protein | The lack of a mutant phenotype in homozygous mutant individuals’ due to compensatory gene expression triggered upstream of protein function has been identified as genetic compensation. Whilst this intriguing process has been recognized in zebrafish, the presence of homozygous loss of function mutations in healthy human individuals suggests that compensation may not be restricted to this model. Loss of skeletal α-actin results in nemaline myopathy and we have previously shown that the pathological symptoms of the disease and reduction in muscle performance are recapitulated in a zebrafish antisense morpholino knockdown model. Here we reveal that a genetic actc1b mutant exhibits mild muscle defects and is unaffected by injection of the actc1b targeting morpholino. We further show that the milder phenotype results from a compensatory transcriptional upregulation of an actin paralogue providing a novel approach to be explored for the treatment of actin myopathy. Our findings provide further evidence that genetic compensation may influence the penetrance of disease-causing mutations.
| Many healthy individuals carry loss of function mutations in essential genes that would normally be deleterious for survival. Intriguingly, it may be the presence of the genomic lesion itself in these individuals that triggers the compensatory pathways. It is not known how widespread this phenomenon is in vertebrate populations and how genetic compensation is activated. We have shown that knockdown of actin causes nemaline myopathy as indicated by the formation of nemaline bodies within the skeletal muscle and reduced muscle function which, remarkably, we did not observe in an actin genetic mutant. We have identified that protection from the disease phenotype results from transcriptional upregulation of an actin paralogue restoring actin protein in the skeletal muscle. This study demonstrates that genetic compensation may be more prevalent than previously anticipated and highlights phenotypic differences resulting from genetic mutations versus antisense knockdown approaches. Furthermore, we suggest that activating compensatory pathways may be exploited as a potential novel therapeutic approach for human disorders caused by loss of function mutations.
| Genetic compensation exists as a mechanism to buffer the organism against gene loss that would otherwise be deleterious to survival. Whilst this term has been used to describe dosage compensation, evolution resulting in reversion to ancestral phenotypes, and gene duplication compensating for mutation, in the present study we refer to genetic compensation as altered gene expression resulting in a normal phenotype in a homozygous mutant individual. Studies in plants [1], worms [2], and yeast [3,4] have demonstrated that genetic robustness can be achieved by the presence of duplicate gene copies, which have retained a similar biological function, or through the presence of redundant pathways. A single case of genetic compensation in zebrafish has been shown to result in the activation of a network of gene expression resulting from the presence of a mutation in egf17, but not from gene knockdown [5]. Intriguingly, data from widespread application of whole genome sequencing demonstrates that homozygous mutations predicted to cause a loss of function, that would normally cause disease, may be present in healthy individuals in the human population [6], suggesting that compensation may be more widespread than previously thought. However, the lack of functionally characterized examples raise the question of whether this is an isolated case, or if genetic compensation may be a common mechanism contributing to genetic robustness in vertebrates.
In the skeletal muscle, expression of cardiac muscle α-actin (ACTC1) can partially compensate for loss of ACTA1 to ameliorate the loss of function phenotype [7]. α-actin is an essential component of the thin filament, with its mutation resulting in a range of skeletal muscle disorders, including nemaline myopathy [8]. ACTA1 differs from ACTC1 by only four amino acids and both are co-expressed in the skeletal muscle and heart during development [9–11]. In vertebrates, ACTC1 is the predominant actin isoform in fetal skeletal muscle, however, after birth ACTC1 is down regulated [11] and by adulthood it comprises <5% of adult skeletal muscle [12], with ACTA1 becoming the predominant form.
Most ACTA1 mutations are de novo dominant mutations with only 10% of patients’ carrying recessively inherited loss of function mutations [8]. Patients displaying a complete absence of ACTA1 typically show retention of ACTC1 in their skeletal muscle, with the level of retention determining the level of clinical severity [7]. Transgenic overexpression of ACTC1 was also shown to rescue the early lethality observed in recessive ACTA1-/- knockout mice strains [13], further demonstrating the potential for the levels of ACTA1 paralogues to influence disease severity.
Here we uncover a novel case of genetic compensation in zebrafish within the highly conserved actin gene family. We [14] and others [15,16] have shown that loss or knockdown of skeletal α-actin is catastrophic for muscle structure and function. Antisense Morpholino (MO) knockdown of Actc1b results in the formation of nemaline bodies and reduced skeletal muscle performance [14]. In contrast, we observe very mild defects in a genetic actc1b mutant. We determine that this is due to a compensatory transcriptional upregulation of an α-actin paralogue in the skeletal muscle buffering the loss of Actc1b function. Our study not only supports the existence of genetic compensation as a phenomenon affecting phenotypic diversity in vertebrates, but also identifies a genetic process that may have therapeutic implications for nemaline myopathy and other actin-related diseases.
To investigate α-actin function in zebrafish we initially examined the expression of the α-actin paralogues. Four α-actin genes (acta1a, acta1b, actc1a, and actc1b) have been identified in the zebrafish [14], and all are expressed in both skeletal and cardiac muscle during early zebrafish development (S1 Fig). qRT-PCR analyses showed that, whilst all genes are expressed during early embryogenesis, actc1b is expressed at much higher levels than the other paralogues and that by 180 dpf (days post-fertilization) actc1a and actc1b are the predominant α-actin isoforms in the heart and skeletal muscle respectively (Fig 1, S1 Table, and S2 Table).
We were surprised to find that the two cardiac α-actin genes were the most highly expressed in the muscle and therefore analyzed the sequence similarity between the α-actin isoforms. An analysis of the surrounding genomic regions shows conserved synteny between human ACTA1 and zebrafish acta1a and acta1b genes, and similarly between human ACTC1 and zebrafish actc1a and actc1b genes (S2 Fig) suggesting that indeed the zebrafish genes are orthologous to their respective human genes. Interestingly, a maximum likelihood phylogenetic analysis predicts that, following an initial duplication, the ACTA1 and ACTC1 copies separate into distinct clades except in zebrafish where actc1a, acta1b, actc1a, and actc1b all form a clade together with ACTA1 (S3 Fig). Although this suggests that the zebrafish actc1 genes may have evolved to become more similar to ACTA1 than to their respective orthologue, the low bootstrapping values reflect a very high level of nucleotide sequence conservation between all genes ranging from 85–90% identity.
To analyze Actc1b function in zebrafish we examined the phenotype of actc1b mutants (actc1bsa12367, referred to hereafter as actc1b-/-). This mutant was generated using ENU mutagenesis and the presence of a nonsense mutation at amino acid 5 was verified by competitive allele specific PCR (KASP) genotyping [17]. We verified the mutation in the actc1b-/- mutant strain and confirmed that the MO binding sites were intact (S4 Fig). Surprisingly, we found no difference in the appearance of the muscle fibers between actc1b-/- embryos and their wildtype siblings (Fig 2A) and only a small, but significant, reduction in swimming capabilities of actc1b-/- embryos compared to wildtype siblings (actc1b+/+ mean 1.0±0.1SEM and actc1b-/- mean 0.77±0.12SEM; p 0.0128; Fig 2B). Whilst actc1b-/- fish show only a 23% reduction in distance swum, in line with our previous work, Actc1b ex2 morphants show a dramatic reduction in swimming performance compared to Standard Control morphants, UTR MO morphants, and uninjected controls (median values Actc1b ex2 MO 0.23, UTR MO 0.89, Standard Control MO 0.98, and uninjected 1.03; p<0.0001 for ex2 MO against all other conditions; Fig 2C).
To determine whether the phenotypic differences observed were due to MO off-target effects we injected the Actc1b ex2, UTR MO, or Standard Control MO into an incross of actc1b+/- zebrafish and assessed phenotypic severity in their offspring in three independent experiments. The phenotypes were classed as either wildtype, mild (slight disruption to the muscle fibers), or severe (large disruption to the muscle fibers and Actinin2 aggregates at the myosepta) (Fig 3A). Remarkably, we found that actc1b-/- mutants never displayed a severe phenotype when injected with either the Actc1b ex2 MO (21 wildtype, 25 mild, 0 severe) or Actc1b UTR MO (14 wildtype, 13 mild, 0 severe) similar to the Standard Control MO (16 wildtype, 9 mild, 0 severe). However, severe phenotypes were observed in actc1b+/+ and actc1b+/- siblings injected with either Actc1b MO (actc1b+/+ injected with the Actc1b ex2 MO: 0 wildtype, 2 mild, 39 severe or Actc1b UTR MO: 0 wildtype, 59 mild, 17 severe and for actc1b+/- injected with the Actc1b ex2 MO: 0 wildtype, 66 mild, 29 severe or Actc1b UTR MO: 25 wildtype, 55 mild, 1 severe) which were not observed following injection with the Standard Control MO (for actc1b+/+: 29 wildtype, 0 mild, 0 severe and for actc1b+/-: 35 wildtype, 0 mild, 0 severe) (Fig 3B). The change in proportions of phenotype classes was significant for both morpholinos into actc1b+/+ and actc1b+/- compared to control MO (p<0.0001, Chi-square test). The insensitivity of the actc1b-/- mutants to Actc1b MO knockdown demonstrates that the severe phenotypes are not due to off-target effects.
In addition, we measured the locomotion of actc1b-/- mutants and their wildtype siblings injected with either the Actc1b ex2, Actc1b UTR, or Standard Control MO at 6 dpf. We observed a significant reduction in distance travelled by actc1b+/+ and actc1b+/- siblings injected with an Actc1b ex2 MO (median 0.14 for actc1b+/+, p<0.0001; and 0.19 for actc1b+/-, p<0.0001) and for actc1b+/+ siblings injected with an Actc1b UTR MO (median 0.83, p0.046) compared to the Standard Control MO (1.0 for actc1b+/+ and 0.95 for actc1b+/-) (Fig 3C). In contrast, actc1b-/- mutants injected with either an Actc1b ex2 (median 0.90) or an Actc1b UTR MO (median 0.95) show comparable locomotion to those injected with a Standard Control MO (median 0.83) confirming that actc1b-/- mutants are indeed unaffected by Actc1b knockdown (Fig 3C).
To determine why there was not a loss of function phenotype in actc1b-/- mutants we first measured α-actin levels in the skeletal muscle of actc1b-/- mutants, wildtype siblings, and Actc1b morphants by western blot analysis. As previously observed (Sztal et al, 2015), Actc1b morphants display decreased total actin levels compared to siblings injected with a control morpholino (Fig 4A and S5 Fig). Quantification and analysis of western blots by two-way ANOVA shows that actc1b+/+ and actc1b+/- siblings injected with either an Actc1b UTR MO (mean 0.94 SD 0.19 for actc1b+/+ (p0.0048) and mean 1.15 ±0.18SD for actc1b+/- (p0.0077)) or Actc1b ex2 MO (mean 0.51 ±0.03SD for actc1b+/+ (p<0.0001) and mean 0.72 ±0.23SD for actc1b+/- (p<0.0001)) show significantly reduced α-actin compared to Control MO injected actc1b+/+ and actc1b+/- siblings (mean 1.51 ±0.27SD for actc1b+/+and mean 1.63 ±0.09SD for actc1b+/-) (Fig 4B and S5 Fig). However, actc1b-/- mutants injected with either an Actc1b UTR MO (mean 1.23 ±0.25SD) or Actc1b ex2 MO (mean 0.77 ±0.09SD), show no decrease in α-actin levels compared to or Standard Control MO (mean 1.08 ±0.11SD; p0.5779 and p0.0955 respectively) explaining the lack of phenotype compared to actc1b+/+ and actc1b+/- morphants. The analysis also identified a significant interaction between genotype and MO treatment (p0.0088) confirming a genotype specific effect of the morpholinos.
Given the normal levels of actin in actc1b-/- fish we hypothesized that other α-actin paralogues were compensating for the loss of Actc1b. To determine the expression of the four α-actin paralogues we assayed the skeletal muscle of actc1b+/+, actc1b+/-, and actc1b-/- embryos using qRT-PCR. We found a significant decrease in RNA levels of actc1b in Actc1b morphants (UTR MO mean 1457 ±70SD, ex2 MO mean 1160 ±237SD, compared to Standard Control MO mean 5418 ±505SD, p<0.0001 for both) as previously shown in Sztal et al (2015) as well as in actc1b-/- mutants (mean 563 ±142SD compared to actc1b+/+ mean 2243 ±848SD, p0.0172), suggesting the actc1b gene product is degraded by nonsense mediated decay (Fig 4C & 4D). Interestingly, actc1a expression was significantly increased in actc1b-/- mutants compared to actc1b+/+ siblings (mean 15334 ±742SD and mean 374 ±80SD respectively, p0.0346, Fig 4C). Conversely, Actc1b morphants did not display significant changes in the expression of any of the other α-actin isoforms compared to siblings injected with a Standard Control MO (Fig 4D).
To confirm that the compensatory upregulation of actc1a was responsible for the milder phenotype in actc1b-/- mutants, we reasoned that if Actc1a was reduced in actc1b-/- mutants, they would display a more severe phenotype, comparable to Actc1b morphants. We used a MO targeting the splice donor site of exon 2 (Actc1a ex2 MO) to knockdown Actc1a. To determine a subphenotypic dose for the Actc1a ex2 MO we injected 0.5, 1.0, or 2.0ng of MO into wildtype embryos and performed both RT-PCR and western analyses to determine MO efficiency. We were able to detect a small decrease in α-actin by western blot and RT-PCR analyses revealed two amplicons, a smaller product observed in all of the samples including controls which was reduced in the MO injected embryos, and a larger band, only observed in 1.0 and 2.0ng injected samples (S6A Fig). We sequenced the larger band and confirmed that it corresponds to the inclusion of intron 2, resulting from mis-splicing of exon 2 and 3, causing the addition of three amino acids and a stop codon which would undoubtedly disrupt Actc1a function. Mutations in actc1a have been shown to cause heart defects resulting in decreased cardiac contractility and altered blood flow [18]. Although the skeletal muscle appeared unaffected by MO injections (S6E Fig), we observed a slightly dilated heart in embryos injected with a 1ng MO concentration which became more severe as the MO dose increased (S6D Fig). Based on these observations, we selected a 1ng MO dose to use in further experiments.
We then injected the Actc1a ex2 MO (or corresponding dose of a Standard Control MO) into an incross of actc1b+/- zebrafish and assessed phenotypic severity in their offspring in three independent experiments. The phenotypes were classed as either wildtype, mild (slight disruption to the muscle fibers), or severe (large disruption to the muscle fibers and Actinin2 aggregates at the myosepta) (Fig 5A). In actc1b+/+ and actc1b+/- siblings injected with either the Standard Control (actc1b+/+: 17 wildtype, 0 mild, 0 severe and actc1b+/-: 55 wildtype, 2 mild, 0 severe) or Actc1a MO (actc1b+/+: 28 wildtype, 0 mild, 0 severe and actc1b+/-: 59 wildtype, 0 mild, 0 severe) we only observed a wildtype or mild phenotype. However, when we injected the Actc1a MO into actc1b-/- mutants we observed a severe phenotype in approximately 40% of fish (3 wildtype, 15 mild, 11 severe) which was not observed in Standard Control MO injected actc1b-/- mutants (4 wildtype, 14 mild, 0 severe, change in phenotype proportions p0.0106 Chi-square test). Taken together these results demonstrate that upregulation of the actc1a paralogue is protective in actc1b-/- mutants.
We have identified compensation triggered by a mutation in actc1b but not following morpholino mediated knockdown of Actc1b. There has been considerable debate recently as a result of phenotypic differences between mutant lines and morpholino-mediated knockdown [19–22]. The study by Rossi et al (2015) provided an example where this difference in phenotype was due to compensation, rather than morpholino off target effects as previously suggested. The data presented in the current study identifies another example of the phenomenon of genetic compensation, and, rather than compensation in egf17 mutants being an isolated case, suggests that this process may be more widespread. It also suggests that in some cases, rather than the phenotypic differences between mutant and knockdown animals being due to off-target or non-specific effects, genetic compensation may influence the phenotypic penetrance of deleterious mutations. Morpholino knockdown may therefore potentially reveal the phenotype resulting from reduction of the protein, without any compensatory transcription upregulation of paralogues or alternative pathways [23].
In zebrafish actc1a is expressed in the skeletal muscle during early embryogenesis, but is downregulated in the skeletal muscle as development proceeds. However, upregulation of actc1a is sufficient to compensate for the loss of Actc1b mimicking the upregulation of ACTC1 in patients suffering from recessive nemaline myopathy caused by mutations in ACTA1 [7]. In this situation, patients have a complete absence of skeletal α-actin but instead cardiac α-actin is upregulated leading to a milder disease phenotype than patients with dominant mutations in ACTA1 [7,15]. The levels of cardiac α-actin in these patients determines the clinical severity of the disease [7].
In contrast to the compensation we have identified, the specific response in the skeletal muscle in individuals with recessive ACTA1 mutations is not sufficient to prevent disease and the majority of individuals die within 6 months of birth [7]. Transgenic expression of ACTC1 in the skeletal muscle is, however, able to rescue both recessive ACTA1-/- and dominant ACTA1D286G mutations in mice [13,16], consistent with our findings that upregulation of actin paralogues can prevent a disease phenotype. The absence of a similar compensatory mechanism resulting from reduction of Actc1b following morpholino antisense mediated knockdown suggests that compensation is not induced by the reduction in Actc1b, but at a step prior to protein formation.
While the mechanisms of genetic compensation remain unclear, two different models have been recently proposed suggesting that compensation may be activated through either a DNA damage response or by the degradation of mutant RNA and subsequent activation of common microRNAs or ribosomal binding proteins to stabilize compensatory interactions [24]. In the case of actc1b, we have shown that there are no alternative transcripts produced in actc1b-/- mutants and that the resulting mRNA is likely to be non-functional and degraded by nonsense mediated decay, which may activate compensation. However, we cannot rule out the possibility that it is either the DNA lesion itself, or the presence of the mutant mRNA that may trigger compensation. Recent studies have shown that a missense mutation in actc1a (actc1as434) causes severe defects in cardiac contractility and altered blood flow resulting from the loss of polymerized cardiac actin [18]. Injection of an Actc1a MO mimicked the heart edema and lack of endocardial cushion formation observed in actc1a mutants [18] suggesting that compensation does not play a role in actc1as434 mutants. However, it may be that if an alternative mutation, such as a nonsense mutation, is introduced or nonsense mediated pathways are activated that compensatory α-actin paralogues responses would be induced. Nevertheless, identifying the specific factors that trigger the compensatory upregulation of cardiac α-actin in the skeletal muscle tissues is the next challenge and could have therapeutic applications to ameliorate ACTA1 skeletal muscle diseases.
Genetic robustness against null mutations appears to be a universal phenomenon in all organisms, however, the mechanisms determining compensation may differ. The existence of duplicate gene copies to compensate for the loss of an essential gene has been previously observed in yeast [4] and worms [2] with genome-wide deletion experiments revealing a significantly lower percentage of duplicates compared to singletons are essential for viability and fertility. Inherited mutations have also been shown to have variable consequences in different individuals [25], which may be due to a plasticity of genetic compensatory responses masking the phenotypic effect of deleterious alleles. Our study verifies the existence of compensatory mechanisms, leading to a milder phenotype in ACTA1 recessive myopathy. More importantly, we suggest that similar compensatory responses may underline phenotypic differences in disease penetrance in the human condition.
Fish maintenance and handling was carried out as per the standard operating procedures approved by the Monash Animal Services Animal Ethics Committee under breeding colony license MARP/2015/004/BC. Fish were anaesthetized using Tricaine methanesulfonate.
Zebrafish were maintained according to standard protocols [26]. The Actc1b ex2 (5’ TGCAGTGTTTTTTTCACCTGGTGAC 3’) Actc1b UTR (5’ GGTCAAGTTGTTATCACAAGACTGA 3’), Actc1a ex2 (5’ TACATGCTTTAGAAGCCCACCTGGT 3’) and Standard Control (5’ CCTCTTACCTCAGTTACAATTTATA 3’) MOs (GeneTools) were diluted in distilled water and co-injected with Cascade Blue labeled dextran (Molecular Probes) into one- to two-cell embryos MO concentrations were calibrated according to [27] at the indicated amounts (2.0 ng for the Actc1b ex2 and UTR MOs corresponding to a concentration of 0.5mM; 0.5, 1.0, or 2.0ng for the Actc1a ex2 MO corresponding to concentrations of 0.125mM, 0.25mM, and 0.5mM; and 1.0 or 2.0ng for the Standard Control MO corresponding to concentrations of 0.25mM and 0.5mM). At 1 dpf, the embryos were sorted for Cascade Blue labeling. The actc1b mutant line (sa12367) was obtained from the Zebrafish International Resource Centre [17]. Allele specific PCR KASP technology (Geneworks) was used for genotyping.
Whole-mount in situ hybridization was carried out as described previously [28]. Probes were constructed using specific gene primers (acta1a F: CAACATCCTATCATTGCCTCCT and R: CATGTTCAGTTTTATTTGTCTGTTGA; acta1b F: ATTCATCGGCTGCATCTGTC and R: TTAACACATATGCGTCACAAAAA; actc1a F: CCAGCACAATGAAGATCAAG and R: CCAGCACAATGAAGATCAAG; actc1b F: TGACCGTATGCAGAAGGAGAT and R: TCTTATCACTTATCTGTTT). Imaging was performed with an Olympus SZX16 stereomicroscope.
Total RNA was extracted using TRIzol reagent (Invitrogen Life Technologies). RNA samples were treated with RQ1 RNase-free DNase (Promega). cDNA was synthesized from 1μg of each RNA sample in a 20ml reaction using Protoscript first strand cDNA synthesis kit (New England Biosciences) and oligo(dT)20 primer following the supplier’s instructions. Quantitative PCR was performed on a Roche Lightcycler instrument and normalized against β-actin and RPS18 [29] as reference genes. Primers for quantitative PCR are listed in S3 Table.
The human actin protein sequences were used as a query against representative databases from mouse, chicken and zebrafish genomes using a BLASTp search. Corresponding orthologues to all six actin isoforms were identified and aligned using ClustalX [30]. After using the MEGA 5.05 [31] program to determine the best–fit model for the analysis, a neighbor-joining tree (JTT, bootstrapping = 1000) was compiled using MEGA 5.05 [31]. A yeast ACT1 (Genbank ID: 850504) protein sequence was used as an outgroup. To analyze ACTA1 and ACTC1 synteny, orthologous genes were identified in Ensembl (http://asia.ensembl.org/index.html) and the flanking genomic regions were annotated.
Locomotion assays were performed on 6 dpf zebrafish as per [32]. An inactivity threshold of 6 mm/s, detection threshold of 25 mm/s and maximum burst threshold of 30 mm/s were used. The total distance swum above the inactivity threshold and below maximum burst threshold in a 10-min period were extracted using the ZebraLab software (ViewPoint Life Sciences). Blinding of treatments groups was used in combination with randomization of experimental replicates to remove any bias. Once the testing and genotyping was completed the treatments groups were uncovered.
Immunofluorescence was performed on 2 dpf zebrafish as per [14] using an anti-Actinin2 antibody (Sigma clone A7811, 1:200), AlexaFluorTM-labelled-596 secondary antibody (Molecular Probes, 1:200) and rhodamine-tagged phalloidin (Molecular Probes, 1:200). For phenotypic experiments, samples were blinded during analyses and genotypes were revealed one all samples were scored. For western blot assays, the head and tails were separated from 2 dpf zebrafish from each condition for three independent biological replicates. The heads were used for genotyping and 20 tails per condition were used for protein lysates as per [33] and quantified using the Qubit fluorometric quantification (Thermo Fisher Scientific). 10–20μg of each sample, along with reducing agent (Life Technologies) and protein loading dye (Life Technologies), was heated at 70°C for 10 min and separated by SDS-PAGE on NuPAGE 4–12% Bis-Tris gels. The protein was transferred onto PVDF membrane (Millipore), following which, the membrane was blocked with 5% skimmed milk in PBST and subsequently probed with anti-Actin (Sigma, A2066, 1/1000), washed and incubated with HRP-conjugated secondary antibody (Southern Biotech, 1:10 000). Immunoblots were developed using ECL prime (GE healthcare) and imaged using a chemiluminescence detector (Vilber Lourmat). The membrane was subsequently stripped, reprobed with anti-β-tubulin antibody (Abcam, ab6046, 1/5000), incubated in HRP-conjugated secondary antibody and developed as above. The membrane was subsequently stripped and stained with Direct Blue 71 (Sigma) to identify total protein. Fiji was used to quantify protein intensity.
For swimming analyses, all values were normalized to the average actc1b+/+ siblings injected with a control MO, or actc1b+/+ siblings. Normality of data was determined using a D’Agostino and Pearson test for normality. Normal data (Fig 2B) was analysed by one-way ANOVA using Dunnett’s correction for multiple comparisons. For data failing the normality test (Fig 2C and Fig 3C), the test was repeated after the outliers were removed by the ROUT method (Q = 1%) or the data was logtransformed. In neither case did this result in a normal distribution of data. Therefore, in these cases the data from the three replicates was pooled and a Kruskal-Wallis test was performed and correction for multiple comparisons conducted using Dunn’s test. For phenotypic analyses (Fig 3B and Fig 5), the results of the three replicates were used to determine the mean percentage of each phenotype and to plot the graphs. The proportion of the phenotypes was determined by pooling the data from all three replicates and conducting a Chi-square test for each treatment against its respective control. For qRT-PCR data (Fig 4C and 4D) a one-way ANOVA was conducted for each gene comparing actc1b+/- and actc1b-/- to actc1b+/+ or UTR MO and ex2 MO to Standard Control MO using Dunnett’s test for multiple comparisons. All statistical analyses were conducted using GraphPad Prism 7.
|
10.1371/journal.ppat.1006026 | Transcriptomic Analysis Implicates the p53 Signaling Pathway in the Establishment of HIV-1 Latency in Central Memory CD4 T Cells in an In Vitro Model | The search for an HIV-1 cure has been greatly hindered by the presence of a viral reservoir that persists despite antiretroviral therapy (ART). Studies of HIV-1 latency in vivo are also complicated by the low proportion of latently infected cells in HIV-1 infected individuals. A number of models of HIV-1 latency have been developed to examine the signaling pathways and viral determinants of latency and reactivation. A primary cell model of HIV-1 latency, which incorporates the generation of primary central memory CD4 T cells (TCM), full-length virus infection (HIVNL4-3) and ART to suppress virus replication, was used to investigate the establishment of HIV latency using RNA-Seq. Initially, an investigation of host and viral gene expression in the resting and activated states of this model indicated that the resting condition was reflective of a latent state. Then, a comparison of the host transcriptome between the uninfected and latently infected conditions of this model identified 826 differentially expressed genes, many of which were related to p53 signaling. Inhibition of the transcriptional activity of p53 by pifithrin-α during HIV-1 infection reduced the ability of HIV-1 to be reactivated from its latent state by an unknown mechanism. In conclusion, this model may be used to screen latency reversing agents utilized in shock and kill approaches to cure HIV, to search for cellular markers of latency, and to understand the mechanisms by which HIV-1 establishes latency.
| The major hindrance to an HIV cure is the ability of the virus to persist in a latent state despite antiretroviral therapy. It is difficult to study this latent state in the HIV-infected patient because only a small proportion of cells in the body are affected and current technologies are not able to identify these cells. Therefore, models in the laboratory have been developed to study HIV latency. However, these models have not been adequately characterized with the latest genomic technologies. We have characterized our model of HIV latency using global gene expression analysis (i.e., RNA-Seq). Our model aims to reflect HIV latency in patients by using primary central memory CD4 T cells, wild type virus, and antiretroviral therapy. Our main finding was that signaling through the p53 protein characterized the latent state, and may be important in its establishment. This has implications for a better understanding of HIV latency which may lead to new therapies. In a broader context, we validated the latent state of our model of HIV latency, which can now be used with confidence to evaluate compounds used in strategies to cure HIV, search for markers of HIV latency, and further investigate the mechanisms leading to the establishment of latency.
| A major obstacle to the eradication of HIV-1 is the persistence of the latent viral reservoir. While antiretroviral therapy (ART) has been extremely effective at suppressing viral replication, it has not eradicated this reservoir [1]. Upon the removal of ART, HIV-1 emerges from the latent state and replicates to levels akin to an acute infection that leads to disease progression [2,3]. The low frequency of latently infected cells within the HIV-1 patient (between 1 and 60 in 106 resting CD4 T cells) complicates the study of this viral reservoir in vivo [4,5,6]. This has prompted the development of models of HIV-1 latency based on chronically infected cell lines and primary human CD4 T cells [7]. To obtain an accurate representation of HIV-1 latency in vivo, it is essential to fully characterize these different models.
Transcriptome profiling by microarrays or RNA-Seq allows the simultaneous evaluation of transcriptional activity of the entire genome within a sample, thus providing a comprehensive analysis of the biological condition of the cell population at a given time. These technologies are becoming important for the study of HIV-1 latency, particularly in the search for biomarkers of HIV-1 latency [8,9] and for evaluating the effects of latency reversing agents [10,11,12,13]. Krishnan and Zeichner [14], utilized cDNA spotted microarrays to compare gene expression in cell lines chronically infected with HIV-1 (i.e., ACH-2, J1.1, U1) and their parental uninfected lines to identify 32 genes that were consistently differentially regulated. The laboratory of Fabio Romerio used Agilent microarrays to profile latently infected and uninfected conditions from four donors using their primary CD4 T cell latency model, and a gene encoding a surface receptor, CD2, was identified to be enriched in latently infected cells [9].
RNA-Seq is the current state-of-the-art technology with respect to transcriptomics and is thought to exhibit greater specificity and dynamic range than microarrays [15]. The first RNA-Seq study of a primary CD4 T cell model of latency incorporated a GFP expressing virus [16]. When samples from a single donor were profiled over time, a large number of genes were identified as dysregulated during the latent phase (N = 227) and were associated with chemokine receptors, cytokine signaling, and general immune responses. We previously used RNA-Seq to profile latently infected and uninfected samples from 4 donors [17] from the first iteration of a cultured primary central memory CD4 T cell (TCM) model [18,19]. This study demonstrated that the defective vectors used to seed the latent reservoir were recombining to reconstitute actively replicating HIV-1. This observation led to the revision of the cultured TCM model of HIV-1 latency by incorporating wild type virus (HIVNL4-3) and ART to suppress virus replication [20].
The purpose of the present RNA-Seq study was primarily to identify mechanisms involved in the establishment of HIV-1 latency but also to characterize this modified cultured TCM model of latency [20] to confirm that the model reflects a latent state. Comparison of latently infected to uninfected cells identified differential expression of genes in the p53 signaling pathway. Treatment with pifithrin-α, an inhibitor of the transcriptional activity of p53 [21], reduced the ability of latently infected cells to be reactivated in the cultured TCM model. The main finding from these results suggests a direct effect of p53 on the establishment and ability to reactivate the latent reservoir.
To dissect the viral and cellular status of the cultured TCM model of latency [20], gene expression profiles were generated by RNA-Seq for a total of 16 samples from 4 different donors representing the following 4 conditions: uninfected (UI), latently infected (LI), uninfected and activated (UIA), and latently infected and activated (LIA). Briefly, naïve CD4 T cells from four HIV-1 negative individuals were isolated and activated with αCD3/αCD28 beads for Three days in conditions that generate central memory CD4 T cells (Fig 1A) [18]. After activation, cells were allowed to expand in the presence of IL-2 and infected by spinoculation at day 7 with HIV-1NL4-3 at a low MOI that rendered 3–7% of cells infected at day 10 (Fig 1B, day 10). Cells were then seeded for an additional three-day period in a 96 well (round bottom) plate to increase the efficiency of virus transmission (Fig 1B, day 13). At that time point, cells were diluted and cultured for 4 extra days in the presence of IL-2 and ART (Raltegravir plus Nelfinavir). At day 17, CD4 positive T cells were isolated using magnetic bead sorting. This strategy was chosen because productively infected cells downregulate CD4 expression on the cell surface due to the expression of the accessory genes Nef and Vpu (Fig 1B, day 17) [22,23]. This procedure largely eliminates productively infected (p24 positive) cells, as well as CD4 negative cells present in the culture. Therefore, cells in the UI and LI conditions were CD4 positive cells that were not expressing detectable levels of viral antigens. UI and LI cells were further activated with αCD3/αCD28 beads in the presence of ART for 48 hours to generate cells for the UIA and LIA conditions.
Initially, the resting and activated conditions were compared (i.e., UI vs. UIA and LI vs. LIA) to validate the phenotype of the resting cells. When comparing RNA content between these conditions, transcriptional amplification had occurred (S1 Fig). Traditional normalization procedures for transcriptomic data do not account for transcriptional amplification [24,25]. Therefore, Biological Scaling Normalization (BSN) using ERCC spike-in control RNAs was used to normalize RNA-Seq data to allow comparison of resting and activated conditions [26] (See Materials and Methods). A number of gene expression markers of CD4 T cell activation were modulated following activation in both the LIA and UIA conditions (Fig 2A). Notably, IL2 and components of its receptor (IL2RA, IL2RB, IL2RG) were upregulated upon activation, as were members of the NFκB complex (NFKB1, NFKB2, REL, RELA, RELB), and CD28 itself. The KLF2 gene, which is highly upregulated in quiescent CD4 T cell lymphocytes, but repressed during activation [27,28], was significantly downregulated as expected. The modulation of IL2 and KLF2 upon T cell activation was further confirmed by RT-qPCR (Fig 2B). In summary, resting cultured TCM cells have the phenotypic characteristics of a quiescent T cell and stimulation with αCD3/αCD28 beads modulates known markers of CD4 T cell activation.
Next, the effect of antigen stimulation mimicked by αCD3/αCD28 beads on HIV-1 transcription was evaluated. HIV-1 transcription in the resting (LI) and activated (LIA) states was compared after BSN. Treatment with αCD3/αCD28 beads induced global upregulation of total HIV-1 reads from the resting to the activated conditions (average 6.6 fold change, s.d. ±3.6, t-test p-value = 0.04) with an increase in all major splicing groups: unspliced (US), singly spliced (SS), and a significant increase (p = 0.015) in multiply spliced (MS) (Fig 3A). A significant increase of US, MS and total polyadenylated HIV-1 transcripts upon activation was confirmed by RT-qPCR (Fig 3B) with a concomitant increase in HIV-1 p24 protein (Fig 3C). The fold change increase in polyadenylated transcripts appears much more variable than US or MS transcripts (Fig 3B). It is unclear what is driving this variation but measurements of polyadenylated transcripts reflect fully elongated and correctly processed HIV-1 mRNA, which relies on the host transcriptional machinery. It is possible that the efficiency of polyadenylation varies across donors, whereas US and MS are less variable because they measure HIV transcripts, whether polyadenylated or not. In support of this, single nucleotide polymorphisms that vary between donors and effect post-transcriptional processing and subsequent gene expression have previously been identified [29]. In summary, examination of HIV-1 US, MS and polyadenylated transcripts further confirmed that the LIA condition of the cultured TCM model reflected activation of transcription from an HIV-1 latent state (LI).
The UI and LI samples were compared to identify 826 differentially expressed genes (DEGs, S1 Table), 275 of which were downregulated and 551 upregulated (false discovery rate [FDR] corrected p-value < 0.05). The top 100 DEGs presented in the heatmap (S2 Fig) demonstrated the consistency of gene expression across donors. Although a large signal of differentially expressed genes was identified between the UI and LI conditions, it should be noted that this signal may be confounded by low proportions of latently infected cells and bystander effects. Specifically, a small proportion of latently infected cells in a background of uninfected cells is being compared to a population of 100% uninfected cells. The possible impact of this potential limitation is fully expanded upon in the Discussion section of this manuscript. The DEGs identified in the analysis were compared to other primary CD4 T cell [9,16] and cell line models of latency [14]. Although no up- or downregulated DEGs were identified in common across all models, (Fig 4A & 4B), greater overlap was identified when comparing primary cell models in a pairwise fashion. In particular, there were 65 genes upregulated during latency in common between our cultured TCM model [20] and the model used by Iglesias-Ussel et al. [9] (S2 Table). Unfortunately, the upregulation of CD2 during latent infection previously identified by Iglesias-Ussel et al. [9] was not confirmed in our TCM model. One explanation to this result is that in the model used by Iglesias-Ussel et al. [9], the cells isolated to perform the analysis were expressing intracellular p24. Our cultured TCM model of HIV-1 latency largely eliminates such cells.
In order to develop a better understanding of genes perturbed during latency in the cultured TCM model, the 826 genes differentially expressed between the LI and UI conditions were subjected to pathway and protein interaction (PIN) analysis. The only Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway that attained significance for over-representation of DEGs was the "p53 signaling pathway" (FDR corrected p-value = 5.5E-06). The results of differential gene expression were overlaid on the p53 signaling pathway and revealed that multiple threads related to apoptosis and DNA repair and damage prevention were upregulated in latently infected cells [30] (Fig 5). Of note, a number of p53 related genes (ACTA2, BBC3, DDB2, DRAM1, FDXR, GADD45A, and TNFRSF10B) were present in the 65 upregulated genes (S2 Table) in common between our study and that of Iglesias-Ussel et al. [9].
A PIN constructed using genes differentially expressed during latency complemented KEGG pathway analysis by confirming the importance of genes related to p53 activity (Fig 6A and Table 1). This PIN contained two major hubs (AR and MDM2) and one minor hub (TNFRSF10B a.k.a. DR5 or TRAIL-R2), which are all related to p53 activity. For example, MDM2 facilitates negative feedback to the p53 signaling pathway by degrading p53 and the upregulation of this gene in the LI condition may be indicative of prior p53 activity [31]. AR is negatively regulated by p53 signaling [32], and correspondingly, is downregulated in the LI condition. TNFRSF10B, a gene involved in programmed cell death, is upregulated and known to be induced by p53 in response to DNA damaging agents [33]. Therefore, the PIN extended KEGG pathway analysis by identifying p53 related hub genes (e.g. AR) and their targets that were not curated into the KEGG p53 signaling pathway.
Several genes (BBC3, FAS, GADD45A, HEXIM1, MDM2, TNFRSF10B and TP53I3) selected from the p53 signaling pathway and the PIN were subjected to RT-qPCR analysis (Fig 6B). The upregulation of transcripts of BBC3, GADD45A, MDM2, TP53I3, and downregulation of HEXIM1 during latent infection was confirmed as significant by RT-qPCR. The direction of fold change was validated by RT-qPCR for the cell surface markers FAS and TNFRSF10B (Fig 6B), which were previously noted as significantly upregulated in the RNA-Seq data (Fig 6C). The cell surface expression of these markers was interrogated in an independent donor set by flow cytometry (Fig 6D) but only FAS (CD95) was confirmed as significantly upregulated on the surface of T cells (fold change of 1.47 ± 0.20, Fig 6E). In summary, functional analysis of genes perturbed during latency using pathway and PIN analyses identified dysregulation of genes associated with p53 activity.
Pifithrin-α, an inhibitor of p53 transcriptional activity [21], was added to CD4 T cells from five additional donors after the initial infection at day 10 and again at day 13 in the cultured TCM model to investigate the possible role of p53 signaling in HIV-1 latency (Fig 7A). Inhibition of p53 had no effect on HIV-1 replication since no difference was observed in the levels of p24 between treated and untreated samples at days 13 and 17 (Fig 7B and 7C). However, inhibition of p53 transcriptional activity using pifithrin-α resulted in a reduction of the number of cells producing p24 (average reduction of 32%, s.d. ±8%) after reactivation with αCD3/αCD28 beads (Fig 7B and 7D). Interestingly, when pifithrin-α was added only during the reactivation step (day 17), there was no effect on viral reactivation (Fig 7E). These results suggest that inhibition of p53 during the active replication phase of HIV-1 may have an effect on the establishment or maintenance of HIV-1 latency. To test this hypothesis, integrated HIV-1 was analyzed by Alu-PCR in LI cells treated or not treated with pifithrin-α. There was a trend towards the reduction of integrated HIV-1 in pifithrin-α treated samples (Fig 7F, p = 0.056). However, the reduction on integration (average reduction of 9%, s.d. ±5%) does not fully explain the reduction in p24 observed after reactivation with αCD3/αCD28 beads (Fig 7D, average reduction of 32%, s.d. ±8%). Therefore, it is possible that in addition to reducing the number of integrated copies of HIV that pifithrin-α may be further inactivating the integrated provirus, albeit through an unknown mechanism. To support this, integrated copies of the provirus were correlated with the percentage of p24 producing cells following reactivation independently for pifithrin-α treated and untreated cells (Fig 7G). The correlation lines in this plot are parallel and shifted to the left for pifithrin-α treatment suggesting that the integrated virus in cells treated with pifithrin-α may be less prone to reactivation with αCD3/αCD28 beads. To compare both populations, we calculated the reactivation index measured as the percentage of cells expressing p24 after αCD3/αCD28 reactivation divided by the number of integrated copies before reactivation. This index compares the ability of an integrated copy to be reactivated by αCD3/αCD28 beads. Interestingly, cells that were treated with pifithrin-α have a lower ability to reactivate latent HIV-1 (Fig 7H). In summary, these studies confirmed that p53 signaling may play an important role in the establishment and maintenance of HIV-1 latency in this cultured TCM model.
In this study, RNA-Seq was utilized to characterize a cultured TCM model of HIV-1 latency [20], which incorporates a replication-competent virus (HIV-1NL4-3) and ART to suppress HIV-1 replication. RNA-Seq analysis demonstrated that the resting condition in this model (LI) reflects a quiescent and a latent state when compared to the activated state (LIA) both at the level of host and virus transcription. Notably, an increase in HIV-1 transcription was observed in all donors after activation (Fig 3A), indicative of a departure from a latent state [49].
The result of subtracting MS from US reads in the LI condition (mean difference 4,134, s.d. ±5448) was significantly different (p = 0.03, paired t-test) than this subtraction in the LIA condition (mean difference -10,928, s.d. ±3,408). This demonstrates a significant shift from US to MS reads upon activation suggesting an increase in early HIV-1 transcripts (Vpr, Tat, Rev, and Nef) after activation. The greater numbers of US versus MS reads in the LI condition of the TCM model is supported by previous reports that consistently detect a greater signal for US over MS transcripts in the resting CD4 T cells isolated from HIV-1 infected patients on ART [50,51]. The increase in HIV-1 transcription was not solely due to abortive transcripts since polyadenylated viral transcripts, which reflect fully elongated and correctly processed HIV-1 mRNA, were significantly upregulated after activation.
By comparing expression profiles from UI to those from LI cells, a total of 826 differentially expressed genes were identified (S2 Table). While there was only minimal overlap in genes dysregulated during latency when comparing this cultured TCM model to published data from three additional models of HIV-1 latency (Fig 4A & 4B), there appeared to be greater overlap between primary CD4 T cell models compared to models based on cell lines (S2 Table). Several reasons may account for these differences. First, this overlap might have been greater if the same transcriptomic technologies had been utilized, e.g. Iglesias-Ussel et al. [9] used the Agilent-012391 Whole Human Genome Oligo Microarray (G4112A) to profile gene expression compared to RNA-Seq in this study. Second, the RNA-Seq study of Mohammadi et al. [16] analyzed samples from only a single donor. Greater overlap between primary HIV-1 latency models may occur when better powered transcriptomic studies (i.e., multiple donors) are performed using the state-of-the-art technology (i.e., RNA-Seq).
Functional analysis of DEGs identified between the LI and UI conditions clearly implicated the transcriptional activation of genes by p53 (Figs 5 and 6, and Table 1). For example, the genes ACTA2, BBC3 (a.k.a. Puma), DDB2, DRAM1, FAS, FDXR, GADD45A, PRDM1, RHOC, TNFRSF10B, and TP53I3 are known to be induced by p53 [33,36,39,43,45,48,52,53,54], and were upregulated in the LI condition, which is in agreement with previous studies showing the activation of the p53 pathway mediated by HIV-1 through type I IFN signaling [55,56,57,58]. Of these genes, FAS (CD95) was further confirmed at the protein level (Fig 6E). In addition to upregulated p53 related genes, activation of p53 by genotoxic stress has been shown to result in downregulation of AR expression [32], also downregulated in this TCM model (Fig 6A). The p53 protein itself is highly regulated and its activity must be tightly controlled to allow normal cellular functioning [59]. To maintain homeostasis, p53 will not only activate genes that enhance and stabilize its activity but also genes that repress its activity through negative feedback loops. Further evidence of prior p53 signaling activity was demonstrated by the upregulation of genes which act to degrade p53 following activation (Table 1). For instance, the protein product of MDM2 mediates ubiquitination and breakdown of p53 resulting in the inhibition of p53-mediated apoptosis [31], and was upregulated in the LI condition (fold change = 1.503). Several other genes also involved in the degradation or inhibition of p53 [31,34,35,44,60,61], were upregulated in the LI condition: BCL3, CCNG1, HEXIM1, LIF, NR4A1, and PTK2. In summary, the latently infected (LI) condition in the TCM model of HIV-1 latency exhibited evidence of prior p53 activation and subsequent negative feedback of this signaling pathway.
The identification of p53 related genes modulated during HIV-1 infection led to the hypothesis that this pathway may be important for the establishment of latency. Experiments with the p53 inhibitor pifithrin-α demonstrated that inhibition of this pathway did not affect viral replication or cell death, but did limit the number of cells that could be reactivated from latency (S3 Fig, Fig 7C, 7D and 7E). A number of explanations can account for these results. First, Alu-PCR analysis suggested that p53 may be required for successful integration of HIV-1 (Fig 7F). The p53 protein is not only involved in apoptosis and cell cycle arrest, but also in the activation of DNA repair mechanisms [62]. A number of p53-responsive genes identified in our study (BBC3, DDB2, GADD45A, FDXR, PCNA and XPC) are related to radiation-induced DNA damage [63] and point towards involvement of the nucleotide excision repair (NER) pathway, a process that recognizes and removes helix-distorting DNA lesions from the genome [64,65]. For example, GADD45A binds to UV-damaged DNA where it is believed to modify DNA accessibility within chromatin [66]. Although prior studies have primarily implicated other DNA repair pathways such as non-homologous end joining and base excision repair in HIV-1 DNA integration [67,68], the results of the present study suggest the p53-responsive genes that are components of the NER pathway may also play a role in the establishment of latency. Interestingly, a previous siRNA screen to characterize DNA repair factors involved in HIV-1 integration demonstrated that siRNA knockdown of DDB2, a component of the NER pathway, resulted in a large reduction (71.3%) of HIV-1 integration [67].
A second hypothesis to explain the effects observed by pifithrin-α could be the silencing of the HIV-1 provirus. It should be noted, that despite approaching significance and in the same direction for each donor, the reduction in HIV integration by pifithrin-α is modest (Fig 7F), and does not correspond entirely to the magnitude of the difference in p24 following reactivation (Fig 7D). Therefore, in addition to reducing the establishment of latency, pifithrin-α may be inactivating the integrated provirus, albeit through unknown mechanisms. Such mechanisms might include a shift in integration sites, inactivated provirus, and/or silencing through alterations in DNA methylation or histone modification at the HIV promoter. To support this, correlating the number of integrated copies of the provirus with the percentage of p24 producing cells following reactivation independently for pifithrin-α treated and untreated cells demonstrates parallel lines but shifted to the left for pifithrin-α treated cells (Fig 7G). Furthermore, there was a significant drop in the reactivation index between pifithrin-α treated and untreated cells (Fig 7H). These data suggest that pifithrin-α may induce changes to the integrated provirus, either directly or at the level of the epigenome, resulting in less efficient reactivation with αCD3/αCD28 beads. This is not without precedent since it has been demonstrated that HIV DNA synthesis by the virus reverse transcriptase is more accurate in the presence of p53, which has exonucleolytic proofreading capabilities [69]. Therefore, inhibition of pifithrin-α in the TCM model of HIV-1 latency may lead to more error prone HIV DNA synthesis during the expansion phase resulting in greater numbers of dysfunctional provirus. Further studies of this pathway will be needed to completely understand the role of p53 in the establishment of HIV-1 latency in cultured TCM.
The importance of p53 signaling may not be confined to only the TCM model analyzed here as several genes related to the p53 pathway were also identified in the overlap with the primary CD4 T cell model examined by Iglesias-Ussel and colleagues [9]. Specifically, the p53 related genes ACTA2, BBC3, DDB2, DRAM1, FDXR, GADD45A, and TNFRSF10B were significantly upregulated in both models (S2 Table). It will be interesting to determine if inhibiting the p53 pathway in other models of HIV latency also affects the establishment of latency. This may suggest common mechanisms involved across primary CD4 T cell models. Finally, the contribution of p53 to the establishment of latency in vivo needs to be evaluated. Interestingly, Castedo and colleagues [70] have previously shown that activation of p53 can be detected in HIV-1 infected patients in both peripheral blood mononuclear cells (PBMCs) as well lymph nodes. Moreover, the authors demonstrated that p53 activation correlates with viral load. These results suggest that p53 may also play a role in the establishment of latency in vivo.
The present study allowed the comparison of gene expression between the LI and UI conditions of this TCM model of HIV-1 latency, which may represent potential biomarkers of latency. However, the search for these biomarkers in this study was somewhat limited by the relatively low proportion of latently infected cells in the LI condition (mean 2.92%, s.d. ±0.71%) that were in a large background of uninfected cells. Therefore, a gene must be upregulated greater than 30-fold in individual latently infected cells in order for it to be identified as being upregulated by 2-fold when comparing the LI and UI conditions. Furthermore, bystander effects could be contributing to the signal of differential gene expression between the LI and UI conditions, whereby signals emanating from previously infected cells (e.g., cytokines and chemokines) may be perturbing gene expression in uninfected cells in the LI condition. Similarly, it is plausible that these results may be confounded by the triggering of innate immune pathways in TCM cells exposed but not latently infected by HIV particles. However, robust triggering of innate immune responses is unlikely, since only 11 interferon stimulated genes with known antiviral properties [71] were found to be differentially expressed between the LI and UI conditions, out of a total of 826 genes, and only 2 of these 11 genes were in the top 100 DEGs (S2 Fig). Therefore, it appears that TCM cells in this model were cultured for a sufficient period of time following virus exposure to allow innate immune responses to recede. Regardless of these limitations, gene expression markers of HIV-1 latency undoubtedly exist within the 826 DEGs identified between the UI and LI conditions (S1 Table) and will need to be verified in future work.
In summary, the primary finding from this RNA-Seq analysis of the cultured TCM model [20] was that p53 signaling may be involved in the establishment of HIV latency. Furthermore, the RNA-Seq data was used to demonstrate that this model was truly reflective of a latent state and thus suitable for screening latency reversing agents for shock and kill approaches to an HIV-1 cure [72,73], further investigating mechanisms associated with establishing a latent state, and identifying gene expression biomarkers of HIV-1 latency. It would be beneficial to subject all primary CD4 T cell models of HIV-1 latency [7] to RNA-Seq analysis in statistically powered studies (i.e. >3 donors) so that similarities and differences across models may be dissected. Finding similar genes across models will lead to the identification of gene expression biomarkers of HIV-1 latency that may be used to isolate latently infected cells from HIV-1 infected subjects and utilized in innovative cure strategies (e.g., radioimmunotherapy [74]) or killing latently infected cells by means of immunotoxins [75]. The profiling of other HIV-1 latency models on the omics scale will lead to the validation or modification of these models, which will undoubtedly result in a better understanding of both in vivo latency and the therapies that can be used to facilitate a cure for HIV.
Nelfinavir was obtained through the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH; raltegravir from Merck & Company, Inc.; human IL-2 from Dr. Maurice Gately, Hoffmann-La Roche Inc. [76]; HIV-1NL4-3 from Dr. Malcolm Martin [77]. Pifithrin-α, p-Nitro, Cyclic [78,79], a more potent analogue of pifithrin-α with a longer half-life, was obtained from Santa Cruz Biotechnology.
Sample preparation and generation of infected TCM cells was fully described previously [20]. Briefly, PBMCs were isolated from HIV-1 negative individuals or obtained from the Gulf Coast Regional Blood Center (Houston, TX). Naïve CD4 T cells were isolated and then activated using human αCD3/αCD28-coated magnetic beads (one bead per cell, Thermo Fisher Scientific, Cat. #11131D) in the presence of αIL-4, (1 μg/mL; Peprotech; Cat. No. 500-P24) αIL-12 (2 μg/mL; Peprotech, Cat No. 500-P154G), and tumor growth factor (TGF)-β1 (10 μg/mL; Peprotech; Cat. No. 100–21) for 3 days at a density of 500,000 cells/ml in 96 well round bottom plates. Cells were expanded in medium containing human IL-2 (30 IU/ml) for additional 4 days. IL-2 and media were replaced at day 4 and day 5. At day 7, cells were infected (or mock-infected) with HIV-1NL4-3 by spinoculation at 2900 rpm at 37C for 2 hours at a multiplicity of infection of 0.1. After infection, cells were further cultured in medium containing IL-2 for 3 days, subjected to crowding in round bottom plates in the presence of IL-2 for another 3 days, and then cultured for a further 4 days in the presence of IL-2 and ART in a cultured flask (nelfinavir, 0.5 μM; raltegravir, 1.0 μM). Every time that media and IL-2 were replaced, cells were kept at a density of 106 cell/ml. At day 17, any remaining productively infected cells were removed by magnetic isolation of CD4+ cells using the Dynabeads CD4 Positive Isolation Kit (Thermo Fisher Scientific, Cat. No. #11551D) following the manufacturer’s instructions with a minor change, i.e. 75 μl per 107 cells was used instead of 25 μl to increase the recovery of CD4 positive cells. CD4 beads were removed from the cells following the manufacture instructions. At this stage, samples were taken for the latently infected (LI) condition. Uninfected (UI) cells were cultured under the same conditions and collected at the same time as LI cells. Additional cell aliquots were subjected to reactivation with αCD3/αCD28-coated magnetic beads in the presence of ART for 2 days for the uninfected activated (UIA) and latently infected activated (LIA) conditions.
To address whether p53 transcriptional activation plays a role in the establishment of latency, 7.5 μM pifithrin-α was added 3 days post infection during the crowding stage (day 10), and then washed out with replenishment at day 13 and thus maintained in culture until just prior to reactivation (day 17) when it was washed out. Production of p24 was then assessed at day 19 after two days of reactivation with αCD3/αCD28-coated magnetic beads. In a separate experiment, to confirm that pifithrin-α was not affecting the reactivation process, pifithrin-α was added only during the reactivation step (day 17 to 19) and p24 production similarly assessed at day 19.
Total RNA was extracted from 16 TCM cell samples from 4 donors for the 4 conditions (UI, LI, UIA and LIA) using the RNeasy Plus Mini Kit (QIAGEN, Cat. No. 74134) according to manufacturer’s instructions, with the addition of an on-column DNase treatment (RNase-free DNase Set, QIAGEN, Cat. No. 79254). RNA integrity (RIN) values of samples were on average 9.9 (s.d. ±0.1) as determined using a Bioanalyzer 2100 (Agilent Technologies, CA, USA) and RNA concentration was measured by Nanodrop 2000 (Thermo Fisher Scientific). To account for transcriptional amplification, synthetic RNA standards from the External RNA Controls Consortium (ERCC RNA Spike-In Mix 1, Ambion, CA, USA) were spiked into total RNA isolated from each sample. TCM cells in each sample were counted in quadruplicate and ERCC spike-ins were added at 1 μl of 1:100 dilution per million cells. RNA-Seq libraries were prepared from 100ng of total RNA using the TruSeq Stranded Total RNA Library Prep kit (Illumina, CA, USA) after depletion of cytoplasmic and mitochondrial ribosomal RNA with Ribo-Zero Gold (Epicentre, WI, USA). All libraries were sequenced to a read depth of >75 million reads using the Illumina HiSeq2000 to generate 50bp paired-end reads (100 bp total read length).
FASTQ files for each sample were mapped to the human genome (hg38) using Tophat (version 2.0.13) [80] and counted against the human GENCODE [81] annotation (v21) with HTSeq [82]. FASTQ files are available through GEO accession number (GSE81810). All reads were then aligned to the HIVNL4-3 genome (Genbank accession number AF324493) using TopHat and counted using HTSeq [82]. Levels of HIV-1 US, SS, and MS transcripts were estimated by the method of Mohammadi and colleagues [16] which counts reads that pass through the two major HIV-1 splice sites D1 and D4. Finally, all reads were mapped to the 92 ERCCs, using Bowtie (version 1.1.1) [83] and then counted against individual ERCCs using HTSeq. When identifying differences in host and HIV-1 gene expression between resting (UI and LI) and activated (UIA and LIA) conditions, BSN was utilized to account for transcriptional activation [26]. Briefly, the expression of ERCC spike-in controls were used to estimate a scaling factor between resting and activated conditions which was then used to adjust the expression levels of host genes and HIV-1 reads. When identifying differences in host gene expression between the UI and LI conditions, RUVSeq [84] was used to normalize the data since transcriptional amplification was not an issue in this comparison. Following normalization, differentially expressed genes were identified with EdgeR [85] (FDR corrected p-value < 0.05). Please refer to (S1 Supplemental Methods) for more details.
Pathway analysis was performed with ToppGene [86] using the functional analysis enrichment tool, ToppFun, with the KEGG pathways selected. Pathway images were generated from the KEGG Pathway Database [30]. Fold change data were log2 transformed, colored, and overlaid upon the p53 signaling pathway. A protein interaction network (PIN) was generated with MetaCore and visualized through Cytoscape v2.8.3 [87]. MetaCore draws connections between the protein products of differentially expressed genes if they have protein-protein or protein-DNA interaction confirmed from the literature record. The advantage of this approach is that the PIN often reveals biological associations that have not been curated in KEGG pathways. For PIN construction, genes were filtered using a log2 fold change of 0.5 between latently infected and uninfected cells. Read pileup figures were generated with the Integrated Genome Browser [88]. Venn diagrams were constructed using Venny 2.1.0 [89].
RT-qPCR validation of the expression of host and virus genes identified by RNA-Seq was performed using TaqMan Gene Expression Assays (Thermo Fisher Scientific) as previously described [90,91,92]. Changes in host and virus gene expression were calculated using the 2-ΔΔCT method with the spike-in ERCC control ERCC_00130, as the normalizer. Please refer to (S1 Supplemental Methods) for more details on RT-qPCR analysis.
For the detection of surface FAS (CD95) expression, cells were stained with FAS/CD95 Antibody (DX2), FITC conjugate (Molecular Probes). For the detection of surface TNFRS10B (DR5) expression, cells were stained with CD262 (DR5, TRAIL-R2) Antibody (DJR2-4 (7–8)), APC conjugate (Biolegend). Cells were also stained with a viability dye (Fixable Viability Dye eFluor 450, Affymetrix, eBioscience, San Diego, CA). For the dual detection of CD4 and HIV-1 p24 Gag, cells were first stained with the viability dye (Fixable Viability Dye eFluor 450), followed by staining with CD4 antibody (S3.5), APC conjugate (Molecular Probes). After staining, cells were fixed, permeabilized, and stained for HIV-1 p24 Gag as previously described [20]. In all experiments, CD4 positive HIV-1 p24 Gag negative staining regions were set with uninfected cells treated in parallel. Flow cytometry was performed with a BD FacsCanto II flow cytometer using FACSDiva acquisition software (Becton Dickinson, Mountain View, CA). Data were analyzed with Flow Jo (TreeStar Inc, Ashland, OR).
DNA from 2x106 cells was isolated using DNeasy Blood and Tissue Kit (Qiagen). DNA was quantified using NanoDrop 1000 (Thermo Fisher Scientific). Genomic DNA was subjected to nested quantitative Alu-LTR PCR for integrated provirus as previously described [93], with modifications. For the first reaction, 250 ng of total DNA was amplified using Platinum Taq DNA polymerase (Invitrogen). Reactions were carried with 1.5 mM of MgCl2, 200 μM dNTPs, 400 nM of Alu164 primer (5’-TCCCAGCTACTCGGGGAGGCTGAGG-3’) and 400 nM of PBS primer (5’-TTTCAAGTCCCTGTTCGGGCGCCA-3’). Amplifications were performed in a MultiGene Optimax (Labnet International, Inc) with the following parameters: 1) 94C 5 min; 2) 18x 94C 30 sec, 66C 30 sec, 72C 5 min; 3) 72C 10 min. PCR samples were subject to a 1/10 dilution in water, then 2 μl of the diluted sample was subject to qPCR reactions in a LightCycler 480 (Roche) using PCR Master Mix (2X) (Thermo Fisher Scientific). Final concentration of primers (AE989-2 5’-CTCTGGCTAACTAGGGAACCCAC-3’; AE990-2 5’-CTGACTAAAAGGGTCTGAGGGATCTC-3’) and probe (5’-FAM-TTAAGCCTCAATAAAGCTTGCCTTGAGTGC-BHQ1-3’) were 400 nM and 200 nM respectively. A serial dilution of pcDNA3.1-LTR was used for a molecular standard curve. pcDNA3.1-LTR was generated by cloning the 3’ LTR from NL43 into pcDNA3.1.
PBMCs were isolated from HIV-1 negative individuals following IRB-approved protocol no. #67637 (University of Utah) or unidentified source leukocytes designed “for research only” were purchased from the Gulf Coast Regional Blood Center (Houston, Texas). All HIV-1 negative individuals provided written informed consent.
|
10.1371/journal.pcbi.1002054 | A Role for Both Conformational Selection and Induced Fit in Ligand Binding by the LAO Protein | Molecular recognition is determined by the structure and dynamics of both a protein and its ligand, but it is difficult to directly assess the role of each of these players. In this study, we use Markov State Models (MSMs) built from atomistic simulations to elucidate the mechanism by which the Lysine-, Arginine-, Ornithine-binding (LAO) protein binds to its ligand. We show that our model can predict the bound state, binding free energy, and association rate with reasonable accuracy and then use the model to dissect the binding mechanism. In the past, this binding event has often been assumed to occur via an induced fit mechanism because the protein's binding site is completely closed in the bound state, making it impossible for the ligand to enter the binding site after the protein has adopted the closed conformation. More complex mechanisms have also been hypothesized, but these have remained controversial. Here, we are able to directly observe roles for both the conformational selection and induced fit mechanisms in LAO binding. First, the LAO protein tends to form a partially closed encounter complex via conformational selection (that is, the apo protein can sample this state), though the induced fit mechanism can also play a role here. Then, interactions with the ligand can induce a transition to the bound state. Based on these results, we propose that MSMs built from atomistic simulations may be a powerful way of dissecting ligand-binding mechanisms and may eventually facilitate a deeper understanding of allostery as well as the prediction of new protein-ligand interactions, an important step in drug discovery.
| Protein-ligand interactions are crucial to chemistry, biology and medicine. Many studies have been conducted to probe the mechanism of protein-ligand binding, leading to the development of the induced fit and conformational selection models. Unfortunately, experimentally probing the atomistic details of protein-ligand binding mechanisms is challenging. Computer simulations have the potential to provide a detailed picture of molecular recognition events. In this study, we construct kinetic network models from atomistic simulations to elucidate the mechanism by which the LAO protein binds to its ligand. Because the LAO protein completely encompasses its substrate in the bound state, it has generally been assumed that it operates via an induced fit mechanism. We find that both the conformational selection and induced fit mechanisms play important roles in LAO binding. Furthermore, we have identified a number of parallel pathways for binding, all of which pass through a single gatekeeper state, which we refer to as the encounter complex state because the protein is partially closed and only weakly interacting with its substrate.
| Molecular recognition plays important roles in many biological processes. For example, enzymes must recognize their substrates and drugs must be designed to have specific binding partners. Unfortunately, our understanding of how ligand binding occurs remains incomplete. In particular, the role that protein dynamics play in protein-ligand binding is unclear.
Two popular models for protein-ligand binding are the induced fit and conformational selection mechanisms. Both attempt to explain how a protein could transition from an unbound conformation to a bound conformation in complex with a ligand. In the induced fit model—introduced by Koshland [1]—the ligand first binds to the protein in its unbound conformation and this binding event induces the protein to transition to the bound state. Such models have been applied to many protein-protein and protein-DNA/RNA binding systems [2], [3], [4]. The conformational selection (or population shift) model [5], [6], [7], [8], [9], [10], [11], [12] is a popular alternative to the induced fit mechanism. In this model, the intrinsic dynamics of the protein lead it to constantly transition between a stable unbound conformation and a less stable bound conformation. The ligand can then bind directly to the bound conformation, thereby stabilizing the bound state and increasing its population relative to the unbound state. The conformational selection model has recently gained popularity in antibody or small ligand binding systems [10], . Some docking studies have also tried to exploit conformational selection by generating an ensemble of protein structures and docking small molecules against each of them in the hopes of identifying a transiently populated bound conformation that will be stabilized by the ligand [13].
Many recent studies have attempted to determine whether a variety of systems under different conditions can be best described by the induced fit or conformational selection model [14], [15], [16], [17], [18], [19], [20], [21]. For example, Okazaki et. al. [20] have found that strong and long range protein-ligand interactions favor the induced fit model, while weak and short range interactions favor the conformational selection model. Based on an analytic model, Zhou has suggested that the determining factor in ligand binding is the timescale for transitioning between the unbound and bound states with and without the ligand [18]. He found that conformational selection dominates when transitioning between the unbound and bound states is slow, while the induced fit mechanism dominates when this transition is fast. Many studies have proposed that conformational selection and induced fit are not mutually exclusive; instead, a blend of these two models may best describe most realistic systems [15], [17], [18], [20], [22], [23]. For example, Zagrovic and coworkers [17] have suggested that conformational selection and induced fit play equal roles in ubiquitin binding based on their analysis of NMR structures. However, in many cases, it is still difficult to dissect the chemical details of binding mechanisms. While it is clear that the bound and unbound states of a protein and their respective interactions with a ligand molecule are of great importance [15], [18], [20], [21], it may also be important to take other conformational states into account. Protein dynamics are ultimately determined by their underlying free energy landscapes, whose ruggedness frequently gives rise to numerous metastable regions-sets of rapidly mixing conformations that tend to persist for extended periods of time.
In this work, we use Markov state models (MSMs) to map out the relevant conformational states in LAO binding and describe mechanistic details of this process. MSMs are a kinetic network model and a powerful approach to automatically identifying metastable states and calculating their equilibrium thermodynamics and kinetics [24], [25], [26], [27]. MSMs focus on metastable regions of phase space, while there also exist other kinetic network models to study transition state [28]. MSMs partition conformational space into a number of metastable states; such that intra-state transitions are fast but inter-state transitions are slow. This separation of timescales ensures an MSM is Markovian (i.e. that the probability of transitioning from state i to state j depends only on the identity of i and not any previously visited state) and allows MSMs built from short simulations to model long timescale events. Many recent studies have demonstrated how MSMs can provide insight into drastic conformational changes like protein and RNA folding [26], [29], [30], [31], [32]. Here we demonstrate that MSMs built with a hierarchical clustering algorithm [30] can capture the mechanism by which the Lysine-, Arginine-, Ornithine-binding (LAO) protein, one of Periplasmic Binding Proteins (PBPs), binds to arginine. The LAO protein has a high binding affinity and undergoes large-scale domain rearrangements from an open to a closed state upon ligand binding [33], [34], [35], [36] (see Fig. 1), making it a valuable model system for probing the coupling between protein conformational changes and binding.
Many have assumed that PBP binding occurs via the induced fit mechanism because the ligand is completely encapsulated by the protein in the bound state (see Fig. S1). Experimental studies of many PBPs support the induced fit mechanism, where the closure of the domains is triggered by the binding of the ligand [33], [34], [36], [37], [38], [39], [40]. However, a few experimental studies indicate that some PBPs (including GGBP [41] and ChoX [42]) are able to reach the closed conformation in the absence of the ligand. This has been suggested as a sign of the conformational selection mechanism [38]. Furthermore, recent NMR studies with paramagnetic relaxation enhancement (PRE) of maltose-binding protein (MBP) identified a minor (∼5%) un-liganded partially closed form. This partially closed state is in equilibrium with the open state and, therefore, is available for the binding of the ligand, which may further facilitate the transition to the bound state. This work suggests a more complex binding mechanism where both conformational selection and induced fit play significant roles [43], but since the ligand was not present during the experiments, it is unclear exactly what roles the two mechanisms may play.
With our MSM, we can directly monitor the mechanism of LAO binding and assess the role of both conformational selection and induced fit. Our model suggests that three dominant states need to be considered to adequately describe LAO binding and that both conformational selection and induced fit play important roles in the transitions between these states. The third dominant state in our model—besides the open and closed states—is only partially closed and weakly bound to the ligand; therefore, we refer to it as the encounter complex state. The ligand can induce the protein to transition from the open state to the encounter complex; however, the ligand-free protein can also transition to the encounter complex state, indicating an important role for conformational selection. In contrast, on our dataset the ligand-free protein never sampled the closed state, this suggest that that the closed state in the absence of the ligand may represent a very high free energy state and that once the ligand reaches the binding site an induced fit mechanism is responsible for transitions from the encounter complex to the closed state.
Before drawing system-specific conclusions from a simulation study, it is important to first test the model against existing experimental data. MSMs built using the Super-level-set Hierarchical Clustering (SHC) algorithm [30] greatly facilitate this task by decomposing a system's conformational space into its constituent metastable regions and describing the thermodynamics and kinetics of each. For instance, one can easily extract representative conformations from each state, determine the equilibrium probability of each state, or calculate the rates of transitioning between sets of states and compare to experimental results. In this study, we describe protein conformations by the opening and twisting angles between their two domains [44] and the location of the ligand because these degrees of freedom describe the slowest dynamics of the system (see Fig. S2). We then construct a 54-state MSM using SHC (See Methods for details of MSM construction). The dominant conformational states in our model are displayed in the Fig. S3.
Fig. 2 demonstrates that our model is capable of ab initio prediction of the bound state. As described in the Methods section, no knowledge of the bound state was included at any stage of our simulations or model construction. Based on the high binding affinity measured in experiments (Kd ∼14 nM) [45] we postulated that the bound state should be the most populated state in our model. Indeed, representative conformations from our most populated state (having an equilibrium population of 74.9%) agree well with the crystal structure of the bound state, with an RMSD to the crystal structure of the binding site as little as 1.2 Å, as shown in Fig. 2A. Moreover, Figs. 2B and 2C show that the crystal structure of the bound state lies within the minimum of the most populated free energy basin. These figures also show that our model's bound state covers a relatively large region of phase space, suggesting that it is flexible, possibly to accommodate favorable interactions with all four of LAOs binding partners (L-lysine, L-arginine, L-ornithine and L-histidine). The structural properties of the remaining states are also consistent with experiments (see the Text S1 for more details). For example, many of the unbound states also contain partially closed protein conformations, consistent with NMR experiments on another PBP protein: MBP [43].
Our model is also in reasonable agreement with the experimentally measured binding free energy and association rates. For example, from the MFPT from all unbound states to the bound state, our model predicts an association timescale of 0.258 ± 0.045 µs (see Methods for calculation details). Since rates are proportional to the exponential of the free energy barrier, an 8-fold difference in rates roughly corresponds to a 2 kT difference in the height of the free energy barrier. Therefore, our result is in reasonable agreement with the experimental value of ∼2.0 µs found in the highly homologous HisJ protein [46] (see Methods for similarity between LAO and HisJ protein). We also estimate a binding free energy of −8.46 kcal/mol using the algorithm introduced by van Gunsteren and co-workers [47], which is also in reasonable agreement with the experimental value of −9.95 kcal/mol for the LAO protein (see Methods for calculation details). Together, this agreement between theory and experiment suggests that our model is a sufficiently good reflection of reality to make hypotheses about details of the binding mechanism.
Arriving at these conclusions with a single long simulation would have been quite difficult due to the slow timescales involved. For example, transitioning from the bound state to an unbound state takes 2.15±0.51 µs on average. Therefore, observing enough transitions to gather statistics on the binding and unbinding rates in a single simulation would require that it be tens of µs long. Such simulations are now possible [31], [48] but are still challenging to perform. Moreover, scaling the long simulation approach to millisecond timescales is still infeasible. MSMs built from many µs timescale simulations, however, have already proven capable of capturing events in a 10 millisecond timescale [49] and can likely scale to even slower processes.
In addition to predicting experimental parameters, MSMs are also useful for mapping out conformational transitions like protein-ligand binding. For example, Figs. 3 and 4 show the 10 highest flux pathways from any of the unbound states in our model to the bound state. All ten pathways pass through an obligatory, gatekeeper state (state 11) that we refer to as the encounter complex state because the protein is partially closed and only weakly interacting with the ligand (see Figs. 3, 4 and State 11 in Fig. S3). In the encounter complex (see Fig. 5) the two lobes of the LAO protein are structurally very similar to those in both the apo and bound X-ray structures (with RMSD less than 2 Å, see Table S1). Therefore the conformational change between crystal structures and the encounter complex could be achieved through a rigid body rotation. We also found that in the encounter complex the ligand was stacked between the lobe I Tyr14 and Phe52 and protrudes upward to interact with the lobe II Thr121. These contacts are also observed in the X-ray bound structure. (see Fig. S4). To further support our conclusion that the encounter complex state is an obligatory step in ligand binding, we have calculated that the average timescale for transitioning from the unbound states to the encounter complex state is 0.190±0.037 while the average timescale for transitioning from the unbound states to the bound state is 0.258 ± 0.045 µs. The average timescale for transitioning from the encounter complex state to the bound state is 0.090±0.015 µs (see Methods for calculation details). Thus, the unbound protein will typically transition to the encounter complex before reaching the bound state.
The top ten paths from the unbound states to the encounter complex can be divided into two sets, one that is best described by conformational selection and one that is better described by the induced fit mechanism. For example, the pathway from state 45 directly to 11 operates through conformational selection (see green arrow in Fig. 4): in the unbound state 45 the protein and ligand are not interacting but the protein conformations are very similar to those in the encounter complex. Since the protein adopts similar conformations in these two states, the ligand can always bind to a pre-existing encounter-complex-like (state 11 like) protein conformation (the conformational selection mechanism). The binding kinetics of this conformational selection pathway is quite rapid, having a mean first passage time for transitioning from the unbound state 45 to the encounter complex state 11 of 0.220±0.054 µs, and this pathway accounts for ∼45% of the flux of the top ten pathways from unbound states to the bound state.
The second group of pathways to the encounter complex, which together account for ∼55% of the flux may be better described by the induced fit mechanism. In general, these pathways start off in conformations that are much more open or twisted than the encounter complex. Next, the system transitions to one or more intermediate states where the ligand is interacting with the protein at (or near) its binding site, though the protein is still quite open or twisted. Finally, the protein-ligand interactions induce a transition to the encounter complex state. For example, the pathway starting from state 47, passing through state 14, and ending at state 11 falls into this category (see Fig. 4).
Transitions from the encounter complex to the bound state are best described by the induced fit mechanism. When the system enters the encounter complex state, the protein is generally in a relatively open conformation (opening angle within 20° to 70°, see Fig. 6). However, when the system leaves the encounter complex state to enter the bound state, the protein is mostly in a more closed conformation (opening angle smaller than 30°, see Fig. 6). Thus, it appears interactions with the ligand induce the protein to close. Furthermore, our model predicts that the encounter complex-to-bound transition (0.090±0.015 µs) is much faster than the encounter complex-to-unbound transition (1.927±0.499 µs), so the encounter complex is not likely to diffuse back to the unbound state instead of converting into the bound state. In addition, the protein never samples fully-closed conformations in the absence of the ligand in our simulations. Together, these observations indicate that the induced fit mechanism should dominate transitions from the encounter complex to the bound state.
We have also identified a number of metastable mis-bound states, like states 4 and 20 in Fig. S3. In these states, the ligand interacts with the protein outside of the binding site. For example, in state 20 (population ∼0.4%) the ligand is bound to the hinge region between the two domains of the protein. Transitioning from a mis-bound state to the bound state generally requires passing through an unbound state (see Fig. S5). Therefore, these states are mostly off pathway and likely slow down the overall binding kinetics.
The LAO protein is a member of the PBP family, which is responsible for transporting low molecular weight ligands from the outer to the inner membrane in the ABC transport mechanism of Gram-negative bacteria [34], [50]. Crystal structures for this system have shown that the binding site is completely closed-off in the bound state [33], [37], making it impossible for the ligand to enter the binding site after the protein has adopted the closed conformation (see Fig. S1). Therefore, it has often been proposed that LAO binding occurs through an induced fit mechanism [51], [52], [53].
Our observations of LAO binding indicate that, like protein folding, ligand binding is a multi-state process with parallel pathways. All top 10 pathways pass through a gate-keeper state that we refer to as the encounter complex state because the protein is partially closed and only weakly interacting with the ligand. The system can reach this state through either the induced fit mechanism or conformational selection. Rather than being a transient state, this encounter complex is quite metastable. Once in the encounter complex state, the ligand is able to quickly induce protein conformational changes that lead to a transition to the fully closed, bound state.
Other systems may operate through a similar mechanism such as other PBPs. Indeed, our model is consistent with Tang et. al.[43]'s findings for MBP. Specifically, they discovered a partially closed state in equilibrium with the open state for the apo protein. Thus, this state is available for the binding of the ligand to form the encounter complex through the conformational selection mechanism. Next, the binding could facilitate the transition to the bound state via the induced fit mechanism. Their model was proposed mainly based on NMR experiments in the absence of the ligand, but our simulations directly observed this interplay at atomic resolution. Furthermore, our results suggest that transitions from the open to the partially closed state occur via a combination of conformational selection and induced fit mechanisms. For several PBPs including MBP and HisJ, the rate constant measured by the stopped flow experiments is proportional to the ligand concentration, indicating a simple two-state binding mechanism [46]. However, as discussed by Tang et. al. [43], these stopped flow measurements may not be able to capture the intermediate encounter complex state because the overall binding timescale is extremely rapid, e.g. a few hundred nanoseconds for LAO.
More broadly, there may also exist other proteins with closed active sites that have metastable encounter complex states. Sullivan et. al. [16] suggested that in such encounter complex state substrate-enzyme interactions are almost identical to the active state, while the enzyme has not yet reached the active form. Furthermore, the enzyme must operate by an induced fit mechanism to reach the active form because of the closure of the enzyme would prevent the substrate from entering the active site. This model is consistent with our findings for the LAO protein. However, in order to reach this encounter complex state, we suggest that both induced fit and conformational selection may play important roles. In general, other proteins with closed active sites may also make use of both conformational selection and induced fit to reach the encounter complex. However, the relative contributions of these mechanisms may vary depending on factors like the relative strength of the protein-ligand interactions [20].
The ability to map out the details of LAO binding using MSMs is an important step towards a deeper understanding of protein-ligand interactions for this system. Future application of these methodologies to other systems could even lead to the identification of general principles of protein-ligand interactions and allostery. This knowledge may also greatly aid in computational drug design. For example, it may not always be possible to identify all the relevant states via other structural methods, like crystallography. Using MSMs, however, one can hope to identify the most important relevant states and design small-molecules to specifically stabilize one or more states over the others. Future work with improved force fields and greater sampling could also greatly enhance our understanding of protein-ligand interactions. However, we stress that the present work lays out the methodology that would be employed in such future research.
In this study, we demonstrate the power of MSMs for understanding protein-ligand interactions using the LAO protein as a model system. Our results indicate that LAO binding is a two-step process involving many states and parallel pathways. In the first step, the ligand binds to a partially closed protein to form an encounter complex. Both the conformational selection and induced fit mechanisms play significant roles in this step. In the second step, the system transits from the encounter complex state to the bound state via the induced fit mechanism. This two-step binding mechanism (see Fig. 7 for schematic diagram of the binding mechanism) may also be used by other systems, such as other PBP proteins, enzymes with closed active sites, and systems where the apo protein dynamics rarely visits the bound conformation. Future applications of MSMs with improved force fields, greater sampling, and to other protein-ligand interactions will reveal how general this mechanism is, aid in understanding allostery, and lay a foundation for improved drug design.
We have performed 65 molecular dynamics simulations, each 200 ns long, of the LAO protein from the organism Salmonella typhimurium and one of its ligands, L-arginine. Ten simulations were started from the open protein conformation (PDB ID: 2LAO [37]) with the ligand at more than 25 Å away from the binding site. The other simulations were initialized from conformations randomly selected from the first ten simulations. We saved conformations every 20ps with a total of more than 650,000 conformations. Among these MD simulations, we have observed multiple binding events and unbinding events (see Fig. S6 and Text S2). The protein was solvated in a water box with 11,500 SPC waters [54] and 1 Na+ ion. All the simulations were performed using the GROMACS 4.0.5 simulation package [55] with the GROMOS96 force field [56]. The simulation system was minimized using a steepest descent algorithm, followed by a 250 ps MD simulation applying a position restraint potential to the protein heavy atoms. The simulations were performed under isothermal-isobaric conditions (NPT) with P = 1 bar and T = 318 K, using Berendsen thermostat and Berendsen barostat with coupling constants of 0.1 ps−1 and 1 ps−1 respectively [57]. A cutoff of 10 Å was used for both VDW and short-range electrostatic interactions. Long-range electrostatic interactions were treated with the Particle-Mesh Ewald (PME) method [58]. Nonbonded pair-lists were updated every 10 steps. Waters were constrained using the SETTLE algorithm [59] and all protein bonds were constrained using the LINCS algorithm [60]. Hydrogen atoms were treated as virtual interaction sites, enabling us to use an integration step size of 5 fs [61].
We used MSMBuilder [27] and SHC [30] to construct the state decomposition for our MSM for LAO binding.
We first used the k-centers algorithm in MSMBuilder [27] to cluster our data into a large number of microstates. The objective of this clustering was to group together conformations that are so geometrically similar that one can reasonably assume (and later verify) that they are also kinetically similar.
Because we had to account for both the protein and ligand, we performed two independent clusterings; one based on the opening and twisting angle of the protein and one based on the relative position of the ligand (see Fig. S2). We then combined the two clusterings by treating them as independent sets. For example, M protein-based clusters and N ligand-based clusters would lead to a total of M×N clusters.
For the protein-based clustering, we created 50 clusters using the Euclidean distance between a vector containing the protein opening and twisting angles. The opening angle (see Fig. S2a) was defined as the angle between the normal vectors of the two planes defined by the center of masses of the following groups of Cα atoms:
The twisting angle (see Fig. S2b) is the angle between the following two planes:
The strong correlation between opening and twisting angles of the protein and the two slowest eigenvectors from Principle Component Analysis (PCA) analysis of a 20 ns MD simulation started from the apo structure in the absence of the ligand demonstrates that they are a reasonable descriptor of the protein's conformation (see Figs. S2c and S2d). As a reference point, we note that the holo X-ray structure (PDB ID: 1LAF) [33] has both opening and twisting angles equal zero, while the apo X-ray structure (PDB ID: 2LAO) [37] has opening angle = 38.2° and twisting angle = −26.2°.
For the ligand-based clustering, we created 5000 clusters using the Euclidean distance between all heavy-atoms.
We then had to modify our clustering to account for the fact that the ligand dynamics fall into two different regimes (see Fig. S7): one where the ligand moves slowly due to interactions with the protein and one where the ligand is freely diffusing in solution. The existing clusters are adequate for describing the first regime. However, when the ligand is freely diffusing (more than 5 Å from the protein), the procedure outlined above results in a large number of clusters with poor statistics (less than ten transitions to other states). Better sampling of these states would be a waste of computational resources as there are analytical theories for diffusing molecules and a detailed MSM would provide little new insight. Instead, we chose to re-cluster these states using the same protein coordinates and the Euclidean distance between the ligand's center of mass (as opposed to the Euclidean distance between all ligand heavy-atoms). For this stage, we created 10 new protein clusters and 100 new ligand clusters.
After dropping empty clusters, this procedure yielded 3,730 microstates. Of these, 3,290 microstates came from the initial high resolution clustering and 440 came from the data that was reclustered at low resolution. To verify that the final microstate model is valid (Markovian) we plotted the implied timescales and found that they level off at a lag time between 2 and 6 ns (see Text S3 and Fig. S8a), implying that the model is Markovian for lag times in this range. Therefore, we can conclude that the microstates are sufficiently small to guarantee that conformations in the same state are kinetically similar.
We then lumped kinetically related microstates into macrostates using the SHC algorithm [30]. This is a powerful lumping method that efficiently generates more humanly comprehensible macrostate models (i.e. ones with fewer small macrostates arising from statistical error) than the PCCA algorithm currently implemented in MSMBuilder.
In SHC, one performs spectral clustering hierarchically using super level sets (or density levels) starting from the highest density level, thereby guaranteeing that highly populated meta-stable regions are identified before less populated ones. For SHC, we selected density levels Lhigh = [0.50, 0.55, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95, 0.99] and Llow = [0.4, 0.95], for the high and low-density regions respectively. The low and high resolution states were lumped separately because the states in each set have different sizes, so it is difficult to compare their densities. We then combined these two sets of macrostates to construct an MSM with 54 macrostates. Once again, we used the implied timescales test to verify that the model is Markovian and found that a 6 ns lag time yields Markovian behavior (see Fig. S8b).
To calculate a transition matrix using the above state decomposition we first counted the number of transitions between each pair of states at some observation interval (the lag time) to generate a transition count matrix, where the entry in row x and column y gives the number of transitions observed from sate x to state y. In particular, we use a sliding window of the lag time on each 200 ns trajectory with a 20 ps interval between stored conformations (i.e. each trajectory contains 10,000 conformations) to count the transitions. Because we use a hard cutoff between states, simulations at the tops of barriers between states can quickly oscillate from one state to the other, leading to an over-estimate of the transition rate between states [62]. To mitigate the effect of these recrossing events, we only counted transitions from state x to state y if the protein remained in state y for at least 300 ps before transitioning to a new state. To generate the transition probability matrix (where the entry in row x and column y gives the probability of transitioning from state x to state y in one lag time), we normalized each row of the transition count matrix.
We followed the procedure in Ref [63] to compute the mean first passage time (MFPT) from initial state i to final state f, i.e. the average time taken to get from state i to state f. In particular, the MFPT (Xif) given that a transition from state i to j was made first is the time it took to get from state i to j plus the MFPT from state j to f. MFPT (Xif) can be defined as,(1)where tij = 6 ns is the lag time of the transition matrix T. The boundary condition is:(2)
A set of linear equations defined by Equation (1) and (2) can be solved to obtain the MFPT Xif. We used bootstrapping to put error bars on MFPTs. That is, one-hundred new data sets were created by randomly choosing trajectories 130 times with replacement. We then calculated the MFPTs for each data set and reported their means and standard deviations. In MFPT calculations, the encounter complex was considered to contain state 11 and state 5, because state 5 also has features of the encounter-complex, though it plays a significantly smaller role (refer to the Text S1 for details).
To compute the binding free energy ΔG from our simulations, we use the method introduced by van Gunsteren and co-workers [47]: (3)where αbound and αfree are the fractions of bound and free species respectively. c0 is the overall concentration of ligand. In our simulations, , where NA is Avogadro's number, and Vbox is the volume of the simulation box. In our system, c0 = 0.0049 mol/L, T = 318K. We consider all the unbound states as free species, so that αfree = 1.75%. Thus, the bound species αbound = 1−αfree = 98.25%, which contains all the states where the protein and ligand are in close contact. From Eq. (3), we have:
We can also derive the association rate. Given a system in equilibrium described by the protein P, ligand L and the protein-ligand complex P•L, the protein-ligand binding reaction can be written as:(4)
Rate equation can be written as(5)Where kon and koff are forward and backward rate constants respectively. Since the forward reaction (i.e. association) only depends on kon, the rate equation for the forward reaction can be written as:(6)The total concentrations of protein and ligand in the system are constant(7)(8)Thus only one concentration among [P], [L], and [P•L] is independent. If we choose [P] as the independent concentration and then rate equation for forward reaction can be rewritten as:(9)If we consider the condition , which is the case for our simulations:(10)We can solve Eq (10) with the initial condition at t = 0:(11)We define the association timescale () as the time when half of the protein () has associated with the ligand:(12)
Since there is no experimental kon rate constant available for the LAO protein, we choose for comparison the kon from the Histidine binding (HisJ) protein. The LAO and HisJ proteins have considerable similarity both in structure and function. For example, both proteins are the same size (238 a.a.) and have a 70% sequence identity. In fact, if conservative mutations are taken into account the sequence identity increases to 83% [64]. These homologous proteins also bind to the same membrane receptor (HisQ/HisM/HisP) [37] and the same ligands (cationic L-amino acids) [65]. The RMSD between the X-ray crystal structures of holo LAO and HisJ bound to histidine (1LAG and 1HBP) is also quite small (Cα RMSD as low as 0.62 Å). The binding affinities of these proteins to their ligands are also similar (all about a nanomolar, though the binding affinities are not exactly the same [Histidine binds to HisJ most strongly, but binds to LAO most loosely]). The similarity between LAO and HisJ has also been discussed in detail in a previous study by Oh et. al.[65]. Therefore, we think it is a reasonable assumption that the LAO and HisJ proteins have similar binding kinetics.
For the Histidine binding protein, . In our simulation, the initial concentration of protein is , thus, the association timescale:
The experimental association timescale is about eight times slower than that computed from our simulations. However, we note that the only available experimental kon was measured at 293K [46], while our simulations were performed at a higher temperature (318K) with faster kinetics. Thus, the difference between the experimental and simulation rates will be smaller at the same temperature. For the binding free energy, the experimental measurement was at an even lower temperature (277K) [45]. Thus, the experimental binding free energy at the temperature we simulated should be closer to our calculated value.
|
10.1371/journal.pntd.0007417 | Oral immunization with a probiotic cholera vaccine induces broad protective immunity against Vibrio cholerae colonization and disease in mice | Oral cholera vaccines (OCVs) are being increasingly employed, but current killed formulations generally require multiple doses and lack efficacy in young children. We recently developed a new live-attenuated OCV candidate (HaitiV) derived from a Vibrio cholerae strain isolated during the 2010 Haiti cholera epidemic. HaitiV exhibited an unexpected probiotic-like activity in infant rabbits, preventing intestinal colonization and disease by wild-type V. cholerae before the onset of adaptive immunity. However, it remained unknown whether HaitiV would behave similarly to other OCVs to stimulate adaptive immunity against V. cholerae. Here, we orally immunized adult germ-free female mice to test HaitiV’s immunogenicity. HaitiV safely and stably colonized vaccinated mice and induced known adaptive immune correlates of cholera protection within 14 days of administration. Pups born to immunized mice were protected against lethal challenges of both homologous and heterologous V. cholerae strains. Cross-fostering experiments revealed that protection was not dependent on vaccine colonization in or transmission to the pups. These findings demonstrate the protective immunogenicity of HaitiV and support its development as a new tool for limiting cholera.
| Oral cholera vaccines are increasingly used as public health tools for prevention of cholera and curtailing the spread of outbreaks. However, current killed vaccines provide minimal protection in young children, who are especially susceptible to this diarrheal disease, and require ~7–14 days between vaccination and development of protective immunity. We recently created HaitiV, a live-attenuated oral cholera vaccine candidate derived from a clinical isolate from the Haiti cholera outbreak. Unexpectedly, HaitiV protected against cholera-like illness in infant rabbits within 24 hours of administration, before the onset of adaptive immunity. However, HaitiV’s capacity to stimulate adaptive immune responses against the cholera pathogen were not investigated. Here, we report that HaitiV induces immunological correlates of protection against cholera in adult germ-free mice and leads to protection against disease in their offspring. Protection against disease was transferable through the milk of the immunized mice and was not due to transmission or colonization of HaitiV in this model. Coupling the immunogenicity data presented here with our previous observation that HaitiV can protect from cholera prior to the induction of adaptive immunity, we propose that HaitiV may provide both rapid-onset short-term protection from disease while eliciting stable and long-lasting immunity against cholera.
| The bacterial pathogen Vibrio cholerae causes the severe human diarrheal disease cholera, a potentially fatal illness characterized by rapid-onset of fluid loss and dehydration. Recent estimates place the global burden of cholera at ~3 million cases per year, and over 1.3 billion people are at risk of this disease [1]. V. cholerae proliferates in the small intestine and produces cholera toxin (CT), which leads to water and electrolyte secretion into the intestinal lumen [2].
The O1 serogroup of V. cholerae causes virtually all epidemic cholera. This serogroup includes two serotypes, Inaba and Ogawa, whose LPS structures differ by a single methyl group on the terminal O-antigen sugar [3]. Serologic and epidemiologic studies have established the existence of extensive serotype cross-reactivity and -protectivity, although immunogenicity and protection is highest to the homologous serotype [4–7]. Toxigenic O1 strains are divided into two major biotypes, classical and El Tor, but the former has not been isolated in over a decade and is thought to be extinct [8]. Ongoing evolution of El Tor V. cholerae has given rise to variant El Tor strains, which are distinguishable from earlier strains by a variety of features, including the expression of non-canonical ctxB alleles that may impact disease severity in afflicted patients [4,9,10]. These contemporary strains, such as the ctxB7-expressing V. cholerae strain responsible for the 2010 Haitian cholera epidemic, are thought to be the globally dominant cause of cholera [10–12]. Currently, serogroup O139 isolates only cause sporadic disease [13]. Notably, antibodies (or immune responses) targeting the O1 O-antigen do not protect against O139 challenge and vice versa [14–16].
Oral cholera vaccines (OCVs) have recently become widely accepted as a tool for cholera control [17]. Vaccines are a potent method to combat cholera due to their ability to both directly and indirectly reduce disease and transmission [18]. Killed multivalent whole-cell OCVs, such as Shanchol, have shown promise both to prevent disease in endemic regions and as reactive agents to limit cholera during epidemics [19]. However, killed OCVs tend to be less effective at eliciting protective immunity in young children (<5 years old), who are most susceptible to cholera [20,21]. Additionally, these vaccines typically require two doses over the span of several weeks, although recent studies suggest that a single dose may still lead to moderate protection [20,22,23].
There is no live-attenuated OCV licensed for use in cholera-endemic regions. The only clinically available live-attenuated OCV is Vaxchora (CVD103-HgR), which is derived from a classical O1 Inaba V. cholerae strain and was approved by the US FDA in 2017 for use in travelers [24]. In contrast to killed OCVs, live vaccines, such as CVD103-HgR and the El Tor-derived vaccine Peru-15, elicit more potent immune responses in young children [25,26], potentially because they more closely mimic natural infection than killed OCVs. In particular, live vaccines can produce antigens in vivo that are not expressed in the in vitro growth conditions used to prepare killed vaccines; furthermore, the inactivation processes used to formulate killed vaccines can destroy antigenic epitopes [27].
In addition to the requirement for multiple doses of some OCVs for optimal protection, all current live and killed OCVs are thought to be accompanied by a post-vaccination lag in protection during induction of anti-V. cholerae adaptive immunity. The shortest reported time to protective efficacy is 8 days post-vaccination, a delay that could hamper reactive vaccination campaigns designed to limit the spread of cholera outbreaks [28]. We recently created HaitiV, a new live-attenuated OCV candidate derived from a variant El Tor O1 Ogawa V. cholerae clinical isolate from the 2010 Haiti cholera outbreak. HaitiV harbors many genetic alterations that render it avirulent and resistant to reversion while preserving its robust capacity for colonization of the small intestine [29]. In an infant rabbit model of cholera [30], intestinal colonization with HaitiV conferred protection against lethal wild-type (WT) V. cholerae challenge within 24 hours of vaccination, a timescale inconsistent with the development of adaptive immunity and suggestive of a “probiotic”-like mechanism of protection. Here, using a mouse model of V. cholerae intestinal colonization, we show that oral administration of HaitiV to female mice elicits serum vibriocidal antibodies and protects their pups from lethal challenge with virulent V. cholerae. Thus, HaitiV has the potential to provide rapid probiotic-like protection as well as to elicit long-lasting immune protection from cholera.
All bacteria were grown in Luria-Bertani (LB) broth supplemented with the relevant chemicals at the following concentrations: streptomycin (Sm, 200μg/mL), kanamycin (Km, 200μg/mL), carbenicillin (Cb, 50μg/mL), sulfamethoxazole/trimethoprim (SXT, 80 and 16μg/mL) and 5-bromo-4-chloro-3-indolyl-β-d-galactopyranoside (X-gal, 60μg/mL). For growth on plates, LB + 1.5% agar was used. All V. cholerae strains in this study, except for PIC158 and PIC018, were spontaneous SmR derivatives of the wild-type. Bacteria were stored as -80°C stocks in LB with 35% glycerol. S1 Table lists the strains used in this study.
The CT deletion strain in the H1 V. cholerae background (HaitiWT) was generated by allelic exchange as previously described, with an additional selection step to enhance the efficiency of obtaining a stable single crossover strain [29]. Briefly, HaitiWT was conjugated with SM10λpir E. coli bearing the suicide plasmid pCVD442-ctxAB-KmR, containing sacB as well as a kanamycin resistance cassette from pKD4 sandwiched by homology arms targeting the ctxAB operon (locus tags N900_RS07040 –N900_RS07045). Single crossovers were selected on LB+Sm/Cb/Km agar plates. To select for a double crossover, verified single crossovers were grown in LB + Cb/Km for 4 hours at 37°C and then passaged in LB+10% sucrose overnight at room temperature. Sucrose-resistant (sacB-negative), KmR and CbS colonies were then conjugated with SM10λpir E. coli bearing pCVD442-ctxAB (no KmR cassette) and clean KmS double crossovers generated via an identical protocol. The ΔctxAB deletion was verified by colony PCR with internal and flanking primers.
4-week old germ-free (GF) female C57BL/6 (Massachusetts Host-Microbiome Center) or Swiss-Webster (Taconic Farms) mice were housed in a BL-2 facility for the duration of the experiment. On Day 0, 2, 4, 6, 14, 28, 42 and 56, mice were anesthetized with isoflurane and orally gavaged with 109 CFU of an overnight culture of either HaitiV or CVD103-HgR in 100μL 2.5% Na2CO3. Mice were weighed at every immunization and once every 4–5 days between boosts. At each weighing, fresh fecal pellets were plated on LB + Sm to enumerate shed bacteria. At Day 7, 14, 28 and 42 post first immunization, blood samples were obtained from each mouse by tail vein incision. A Day 1 blood sample was collected from the Swiss-Webster cohort and the single-dose C57BL/6 cohort. Blood was clotted at room temperature for 1 hour, centrifuged at 20000 x g for 5 minutes and the supernatant (serum) stored at -20°C for analysis.
Vibriocidal antibody quantification was performed by complement-mediated cell lysis using PIC018 (Inaba) or PIC158 (Ogawa) V. cholerae as the target strain as previously described [31]. Seroconversion was defined as ≥4x increase in titer over the baseline measurement. The characterized mouse monoclonal antibody 432A.1G8.G1.H12 targeting V. cholerae O1 OSP was used as a positive control for the vibriocidal assay. Titers are reported as the dilution of serum causing a 50% reduction in target optical density compared to no serum control wells.
Anti-cholera toxin B subunit (CtxB) and anti-O-specific polysaccharide (OSP) responses were measured by previously described isotype-specific ELISAs [32,33]. Briefly, 96-well plates (Nunc) were coated with 1 μg/mL solution of bovine GM1 monosialoganglioside (Sigma) in 50mM carbonate buffer overnight. Next, 1μg/mL CtxB in 0.1% BSA/PBS purified from the classical Inaba strain 569B (List Biological Laboratories) was layered onto the GM1-coated wells. Wells were blocked with a 1% BSA/PBS mixture after which 1:50 dilutions of the mouse serum samples were loaded into each well. Goat anti-mouse IgA, IgG or IgM secondaries conjugated to HRP (Southern Biotechnology) were then added at a concentration of 1μg/mL in 0.1% BSA/0.05% Tween/PBS and incubated for 90 minutes. Detection was performed by adding an ABTS/H2O2 mixture to the wells and taking an absorbance measurement at 405nm with a Vmax microplate kinetic reader (Molecular Devices Corp., Sunnyvale, CA). Plates were read for 5 min at 30 s intervals, and the maximum slope for an optical density change of 0.2 U was reported as millioptical density units per minute (mOD/min). Results were normalized using pooled control serum from mice previously immunized against V. cholerae and reported as ELISA Units as previously described [32]. Anti-OSP responses were measured and reported similarly to anti-CtxB responses, only instead of CtxB, purified OSP:BSA from either PIC018 or PIC158 (1 μg/mL) was used to coat plates as previously described [34]. Additionally, OSP ELISAs were carried out with 1:25 dilutions of the serum samples.
The infant mouse survival challenge was adapted from previous reports to optimize the dosage for HaitiWT and to include more frequent monitoring intervals [31,32]. Pregnant dams were singly housed at E18-19 for delivery. At P3 (third day of life), pups were orally inoculated with 50μL LB containing 107 CFU of a directly diluted 30°C overnight culture of V. cholerae and returned to their dam. Infected pups were monitored every 4–6 hours for onset of diarrhea and reduced body temperature. Once symptoms appeared, monitoring was increased to every 30 minutes until moribundity was reached, at which point pups were removed from the nest and euthanized by isoflurane inhalation followed by decapitation for dissection and CFU plating of the small intestine on LB + Sm/X-gal. Pups that were alive at 48 hpi were deemed protected from the challenge. Cross-fostering was performed by transferring up to half of a litter between dams on the first day of life (P1). Fostering was maintained for at least 48 hours before infection to fully replace the milk from the original dam. We excluded rejected pups from analyses due to our inability to attribute mortality to infection alone.
Statistical analyses were performed with Prism 8 (Graphpad). To analyze whether immune responses were significantly changed over time, a one-way ANOVA was performed. Due to missing values from paired measurements as a result of insufficient serum sample volumes, antibody titers were analyzed with a mixed-effects model one-way ANOVA using the earliest sample (Day 1 or Day 7) as the control and post hoc tests performed with a Dunnett’s multiple comparison test. Survival curves were analyzed with the log-rank (Mantel-Cox) test and CFU burdens were compared with the Mann Whitney U test. A p-value <0.05 was considered statistically significant.
This study was performed in accordance with the NIH Guide for Use and Care of Laboratory animals and was approved by the Brigham and Women’s Hospital IACUC (Protocol #2016N000416). Infant (P14 or younger) mice were euthanized by isoflurane inhalation followed by decapitation. When required, adult mice were euthanized by isoflurane inhalation followed by cervical dislocation.
While the infant rabbit model enables investigation of the progression of a V. cholerae-induced diarrheal disease that closely mimics human cholera [30], it is not appropriate to study vaccine immunogenicity because newborn animals lack a fully developed immune system. Instead, we used adult GF mice to study HaitiV immunogenicity. In contrast to normal adult mice, which are resistant to V. cholerae intestinal colonization, oral inoculation of GF mice with V. cholerae results in stable intestinal colonization without adverse effects [35–37]. In the GF model, serum markers of immunity, such as vibriocidal titers, can be measured, but challenge studies are not possible due to the persistent colonization of the vaccine strain and the resistance of adult mice to diarrheal disease. Here, we further developed a variation of the GF model [38]. Besides measuring serum markers in the orally vaccinated adult mice, neonatal pups (which are sensitive to V. cholerae induced diarrheal disease) born to these mice were subjected to challenge studies to evaluate vaccine protective efficacy.
We established two cohorts of orally immunized adult female GF mice. In the first cohort, a small pilot study was set up to compare the immunogenicity of HaitiV and a streptomycin-resistant derivative of CVD-103HgR. This cohort consisted of 4-week-old Swiss-Webster GF mice that were immunized with either vaccine strain (n = 3 per group). Cohort 2 consisted of a set of 4-week-old C57BL/6 mice that were all immunized with HaitiV (n = 7). We generally followed the multi-dose oral immunization scheme previously used in this model, which included eight doses of 1x109 CFU vaccine over eight weeks [36,37]. After this vaccination regimen, the mice in cohort 2 were mated and vaccine-induced protective immunity was assessed in the progeny (Fig 1A).
Based on fecal CFU, all animals in both cohorts were stably colonized with high levels of either vaccine strain (Fig 1B). No adverse effects of long-term colonization with HaitiV or CVD103-HgR were noted, and all mice gained weight over the course of the study (Fig 1B). Fecal shedding and presumably intestinal colonization of HaitiV in cohort 2 was eliminated after these dams were used to cross-foster pups born to specified-pathogen free (SPF) control mice (described below), suggesting that a normal microbiota can outcompete HaitiV.
Serum samples from the immunized mice were used to quantify antibodies targeting several V. cholerae factors thought to play roles in protection from cholera. One of these metrics, the vibriocidal antibody titer, is a validated correlate of protection in vaccinated humans [39–42]. In cohort 1, all mice immunized with HaitiV or CVD-103HgR seroconverted within 2 weeks and developed vibriocidal titers consistent with those reported in human studies for live OCVs (Fig 2) [41,43]. Furthermore, HaitiV and CVD-103HgR elicited comparable vibriocidal titers.
In cohort 2, HaitiV immunization of C57BL/6 mice also induced high vibriocidal titers to Ogawa and Inaba target strains (Fig 2C). Isotype-specific levels of antibodies targeting Ogawa and Inaba OSP, and CtxB were also measured since they also likely contribute to immunity to cholera [39]. Although we did not measure Day 1 titers in cohort 2, measurements from naïve GF C57BL/6 mice and baseline measurements from cohort 1, and Day 1 of HaitiV-inoculated C57BL/6 mice in a later cohort (S2 Fig) showed undetectable levels of vibriocidal antibodies (Figs 2A and S2). The cohort 2 mice developed strong anti-Ogawa and anti-Inaba OSP responses (Fig 3, S2 Table). The anti-Ogawa OSP titers were generally higher than those targeting Inaba OSP, likely reflecting the fact that HaitiV is an Ogawa strain. All mice in cohort 2 also developed high levels of anti-CtxB IgA, IgG and IgM antibodies (Fig 4, S1 Table). The 100% seroconversion rate and general increase over time of all three humoral immune responses measured (vibriocidal, anti-CtxB and anti-OSP antibodies) reveals that orally delivered HaitiV can elicit V. cholerae-specific immune responses.
To assess the protective efficacy of HaitiV in this model, we challenged the neonatal progeny of HaitiV-immunized or control dams with lethal doses of different wild type V. cholerae strains. This assay has been used to study passive immunity elicited by cholera vaccines, but has not been characterized in vaccinated GF mice [31,32,44]. Initially, we optimized this assay with litters from SPF C57BL/6 control mice. Three or four-day old pups were inoculated with 107 or 108 CFU of HaitiWT, the virulent strain from which HaitiV was derived, and returned to their dams for monitoring (Fig 5A). Infected pups from both groups rapidly developed signs of dehydrating diarrheal disease, including accumulation of nest material on their anogenital regions, lethargy, skin tenting and hypothermia. All infected pups died by 48 hours post inoculation (hpi), with a median time to moribundity of ~23–26 hpi (Fig 5B). At the time of death, all pups were heavily colonized, with >107 CFU/small intestine (Fig 5C), and had swollen ceca, another hallmark of productive cholera infection in mammalian models [30,45]. Since there were no significant differences in survival or bacterial loads in mice challenged with either 107 or 108 CFU, the smaller dose was used in subsequent experiments (Fig 5B). Diarrhea and death in this model were entirely dependent on CT; infant mice inoculated with 107 CFU HaitiWT ΔctxAB or HaitiV were completely healthy at 48 hpi, despite sustained intestinal colonization (Fig 5C).
We next mated HaitiV-immunized animals from cohort 2 with age-matched GF male mice, thereby preserving their colonization with HaitiV. When challenged with HaitiWT, none of the 16 pups born to HaitiV-immunized dams developed signs of diarrhea or died by 48 hpi; in stark contrast, all pups born to non-immunized dams died within ~30 hpi (Fig 6A, left). There was a marked ~5,000-fold reduction in the intestinal load of HaitiWT in pups born to immunized versus control dams (Fig 6A, right). The pups of the immunized dams remained healthy for at least 2 weeks post-challenge, even though there were still detectable but very low levels of HaitiWT in their intestinal homogenates (S1 Fig). Thus, oral immunization with HaitiV elicits an immune response that provides potent protection in nursing pups from diarrheal disease, death and V. cholerae intestinal colonization.
Pups of HaitiV-immunized dams were similarly challenged with heterologous V. cholerae strains, to test the serotype and serogroup specificity of protection engendered by oral immunization with HaitiV. The additional challenge strains included an O1 Inaba strain (N16961) that has been used as the challenge strain in several human volunteer cholera studies [43,46] and the serogroup O139 strain MO10, which was isolated during the 1992 O139 outbreak in India. Most pups from HaitiV-immunized dams were protected from N16961 V. cholerae challenge (7/10 survival at 48 hpi, Fig 6B). Despite the clinical protection, there was a much less dramatic reduction in the intestinal burden of N16961 (~20-fold) compared to that observed with HaitiWT challenge, indicating serotype-specific responses play an important role in limiting colonization. Surprisingly, pups challenged with MO10 also exhibited some protection, but there was no concomitant reduction in the intestinal burden of this O139 strain (Fig 6C). Together, these observations demonstrate that animals can exhibit protection from death despite relatively robust colonization, suggesting that protection from disease may result from immunity targeting factors such as CtxB, in addition to those that impede colonization.
Since our earlier studies indicated that HaitiV itself can mediate rapid protection against cholera independent of an adaptive immune response, it was important to investigate whether pups nursed by HaitiV immunized dams were colonized with the vaccine strain. Extensive plating of intestinal samples from the >50 pups used for survival assays (limit of detection = 50 CFU/small bowel) did not reveal any HaitiV CFU in the pups reared by HaitiV-shedding dams. Thus, vaccine strain transmission and its probiotic effects are almost certainly not the explanation for the potent protection observed in nursing pups.
Cross-fostering experiments were undertaken to investigate the likely passive nature of the protection. P1 pups born to SPF dams were transferred to and reared by HaitiV-immunized dams and then challenged 2 days later with HaitiWT (Fig 5A, between P1-P3). All pups crossed-fostered by immunized dams were protected (100% survival at 48 hpi) and nearly all had marked reductions (~1,000 fold) in their intestinal HaitiWT burdens (Fig 7A). These observations mirror the challenge studies presented above (Fig 6A), indicating passive immunity from milk accounts for the protection that HaitiV-immunized dams bestow to their progeny. Conversely, when pups born to HaitiV-immunized dams were cross-fostered by SPF (non-vaccinated) dams, all succumbed to HaitiWT challenge, albeit with an increase in median survival time (by ~6-hour) and had high HaitiWT intestinal burdens (Fig 7B). The modest extension in survival time in these mice may be due to trans-placentally derived immunity or residual milk from the HaitiV-immunized dam.
Although prior work with OCVs in GF mice suggested that multiple boosts were required to maximally induce immune responses, our observation of the prolonged colonization in multiply-vaccinated animals led us to test whether a single oral dose of HaitiV could also stimulate protective immune responses [37]. A singly-vaccinated group of female GF C57BL/6 mice (n = 4) was established to investigate this possibility. Like our studies with the multi-dose regimen, a single dose of HaitiV led to sustained colonization in the mice (S2 Fig). HaitiV induced vibriocidal antibody titers comparable in magnitude to those from serially immunized mice (Fig 2C). Litters from singly-immunized mice were also completely protected from disease resulting from HaitiWT challenge, phenocopying pups from the first cohorts (S2 Fig).
Vaccines for cholera are being increasingly embraced as public health tools for prevention of endemic cholera and limiting the spread of cholera epidemics [17]. Killed OCVs are efficacious in endemic populations, but live OCVs promise to be more potent, particularly in young children [19]. Here, we showed that the live OCV candidate HaitiV induces vibriocidal antibodies and other immunological correlates of protection against cholera in GF mice and leads to protection against disease in their offspring. Protection in this model was dependent on passively acquired factors in the milk of immunized dams and not transmission or colonization of HaitiV. Although our relatively small cohort sizes precluded rigorous statistical comparisons of immune responses in the immunized mice, oral administration of even a single dose of HaitiV elicited detectable vibriocidal antibodies in all animals. These observations provide strong data establishing HaitiV’s immunogenicity. Additionally, the comparable vibriocidal titers elicited by HaitiV and CVD-103HgR, a live OCV licensed by the FDA for travelers, bodes well for HaitiV’s immunogenicity in humans. Combining the immunogenicity data presented here with our finding that HaitiV can protect from cholera prior to the induction of adaptive immunity [29], suggests that HaitiV may function as a dual-acting agent, providing both rapid-onset short-term protection from disease while eliciting stable and long-lasting immunity against cholera.
Data from the challenge experiments (Fig 6) are consistent with the prevailing notion that serogroup, and to a lesser extent serotype are major determinants of protection against V. cholerae challenge [39,47]. Although it is thought that exposure to Inaba strains is more cross-protective than exposure to Ogawa strains, the relative potency of Inaba versus Ogawa vaccines in eliciting dual protection against both O1 serotypes requires further definition, as it has been suggested that both Ogawa and Inaba vaccine strains are good candidates for development [4–7]. A mixture of Ogawa and Inaba serotypes either as distinct strains or one bivalent strain (serotype Hikojima) may be beneficial in broadening the breadth of the immune response to HaitiV [48,49].
The modest protection that HaitiV immunization provided against V. cholerae O139 was unexpected. The epidemiology of the original O139 outbreak and experimental studies in rabbits demonstrate a lack of cross-protection between the two serogroups [14,15,39]. Notably, although pups born to HaitiV-immunized dams and challenged with MO10 survived longer than pups born to non-immunized dams, there was little difference in the MO10 intestinal colonization between these groups (Fig 6C). The discrepancy between clinical protection and relatively robust colonization suggests that HaitiV stimulates immune responses to V. cholerae factors, like CT, that may contribute to disease but not directly to colonization. The capacity of live OCVs to induce immune responses to in vivo-expressed antigens, including CtxB, is a property that heightens the appeal of live vs killed OCVs [27,50].
Although GF mice enabled us to test the protective efficacy of a candidate live OCV, their absence of the microbiota and resulting improper immune development are important caveats to consider. The GF model does not recapitulate the competitive microbial environment that live OCVs will encounter in the human host. We observed similar prolonged shedding patterns for both CVD-103HgR and HaitiV in the GF mice (Fig 1), yet CVD-103HgR is known to be shed by human volunteers at a low frequency for a short period [25,51]. Thus, our findings likely overestimate HaitiV’s capacity to colonize the human intestine. The observation that exposure of HaitiV-immunized dams to SPF-derived pups during the cross-fostering experiments led to the elimination of detectable HaitiV in feces supports the prediction that this vaccine will not stably colonize humans. Since we employed a multi-dose vaccination schedule here, it remains an open question whether transient exposure of naïve mice to HaitiV will also stimulate protective immunity, as has been shown in the context of vaccination with V. cholerae outer membrane vesicles [52,53]. The streptomycin-treated mouse model of V. cholerae colonization, which allows for temporary intestinal colonization, may also be useful to investigate the duration of colonization required for immunity [54]. Ultimately, the capacity of HaitiV to colonize the intestine and the relationship between colonization and protective immunity will need to be defined in human volunteers.
The immunogenicity of live OCVs in mice has only been investigated in GF animals because adult mice with intact microbiota are refractory to intestinal V. cholerae replication and colonization. However, previous studies of live OCVs in GF mice only analyzed immune correlates of protection and not protection against challenge [36,37]. The combination of the neonatal survival assay with the oral GF vaccination model builds on existing knowledge of these mice to assay both the immunogenicity and protective efficacy of live OCV candidates [38]. This model may be a useful addition to existing approaches that probe the molecular bases of vaccine-mediated mucosal protection against pathogens, a topic with significant translational potential that remains poorly understood [53,55]. A recent report employing a similar maternal-infant transmission model in the context of intraperitoneally-delivered heat-killed Citrobacter rodentium highlights the versatility of assessing vaccine protective efficacy using the infant progeny of immunized animals as readouts [56]. The broad availability of genetically engineered mice and the relative ease of GF-derivation provides a powerful opportunity to leverage both host and bacterial genetics to explore how live-OCVs can be optimized to better defend against this ancient pathogen.
|
10.1371/journal.pntd.0005123 | Blood Transcriptional Profiling Reveals Immunological Signatures of Distinct States of Infection of Humans with Leishmania infantum | Visceral leishmaniasis (VL) can be lethal if untreated; however, the majority of human infections with the etiological agents are asymptomatic. Using Illumina Bead Chip microarray technology, we investigated the patterns of gene expression in blood of active VL patients, asymptomatic infected individuals, patients under remission of VL and controls. Computational analyses based on differential gene expression, gene set enrichment, weighted gene co-expression networks and cell deconvolution generated data demonstrating discriminative transcriptional signatures. VL patients exhibited transcriptional profiles associated with pathways and gene modules reflecting activation of T lymphocytes via MHC class I and type I interferon signaling, as well as an overall down regulation of pathways and gene modules related to myeloid cells, mainly due to differences in the relative proportions of monocytes and neutrophils. Patients under remission of VL presented heterogeneous transcriptional profiles associated with activation of T lymphocytes via MHC class I, type I interferon signaling and cell cycle and, importantly, transcriptional activity correlated with activation of Notch signaling pathway and gene modules that reflected increased proportions of B cells after treatment of disease. Asymptomatic and uninfected individuals presented similar gene expression profiles, nevertheless, asymptomatic individuals exhibited particularities which suggest an efficient regulation of lymphocyte activation and a strong association with a type I interferon response. Of note, we validated a set of target genes by RT-qPCR and demonstrate the robustness of expression data acquired by microarray analysis. In conclusion, this study profiles the immune response during distinct states of infection of humans with Leishmania infantum with a novel strategy that indicates the molecular pathways that contribute to the progression of the disease, while also providing insights into transcriptional activity that can drive protective mechanisms.
| Infections of humans with the protozoan parasites L. donvani and L. infantum can lead to the development of the disease visceral leishmaniasis, but also to an asymptomatic status. However, the mechanisms that result in these clinical outcomes after infection are poorly understood. In this study, we applied a data-driven approach to obtain insights into the immunological processes linked to the progression of the disease or to protective mechanisms. For this purpose, we evaluated the patterns of expression for genes that code proteins from the entire human genome in the peripheral blood from patients with visceral leishmaniasis, from individuals who remained asymptomatic after infections with L. infantum, from patients who were recovering from disease after treatment and from uninfected individuals. By employing computational analysis to evaluate the blood transcriptional activity of each group, we identified transcriptional signatures that correlate with previous findings obtained through different analytical methods. Moreover, our analyses uncovered hitherto unidentified molecular pathways and gene networks associated with the transcriptional profiles of individuals recovering from disease or that did not develop symptoms after infection. This suggests that activation of protective responses can be useful targets for the development of new therapies for visceral leishmaniasis.
| Infections with the protozoan parasites Leishmania donovani or L. infantum (chagasi) result in clinical outcomes that range from asymptomatic infection to active visceral leishmaniasis (VL). When disease occurs, symptoms often include fever, hepatosplenomegaly, cachexia, pancytopenia and hypergammaglobulinemia [1], while the lethality of VL correlates with severe symptoms such as secondary infections, hemorrhage, liver failure and cardiotoxicity due to treatment [2].
Depressed cellular immunity is considered a hallmark of VL, which is evidenced by the inability of VL patients to develop a positive delayed type hypersensitivity (DTH) in Montenegro skin tests in spite of infection [3], and the absence of IFN-γ in cultures of peripheral blood mononuclear cells stimulated with leishmanial antigens [4]. On the other hand, whole blood assays showed that VL patients do not lack the ability to mount Leishmania specific IFN-γ responses [5]. Furthermore, peripheral blood or splenic CD4+ T lymphocytes from VL patients produce IFN-γ in response to leishmanial antigens, which is also crucial to limit parasite replication in splenic aspirate cultures [6]. These findings indicate that progression of VL involves other molecular mechanisms besides failures in activation and differentiation of CD4+ T lymphocytes. Development and severity of VL have been associated with several pro-inflammatory and immunoregulatory factors such as cytokines [7,8], lipopolysaccharide [9], mannan-binding lectin [10], C reactive protein and patterns of IgG Fc N-glycosylation [8]. In addition, studies addressing features of infected asymptomatic individuals point towards a fine regulation of several immune compartments thought to control parasites without damage to the host [8,11,12]. Thus, particular clinical outcomes after infections with L. infantum are influenced by complex multi-factorial immunological processes.
Re-circulation between central and peripheral lymphoid organs has a major impact on effective immune responses and infections and inflammation cause cell migration via lymphatic and circulatory systems [13]. During physiological or pathological events in which factors are released systemically, features of peripheral cell re-circulation provide an informative platform to study the human immune system with molecular methods of genomic scale, which have been used to investigate blood transcriptional and immunological profiles during human infections, including parasitic diseases [14–16].
Genome-wide profiling strategies have been employed to evaluate in vitro systems of infection with Leishmania and in vivo models of VL [17–20], while studies in humans are limited to biopsies from patients with cutaneous leishmaniasis [21,22]. We hypothesized that a global overview of gene expression in the peripheral blood of humans presenting with distinct states of infection with L. infantum could reveal unappreciated immunological features that account for pathological or protective responses. To address this issue, we undertook a series of molecular approaches and functional analyses to uncover the transcriptional activity of the immune response that extend the understanding and provide new insights into the immunobiology of human VL.
This study was conducted as per protocols approved by the Research Ethics Committee of the Clinics Hospital of the Ribeirão Preto Medical School—USP (protocol 2347/2012). All the methods were carried out in accordance with approved guidelines. Informed written consent was obtained from all of the participants or their parents or legal guardians. Whole peripheral blood was collected from patients with symptoms of VL admitted to Natan Portella Institute of Tropical Diseases, UFPI, Teresina-PI, Brazil. Diagnosis was confirmed by identification of Leishmania amastigotes in Giemsa-stained smears of bone marrow aspirate, and patients diagnosed with VL received treatment according to Brazilian guidelines [23]. Additionally, whole peripheral blood was collected from a distinct group of VL patients at 2 to 5 months after the beginning of therapy with pentavalent antimonial, which were under remission of the disease (Table 1). Study subjects also included healthy individuals living in the same areas and considered to be asymptomatically infected with L. infantum, who were identified by a positive delayed type hypersensitivity (DTH) to leishmanial antigens (Table 1). Controls included individuals from different regions of Brazil (Teresina-PI and Ribeirão Preto-SP) who presented a negative DTH to leishmanial antigens (Table 1). The groups did not present significant differences with respect to age (ANOVA P value = 0.370) or sex distribution (Chi-square P value = 0.4181). Whole peripheral blood samples were stabilised in PAXgene Blood RNA tubes (PreAnalitiX, Hombrechtikon, Switzerland) and stored at -80°C.
Isolation and purification of total RNA was performed using the PAXgene Blood RNA Kit (PreAnalytix) according to the manufacturer’s instructions. RNA concentration was verified with NanoDrop 1000 spectrophotometer (NanoDrop Technologies, Wilmington, DE, USA) and the RNA integrity was determined using an Agilent 2100 Bioanalyzer (Agilent Technologies, Foster City, CA, USA). The RNA samples were submitted to microarray hybridization at the Functional Genomics Unit of the Roy J. Carver Biotechnology Center, University of Illinois, Urbana-Champaign, Illinois, USA. All procedures were performed according to the manufacturer’s instructions. Briefly, cRNA amplification and labelling was carried out on 1 ug of total RNA by using an Illumina TotalPrep Amplification kit (Ambion, Austin, TX, USA). The samples were then hybridized onto on Illumina HumanHT-12 v4 Expression BeadChips that were scanned with an Illumina iScan System (Illumina, San Diego, CA, USA). Illumina´s Beadstudio software was used to generate signal intensity values from the scans.
Raw data were processed using the R Language and Environment for Statistical Computing (R) 3.2.0 [24] in association with Bioconductor 3.1 [25]. The lumi package for R [26] was used to perform quality control, log2 transformation and normalization with robust spline normalization (RSN) method. This processing pipeline was based on the comparison and variation of transformation and normalization methods and optimized according to the number of samples, as well as the array technology [27]. Data was filtered to remove unexpressed genes based on detection call p-values computed for each probeset of the > 47,000 probes present on the Illumina HumanHT-12 v4 array and 17,015 probes were retained for further analysis. Probe-level expression data files were deposited at the Gene Expression Omnibus (GEO) repository under accession number GSE77528.
The patterns of differential gene expression between the study groups were evaluated by generating linear models and moderated t-statistic or ANOVA with the package Limma for R [28]. P values were adjusted with Benjamini-Hochberg false discovery rate (FDR) correction, whereby differentially expressed probes were identified by a FDR <0.01 and mean fold-difference ≥ 1.5 between VL patients and controls or asymptomatic individuals; or mean fold-difference ≥ 1.3 between patients under remission and VL patients, controls or asymptomatic individuals. Two different cut-offs of fold-differences were chosen in order to avoid over-estimating differentially expressed probes or penalizing particular transcriptional profiles in pathway analyses. Differentially expressed probes were collapsed into genes using the function collapseRows() and “Max-Mean” method [29] from the WGCNA package for R [30]. On the basis of differentially expressed genes (DEGs) between the study groups, a heat map was generated to visualize the resulting hierarchical clustering of expression data performed with Euclidian distance and complete algorithm linkage. DEGs lists were incorporated to the GeneGo MetaCore pathway analysis tool (Thomson Reuters, NY) and used to identify genes that overlap within curated biological processes and pathways at a higher frequency than would normally be expected to occur for a randomly selected set of genes. A FDR <0.05 was used as a threshold to determine whether a process or pathway was statistically represented by DEGs.
Co-expressed genes across the whole data set (n = 45 samples in the same analysis) were selected using the weighted gene co-expression network analysis (WGCNA) package for R [30]. Log-transformed, normalized expression data were filtered by the 3700 most variant genes. A soft threshold power beta was chosen based on the scale-free topology criterion [31]. Constructed gene networks were then used to identify modules from the topological overlap matrix with the functions cutreeDynamic() and mergeCloseModules() and imported for network visualization into Cytoscape v 3.2.1.
Gene Set Enrichment Analysis (GSEA) [32] was used to determine significant associations between blood transcriptional patterns of each study group and the modules identified by WGCNA, which were loaded as gene sets. In addition, we also implemented GSEA based on a framework of Blood Transcriptional Modules (BTM) which was previously constructed from over 30,000 human blood transcriptomes derived from more than 500 studies available in public databases [33]. GSEA parameters included weighted enrichment statistic and Signal2Noise metric, with 1,000 permutations.
To estimate relative abundance of cell subsets from whole blood expression profiles we implemented the meanProfile method with the CellMix package for R [34]. We applied this method using previously published signatures for erythroblasts, megakaryocytes, granulocytes, monocytes, NK cells, CD4+ T lymphocytes, CD8+ T lymphocytes and B lymphocytes [35].
Reverse Transcription followed by quantitative PCR (RT-qPCR) was performed using Complementary DNA was synthesized starting from 200 ng of RNA using High Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Foster City, CA). SYBR Green real-time amplifications were performed on a Rotor-Gene 6000 instrument (Corbett Life Science, Valencia, CA, USA) using a set of designed primers (S1 Table). Samples were analyzed in duplicate and values obtained by RT-PCR for target genes were normalized to the average of cycling threshold from "housekeeping" genes ACTB, B2M, RNA18S5 and PPIA. Fold changes were calculated according to the 2(-ΔΔCt) method.
Data analysis was performed with GraphPad Prism V 5.0. One-way ANOVA with Bonferroni´s multiple-comparison test or one-sample t test were used to evaluate differences among independent groups. Spearman’s rank correlation was applied to assess nonparametric associations. P values less than 0.05 were considered significant.
To determine transcriptional signatures associated with distinct states of infection with L. infantum, we evaluated the patterns of gene expression of whole blood from active VL patients, from patients that received treatment and were considered to be under remission of disease, from healthy individuals that exhibited a positive delayed type hypersensitivity reaction to leishmanial antigens and that were considered to be asymptomatically infected with L. infantum, and control individuals that exhibited a negative delayed type hypersensitivity to leishmanial antigens and were considered to be uninfected (Table 1). Principal component analysis (PCA) of the 17,105 annotated probe sets (12,491 genes) resulted in a consistent pattern of clustering for half of the VL patients, whereas PCA based on expression data from uninfected controls and asymptomatic individuals indicates similar global transcriptional profiles between those subjects (Fig 1A). In addition, the transcriptional profile of patients under remission of disease exhibited an intermediary pattern of clustering between VL patients and uninfected controls or asymptomatic individuals (Fig 1A). Linear model-based statistical analysis with a FDR < 0.01 identified 817 or 799 differentially expressed genes (DEGs) between VL patients and uninfected controls or asymptomatic individuals, respectively (Fig 1B—upper left and middle panels—and S1 Data). Applying this statistical method, we observed that asymptomatic individuals did not present significant differences in whole blood gene expression when compared to uninfected controls (Fig 1B—upper right panel). Further analysis resulted in 324, 459 or 528 DEGs between expression data from patients under remission and VL patients, uninfected controls or asymptomatic individuals, respectively (Fig 1B—lower panels—and S1 Data).
To evaluate whether transcriptional signatures based on differential expression analysis could segregate subjects from distinct states of infection, we applied an unsupervised hierarchical clustering on expression data from highly significant DEGs between all groups (ANOVA P < 0.001, 2232 genes), shown in Fig 1C. The analysis resulted in two main clusters of individuals. The first cluster was comprised only by VL patients (Fig 1C). The second cluster resolved into two sub-clusters; one composed mainly by uninfected controls, but which also contained asymptomatic individuals; and a second sub-cluster formed by patients under remission of disease, asymptomatic individuals and uninfected controls (Fig 1C). Of interest, most patients under remission of disease clustered together into a unique group within this sub-cluster (Fig 1C). Taken together, these results demonstrate that infections with L. infantum induce significant changes in the abundance of blood transcripts and that the patterns of gene expression depend on the clinical status after infection or activity of the disease.
To understand the biological processes reflected by the identified transcriptional signatures, we used the GeneGO Metacore platform for functional analysis to retrieve the ontology of immunity related genes that were differentially expressed, whereby their expression changed according to each of the comparisons between the states of infection (Fig 2A). Relative to uninfected controls or asymptomatic individuals, the expression of genes annotated into processes such as leukocyte chemotaxis (CCR1, CCR3, CXCR1, CXCR4, CXCL16, CXCL8) or neutrophil activation (CXCL8, FPR1, C5AR1) were down-regulated in VL patients (Fig 2A). On the other hand, up-regulated genes were mainly enriched into network processes such as NK cell cytotoxicity (GZMA, GZMB, PRF1) or TCR signaling (CD3D, CD3G, CD8A, LAT) (Fig 2A). Compared to uninfected controls, VL patients exhibited a wide modulation of genes enriched into the interferon signaling network process (IDO1, IFI35, IFIT1, IFITM2, IFNG, SOCS1, STAT1, STAT2) (Fig 2A and 2B). Yet, compared to asymptomatic individuals, GBP2, IDO1, IFI35, STAT2, TAP1 were not differentially expressed in VL patients, indeed most of the interferon signaling related genes were down-regulated in this group of individuals (Fig 2A and 2B). The majority of genes enriched for the BCR-pathway were down-regulated in VL patients when compared to both uninfected controls and asymptomatic individuals (Fig 2A). However, compared to other clinical-epidemiologic groups analyzed herein, genes enriched for the BCR-pathway (BTK, CD19, CD72, CD79A, CD79B, LYN) were up-regulated in patients under remission of disease (Fig 2A).
To obtain insights into the regulation of canonical pathways reflected by the transcriptional profiles from distinct states of infection with L. infantum, we analyzed the enrichment of up-regulated or down-regulated DEGs on pathway maps annotated in the GeneGO Metacore database (Fig 2C). Compared to uninfected controls or asymptomatic infected individuals, up-regulated DEGs from VL patients were enriched in pathways such as: "antigen presentation by MHC class I" (S1 Fig), "differentiation and clonal expansion of CD8+ T cells" and “Granzyme A signaling” (Fig 2C—left panel). Moreover, only when compared to uninfected controls, up-regulated DEGs from VL patients were enriched in pathways as: "T regulatory cell-mediated modulation of antigen-presenting cell functions", "IFN alpha/beta signaling" and "antiviral actions of interferons" (Fig 2C—left panel). Those results suggest that VL patients exhibit an increased activation of cytotoxic T lymphocytes, which is in agreement with the up-regulation of genes related to TCR signaling (Fig 1A). Moreover, these results also represent the first evidence of an increased activity of type I interferon signaling in humans infected with L. infantum. Compared to VL patients, up-regulated DEGs from patients under remission were enriched in pathways such as: "integrin inside-out signaling in neutrophils", and "Notch signaling pathway" (Fig 2C—left panel). Moreover, compared to uninfected controls or to asymptomatic individuals, up-regulated DEGs from patients under remission were enriched into metabolic pathways (Fig 2C—left panel).
Compared to uninfected controls or asymptomatic individuals, down-regulated DEGs from VL patients were enriched for pathways such as: "integrin inside-out signaling in neutrophils" (S2 Fig), chemokine and cytokine signaling and immune receptor signaling as shown in the right panel of Fig 2C. Compared to VL patients, down-regulated DEGs identified for patients under remission were significantly enriched into pathways such as: "T regulatory cell-mediated modulation of antigen-presenting cell functions", "initiation of T cell recruitment in allergic contact dermatitis" (Fig 2C—right panel). Furthermore, compared to uninfected controls or asymptomatic individuals, down-regulated DEGs from patients under remission exhibited significant enrichments in pathways previously associated with expression data from VL patients, except for: “lipoxin inhibitory action on formyl-Met-Leu-Phe-induced neutrophil chemotaxis”, “chemokine signaling (CCL2, CCR3, CXCL16, CXCR4)”, “cytokine signaling (IL-6, IL-8, IL-9, IL-10) “and “Fc gamma R-mediated phagocytosis in macrophages” (Fig 2C—right panel).
Overall, those results indicate that upon development of VL, several pathways related to the immune response are subjected to profound perturbations and suggest that the innate immune response is mainly down-regulated. In contrast, treatment might trigger the activation of pathways as Notch signaling (Fig 2C—left panel) or even down-regulate the transcriptional activity of pathways as “T regulatory cell-mediated modulation of antigen presenting cell functions” or “NETosis in SLE” (Fig 2C—right panel).
Although analysis at the level of single genes has been widely used for interpretation of expression data and discovery of biomarkers, the large number of comparisons are permissive to noise and may lack power to detect subtle, but important features of gene expression datasets [32]. Therefore, in order to obtain an additional perspective about the nature of responses reflected by transcriptional profiles from the subjects evaluated herein, we performed a weighted gene co-expression network analysis (WGCNA), which is based on coordinately expressed genes for the identification of gene modules. First, we detected the 3,700 most variable genes from the study population, which included all forty-five samples irrespective of state of infection or treatment. Next, a hierarchical clustering was applied to expression data from those most variable genes, which identified thirteen color-coded co-expression modules (Fig 3A—Merged dynamic). Eleven modules could be annotated with GeneGo Metacore and were enriched in network processes and/or pathway maps described in Table 2; genes composing specific modules are detailed in S2 Data. Demonstrative networks of genes clustered into the cyan module (Type I interferon) or light-green module (antigen presentation by MHC class I) are depicted in S3 Fig.
WGCNA relies entirely on a data-driven process, which reflects fluctuations in blood transcript abundance measured across an entire population irrespective of state of infection or treatment. Therefore, the transcriptional profiles of the study subjects were graphically represented for individual modules (Fig 3B). Those results demonstrate coordinated expression of genes retained in specific clusters (Fig 3A) and also suggest differential activity of gene modules among distinct clinical-epidemiologic groups (Fig 3B). To further address this question, we conducted gene set enrichment analysis (GSEA) using WGCNA modules as customized gene sets and a FDR <0.05 (Fig 3C). Compared to uninfected controls and asymptomatic individuals, we highlight that the transcriptional profiles of VL patients were associated with a positive regulation of gene modules annotated as: "TCR signaling and antigen presentation" (blue module) and "antigen presentation" (light-green); at the same time, transcriptional profiles of VL patients were associated with a negative regulation of gene modules annotated as: "cell adhesion and neutrophil migration" (brown module), "Notch signaling pathway" (gray60 module) and "cell adhesion and LTBR1 signaling" (light-cyan module) (Fig 3C). Of note, a negative regulation of the module annotated as the "B lymphocyte related module" (salmon) was associated with the transcriptional profile of VL patients only when compared to uninfected controls (Fig 3C). Noteworthy is the fact that a positive regulation of the cyan module (type I interferon, S3A Fig) was also associated with the transcriptional profile of VL patients when compared to uninfected controls, however the same module was negatively regulated in VL patients when compared to asymptomatic individuals (Fig 4C). Overall, compared with VL patients, uninfected controls or asymptomatic individuals, the transcriptional profiles of patients under remission were associated with a positive regulation of modules annotated as: "TCR signaling and antigen presentation" (blue module), "Notch signaling pathway" (Gray60 module) and "B lymphocyte related module" (salmon module) (Fig 3C). Yet, the regulation of several modules associated with transcriptional profiles of patients under remission was dependent on specific comparison with the other clinical-epidemiologic groups (Fig 3C). We also evaluated differences between the transcriptional profiles of asymptomatic individuals in comparison to those of uninfected controls using this same approach. We highlight associations with positive regulations of modules annotated as: "type I interferon signaling" (green-yellow module), "Notch signaling pathway" (gray60 module), "cell adhesion and LTBR1 signaling" (light-cyan module) and "antigen presentation" (light-green module) (Fig 3C); Moreover, asymptomatic individuals exhibited negative regulation of modules as: "TCR signaling and antigen presentation" (blue module) and "B lymphocyte related module" (salmon) (Fig 3C).
Using a different strategy of analysis, we found that results from WGCNA are highly correlated with those from single gene level, capturing significant perturbations of TCR signaling and antigen presentation, as for cell adhesion and neutrophil-related modules in VL patients. Those results also reinforce the fact that treatment of VL patients triggers the activation of Notch signaling pathway and increases the transcriptional activity of B lymphocytes. Of note, using this strategy we identified that, independently of the clinical outcome, infection with L. infantum induces a transcriptional signature of type I interferon. However, this response exhibits a degree of association with distinct statuses of infection, whereby asymptomatic individuals presented with the strongest associations with activation of this module, followed by VL patients and then by patients under remission of disease.
We also employed GSEA (FDR <0.05) with a previously constructed framework of Blood Transcription Modules (BTMs) [33] to expand the modular analyses obtained by WGCNA and further evaluate the association of transcriptional profiles with distinct status of infection with L. infantum. Thus, transcriptional profiles of VL patients were associated with a positive regulation of several modules mainly enriched in NK cells (Fig 4A), T lymphocytes (Fig 4B) and type I interferon response (Fig 4C); and cell cycle, synthesis and metabolism (Fig 4D). On the other hand, the transcriptional profiles of VL patients were associated with a negative regulation of several modules related to myeloid cells (Fig 4A), B lymphocytes and effector responses such as cell adhesion and chemotaxis, immune activation with innate sensing and signaling (Fig 4C). Those results are in agreement with data from previous sections, demonstrating up-regulation of genes related to NK cells (Fig 2A) or activation of T lymphocytes (Figs 2 and 3), as well as for an overall down-regulation of the transcriptional activity of the innate immune response (Figs 2 and 3). Compared to VL patients, transcriptional profiles of patients under remission were mainly associated with a positive regulation of modules related to monocytes and neutrophils (Fig 4A), B lymphocytes (Fig 4B), as well as with a few effector pathways such as coagulation and complement systems (Fig 4C). However, relative to uninfected controls and/or asymptomatic individuals, several of those same modules were actually down-regulated, whereby positive associations were found mainly for modules related to NK cells (Fig 4A), both B and T lymphocytes (Fig 4B) and to cell cycle, synthesis and metabolism (Fig 4D). Taken together, those data suggest that treatment induces significant recovery of circulation of neutrophils and monocytes, however not comparable to that seen in healthy individuals. Furthermore, the results suggest that after treatment, there is a more substantial circulation of both B and T lymphocytes, indicating that treatment might function by restoring a balance to adaptive responses (Fig 4B). Additionally, compared to uninfected controls, transcriptional profiles of asymptomatic individuals were associated with a positive regulation of several innate immune cells, including those related to dendritic cells (Fig 4A) whereas modules related to lymphocytes were mainly down-regulated (Fig 4B). Furthermore, several modules related to effector and regulatory pathways were up-regulated in asymptomatic individuals and we highlight the up-regulation of all modules related to type I interferon response, which support the finding that those individuals indeed exhibit the strongest type I interferon response among distinct statuses of infection with L. infantum. Collectively, those data point to significant differences between blood transcriptional profiles which can reflect molecular mechanisms associated with pathogenic or protective responses during infections with L. infantum.
Modular analyses are highly informative for capturing differences in immune-related processes during disease, however, recent work demonstrates that gene modules are not independent and are subjected to higher coordinated regulation [36]. In view of that we sought to understand the relationship among the BTMs that were associated to the transcriptional profiles evaluated in this study. Using PCA, we extracted scores from the principal component 1 (PC1) for each of the 101 modules depicted in Fig 4 and performed a hierarchical clustering for coefficients of correlation among PC1 scores, which resulted in 5 main meta-modules (S4 Fig). We highlight meta-module II, which was highly enriched for modules depicting type I interferon signaling, dendritic cells and innate immune activation. Those results corroborate GSEA with BTMs. As an example, asymptomatic individuals indeed exhibited positive associations with several modules involving dendritic cells, innate immune activation and all modules related to type I interferon signaling (Fig 4A and 4C), suggesting a role for dendritic cells in the strong type I interferon signature observed in asymptomatic infection. Indeed, infections with L. major induce a type I interferon signature in human dendritic cells, which is required for production of IL-12 [37]. In addition, meta-module IV was highly enriched for modules involving B and T lymphocytes, as well as cell cycle (S4 Fig), which is in agreement with the proliferative characteristics of those cells. VL patients exhibited positive associations with modules depicting T lymphocytes and cell cycle, while patients under remission presented up-regulation of modules related to both B and T lymphocytes, as well as cell cycle (Fig 4B and 4D). As expected, those data indicate that BTMs are correlated and support the concept that infections with L. infantum elicit the coordinated activity of a multi-factorial network of biological processes rather than perturbations in a particular compartment of the immune response.
Whole blood presents a heterogeneous environment composed by numerous distinct, yet interacting cell populations, thus the transcriptional signatures from different states of infection with L. infantum could represent altered proportions of several cellular subsets. In view of these facts, we undertook a cell deconvolution analysis based on previously published cell signatures [35]. The expression signatures from erythroblasts, megakaryocytes, granulocytes, monocytes, NK cells, CD4+ T lymphocytes, CD8+ T lymphocytes and B lymphocytes of each study subject are shown in Fig 5A. Compared to uninfected controls, the relative abundance of erythroblasts increased significantly in VL patients (Fig 5B). In contrast, compared to uninfected controls and asymptomatic individuals, the relative abundance of monocytes and granulocytes decreased significantly in VL patients (Fig 5B). Patients under remission exhibited an increase in the relative abundance of CD8+ T lymphocytes only when compared to asymptomatic individuals (Fig 5B). However, compared to VL patients, uninfected controls or asymptomatic individuals, the relative abundance of B lymphocytes increased significantly in patients under remission (Fig 5B). Of interest, the relative abundance of megakaryocytes, NK cells and CD4+ T lymphocytes did not change among the clinical-epidemiologic groups (Fig 5B). Indeed, those results correlate with those of the modular analyses, which also demonstrate negative associations of myeloid cells with transcriptional profiles from VL patients and patients under remission of disease (Fig 4A), while modules related to B lymphocytes were highly associated with transcriptional profiles of patients under remission of disease (Fig 4B). In view of that, it should be considered that the overall down-regulation of pathways (Fig 2C) or effector and regulatory modules (Figs 3C and 4C) related to the innate immune response observed for VL patients and patients under remission could be driven mainly by decreased proportions of circulating myeloid cells. This decrease, in turn, might be due to entrapment of these cells into the spleen/liver or even be related to defects of the bone marrow and release of cells into the circulation.
To validate the expression obtained by microarray profiling, we also evaluated the expression of a set of genes by RT-qPCR (S1 Table). Relative to the expression of housekeeping genes, fold changes of selected genes were strongly correlated with those obtained with microarray expression profiling (Fig 6A). A detailed analysis of the relative expression of target genes demonstrated that microarray profiling was robust enough to capture significant differences in gene expression of highly modulated blood transcriptional profiles, such as those of VL patients and patients under remission of disease (Fig 6B). Moreover, we found that compared to uninfected controls, asymptomatic individuals exhibited significant modulations on relative expression of the majority of the target genes. Those results support the benefits of combining distinct functional analysis methods in blood transcriptomics and corroborate the findings obtained with WGCNA and GSEA.
The mechanisms that drive progression towards disease or protect individuals from developing symptoms while infected with Leishmania parasites remain poorly understood. Using a genome-wide approach to investigate patterns of gene expression from whole blood, we identified transcriptional profiles that shed light on pathways and/or gene expression modules associated with distinct states of human infections with L. infantum. It is noteworthy that the transcriptional signatures identified in this study discriminated between VL patients from patients under remission of disease and healthy individuals (Fig 1C). Importantly, by assessing the levels of expressions of a set of target genes by RT-qPCR we validated the robustness of the expression data acquired by microarray analysis (Fig 6). Transcriptional signatures from human samples have been shown to be sensitive to factors such as age and sex [38] and sample size [39]. The groups evaluated in this study did not exhibit significant differences in distribution of age or sex. Similar numbers of patients infected with L. braziliensis and controls were evaluated by pioneering studies that not only identified unique transcriptional signatures, but were also able to recapitulate previously described immunopathological responses in lesions of individuals with cutaneous leishmaniasis [21,22,40]. Of note, samples from patients under remission of disease were collected during distinct time points after the beginning of therapy, which could influence their blood transcriptional profiles. Nonetheless, we were able to identify transcriptional signatures that segregated patients under remission of disease from VL patients before therapy (Fig 1C); concomitantly, relative to asymptomatic individuals or uninfected controls, the majority of patients under remission of disease exhibited strong correlations of levels of expression for DEGs (Fig 1C). Furthermore, linear model-based statistical analysis and adjusted P values (FDR) did not detect significant differences between the transcriptional profiles of asymptomatic individuals and uninfected controls (Fig 1B and 1C). However, uncorrected P values retrieved 620 differentially expressed probes (S1 Data), which included probes for genes shown to be differentially expressed by RT-qPCR (Fig 6). These results demonstrate that, compared to uninfected controls, asymptomatic individuals present only a subset of differentially expressed genes, whereby the large number of comparisons between 17,105 probes [12,491 genes) can lead to a type II statistical error and inflate the rate of false negatives [41]. To overcome this issue, we conducted distinct approaches with the ability to estimate the differences between the transcriptional profiles of asymptomatic individuals and uninfected controls; the combination of distinct functional analyses and common features retrieved by them support the robustness of the immunological signatures identified for distinct states of infections with L. infantum. Therefore, we propose that in-depth analysis of transcriptional profiles from such populations, as well as longitudinal studies including patients followed throughout treatment can be useful for the prospection of new biomarkers of VL or asymptomatic infection, as well as for the prognosis after treatment and remission of disease.
Previous analyses demonstrated that the in vitro infection of monocyte-derived macrophages (MDM) and dendritic cells (MDC) with Leishmania elicits both species and cell-specific expression signatures [17]. Moreover, macrophage cultures infected with species of L. donovani complex exhibit an overall suppression of gene expression, suggesting a failure of proper macrophage activation [18,42]. However, the regulation of gene expression in MDM infected with L. chagasi was significantly impacted by the co-culture with autologous Leishmania-naïve T cells [18], suggesting that the inflammatory milieu of complex microenvironments such as the infection foci and peripheral blood influence the transcriptional programs of immune cells. Indeed, gene expression profiling of liver-resident macrophages (Kuppfer cells) from mice infected with L. donovani identified a key transcriptomic network centered around the retinoid X receptor alpha, which was only active in bystander uninfected Kupffer cells exposed to the inflammatory factors in infected livers [19]. As observed for experimental VL [43], our study supports the view of compartmentalized responses, i.e., the dynamics of dominant pathways in specific cells of the spleen, liver, bone marrow and peripheral blood might be differentially associated with pro-inflammatory or regulatory processes during the course of the infection. For instance, there are strong evidences that IL-10 plays a role in the suppression of the response in the spleen of VL patients [44,45], but, despite increased serum concentrations or production of IL-10 in whole blood assays [8,46] and elevated expression of IL-10 found herein by RT-qPCR, we were unable to identify an up-regulation of the "IL-10 signaling" pathway in the peripheral blood of VL patients. On the other hand, we did identify a negative regulation of "IL-10 signaling" pathway in the transcriptional profiles of patients under remission, which might correlate with decreased levels of this cytokine and recovery after therapy [47].
We highlight that, regardless of clinical status, expression data from individuals exposed to L. infantum display positive regulations of pathways and gene modules related to "type I interferon signaling" when compared to uninfected individuals, suggesting that IFN-αβ might play important roles in infections with L. infantum. Although the role of IFN-γ in infections with Leishmania has been extensively explored, the function of type I interferon signaling is not clear [48]. Of note, transcriptomic profiling of lesions from patients infected with L. braziliensis identified a positive regulation of type I interferon signaling [22], while L. major induces a type I transcriptional signature in human dendritic cells [37], indicating common responses from both cutaneous and visceral infections with Leishmania. Of interest, our analyses suggest that the response induced by IFN-αβ signaling might depend on the context and clinical status of infection with L. infantum. In other words, while type I interferon signaling is elicited in both VL patients and asymptomatic individuals, it might present differential regulation of its transcriptional program in these two states of infection. In line with this concept, other work showed that only low doses of IFN-β protected BALB/c mice from progressive cutaneous disease [49]. Furthermore, the "type I interferon signaling" gene module identified herein is composed by some interferon regulatory factors (IRF), in which IRF-7 was shown to exhibit a crucial role for parasite control in mice infected with L. donovani [50]. Thus, a fine regulation of IFN-αβ expression and of the transcriptional programs induced by those cytokines could promote a protective response in asymptomatically infected individuals. In contrast, an unbalanced signaling by these cytokines during chronic VL can play a similar pathological role as that observed in infections with Mycobacterium tuberculosis and Plasmodium [14,51]. Indeed, chronic exposure to type I interferon could impact homeostasis of CD4+ T cells [52] or even counter-regulate signaling by IL-1β [53] and limit protective mechanisms against infections with Leishmania [54]. In addition, IFNAR-deficient mice present enhanced immunity against L. amazonensis, which correlated with a critical role of neutrophils in parasite clearance [55]. The reason for such differences between VL patients and asymptomatic individuals might depend on several factors that include the genetic background of strains of both host and parasite, host nutritional status, history of exposure to vectors, co-infections with other pathogens, among other factors known to influence the magnitude and regulation of the immune response.
Although the expression data of VL patients seems to depict a general suppression of pathways and gene modules associated with innate immune response, a decrease in proportions neutrophils and monocytes was suggested by gene expression modular analyses (Figs 3 and 4) and validated with cell decovolution analysis (Fig 5), prompting careful interpretation. Indeed, neutropenia has been shown to be an independent risk factor for death in children with VL [56], indicating that the low proportion of circulating granulocytes and monocytes translate into the down-regulation of genes coding for chemokine receptors and chemokines as CCR1, CXCR1, CXCL8 or even neutrophil activation receptors as FPR1, C5AR1; in contrast, the up-regulation of IFNG underscores the increased levels of IFN-γ present in serum from VL patients [8], whereas up-regulation of both IFNG and CXCL10 support the activation of T lymphocytes and development of a Th1 response during active disease [6]. Accordingly, we were unable to identify significant differences in the relative proportions of T lymphocytes from VL patients compared to other groups (Fig 5B), indicating that DEGs annotated into processes as TCR signaling were not influenced by the relative proportion of those cells.
Dysfunctional responses during chronic VL might originate from failures in proper antigen presentation and stimulation, indicated by the significant associations between polymorphisms in the HLA class II region and susceptibility to visceral leishmaniasis [57], as well as by the negative regulation of gene modules related to major histocompatibility complex (MHC) class II identified in expression data from VL patients (Fig 4C). Yet, the extensive up-regulation of pathways and gene modules related to TCR signaling and antigen presentation through MHC class I suggests a chronic stimulation of CD8+ T lymphocytes during VL and correlates with previous findings from studies that used different analytical approaches [58,59]. Persistent cross-linking of TCR and MHC class I without appropriate co-stimulation of CD8+ T lymphocytes results in an exhausted cellular phenotype, which is characterized by the expression of the inhibitory receptors PD-1, CTLA-4, LAG3, TIM3 and TIGIT [60]. Indeed, CD8+ T lymphocytes isolated from VL patients exhibit increased expression of inhibitory surface receptors [58]. We also identified increased expression of PD-1, CTLA-4 and LAG3 in expression data from VL patients, while expression of such genes was down-regulated in patients under remission of the disease (S1 Data). Those findings suggest that perturbations in antigen presentation pathways may lead to inefficient activation and differentiation of CD4+ T lymphocytes, promote the exhaustion of CD8+ T lymphocytes and account for parasite evasion from the host response during VL.
Proper antigen presentation is crucial for T cell activation and differentiation, but other factors might impact lymphocyte function during VL. Indeed, T cell-specific deletion of Notch 1 and Notch 2 demonstrated that they are required for efficient development of Th1 immune responses and resistance in mice infected with L. major [61], thus polymorphisms affecting such molecules and transcriptional programs induced by their activation may influence T cell responses during infections with L. infantum. Indeed, a genome-wide association study in mixed-breed dogs with VL identified a marker located between two predicted transcription factor binding sites that regulate the expression of TLE1, a molecule involved in the Notch signaling pathway [62]. Another perspective is given by the demonstration that Notch 1 signaling pathway drives the activation of mouse macrophages into a M1 phenotype through metabolic up-regulation of mitochondrial oxidative phosphorylation and attendant reactive oxygen species [63], molecules that are implicated in the killing of L. braziliensis by human classical monocytes [64]. In the light of those findings, the positive associations between "Notch signaling pathway" with transcriptional profiles of patients under remission of disease and asymptomatic individuals support a protective role of this pathway in human infections with L. infantum.
Cell deconvolution analysis corroborates previous investigations focused on lymphocyte proportions in peripheral blood of VL patients [45,65], whereby strong associations between transcriptional profiles of patients under remission with B lymphocyte-related modules likely reflect higher abundance of those cells in the peripheral blood after therapy [65]. Despite this, hypergammaglobulinemia is frequently observed in VL patients due to polyclonal activation of B lymphocytes [66]. However, modules related to B lymphocytes were mainly down-regulated in VL patients and asymptomatic individuals, which suggests that after infections with L. infantum, activated B lymphocytes undergo differentiation to plasma cells and migrate to specific niches such as the bone marrow [67], while those remaining in the periphery might display unique transcriptional programs and functions. Indeed, follicular T cell-mediated regulation of the B lymphocyte compartment may account for beneficial or pathogenic responses during distinct states of infection with L. infantum [68], a hypothesis that is supported by significant differences in structural and functional features of immunoglobulin G isolated from VL patients and asymptomatic individuals [8].
In conclusion, this is the first attempt to screen for blood transcriptional signatures from distinct states of infection of humans with L. infantum. Future studies including individuals of other populations, as well as investigations focused on specific pathways highlighted by these signatures shall confirm and extend the hypotheses discussed herein. These signatures point to novel directions for studying human immune responses after infections with L. infantum, which can guide development of new strategies of intervention.
|
10.1371/journal.pcbi.1002211 | Neural Dynamics as Sampling: A Model for Stochastic Computation in Recurrent Networks of Spiking Neurons | The organization of computations in networks of spiking neurons in the brain is still largely unknown, in particular in view of the inherently stochastic features of their firing activity and the experimentally observed trial-to-trial variability of neural systems in the brain. In principle there exists a powerful computational framework for stochastic computations, probabilistic inference by sampling, which can explain a large number of macroscopic experimental data in neuroscience and cognitive science. But it has turned out to be surprisingly difficult to create a link between these abstract models for stochastic computations and more detailed models of the dynamics of networks of spiking neurons. Here we create such a link and show that under some conditions the stochastic firing activity of networks of spiking neurons can be interpreted as probabilistic inference via Markov chain Monte Carlo (MCMC) sampling. Since common methods for MCMC sampling in distributed systems, such as Gibbs sampling, are inconsistent with the dynamics of spiking neurons, we introduce a different approach based on non-reversible Markov chains that is able to reflect inherent temporal processes of spiking neuronal activity through a suitable choice of random variables. We propose a neural network model and show by a rigorous theoretical analysis that its neural activity implements MCMC sampling of a given distribution, both for the case of discrete and continuous time. This provides a step towards closing the gap between abstract functional models of cortical computation and more detailed models of networks of spiking neurons.
| It is well-known that neurons communicate with short electric pulses, called action potentials or spikes. But how can spiking networks implement complex computations? Attempts to relate spiking network activity to results of deterministic computation steps, like the output bits of a processor in a digital computer, are conflicting with findings from cognitive science and neuroscience, the latter indicating the neural spike output in identical experiments changes from trial to trial, i.e., neurons are “unreliable”. Therefore, it has been recently proposed that neural activity should rather be regarded as samples from an underlying probability distribution over many variables which, e.g., represent a model of the external world incorporating prior knowledge, memories as well as sensory input. This hypothesis assumes that networks of stochastically spiking neurons are able to emulate powerful algorithms for reasoning in the face of uncertainty, i.e., to carry out probabilistic inference. In this work we propose a detailed neural network model that indeed fulfills these computational requirements and we relate the spiking dynamics of the network to concrete probabilistic computations. Our model suggests that neural systems are suitable to carry out probabilistic inference by using stochastic, rather than deterministic, computing elements.
| Attempts to understand the organization of computations in the brain from the perspective of traditional, mostly deterministic, models of computation, such as attractor neural networks or Turing machines, have run into problems: Experimental data suggests that neurons, synapses, and neural systems are inherently stochastic [1], especially in vivo, and therefore seem less suitable for implementing deterministic computations. This holds for ion channels of neurons [2], synaptic release [3], neural response to stimuli (trial-to-trial variability) [4], [5], and perception [6]. In fact, several experimental studies arrive at the conclusion that external stimuli only modulate the highly stochastic spontaneous firing activity of cortical networks of neurons [7], [8]. Furthermore, traditional models for neural computation have been challenged by the fact that typical sensory data from the environment is often noisy and ambiguous, hence requiring neural systems to take uncertainty about external inputs into account. Therefore many researchers have suggested that information processing in the brain carries out probabilistic, rather than logical, inference for making decisions and choosing actions [9]–[22]. Probabilistic inference has emerged in the 1960’s [23], as a principled mathematical framework for reasoning in the face of uncertainty with regard to observations, knowledge, and causal relationships, which is characteristic for real-world inference tasks. This framework has become tremendously successful in real-world applications of artificial intelligence and machine learning. A typical computation that needs to be carried out for probabilistic inference on a high-dimensional joint distribution is the evaluation of the conditional distribution (or marginals thereof) over some variables of interest, say , given variables . In the following, we will call the set of variables , which we condition on, the observed variables and denote it by .
Numerous studies in different areas of neuroscience and cognitive science have suggested that probabilistic inference could explain a variety of computational processes taking place in neural systems (see [10], [11]). In models of perception the observed variables are interpreted as the sensory input to the central nervous system (or its early representation by the firing response of neurons, e.g., in the LGN in the case of vision), and the variables model the interpretation of the sensory input, e.g., the texture and position of objects in the case of vision, which might be encoded in the response of neurons in various higher cortical areas [15]. Furthermore, in models for motor control the observed variables often consist not only of sensory and proprioceptive inputs to the brain, but also of specific goals and constraints for a planned movement [24]–[26], whereas inference is carried out over the variables representing a motor plan or motor commands to muscles. Recent publications show that human reasoning and learning can also be cast into the form of probabilistic inference problems [27]–[29]. In these models learning of concepts, ranging from concrete to more abstract ones, is interpreted as inference in lower and successively higher levels of hierarchical probabilistic models, giving a consistent description of inductive learning within and across domains of knowledge.
In spite of this active research on the functional level of neural processing, it turned out to be surprisingly hard to relate the computational machinery required for probabilistic inference to experimental data on neurons, synapses, and neural systems. There are mainly two different approaches for implementing the computational machinery for probabilistic inference in “neural hardware”. The first class of approaches builds on deterministic methods for evaluating exactly or approximately the desired conditional and/or marginal distributions, whereas the second class relies on sampling from the probability distributions in question. Multiple models in the class of deterministic approaches implement algorithms from machine learning called message passing or belief propagation [30]–[33]. By clever reordering of sum and product operators occurring in the evaluation of the desired probabilities, the total number of computation steps are drastically reduced. The results of subcomputations are propagated as "messages" or "beliefs" that are sent to other parts of the computational network. Other deterministic approaches for representing distributions and performing inference are probabilistic population code (PPC) models [34]. Although deterministic approaches provide a theoretically sound hypothesis about how complex computations can possibly be embedded in neural networks and explain aspects of experimental data, it seems difficult (though not impossible) to conciliate them with other aspects of experimental evidence, such as stochasticity of spiking neurons, spontaneous firing, trial-to-trial variability, and perceptual multistability.
Therefore other researchers (e.g., [16]–[18], [35]) have proposed to model computations in neural systems as probabilistic inference based on a different class of algorithms, which requires stochastic, rather than deterministic, computational units. This approach, commonly referred to as sampling, focuses on drawing samples, i.e., concrete values for the random variables that are distributed according to the desired probability distribution. Sampling can naturally capture the effect of apparent stochasticity in neural responses and seems to be furthermore consistent with multiple experimental effects reported in cognitive science literature [17], [18]. On the conceptual side, it has proved to be difficult to implement learning in message passing and PPC network models. In contrast, following the lines of [36], the sampling approach might be well suited to incorporate learning.
Previous network models that implement sampling in neural networks are mostly based on a special sampling algorithm called Gibbs (or general Metropolis-Hastings) sampling [9], [17], [18], [37]. The dynamics that arise from this approach, the so-called Glauber dynamics, however are only superficially similar to spiking neural dynamics observed in experiments, rendering these models rather abstract. Building on and extending previous models, we propose here a family of network models, that can be shown to exactly sample from any arbitrary member of a well-defined class of probability distributions via their inherent network dynamics. These dynamics incorporate refractory effects and finite durations of postsynaptic potentials (PSPs), and are therefore more biologically realistic than existing approaches. Formally speaking, our model implements Markov chain Monte Carlo (MCMC) sampling in a spiking neural network. In contrast to prior approaches however, our model incorporates irreversible dynamics (i.e., no detailed balance) allowing for finite time PSPs and refractory mechanisms. Furthermore, we also present a continuous time version of our network model. The resulting stochastic dynamical system can be shown to sample from the correct distribution. In general, continuous time models arguably provide a higher amount of biological realism compared to discrete time models.
The paper is structured in the following way. First we provide a brief introduction to MCMC sampling. We then define the neural network model whose neural activity samples from a given class of probability distributions. The model will be first presented in discrete time together with some illustrative simulations. An extension of the model to networks of more detailed spiking neuron models which feature a relative refractory mechanism is presented. Furthermore, it is shown how the neural network model can also be formulated in continuous time. Finally, as a concrete simulation example we present a simple network model for perceptual multistability.
In machine learning, sampling is often considered the “gold standard” of inference methods, since, assuming that we can sample from the distribution in question, and assuming enough computational resources, any inference task can be carried out with arbitrary precision (in contrast to some deterministic approximate inference methods such as variational inference). However sampling from an arbitrary distribution can be a difficult problem in itself, as, e.g., many distributions can only be evaluated modulo a global constant (the partition function). In order to circumvent these problems, elaborate MCMC sampling techniques have been developed in machine learning and statistics [38]. MCMC algorithms are based on the following idea: instead of producing an ad-hoc sample, a process that is heuristically comparable to a global search over the whole state space of the random variables, MCMC methods produce a new sample via a “local search” around a point in the state space that is already (approximately) a sample from the distribution.
More formally, a Markov chain (in discrete time) is defined by a set of states (we consider for discrete time only the case where has a finite size, denoted by ) together with a transition operator . The operator is a conditional probability distribution over the next state given a preceding state . The Markov chain is started in some initial state , and moves through a trajectory of states via iterated application of the stochastic transition operator . More precisely, if is the state at time , then the next state is drawn from the conditional probability distribution . An important theorem from probability theory (see, e.g., p. 232 in [39]) states that if is irreducible (i.e., any state in can be reached from any other state in in finitely many steps with probability ) and aperiodic (i.e., its state transitions cannot be trapped in deterministic cycles), then the probability converges for to a probability that does not depend on the initial state . This state distribution is called the invariant distribution of . The irreducibility of implies that it is the only distribution over the states that is invariant under its transition operator , i.e.(1)
Thus, in order to carry out probabilistic inference for a given distribution , it suffices to construct an irreducible and aperiodic Markov chain that leaves invariant, i.e., satisfies equation (1). Then one can answer numerous probabilistic inference questions regarding without any numerical computations of probabilities. Rather, one plugs in the observed values for some of the random variables (RVs) and simply collects samples from the conditional distribution over the other RVs of interest when the Markov chain approaches its invariant distribution.
A convenient and popular method for the construction of an operator for a given distribution is looking for operators that satisfy the following detailed balance condition,(2)for all . A Markov chain that satisfies (2) is said to be reversible. In particular, the Gibbs and Metropolis-Hastings algorithms employ reversible Markov chains. A very useful property of (2) is that it implies the invariance property (1), and this is in fact the standard method for proving (1). However, as our approach makes use of irreversible Markov chains as explained below, we will have to prove (1) directly.
Let be some arbitrary joint distribution over binary variables that only takes on values . We will show that under a certain computability assumption on a network consisting of spiking neurons can sample from using its inherent stochastic dynamics. More precisely, we show that the stochastic firing activity of can be viewed as a non-reversible Markov chain that samples from the given probability distribution . If a subset of the variables are observed, modelled as the corresponding neurons being “clamped” to the observed values, the remaining network samples from the conditional distribution of the remaining variables given the observables. Hence, this approach offers a quite natural implementation of probabilistic inference. It is similar to sampling approaches which have already been applied extensively, e.g., in Boltzmann machines, however our model is more biologically realistic as it incorporates aspects of the inherent temporal dynamics and spike-based communication of a network of spiking neurons. We call this approach neural sampling in the remainder of the paper.
In order to enable a network of spiking neurons to sample from a distribution of binary variables , one needs to specify how an assignment of values to these binary variables can be represented by the spiking activity of the network and vice versa. A spike, or action potential, of a biological neuron has a short duration of roughly . But the effect of such spike, both on the neuron itself (in the form of refractory processes) and on the membrane potential of other neurons (in the form of postsynaptic potentials) lasts substantially longer, on the order of to . In order to capture this temporally extended effect of each spike, we fix some parameter that models the average duration of these temporally extended processes caused by a spike. We say that a binary vector is represented by the firing activity of the network at time for iff:(3)
In other words, any spike of neuron sets the value of the associated binary variable to 1 for a duration of length .
An obvious consequence of this definition is that the binary vector that is defined by the activity of at time does not fully capture the internal state of this stochastic system. Rather, one needs to take into account additional non-binary variables , where the value of at time specifies when within the time interval the neuron has fired (if it has fired within this time interval, thereby causing at time ). The neural sampling process has the Markov property only with regard to these more informative auxiliary variables . Therefore our analysis of neural sampling will focus on the temporal evolution of these auxiliary variables. We adopt the convention that each spike of neuron sets the value of to its maximal value , from which it linearly decays back to during the subsequent time interval of length .
For the construction of the sampling network , we assume that the membrane potential of neuron at time equals the log-odds of the corresponding variable to be active, and refer to this property as neural computability condition:(4)where we write for and for the current values of all other variables with . Under the assumption we make in equation (4), i.e., that the neural membrane potential reflects the log-odds of the corresponding variable , it is required that each single neuron in the network can actually compute the right-hand side of equation (4), i.e., that it fulfills the neural computability condition.
A concrete class of probability distributions, that we will use as an example in the remainder, are Boltzmann distributions:(5)with arbitrary real valued parameters which satisfy and (the constant ensures the normalization of ). For the Boltzmann distribution, condition (4) is satisfied by neurons with the standard membrane potential(6)where is the bias of neuron (which regulates its excitability), is the strength of the synaptic connection from neuron to , and approximates the time course of the postsynaptic potential in neuron caused by a firing of neuron with a constant signal of duration (i.e., a square pulse). As we will describe below, spikes of neuron are evoked stochastically depending on the current membrane potential and the auxiliary variable .
The neural computability condition (4) links classes of probability distributions to neuron and synapse models in a network of spiking neurons. As shown above, Boltzmann distributions satisfy the condition if one considers point neuron models which compute a linear weighted sum of the presynaptic inputs. The class of distributions can be extended to include more complex distributions using a method proposed in [40] which is based on the following idea. Neuron representing the variable is not directly influenced by the activities of the presynaptic neurons, but via intermediate nonlinear preprocessing elements. This preprocessing might be implemented by dendrites or other (inter-) neurons and is assumed to compute nonlinear combinations of the presynaptic activities (similar to a kernel). This allows the membrane potential , and therefore the log-odds ratio on the right-hand side of (4), to represent a more complex function of the activities , giving rise to more complex joint distributions . The concrete implementation of non-trivial directed and undirected graphical models with the help of preprocessing elements in the neural sampling framework is subject of current research. For the examples given in this study, we focus on the standard form of the membrane potential (6) of point neurons. As shown below, these spiking network models can emulate any Boltzmann machine (BM) [36].
A substantial amount of preceding studies has demonstrated that BMs are very powerful, and that the application of suitable learning algorithms for setting the weights makes it possible to learn and represent complex sensory processing tasks by such distributions [37], [41]. In applications in statistics and machine learning using such Boltzmann distributions, sampling is typically implemented by Gibbs sampling or more general reversible MCMC methods. However, it is difficult to model some neural processes, such as an absolute refractory period or a postsynaptic potential (PSP) of fixed duration, using a reversible Markov chain, but they are more conveniently modelled using an irreversible one. As we wish to keep the computational power of BMs and at the same time to augment the sampling procedure with aspects of neural dynamics (such as PSPs with fixed durations, refractory mechanisms) to increase biological realism, we focus in the following on irreversible MCMC methods (keeping in mind that this might not be the only possible way to achieve these goals).
Here we describe neural dynamics in discrete time with an absolute refractory period . We interpret one step of the Markov chain as a time step in biological real time. The dynamics of the variable , that describes the time course of the effect of a spike of neuron , are defined in the following way. is set to the value when neuron fires, and decays by at each subsequent discrete time step. The parameter is chosen to be some integer, so that decays back to in exactly time steps. The neuron can only spike (with a probability that is a function of its current membrane potential ) if its variable . If however, , the neuron is considered refractory and it cannot spike, but its is reduced by 1 per time step. To show that these simple dynamics do indeed sample from the given distribution , we proceed in the following way. We define a joint distribution which has the desired marginal distribution . Further we formalize the dynamics informally described above as a transition operator operating on the state vector . Finally, in the Methods section, we show that is the unique invariant distribution of this operator , i.e., that the dynamics described by produce samples from the desired distribution . We refer to sampling through networks with this stochastic spiking mechanism as neural sampling with absolute refractory period due to the persistent refractory process.
Given the distribution that we want to sample from, we define the following joint distribution over the neural variables:(7)
This definition of simply expresses that if , then the auxiliary variable can assume any value in with equal probability. On the other hand necessarily assumes the value if (i.e., when the neuron is in its resting state).
The state transition operator can be defined in a transparent manner as a composition of transition operators, , where only updates the variables and of neuron , i.e., the neurons are updated sequentially in the same order (this severe restriction will become obsolete in the case of continuous time discussed below). We define the composition as , i.e., is applied prior to . The new values of and only depend on the previous value and on the current membrane potential . The interesting dynamics take place in the variable . They are illustrated in Figure 1 where the arrows represent transition probabilities greater than 0.
If the neuron is not refractory, i.e., , it can spike (i.e., a transition from to ) with probability(8)where is the standard sigmoidal activation function and the denotes the natural logarithm. The term is the current membrane potential, which depends on the current values of the variables for . The term in (8) reflects the granularity of a chosen discrete time scale. If it is very fine (say one step equals one microsecond), then is large, and the firing probability at each specific discrete time step is therefore reduced. If the neuron in a state with does not spike, relaxes into the resting state corresponding to a non-refractory neuron.
If the neuron is in a refractory state, i.e., , its new variable assumes deterministically the next lower value , reflecting the inherent temporal process:(9)
After the transition of the auxiliary variable , the binary variable is deterministically set to a consistent state, i.e., if and if .
It can be shown that each of these stochastic state transition operators leaves the given distribution invariant, i.e., satisfies equation (1). This implies that any composition or mixture of these operators also leaves invariant, see, e.g., [38]. In particular, the composition of these operators leaves invariant, which has a quite natural interpretation as firing dynamics of the spiking neural network : At each discrete time step the variables are updated for all neurons , where the update of takes preceding updates for with into account. Alternatively, one could also choose at each discrete time step a different order for updates according to [38]. The assumption of a well-regulated updating policy will be overcome in the continuous-time limit, i.e., in case where the neural dynamics are described as a Markov jump process. In the methods section we prove the following central theorem:
The neural sampling model proposed above was formulated in discrete time of step size , inspired by the discrete time nature of MCMC techniques in statistics and machine learning as well as to make simulations possible on digital computers. However, models in continuous time (e.g., ordinary differential equations) are arguably more natural and “realistic” descriptions of temporally varying biological processes. This gives rise to the question whether one can find a sensible limit of the discrete time model in the limit , yielding a sampling network model in continuous time. Another motivation for considering continuous time models for neural sampling is the fact that many mathematical models for recurrent networks are formulated in continuous time [5], and a comparison to these existing models would be facilitated. Here we propose a stochastically spiking neural network model in continuous time, whose states still represent correct samples from the desired probability distribution at any time . These types of models are usually referred to as Markov jump processes. It can be shown that discretizing this continuous time model yields the discrete time model defined earlier, which thus can be regarded as a version suitable for simulations on a digital computer.
We define the continuous time model in the following way. Let , for , denote the firing times of neuron . The refractory process of this neuron, in analogy to Figure 1 and equation (8)-(9) for the case of discrete time, is described by the following differential equation for the auxiliary variable , which may now assume any nonnegative real number :(12)
Here denotes Dirac’s Delta centered at the spike time . This differential equation describes the following simple dynamics. The auxiliary variable decays linearly with time constant when the neuron is refractory, i.e., . Once arrives at its resting state it remains there, corresponding to the neuron being ready to spike again (more precisely, in order to avoid point measures we set it to a random value in , see Methods). In the resting state, the neuron has the probability density to fire at every time . If it fires at , this results in setting , which is formalized in equation (12) by the sum of Dirac Delta’s . Here the current membrane potential at time is defined as in the discrete time case, e.g., by for the case of a Boltzmann distribution (5). The binary variable is defined to be 1 if and 0 if the neuron is in the resting state . Biologically, the term can again be interpreted as the value at time of a rectangular-shaped PSP (with a duration of ) that neuron evokes in neuron . As the spikes are discrete events in continuous time, the probability of two or more neurons spiking at the same time is zero. This allows for updating all neurons in parallel using a differential equation.
In analogy to the discrete time case, the neural network in continuous time can be shown to sample from the desired distribution , i.e., is an invariant distribution of the network dynamics defined above. However, to establish this fact, one has to rely on a different mathematical framework. The probability distribution of the auxiliary variables as a function of time , which describes the evolution of the network, obeys a partial differential equation, the so-called Differential-Chapman-Kolmogorov equation (see [43]):(13)where the operator , which captures the dynamics of the network, is implicitly defined by the differential equations (12) and the spiking probabilities. This operator is the continuous time equivalent to the transition operator in the discrete time case. The operator consists here of two components. The drift term captures the deterministic decay process of , stemming from the term in equation (12). The jump term describes the non-continuous aspects of the path associated with “jumping” from to at the time when the neuron fires.
In the Methods section we prove that the resulting time invariant distribution, i.e., the distribution that solves , now denoted as it is not a function of time, gives rise to the desired marginal distribution over :(14)where and if and otherwise. denotes Kronecker’s Delta with if and otherwise. Thus, the function simply reflects the definition that if and 0 otherwise. For an explicit definition of , a proof of the above statement, and some additional comments see the Methods section.
The neural samplers in discrete and continuous time are closely related. The model in discrete time provides an increasingly more precise description of the inherent spike dynamics when the duration of the discrete time step is reduced, causing an increase of (such that is constant) and therefore a reduced firing probability of each neuron at any discrete time step (see the term in equation (8)). In the limit of approaching , the probability that two or more neurons will fire at the same time approaches , and the discrete time sampler becomes equal to the continuous time system defined above, which updates all units in parallel.
It is also possible to formulate a continuous time version of the neural sampler based on neuron models with relative refractory mechanisms. In the Methods section the resulting continuous time neuron model with a relative refractory mechanism is defined. Theoretical results similar to the discrete time case can be derived for this sampler (see Lemmata 9 and 10 in Methods): It is shown that each neuron “locally” performs the correct computation under the assumption of static input from the remaining neurons. However one can no longer prove in general that the global network samples from the target distribution .
In the following we present a network model for perceptual multistability based on the neural sampling framework introduced above. This simulation study is aimed at showing that the proposed network can indeed sample from a desired distribution and also perform inference, i.e., sample from the correct corresponding posterior distribution. It is not meant to be a highly realistic or exhaustive model of perceptual multistability nor of biologically plausible learning mechanisms. Such models would naturally require considerably more modelling work.
Perceptual multistability evoked by ambiguous sensory input, such as a 2D drawing (e.g., Necker cube) that allows for different consistent 3D interpretations, has become a frequently studied perceptual phenomenon. The most important finding is that the perceptual system of humans and nonhuman primates does not produce a superposition of different possible percepts of an ambiguous stimulus, but rather switches between different self-consistent global percepts in a spontaneous manner. Binocular rivalry, where different images are presented to the left and right eye, has become a standard experimental paradigm for studying this effect [44]–[47]. A typical pair of stimuli are the two images shown in Figure 4A. Here the percepts of humans and nonhuman primates switch (seemingly stochastically) between the two presented orientations. [16]–[18] propose that several aspects of experimental data on perceptual multistability can be explained if one assumes that percepts correspond to samples from the conditional distribution over interpretations (e.g., different 3D shapes) given the visual input (e.g., the 2D drawing). Furthermore, the experimentally observed fact that percepts tend to be stable on the time scale of seconds suggests that perception can be interpreted as probabilistic inference that is carried out by MCMC sampling which produces successively correlated samples. In [18] it is shown that this MCMC interpretation is also able to qualitatively reproduce the experimentally observed distribution of dominance durations, i.e., the distribution of time intervals between perceptual switches. However, in lack of an adequate model for sampling by a recurrent network of spiking neurons, theses studies could describe this approach only on a rather abstract level, and pointed out the open problem to relate this algorithmic approach to neural processes. We have demonstrated in a computer simulation that the previously described model for neural sampling could in principle fill this gap, providing a modelling framework that is on the one hand consistent with the dynamics of networks of spiking neurons, and which can on the other hand also be clearly understood from the perspective of probabilistic inference through MCMC sampling.
In the following we model some essential aspects of an experimental setup for binocular rivalry with grating stimuli (see Figure 4A) in a recurrent network of spiking neurons with the previously described relative refractory mechanism. We assigned to each of the 217 neurons in the network a tuning curve , centered around its preferred orientation as shown in Figure 4B. The preferred orientations of the neurons were chosen to cover the entire interval of possible orientations and were randomly assigned to the neurons. The neurons were arranged on a hexagonal grid as depicted in Figure 4F. Any two neurons with distance were synaptically connected (neighboring units had distance ). We assume that these neurons represent neurons in the visual system that have roughly the same or neighboring receptive field, and that each neuron receives visual input from either the left or the right eye. The network connections were chosen such that neurons that have similar (very different) preferred orientations are connected with positive (negative) weights (for details see Methods section).
We examined the resulting distribution over the dimensional network states. To provide an intuitive visualization of these high dimensional network states , we resort to a 2-dimensional projection, the population vector of a state (see Methods for details of the applied population vector decoding scheme). Only the endpoints of the population vectors are drawn (as colored points) in Figure 4D,E. The orientation of the population vector is assumed to correspond to the dominant orientation of the percept, and its distance from the origin encodes the strength of this percept. We also, somewhat informally, call the strength of a percept its coherence and a network state which represents a coherent percept a coherent network state. A coherent network state hence results in a population vector of large magnitude. Each direction of a population vector is color coded in Figure 4D,E, using the color code for directions shown on the right hand side of Figure 4F. In Figure 4D the distribution of the network is illustrated by sampling of the network for , with samples taken every millisecond. Each dot equals a sampled network state . In a biological interpretation the spike response of the freely evolving network reflects spontaneous activity, since no observations, i.e., no external input, was added to the system. Figure 4D shows that the spontaneous activity of this simple network of spiking neurons moves preferably through coherent network states for all possible orientations due to the chosen recurrent network connections (being positive for neurons with similar preferred orientation and negative otherwise). This can directly be seen from the rare occurrence of population vectors with small magnitude (vectors close to the “center”) in Figure 4D.
To study percepts elicited by ambiguous stimuli, where inputs like in Figure 4A are shown simultaneously to the left and right eye during a binocular rivalry experiment, we provided ambiguous input to the network. Two cells with preferred orientation and two cells with were clamped to . Additionally four neurons with resp. were muted by clamping to . This ambiguous input is incompatible with a coherent percept, as it corresponds to two orthogonal orientations presented at the same time. The resulting distribution over the state of the 209 remaining neurons is shown for a time span of of simulated biological time (with samples taken every millisecond) in Figure 4E. One clearly sees that the network spends most of the time in network states that correspond to one of the two simultaneously presented input orientations ( and ), and virtually no time on orientations in between. This implements a sampling process from a bimodal conditional distribution. The black line marks a trace of network states around a perceptual switch: The network remained in one mode of high probability – corresponding to one percept – for some period of time, and then quickly traversed the state space to another mode – corresponding to a different percept.
Three of the states around this perceptual switch (, and in Figure 4E) are explicitly shown in Figure 4F. Neurons that fired during the preceding interval of ms (marked in gray in Figure 4G) are drawn in the respective color of their preferred orientation. Inactive neurons are drawn in white, and clamped neurons are marked by a black dot ().
Figure 4G shows the action potentials of the non-clamped neurons during the same trace around the perceptual switch. One sees that the sampling process is expressed in this neural network model by a sparse, asynchronous and irregular spike response. It is worth mentioning that the average firing rate when sampling from the posterior distribution is only slightly higher than the average firing rate of spontaneous activity ( and respectively), which is reminiscent of related experimental data [7]. Thus on the basis of the overall network activity it is indistinguishable whether the network carries out an inference task or freely samples from its prior distribution. It is furthermore notable, that a focus of the network activity on the two orientations that are given by the external input can be achieved in this model, in spite of the fact that only two of the neurons were clamped for each of them. This numerical relationship is reminiscent of standard data on the weak input from LGN to V1 that is provided in the brain [48], [49], and raises the question whether the proposed neural sampling model could provide a possible mechanism (under the modelling assumptions made above) for cortical processing of such numerically weak external inputs.
The distribution of the resulting dominance durations, i.e., the time between perceptual switches, for the previously described setup with ambiguous input is shown for a continuous run of in Figure 4C (a similar method as in [18] was used to measure dominance durations, see Methods). This distribution can be approximated quite well by a Gamma distribution, which also provides a good fit to experimental data (see the discussion in [18]). We expect that also other features of the more abstract MCMC model for biological vision of [17], [18], such as contextual biases and traveling waves, will emerge in larger and more detailed implementations of the MCMC approach through the proposed neural sampling method in networks of spiking neurons.
We have presented a spiking neural network that samples from a given probability distribution via its inherent network dynamics. In particular the network is able to carry out probabilistic inference through sampling. The model, based on assumptions about the underlying probability distribution (formalized by the neural computability condition) as well as on certain assumptions regarding the underlying MCMC model, provides one possible neural implementation of the “inference-by-sampling paradigm” emerging in computational neuroscience.
During inference the observations (i.e., the variables which we wish to condition on) are modeled in this study by clamping the corresponding neurons by strong external input to the observed binary value. Units which receive no input or input with vanishing contrast (stimulus intensity) are treated as unobserved. Using this admittedly quite simplistic model of the input, we observed in simulations that our network model exhibits the following property: The onset of a sensory stimulus reduces the variability of the firing activity, which represents (after stimulus onset) a conditional distribution, rather than the prior distribution (see the difference between panels D and E of Figure 5. It is tempting to compare these results to the experimental finding of reduced firing rate variability after stimulus onset observed in several cortical areas [50]. We wish to point out however, that a consistent treatment of zero contrast stimuli requires more thorough modelling efforts (e.g., by explicitly adding a random variable for the stimulus intensity [35], [51]), which is not the focus of the presented work.
Virtually all high-level computational tasks that a brain has to solve can be formalized as optimization problems, that take into account a (possibly large) number of soft or hard constraints. In typical applications of probabilistic inference in science and engineering (see e.g. [52], [53]) such constraints are encoded in e.g., conditional probability tables or factors. In a biological setup they could possibly be encoded through the synaptic weights of a recurrent network of spiking neurons. The solution of such optimizations problems in a probabilistic framework via sampling, as implemented in our model, provides an alternative to deterministic solutions, as traditionally implemented in neural networks (see, e.g., [54] for the case of constraint satisfaction problems). Whereas an attractor neural network converges to one (possibly approximate) solution of the problem, a stochastic network may alternate between different approximate solutions and stay the longest at those approximate solutions that provide the best fit. This might be advantageous, as given more time a stochastic network can explore more of the state space and avoid shallow local minima. Responses to ambiguous sensory stimuli [44]–[47] might be interpreted as an optimization with soft constraints. The interpretation of human thinking as sampling process solving an inference task, recently proposed in cognitive science [28], [55], [56], further emphasizes that considering neural activity as an inferential process via sampling promises to be a fruitful approach.
Our approach builds on, and extends, previous work where recurrent networks of non-spiking stochastic neurons (commonly considered in artificial neural networks) were shown to be able to carry out probabilistic inference through Gibbs sampling [36]. In [57] a first extension of this approach to a network of recurrently connected spiking neurons had been presented. The dynamics of the recurrently connected spiking neurons are described as stepwise sampling from the posterior of a temporal Restricted Boltzmann Machine (tRBM) by introducing a clever interpretation of the temporal spike code as time varying parameters of a multivariate Gaussian distribution. Drawing one sample from the posterior of a RBM is, by construction, a trivial one-step task. In contrast to our model, the model of [57] does not produce multiple samples from a fixed posterior distribution, given the fixed input, but produces exactly one sample consisting of the temporal sequence of the hidden nodes, given a temporal input sequence. Similar temporal models, sometimes called Bayesian filtering, also underlie the important contributions of [58] and [32]. In [32] every single neuron is described as hidden Markov Model (HMM) with two states. Instead of drawing samples from the instantaneous posterior distribution using stochastic spikes, [32] presents a deterministic spike generation with the intention to convey the analog probability value rather than discrete samples. The approach presented here can be interpreted as a biologically more realistic version of Gibbs sampling for a specific class of probability distributions by taking into account a spike-based communication, finite duration PSPs and refractory mechanisms. Other implementations based on different distributions (e.g., directed graphical models) and different sampling methods (e.g., reversible MCMC methods) are of course conceivable and worth exploring.
In a computer experiment (see Figure 4, we used our proposed network to model aspects of biological vision as probabilistic inference along the lines of argumentation put forward in [16]–[18]. Our model was chosen to be quite simplistic, just to demonstrate that a number of experimental data on the dynamics of spontaneous activity [51], [59], [60] and binocular rivalry [44]–[47] can in principle be captured by this approach. The main point of the modelling study is to show that rather realistic neural dynamics can support computational functions rigorously formalized as inference via sampling.
We have also presented a model of spiking dynamics in continuous time that performs sampling from a given probability distribution. Although computer simulations of biological networks of neurons often actually use discrete time, it is desirable to also have a sound approach for understanding and describing the network sampling dynamics in continuous time, as the latter is arguable a natural framework for describing temporal processes in biology. Furthermore comparison to many existing continuous time neuron and network models of neurons is facilitated.
We have made various simplifying assumption regarding neural processes, e.g., simple symbolic postsynaptic potentials in the form of step-functions (reminiscent of plateau potentials caused by dendritic NMDA spikes [61]). More accurate models for neurons have to integrate a multitude of time constants that represent different temporal processes on the physical, molecular, and genetic level. Hence the open problem arises, to which extent this multitude of time constants and other complex dynamics can be integrated into theoretical models of neural sampling. We have gone one first step in this direction by showing that in computer simulations the two temporal processes that we have considered (refractory processes and postsynaptic potentials) can approximately be decoupled. Furthermore, we have presented simulation results suggesting that more realistic alpha-shaped, additive EPSPs are compatible with the functionality of the proposed network model.
Finally, we want to point out that the prospect of using networks of spiking neurons for probabilistic inference via sampling suggests new applications for energy-efficient spike-based and massively parallel electronic hardware that is currently under development [62], [63].
We first provide details and proofs for the neural sampling models, followed by details for the computer simulations. Then we investigate typical firing statistics of individual neurons during neural sampling and examine the approximation quality of neural sampling with different neuron and synapse models.
The simulation results shown in Figure 2, Figure 3 and Figure 4 used the biologically more realistic neuron model with the relative refractory mechanism. During all experiments the first second of simulated time was discarded as burn-in time. The full list of parameters defining the experimental setup is given in Table 1. All occurring joint probability distributions are Boltzmann distributions of the form given in equation (5). Example Python [64] scripts for neural sampling from Boltzmann distributions are available on request and will be provided on our webpage. The example code comprises networks with both absolute and relative refractory mechanism. It requires standard Python packages only and is readily executable.
In previous sections it was shown that a spiking neural network can draw samples from a given joint distribution which is in a well-defined class of probability distributions (see the neural computability condition (4)). Here, we examine some statistics of individual neurons in a sampling network which are commonly used to analyze experimental data from recordings. The spike trains and membrane potential data are taken from the simulation presented in Figure 3.
Figure 5A,B exemplarily show the distribution of the membrane potential and the interspike interval (ISI) histogram of a single neuron, namely neuron which was already considered in Figure 3B. The responses of other neurons yield qualitatively similar statistics. The bell-shaped distribution of the membrane potential is commonly observed in neurons embedded in an active network [66]. The ISI histogram reflects the reduced spiking probability immediately after an action potential due the refractory mechanism. Interspike intervals larger than the refractory time constant roughly follow an exponential distribution. Similar ISI distributions were observed during in-vivo recordings in awake, behaving monkeys [67].
Figure 5C shows a scatterplot of the coefficient of variation (CV) of the ISIs versus the average ISI for each neuron in the network. The neurons exhibited a variety of average firing rates between and . Most of the neurons responded in a highly irregular manner with a CV . Neurons with high firing rates had a slightly lower CV due to the increased influence of the refractory mechanism The dashed line marks the CV of a Poisson process, i.e., a memoryless spiking behavior. The CV of neuron is marked by a cross. The structure of this plot resembles, e.g., data from recordings in behaving macaque monkeys [68] (but note the lower average firing rate).
The theory of the neuron model with absolute refractory mechanism guarantees sampling form the correct distribution. In contrast, the theory for the neuron model with a relative refractory mechanism only shows that the sampling process is “locally correct”, i.e., that it would yield correct conditional distributions for each individual neuron if the state of the remaining network stayed constant. Therefore, the stationary distribution of the sampling process with relative refractory mechanism only provides an approximation to the target distribution. In the following we examine the approximation quality and robustness of sampling networks with different refractory mechanisms for target Boltzmann distributions with parameters randomly drawn from different distributions. Furthermore, we investigate the effect of additive PSP shapes with more realistic time courses.
We generated target Boltzmann distributions with randomly drawn weights and biases (excitabilities) and computed the similarity between these reference distributions and the corresponding neural sampling approximations. The setup of these simulations is the same as for the simulation presented in Figure 3. As we aimed to compare the distribution sampled by the network with the exact Boltzmann distribution , we reduced the number of neurons per network to . This resulted in a state space of possible network states for which the normalization constant for the target Boltzmann distribution could be computed exactly. The weight matrix was constraint to be symmetric with vanishing diagonal. Off-diagonal elements were drawn from zero-mean normal distributions with three different standard deviations , and , whereas the were sampled from the same distribution as in Figure 3. For every value of the hyperparameter we generated 100 random distributions. For Boltzmann distributions with small weights (), the RVs are nearly independent, whereas distributions with intermediate weights () show substantial statistical dependencies between RVs. For very large weights (), the probability mass of the distributions is concentrated on very few states (usually 90% on less than 10 out of the states). Hence, the range of the hyperparameter considered here covers a range a very different distributions.
The approximation quality of the sampled distribution was measured in terms of the Kullback-Leibler divergence between the target distribution and the neural approximation (21)
We estimated from samples for each simulation trial using a Laplace estimator, i.e., we added a priori to the number of occurrences of each state .
Table 2 shows the means and the standard deviations of the Kullback-Leibler divergences between the target Boltzmann distributions and the estimated approximations stemming from neural sampling networks with three different neuron and synapse models: the exact model with absolute refractory mechanism and two models with different relative refractory mechanisms shown in the bottom and middle row in Figure 2B. Additionally, as a reference, we provide the (analytically calculated) Kullback-Leibler divergences for fully factorized distributions, i.e., with correct marginals but independent variables for .
The absolute refractory model provides the best results as we expected due to the theoretical guarantee to sample from the correct distribution (the non-zero Kullback-Leibler divergence is caused by the estimation from a finite number of samples). The models with relative refractory mechanism provide faithful approximations for all values of the hyperparameter considered here. These relative refractory models are characterized by the theory to be “locally correct” and turn out to be much more accurate approximations than fully factorized distributions if substantial statistical dependencies between the RVs are present (i.e., , ). As expected, a late recovery of the refractory function is beneficial for the approximation quality of the model as it is closer to an absolute refractory mechanism. Figure 6 shows the full histograms of the Kullback-Leibler divergences for the intermediate weights group (). Systematic deviations due to the relative refractory mechanism are on the same order as the effect of estimating from finite samples (as can be seen, e.g., from a comparison with the absolute refractory model which has 0 systematic error). For completeness, we mention that the divergences of the fully factorized distributions of out of the networks with are not shown in the plot.
The theorems presented in this article assumed renewed (i.e., non-additive), rectangular PSPs. In the following we examine the effect of additive PSPs with more realistic time courses. We define additive, alpha-shaped PSPs in the following way. The influence of each presynaptic neuron on the postsynaptic membrane potential is modeled by convolving the input spikes with a kernel :(22)where for and for , and for are the spike times of the presynaptic neuron . The time constant governing the rising edge of the PSPs was set to . The time constant controlling the falling edge was chosen equal to the duration of rectangular PSPs, . The scaling parameter was set such that the time integral over a single PSP matches the time integral over the theoretically optimal rectangular PSP, i.e., . These parameters display a simple and reasonable choice for the purpose of this study (an optimization of , and is likely to yield an improved approximation quality). Figure 7A shows the resulting shape of the non-rectangular PSP. Furthermore the time course of the function caused by a single spike of neuron is shown in order to illustrate that the time constants of and of a PSP are closely related due to the assumption made above. Preliminary and non-exhaustive simulations seem to suggest that the choice yields better approximation quality than setting or ; however it is very well possible that a mismatch between and can be compensated for by adapting other parameters, e.g., the PSP magnitude or a specific choice of the refractory function . Figure 7B shows the results of an experiment, similar to the one presented in Figure 3C , with additive, alpha-shaped PSPs and relative refractory mechanism. While differences to Gibbs sampling results are visible, the spiking network still captures dependencies between the binary random variables quite well.
For a quantitative analysis of the approximation quality, we repeated the experiment of Figure 6 with additive, alpha-shaped PSPs (shown as green bars). The Kullback-Leibler divergence to the true distribution is clearly higher compared to the case of renewed, rectangular PSPs. Still networks with this more realistic synapse model account for dependencies between the random variables and yield a better approximation of than fully factorized distributions.
|
10.1371/journal.pmed.1002223 | Master Regulators of Oncogenic KRAS Response in Pancreatic Cancer: An Integrative Network Biology Analysis | KRAS is the most frequently mutated gene in pancreatic ductal adenocarcinoma (PDAC), but the mechanisms underlying the transcriptional response to oncogenic KRAS are still not fully understood. We aimed to uncover transcription factors that regulate the transcriptional response of oncogenic KRAS in pancreatic cancer and to understand their clinical relevance.
We applied a well-established network biology approach (master regulator analysis) to combine a transcriptional signature for oncogenic KRAS derived from a murine isogenic cell line with a coexpression network derived by integrating 560 human pancreatic cancer cases across seven studies. The datasets included the ICGC cohort (n = 242), the TCGA cohort (n = 178), and five smaller studies (n = 17, 25, 26, 36, and 36). 55 transcription factors were coexpressed with a significant number of genes in the transcriptional signature (gene set enrichment analysis [GSEA] p < 0.01). Community detection in the coexpression network identified 27 of the 55 transcription factors contributing to three major biological processes: Notch pathway, down-regulated Hedgehog/Wnt pathway, and cell cycle. The activities of these processes define three distinct subtypes of PDAC, which demonstrate differences in survival and mutational load as well as stromal and immune cell composition. The Hedgehog subgroup showed worst survival (hazard ratio 1.73, 95% CI 1.1 to 2.72, coxPH test p = 0.018) and the Notch subgroup the best (hazard ratio 0.62, 95% CI 0.42 to 0.93, coxPH test p = 0.019). The cell cycle subtype showed highest mutational burden (ANOVA p < 0.01) and the smallest amount of stromal admixture (ANOVA p < 2.2e–16). This study is limited by the information provided in published datasets, not all of which provide mutational profiles, survival data, or the specifics of treatment history.
Our results characterize the regulatory mechanisms underlying the transcriptional response to oncogenic KRAS and provide a framework to develop strategies for specific subtypes of this disease using current therapeutics and by identifying targets for new groups.
| Outcomes for patients diagnosed with pancreatic cancer are very poor because surgical approaches plus other current treatments are often inadequate to treat this disease.
Previous efforts have been made to subtype the disease in an effort to identify more clinically relevant groups for tailored treatment.
To improve on these “landscape” studies, we focussed on transcriptional changes induced by KRAS mutations to understand perturbed pathways and their effects on patients.
We created a transcriptional signature of oncogenic KRAS in an isogenic mouse ductal cell line.
We then combined this signature with a coexpression network derived from a large collection of pancreatic cancer cases and used a well-validated algorithm to identify the transcription factors (so-called master regulators) responsible for the signature.
The master regulators clustered into three distinct biological groups (Notch, cell cycle, and Hedgehog) characterised by significant differences in clinical survival and mutational load as well as immune cell and stromal infiltration.
Our results provide evidence that distinct modes of transcriptional reprogramming occur following KRAS-mediated transformation
This improved understanding of pancreatic cancer biology may provide novel prognostic and therapeutic opportunities to counter this devastating disease.
| Pancreatic ductal adenocarcinoma (PDAC) is the most lethal human malignancy, with a 5-y survival of 4% [1]. There are very few treatment options, with the only chance of a cure being surgical resection if the tumour is detected at an early, confined stage. Chemotherapeutic options are used in the palliative setting but have toxic side effects and do not extend survival for more than a few months. Pancreatic cancers display vast intratumoural heterogeneity with respect to their mutational profiles [2], but more than 90% of cases have a mutation in the KRAS oncogene, which almost exclusively is located in codon 12 [3].
Other highly recurring mutations in PDAC include INK4/ARF, P53, and SMAD4 [4]. Recent studies [4–6] on the genomic landscape of PDAC have identified alterations in genes related to chromatin remodeling, DNA damage repair, and axon guidance, as well as focal amplifications in druggable genes—including ERBB2, MET, and FGFR1—in a small subset of patients. Several recent studies [7–9] have described the transcriptomic landscape of PDAC and identified different subtypes with a link to survival.
Here, we took advantage of seven existing transcriptomic studies containing in total 560 samples to explore KRAS biology (see S1 Computational Analysis). The role of KRAS in PDAC initiation is well known [10], and cancer growth and survival often directly depend on it [11]. However, the mechanisms by which KRAS contributes to PDAC progression, in particular its interactions with downstream pathways, have not been equally well characterised [12]. The aims of our study were to identify transcription factors determining the transcriptional response to oncogenic KRAS and to explore what impact their activity has on the development of the disease and patient outcome.
This study did not have a protocol or prospective analysis plan. To achieve the first aim, we used master regulator analysis, a well-established network biology strategy [13–16], which combines a transcriptional signature with a coexpression network to identify key transcription factors. For the second aim, we used clustering techniques to identify patient subtypes based on transcription factor activities and characterised these subtypes by survival analysis, integration of mutation data, and methods that infer immune activity from gene expression profiles. For clarity, the Methods description is split into three main sections: defining a transcriptional signature for oncogenic KRAS, identification of master regulators of KRAS response, and characterisation of PDAC subtypes. All code and scripts to reproduce our analysis are available as annotated documents as part of the supplementary information (S1 Computational Analysis, S2 Computational Analysis).
Here, we describe how we derived the gene expression signature of oncogenic KRAS, which provides the focus point of our study.
Here, we describe the data and methods to build a coexpression network for PDAC and how it is used to identify the transcription factors determining the KRAS signature, the so-called master regulators.
The core disease processes define three subtypes, which we comprehensively characterized.
A KRAS-specific transcriptional signature enables us to understand the regulatory patterns induced in the presence of this mutation through downstream network analysis. To facilitate this analysis, we first derived a transcriptional signature of oncogenic KRAS by using a previously reported murine pancreatic ductal cell line with an inducible Lox-Stop-Lox (LSL) cassette in front of the KRASG12D oncogene to regulate transcription [35]. Comparing the transcriptomes of induced KRAS-on cells with KRAS-off control population in six replicates, we identified 667 differentially expressed probe sets (Fig 1A). Using gene set enrichment analysis, we found that the KRAS signature was enriched for many transcript changes in the pathways associated with the MAPK pathway, regulation of cell growth, and regulation of the actin cytoskeleton, which is in line with established functions of KRAS (S1 Table). Given that these cell lines are derived from murine samples, genes in the signature were mapped to their equivalent human orthologue where possible and removed from the signature if no orthologue exists.
To study which transcription factors (TFs) act as master regulators for the observed transcriptional changes, we used 560 gene expression profiles collected from seven independent studies [8, 9, 36–40] to infer a coexpression network between transcription factors and their potential targets. Details of datasets are described in the Methods section. We used different safeguards against biases potentially incurred during data integration. Every dataset was normalized independently and corrected for batch-effects if it contained sample subgroups (S2 Fig). We then used a shrinkage estimate of partial correlations [26] to infer a network between regulators and their targets in each dataset. The shrinkage estimate automatically adapts to differences in sample size and ensures reliable results across studies. A significant partial correlation corresponds to an edge in the network, and the set of all inferred targets of a regulator is called its “regulon.” We confirmed that the regulons were well conserved between the seven networks (S3 Fig). We then converted the p-values of all edges into Z-scores and combined them into a single, integrated network using Stouffer’s meta-analysis score [28], which weights the significance of an edge by the size of each study to ensure that larger cohorts contribute more to the integrated network. Thus, the final integrated network consists of edges showing consistent results across all seven studies.
In the next step, we superimposed the transcriptional KRAS signature onto the regulatory network using the VIPER (Virtual Inference of Protein-activity by Enriched Regulon analysis) algorithm [5, 9, 13, 14, 29, 40]. VIPER tests for significant overlap between genes in the transcriptional signature and genes in each regulon. 55 regulons were significantly enriched for signature genes, and we called the corresponding TFs “master regulators” (MRs) of activated oncogenic KRAS (S4 Fig). Several MRs have been previously implicated in PDAC, such as GLI3, AEBP1, and CASP5, while some of the MRs have not been associated with PDAC before, such as TCF21, TWIST1, and FOXF2.
To identify the core biological processes represented by these master regulators, we used a community detection algorithm on the transcriptional network and identified Notch signaling, Hedgehog/Wnt signalling, and cell cycle as close clusters of 27 of the 55 MRs (S5 Fig). Eight of the 27 MRs had an overall activation in protein gene up-regulation activity, and the other 19 had repressed protein activity (Fig 1D). The Notch and Hedgehog/Wnt clusters cluster closer together in the network, whereas the cell cycle MRs form a distinct group (Fig 1C).
Based on the expression of genes within each regulon, we inferred the activity of the MRs in each patient sample (S1 Computational Analysis). These activity scores clearly defined three disease processes (Fig 2). This result indicates that even though PDAC has KRAS as a dominant oncogene driver, it develops along one of three distinct paths, which can be identified by their underlying biological pathways.
The murine cell line we used to derive the transcriptional signature harboured a G12D mutation, whereas patients in the TCGA and ICGC cohorts also showed a variety of KRAS mutation types. However, we did not find any of our subtypes biased towards one particular KRAS mutation type (S7 Fig).
When comparing our disease regulatory processes to the subtypes identified by Bailey et al. [9], we found that the Hedgehog/Wnt process is overrepresented by samples from the squamous subtype (phyper = 1.1x10−11), the cell cycle process by samples from the immunogenic subtype (phyper = 7.8x10−12), and the Notch by samples from the ADEX (p-value = 8.2x10−8) and pancreatic progenitor (phyper = 6.1x10−8) subtypes (S6 Fig).
In all further analyses, we focused on the two biggest studies in our dataset collection and used the ICGC cohort [9] (n = 242) as a discovery set and the TCGA cohort [40] (n = 178) as an independent validation set. The cohorts are comparable and patients all had their pancreatic cancers resected at early stage disease (typically stage I/II).
To quantify the clinical importance of the different regulatory processes, we compared them to survival data (clinical features of both cohorts are summarized in S2 Table). In discovery and validation cohorts, the Hedgehog/Wnt group has a very poor prognosis, while the Notch group has the best prognosis (Fig 3A) (Cox proportional hazards regression model, Hedeghog/Wnt group HR = 1.73, 95% CI 1.1 to 2.72, p-value = 0.018; Notch group HR = 0.62, 95% CI 0.42 to 0.93, p-value = 0.019; when compared to the cell cycle group and after correcting for gender, age, and tumour stage). Our findings agree with previous observations linking the repression of the Hedgehog pathway to more aggressive pancreatic cancers [9, 41, 42], and this particular subtype also overlapped with the poor prognostic group found in Bailey et al. [9].
For consistency with the ICGC cohort, we substratified the TCGA dataset into treated and untreated samples. The untreated cohort demonstrated identical survival stratification to our discovery cohort, with the Notch group exhibiting favourable outcome and Hedgehog/Wnt exhibiting the worst. Although these results reciprocated the findings in the discovery cohort, the p-value was not significant at a threshold of 0.05, which is in part due to the small sample size and the low number of events.
Using a multivariate Cox proportional hazards regression analysis, we noticed that different pharmaceutical treatments had an effect on survival rates (ANOVA p-value = 0.04; after adjusting for gender, age, and tumour stage). We split the cohorts into “treated” and “nontreated” based on the treatment indicator in the TCGA data and observed that the prognosis for the Hedgehog/Wnt group was worse only if the patients are not treated (HR = 4.12 compared to cell cycle group; 95% CI 1.2 to 13.8, coxPH test p-value = 0.02; after correcting for gender, age, tumour stage, and radiation therapy indicator). In the vast majority of cases, gemcitabine, a known chemotherapy agent, was listed as listed as the “drug name” by itself (57%) or in combination with another chemotherapy drug (e.g. oxaliplatin, irinotecan, Abraxane) or one of the other listed drugs (39%), suggesting that chemotherapy was the basis of the “targeted therapy” indicator. However, it was not possible to obtain information about the specifics of the treatment regime, and we also noted that 12 cases annotated as “not treated” in the TCGA data were also listed as receivers of gemcitabine. Thus, our analyses are exploratory and no final conclusions are possible, yet there is some evidence that targeted therapeutics used in the Hedgehog/Wnt group show a pharmacological benefit to this group with a dismal prognosis.
We next investigated the molecular basis for these survival differences and compared the mutational patterns between the three subtypes (Fig 3C). We found that the cell cycle process had a significantly (ANOVA p-value < 0.01) higher mutational burden than the Hedgehog/Wnt and Notch processes—with an average of 9.4% of altered samples in the cell cycle group, compared to 4.6% and 3.2% in the Hedgehog/Wnt and Notch groups, respectively—when considering a panel of 39 key pancreatic cancer genes and pathways [5]. The same pattern was observed in the data from the TCGA cohort (S8 Fig).
To understand how well the patient subtypes were represented in experimental model systems, we used the Broad Institute’s cell line resource [43] and found that all the cell lines shared the characteristics of the cell cycle process but not of the other two processes (S9 Fig). As cell lines are artificially homogeneous cell populations, we hypothesized that Hedgehog and Notch activity are determined by the tumour’s interaction with its microenvironment.
We thus focused on differences in the tumour microenvironment to explain the differences between regulatory processes and used ESTIMATE [32] to infer stromal and immune cell admixture from gene expression data. The samples in the Hedgehog and Notch processes showed a significantly higher stromal infiltration compared to the cell cycle subtype (Fig 4B) on both the discovery and validation cohort. All three subgroups showed significantly varied levels of immune cell enrichment, with Notch being the most immunogenic and cell cycle being the least. Although Hedgehog and cell cycle samples demonstrate considerable stromal infiltration, Hedgehog is significantly less immunogenic in both cohorts (ICGC: p = 1.7e–05, TCGA: p = 7.8e–08; Wilcoxon rank sum test). By contrast, the difference in stromal content between the Hedgehog and Notch classes is considerably subtler (ICGC: p = 0.0016; TCGA p = 0.97; Wilcoxon rank sum test).
We used complementary strategies to explore the immune content of the three processes. First, we aggregated a collection of 77 immunological pathways from MSigDB. We found significant positive associations between Notch activity and 22 pathways. Of these were enrichment for T cell–related pathways, such as those pertaining to T cell activation, proliferation, and differentiation (Fig 4A). General immune system pathways, including the adaptive immune response and the positive regulation of the immune response, were similarly far more associated with Notch than Hedgehog. No significant associations were observed between T cell–related pathways and Hedgehog activity enrichment across both datasets (S10 Fig).
We then focused on dissecting the heterogeneity of leukocyte populations in the tumour samples using a leukocyte deconvolution method [32]. Both methods use a panel of cell type–specific signatures to identify which cell types are present in a mixed gene expression profile (see Methods). Results between the ICGC and TCGA cohorts correlated well, with both demonstrating the significant prevalence of infiltrating CD8+ T cells in Notch pathway–dominated samples (Fig 4C) and a dominant presence of M2 macrophages in samples exhibiting strong Hedgehog signalling.
Finally, we correlated Notch and Hedgehog activity with the expression of known immune therapy targets [44]: programmed cell death protein 1 (PD-1), programmed death-ligand 1 (PD-L1), cytotoxic T lymphocyte–associated protein 4 (CTLA4), TIGIT, and LAG3. Across both the discovery and validation cohorts, CTLA4, PD-1, and TIGIT showed significant positive correlations with Notch activity, whereas no significant associations were observed with Hedgehog activity (S10 Fig).
The first aim of our study was to identify transcription factors determining the transcriptional response to oncogenic KRAS. To address this aim, we used master regulator analysis to combine a murine gene expression signature of oncogenic KRAS with a coexpression network integrating data from 560 patients. We found that the KRAS-specific master regulators represented three main biological processes: Notch, repressed Hedgehog/Wnt, and the cell cycle. The second aim was to explore the extent to which activity of these biological processes had on the development of the disease and patient outcome. We demonstrated that the functional groupings of the transcription factors represented distinct clinical entities of PDAC, with differences in survival, mutational load, and immune activity. These three patient subtypes represent three routes of PDAC development, characterized by three different transcriptional programs.
Our study has several limitations. Although the number of samples studied here is large compared to those of other publications in PDAC, the size of the patient cohorts prevents strong conclusions about effect sizes and clinical impact. By reanalysing mainly published data, our study is constrained by the quantity and quality of the information provided in these studies. For example, mutational profiles were not available for all patients and the clinical information is often incomplete, which impedes the assessment of treatment effects. In addition, we did not have access to tissue sections for the analysed genomic profiles that could have been used to validate hypotheses about immune activity in the subtypes.
The groups we identified in general agreed with previous patient stratifications, with one surprising exception: the cell cycle group overlapped significantly with Bailey et al.’s immunogenic subtype [9], even though in our reanalysis of the data, these patients show little evidence for immune activity. This indicates that more work needs to be done to consolidate different stratification schemes into consensus subtypes of pancreatic cancer.
We found some evidence demonstrating the potential clinical importance of the groups we identified. Using survival data from two independent cohorts, we demonstrated that the Hedgehog process has the worst prognosis and the Notch process has the best prognosis. We demonstrated that the cell cycle process has the largest mutational burden of all the regulatory processes and has a smaller stromal composition. At the same time, the Notch and Hedgehog/Wnt groups showed substantial stromal contributions. We characterized stromal content and found M2 macrophages enriched in the Hedgehog/Wnt process and CD8+ T cells in the Notch process. Furthermore, analysis of checkpoint markers showed that the T cells collated with CTLA, PD-1, and TIGIT. This indicates that the Notch subtype is potentially amenable to immunotherapy. The Hedgehog/Wnt group, even though it has the worst prognosis, is potentially amenable to newer targeted therapies.
In summary, these examples show how the transcriptional signatures of KRAS-specific master regulators could be used in the clinic to stratify patients and guide therapeutic decisions.
|
10.1371/journal.pmed.1002898 | Early occupational intervention for people with low back pain in physically demanding jobs: A randomized clinical trial | Occupational medicine seeks to reduce sick leave; however, evidence for an add-on effect to usual care is sparse. The objective of the GOBACK trial was to test whether people with low back pain (LBP) in physically demanding jobs and at risk of sick leave gain additional benefit from a 3-month complex intervention that involves occupational medicine consultations, a work-related evaluation and workplace intervention plan, an optional workplace visit, and a physical activity program, over a single hospital consultation and an MRI.
We enrolled people from the capital region of Denmark to an open-label, parallel-group randomized controlled trial with a superiority design from March 2014 through December 2015. In a hospital setting 305 participants (99 women) with LBP and in physically demanding jobs were randomized to occupational intervention (n = 153) or no additional intervention (control group; n = 152) added to a single hospital consultation giving a thorough explanation of the pain (i.e., clinical examination and MRI) and instructions to stay active and continue working. Primary outcome was accumulated sick leave days due to LBP during 6 months. Secondary outcomes were changes in neuropathic pain (painDETECT questionnaire [PDQ]), pain 0–10 numerical rating scale (NRS), Fear-Avoidance Beliefs Questionnaire (FABQ), Roland–Morris Disability Questionnaire (RMDQ), Short Form Health Survey (SF-36) for physical and mental health-related quality of life (HRQoL), and self-assessed ability to continue working (range 0–10). An intention-to-treat analysis of sick leave at 6 months showed no significant difference between groups (mean difference in days suggestively in favor of no additional intervention: 3.50 [95% CI –5.08 to 12.07], P = 0.42). Both groups showed significant improvements in average pain score (NRS), disability (RMDQ), fear-avoidance beliefs about physical activities and work (FABQ), and physical HRQoL (SF-36 physical component summary); there were no significant differences between the groups in any secondary outcome. There was no statistically significant improvement in neuropathic pain (PDQ score), mental HRQoL (SF-36 mental component summary), and self-assessed ability to stay in job. Four participants could not complete the MRI or the intervention due to a claustrophobic attack or accentuated back pain. Workplace visits may be an important element in the occupational intervention, although not always needed. A per-protocol analysis that included the 40 participants in the intervention arm who received a workplace visit as part of the additional occupational intervention did not show an add-on benefit in terms of sick leave (available cases after 6 months, mean difference: –0.43 days [95% CI –12.8 to 11.94], P = 0.945). The main limitations were the small number of sick leave days taken and that the comprehensive use of MRI may limit generalization of the findings to other settings, for example, general practice.
When given a single hospital consultation and MRI, people in physically demanding jobs at risk of sick leave due to LBP did not benefit from a complex additional occupational intervention. Occupational interventions aimed at limiting biopsychological obstacles (e.g., fear-avoidance beliefs and behaviors), barriers in the workplace, and system barriers seem essential to reduce sick leave in patients with LBP. This study indicates that these obstacles and barriers may be addressed by thorough usual care.
Clinical Trials.gov: NCT02015572
| Individuals with low back pain (LBP), especially those in physically demanding jobs, are at risk of taking sick leave.
The resultant loss of productivity adds to the already considerable socioeconomic burden of LBP.
While occupational medicine interventions can help to reduce sick leave, there is limited evidence available on the effect of such interventions in people at risk of sick leave due to LBP who receive a thorough single hospital consultation.
This parallel-group randomized clinical trial compared an early occupational intervention added to usual care with usual care alone in patients with LBP and physically demanding jobs; of note, usual care included a clinical examination and magnetic resonance imaging scan to enable a thorough explanation of the pain, and recommendations to stay active in work.
The trial showed no significant difference between the treatment groups in terms of accumulated sick leave during a 6-month period, although both groups did show improvements from baseline in pain, fear-avoidance beliefs, physical quality of life, and disability at 6 months.
This study indicates that LBP interventions that include an explanation of the pain might be sufficient to limit sick leave in patients with LBP and physically demanding jobs; an approach that could be integrated into usual care and an early occupational intervention would not necessarily have to be carried out by a specialist in occupational medicine.
| The lifetime prevalence of low back pain (LBP) is about 70% [1]. In the US alone, an estimated $87 billion is spent annually on healthcare for individuals with back pain, which has been one of the fastest growing healthcare expenses [2]. Sick leave and productivity loss add to this considerable socioeconomic burden [3]. Despite considerable resources being used to prevent LBP, the incidence curve has still not declined, and this has led to greater attention on tertiary prevention [4,5]. Occupational attachment is associated with physical and mental well-being [6]. Therefore, attachment to the labor market is recommended in the management of patients who have developed LBP [7,8], and cognitive behavioral therapy focusing on biopsychosocial aspects (e.g., fear-avoidance behavior) has proven effective [5,9,10]. Combining workplace interventions with physical exercise seems to reduce LBP disability and sick leave among workers with musculoskeletal disorders [7,11]. Recent studies of interventions for chronic LBP focusing on occupational attachment found not only a decrease in disability but also a better cost-effectiveness when compared with usual care in a primary healthcare setting [10,11]. A similar effect has been seen on sick leave in a secondary healthcare setting [12]. Most guidelines for LBP care advocate the use of reassurance, analgesics, and recommendations to stay active and continue working [5]. A thorough explanation of the pain has been successful in altering fear-avoidance behavior and seems to reduce sick leave in patients with LBP [13,14]. Occupational intervention is usually given as an add-on to usual care. Therefore, the GOBACK trial was conducted to test whether individuals with LBP in physically demanding jobs at risk of sick leave gain further benefit when adding a 3-month complex early occupational intervention to a single hospital consultation, when this consultation includes a thorough explanation of the pain (i.e., clinical examination and MRI) and recommendations to stay active and continue working.
The GOBACK trial protocol has been published elsewhere [15], and the prespecified statistical analysis plan (S1 Text) and the Consolidated Standards of Reporting Trials (CONSORT) checklist (S2 Text) are provided. The GOBACK trial was a 6-month, pragmatic, single-center, open-label, parallel-group randomized clinical trial with a superiority design. It was approved by the local ethics committee (H-3-2013-161) and by the Danish Data Protection Agency (DPA: 2014-41-2673). The study was done in a hospital setting in Frederiksberg, Denmark.
The participants were recruited from an advertisement in a newspaper or referred from general practices or the study hospital’s Department of Rheumatology. Potential participants were contacted by telephone and screened for eligibility. After providing written informed consent, participants were scheduled for hospital consultation including the baseline assessment, followed by random assignment to the additional occupational intervention (intervention group) or no add-on (control group) and a 6-month follow-up assessment. An investigator (RC) not related to the intervention or outcome assessments performed the allocation.
Eligible participants were aged 18 to 65 years with a current episode of 2–4 weeks of LBP and a self-reported physically demanding job, who—independent of sick leave status and previous history of LBP—expressed concerns about their ability to continue working (minimum 30 hours/week). A physically demanding job was defined by the participants “agreeing” or “strongly agreeing” on the question whether their job was physically demanding. Exclusion criteria were pregnancy, severe somatic or psychiatric disease, cancer or metastatic disease, treatment or referral to outside providers (e.g., surgery), or contraindications for having magnetic resonance imaging (MRI).
The clinical examination and an MRI scan of the lumbar spine were included in the baseline assessment, and questionnaires were answered using a validated touchscreen without involvement of the assessors [16]. To further characterize workload, all participants’ job titles were classified using the Danish version of the International Standard Classification of Occupations (DISCO-88) [17]. A junior medical physician (BBH) or a specialist in rheumatology (LEK) performed a clinical examination and reviewed clinically relevant health-related answers, noting if further examinations were necessary. A radiologist evaluated the MRI scans, and the participants were informed of the findings by telephone 2–3 weeks after the baseline assessment. The information was presented in a way to reduce fear of a severe condition causing the back pain in the participants. The physicians who performed the clinical examinations were not allowed to participate in the occupational intervention but were allowed to take action if the physical examination revealed conditions that needed further clinical intervention. In addition, the 6-month follow-up assessment was blinded for allocation by instructing the participants not to mention their allocation.
The primary outcome was accumulated sick leave, in full days, due to LBP during the 6 months from baseline. The participants reported their sick leave weekly in a paper-based diary, and they received a weekly text message with a reminder to fill in the diary. Sick leave due to reasons other than LBP was not included in this study. Secondary outcomes, assessed at baseline and after 6 months, included a screening questionnaire to identify neuropathic components of participants’ back pain (painDETECT questionnaire [PDQ]) [18], LBP severity on a 0–10 numeric rating scale (NRS) [19], and back-pain-related disability by the 24-item Roland–Morris Disability Questionnaire (RMDQ) modified to 23 items [20] and converted to a 0–100 scale [21]. Further secondary outcomes included the Fear-Avoidance Beliefs Questionnaire (FABQ) for physical activity and work [22]; the Short Form Health Survey (SF-36) questionnaire for physical and mental health-related quality of life (HRQoL), expressed in 2 composite scores, a physical component summary score and a mental component summary score, in which the ordinal scale scores were transformed into linear scales ranging from 0 to 100 [23]; and self-assessed ability to continue working on a 0 to 10 NRS, with higher scores indicating better ability to stay in job. Overall satisfaction with the intervention was rated on a 5-point NRS, with the anchors “not at all satisfied” = 0 and “extremely satisfied” = 4.
After the baseline assessment and the consultation with the medical physician, participants were randomized 1:1 to either the additional occupational intervention or no add-on. The randomization was based on a computer-generated list (random permuted block design using block sizes of 2, 4, or 6) generated by an independent statistician and administrated by sealed envelopes. The participants were stratified by age at enrollment (<40 years or ≥40 years) and sex (male/female).
All participants received the single hospital consultation, consisting of a clinical examination and an MRI scan to give an explanation of the pain, and recommendations to stay active and continue working. The additional occupational intervention lasted 3 months and started with an initial consultation with an occupational medicine physician, who performed a work-related evaluation and provided guidance to address biopsychosocial obstacles and fear-avoidance behavior towards work. In collaboration with the participant, a workplace intervention plan was developed, with an optional workplace visit to address ergonomic obstacles. A physical therapist guided the participant to remain as active as possible by performing a 45-minute self-administered physical activity program 3 times weekly. There was a midway consultation after 6 weeks, to ensure that the workplace intervention and training were followed and necessary adjustments to the plan could be made. Finally, there was an end consultation at 3 months with an occupational medicine physician, who evaluated the intervention and provided further guidance to the participant. The participants were contacted weekly during the first month and every second week during the following 2 months to encourage the participant to follow the intervention and training. The intervention has been described in further detail in the protocol [15].
Independent of allocation, participants were able to contact the trial’s physicians in case of an adverse event (e.g., temporarily increased pain or neurological symptoms). All participants received usual care, and no treatment was withheld from the participants during the trial.
The prespecified sample-size calculation found that 127 participants would be required in each group to obtain 80% power to detect a mean difference in sick leave of 6 days between the treatment groups (2-sample pooled t test; P = 0.05), assuming a standard deviation (SD) of 17 days [24]. Expecting some dropouts during the trial period (<20%), we decided to enroll 300 participants in total (≥150 participants in each group).
Baseline characteristics are presented by group. Analyses of primary and secondary outcomes were based on the intention-to-treat population and conducted by analysis of covariance, including treatment group, age, sex, and baseline value of the relevant variable as covariates. Multiple imputation was used for missing observations. To test the robustness of the analyses, we also explored the sensitivity of the overall conclusions to various limitations of the data, assumptions, and analytic approaches to data analysis. These analyses included available case analysis and non-responder imputation using a baseline observation carried forward (BOCF) approach for missing data. Furthermore, a per-protocol analysis was conducted including only participants who received a workplace visit as part of the occupational intervention. A post hoc analysis was also conducted that included only participants who reported having very physically demanding jobs. All statistical tests were 2-sided, with P values < 0.05 considered statistically significant, and were carried out using R 3.0.1 (http://www.R-project.org; R Foundation for Statistical Computing). Additional methodological details are outlined in our statistical analysis plan (S1 Text). This trial is registered at ClinicalTrials.gov (number NCT02015572).
From 7 March 2014 to 17 December 2015, 573 potentially eligible participants were identified. Of those, a majority (n = 556) were recruited via newspaper advertisements, and 326 were assigned for baseline assessment. One participant changed to non-physical work, 7 did not attend the baseline assessment, and 13 withdrew verbal consent. Therefore, randomization assigned 153 participants to the intervention group and 152 participants to the control group (Fig 1). The mean age was 45.5 years (SD 10.3), and, as expected due to the inclusion criteria, more men than women (32.5%) were enrolled (Table 1). Overall, 226 (74.1%) participants reported more than 3 months with back pain, thus characterized as having chronic LBP at baseline [4]. Based on the DISCO-88 job title categorizations, 59.3% of the participants’ jobs were classified as manual labor; 33.8% as office work, technical, or health related (e.g., nursing); and 6.9% as administrative [17]. Mean [SD] FABQ (25.0 [7.4]) and self-assessed ability to continue working (6.24 [2.03]) indicated a high level of fear-avoidance beliefs towards work [25]. The clinical examination resulted in 15 participants being referred to additional outside providers: MRI indicated inflammatory rheumatic spine diseases in 11 participants, who were referred to a rheumatologist; 3 participants were referred to an orthopedic surgeon due to additional symptoms of hip osteoarthritis; and 1 was referred to a spine surgeon.
Thirty-six participants (11.8% of the total participants; 16 in the intervention group and 20 in the control group) dropped out during the trial; 25 gave no reason, 10 cited lack of time due to work responsibilities, and 1 was referred for treatment for a gynecological cancer (Fig 1). All the participants allocated to the additional intervention (the intervention group) attended the initial consultation and the midway consultation. Workplace visits were considered relevant for 55 (35.9%) participants. However, these visits took place in only 40 cases (27.2%), as 15 participants were concerned that the visit would increase the risk of being dismissed from their job. The physical therapist reported that 127 (83.0%) participants adhered to the physical activity plan. Reasons for non-adherence were lack of time (n = 15), back pain (n = 5), or other conditions that needed attention (n = 6). In total, 137 participants (89.5%) attended the end consultation, and 51.8% were very satisfied with the intervention (for more details, see S1 Table).
More than 70% of the participants in each group reported fewer than 7 days of sick leave due to back pain during the 6 months. The intention-to-treat analysis with imputations for non-responders of sick leave at 6 months did not show a significant add-on benefit of the occupational intervention (mean difference between groups in days: 3.50 [95% CI –5.08 to 12.07], P = 0.42; Table 2). This result was also found in an additional analysis based on the available cases at 6 months (mean difference between groups in days without imputation applied: 0.12 [95% CI –7.90 to 8.15], P = 0.97; for more details, see S2 Table).
Both groups showed improvements in average pain score (NRS), disability (RMDQ), fear-avoidance beliefs for physical activities and work (FABQ), and physical HRQoL (SF-36 physical component summary);no statistically significant difference was found between the groups. There was no statistically significant improvement in neuropathic pain (PDQ score), mental HRQoL (SF-36 mental component summary), and self-assessed ability to stay in job in either group, and no difference between the groups. This result was also found with an alternative imputation technique (baseline observation carried forward) and in analyses including available cases (for more details, see S3 Table).
We conducted additional analyses to explore whether workplace visits may be an important element in the occupational intervention and whether the intervention may be beneficial for subgroups. A per-protocol analysis that included the 40 participants who received a workplace visit as part of the additional occupational intervention was conducted. This did not show an add-on benefit in terms of sick leave (available cases after 6 months, mean difference between groups: –0.43 days [95% CI –12.8 to 11.94], P = 0.945; Table 3). A post hoc analysis was performed including only participants who reported their job to be very physically demanding. The analysis included 69 participants receiving the additional occupational intervention and 75 participants receiving no add-on. Again, there was no statistically significant benefit in terms of sick leave from the additional intervention (available cases after 6 months, mean difference between groups: –1.84 days [95% CI –13.48 to 9.79], P = 0.754; Table 4). The baseline characteristics of the 40 participants who received a workplace visit as part of the additional occupational intervention and the 144 participants who reported their job to be very physically demanding can be seen in S4 Table.
Two participants had a claustrophobic attack, and 1 participant had accentuated back pain during the MRI examination. One participant in the intervention group reported worsening thoracic back pain during the intervention; however, additional intervention was not needed.
The objective of the GOBACK trial was to evaluate whether a 3-month complex early occupational intervention, given as an add-on to a single hospital consultation, decreases sick leave among patients with LBP at risk of taking sick leave during a 6-month period. Improvements from baseline to 6 months were observed in pain, fear-avoidance beliefs, physical HRQoL, and disability in both groups, and no statistically significant differences were found between the groups in accumulated sick leave (P = 0.422). A per-protocol analysis that included the 40 participants who received a workplace visit as part of the additional occupational intervention did not show an add-on benefit in terms of sick leave (P = 0.945). A post-hoc analysis was performed including only participants who reported their job to be very physically demanding. Still, there was no statistically significant benefit of the additional intervention in terms of sick leave (P = 0.754). These findings indicate that LBP interventions comprising an explanation of the pain based on a clinical examination and MRI, combined with instructions to stay active and continue working, might be sufficient to keep patients with physically demanding jobs and at risk of sick leave due to LBP out of sick leave.
The findings in this study add to beliefs that an explanation for back pain given by a medical physician may alter fear-avoidance beliefs and behaviors and thereby increase the odds for work participation in patients with LBP [13,14]. Interventions with a focus on limiting biopsychological obstacles (e.g., fear-avoidance beliefs and behaviors), barriers in the workplace, and system barriers seem essential to reduce sick leave in patients with LBP; these are all fundamental elements of occupational interventions [7,8,22]. However, this trial indicates that these elements may be integrated into usual care and do not necessarily have to be carried out by a specialist in occupational medicine, who can focus on sick-listed patients or primary prevention.
The findings of this trial contrast with those of previous studies showing that occupational intervention for patients with LBP seems effective for reducing both short-term (i.e., 3 months) and long-term (i.e., 12 months) sick leave compared with treatment in general practice [12,26]. In the current trial, the occupational intervention was designed to reflect normal clinic practice, and, therefore, the intervention was given as an add-on to a single hospital consultation. It is well established that most patients with acute LBP recover reasonably quickly [27]. There is therefore a substantial risk that the lack of a between-group difference in this study can be explained by the self-limiting nature of this condition. This possibility may also be supported by the small amount of sick leave in both groups during the 6-month observation period. On the other hand, in this study 74.1% of participants reported more than 3 months with back pain, and were thus characterized as having chronic LBP at baseline. In patients with chronic LBP, only about 40% recover within 12 months [27], and in our study only moderate improvements in pain were observed from baseline to 6 months in both groups.
All participants were given an MRI scan of the lumbar spine and subsequently informed of its findings to personalize the explanation of the back pain and to remove fear of severe conditions. An MRI scan is not a recommended routine examination in the diagnostics of non-specific LBP [28], and in the absence of “red flag” symptoms, most guidelines advocate for several weeks of conservative care without any diagnostic imaging [28]. Studies have also indicated that liberal use of imaging in LBP may even worsen long-term outcomes in some patients [29]. In this study, MRI was used as part of the baseline assessment to exclude fear of serious diagnoses in the patients (e.g., cancer) and to highlight the benign nature of degenerative findings. It could be speculated that this “extra attention” given to all the participants explains why we did not find a benefit from the additional intervention. However, an occupational intervention should be beneficial over a single hospital consultation (with MRI), before this became a relevant add-on in the clinic.
There is strong evidence that work-related physical factors, such as manual lifting, increase the incidence of LBP [30]. By offering an early occupational intervention to individuals who reported their job to be physically demanding, we intended to include a subgroup that was likely to benefit from such an intervention [31]. Despite an average pain level of 5.6 on the 0–10 NRS and 23% of the participants using opioids on a weekly basis to manage their pain during work, surprisingly few participants (<30%) had more than 7 days of sick leave due to LBP during the 6-month observation period. To increase sensitivity, we conducted a post-hoc analysis that included only the participants who reported their job as very demanding; however, no add-on benefit in terms of sick leave was found from the additional intervention. These findings seem to support previous findings indicating that workload is not the primary risk factor for sick leave among workers with LBP [32]. Despite this, a Cochrane review found high-quality evidence to support the use of workplace interventions among workers with musculoskeletal disorders to reduce sick leave compared with usual care [7]. It is argued that a good workplace intervention should include both the worker and the employer and that this approach may be cost-effective over usual care [33]. In the current trial, workplace visits were relevant for only 55 (35.9%) participants, which adds to the previous statement that workload is only one risk factor among many for these workers. This finding is further supported by our per-protocol analysis, which included only participants (n = 40) who had a workplace visit as part of their occupational intervention, and in which still no benefit from the intervention was found. Furthermore, 15 participants refused to have a workplace visit, as they were concerned the visit would increase the risk of dismissal from their job.
It may be argued that personal fitness among workers may increase the ability to adapt to the physical demands at work. For workers with LBP, physical exercise seems to reduce the number of recurrences of LBP or prolong the time to recurrence, although no particular type of exercise seems superior [34]. For this reason, our intervention’s physical exercise plan was not standardized but planned individually for each participant to increase adherence. This resulted in 83% of participants following the plan to an acceptable level after 3 months, although this may not represent a long-lasting change in behavior. To address these and other questions further, follow-up and additional analyses are planned.
The strengths of this trial are the randomized clinical trial design, the early occupational intervention to reduce sick leave, the high rate of participation during the 6 months from baseline, having sick leave as the primary outcome, the blinded assessment, and the inclusion of individuals with LBP in physically demanding jobs, which is a highly relevant subgroup for an additional occupational intervention. A significant number of participants had previous history of LBP, and therefore may already have had usual care to a varying degree before enrollment. This may have limited the chances of detecting a benefit of the complex occupational intervention. A major limitation of this trial is the skewness of data and small amount of sick leave in both groups during the 6-month observation period, which increases the risk of a floor effect. The small amount of sick leave may be explained by the inclusion of non-sick-listed patients.
Another reason for the small amount of sick leave in both groups could be that an MRI scan is not a routine examination in the diagnostics of non-specific LBP [28], and, therefore, some participants with a low risk of sick leave may have attended the trial with the purpose of having an MRI scan performed. The trial’s extensive use of MRI and the use of a single consultation in a hospital may limit the generalization of the findings to other settings, for example, general practice.
When given an explanation for the pain based on a clinical examination and an MRI scan, followed by instructions to stay active and continue working, workers in physically demanding jobs at risk of sick leave due to LBP do not benefit from a 3-month complex early additional occupational intervention. This indicates that occupational elements may be integrated into usual care and do not necessarily have to be carried out by a specialist in occupational medicine, who can focus on sick-listed patients or primary prevention.
|
10.1371/journal.ppat.1007802 | Strength of T cell signaling regulates HIV-1 replication and establishment of latency | A major barrier to curing HIV-1 is the long-lived latent reservoir that supports re-emergence of HIV-1 upon treatment interruption. Targeting this reservoir will require mechanistic insights into the establishment and maintenance of HIV-1 latency. Whether T cell signaling at the time of HIV-1 infection influences productive replication or latency is not fully understood. We used a panel of chimeric antigen receptors (CARs) with different ligand binding affinities to induce a range of signaling strengths to model differential T cell receptor signaling at the time of HIV-1 infection. Stimulation of T cell lines or primary CD4+ T cells expressing chimeric antigen receptors supported HIV-1 infection regardless of affinity for ligand; however, only signaling by the highest affinity receptor facilitated HIV-1 expression. Activation of chimeric antigen receptors that had intermediate and low binding affinities did not support provirus transcription, suggesting that a minimal signal is required for optimal HIV-1 expression. In addition, strong signaling at the time of infection produced a latent population that was readily inducible, whereas latent cells generated in response to weaker signals were not easily reversed. Chromatin immunoprecipitation showed HIV-1 transcription was limited by transcriptional elongation and that robust signaling decreased the presence of negative elongation factor, a pausing factor, by more than 80%. These studies demonstrate that T cell signaling influences HIV-1 infection and the establishment of different subsets of latently infected cells, which may have implications for targeting the HIV-1 reservoir.
| Activation of CD4+ T cells facilitates HIV-1 infection; however, whether there are minimal signals required for the establishment of infection, replication, and latency has not been explored. To determine how T cell signaling influences HIV-1 infection and the generation of latently infected cells, we used chimeric antigen receptors to create a tunable model. Stronger signals result in robust HIV-1 expression and an inducible latent population. Minimal signals predispose cells towards latent infections that are refractory to reversal. We discovered that repression of HIV-1 transcription immediately after infection is due to RNA polymerase II pausing and inefficient transcription elongation. These studies demonstrate that signaling events influence the course of HIV-1 infection and have implications for cure strategies. They also provide a mechanistic explanation for why a significant portion of the HIV-1 latent reservoir is not responsive to latency reversing agents which function by modifying chromatin.
| HIV-1 persists in a transcriptionally silent latent state in long-lived memory T cells. Although antiretroviral therapies (ART) suppress HIV-1 replication, interruption of treatment results in rapid viral rebound. Therefore, HIV-1 patients must remain on ART indefinitely, despite long-term side effects, development of treatment resistance, and viral-induced inflammation [1–3]. For this reason, one strategy currently being explored for cure efforts is “shock and kill,” in which latent HIV-1 is reactivated in conjunction with ART using latency-reversing agents (LRAs). Following reactivation, infected cells are predicted to be eliminated by HIV-specific immunity or virally induced apoptosis. However, clinical trials using LRAs have only minimally perturbed the size of the viral reservoir [4–6].
A cure for latent HIV-1 will require a better understanding of the biochemical factors involved in regulating proviral transcription. Latency in chronically infected primary cells and cell lines is regulated by multiple transcriptional mechanisms including NF-κB activation, chromatin accessibility, provirus transcription initiation, Tat availability, P-TEFb sequestration, and transcriptional elongation [7–11]. However, what is not understood is how latency is initially established within a cell and if events at the time of HIV-1 infection influence the transcriptional status of the provirus. These questions are relevant since the latent reservoir is established within the first two weeks of infection [12,13]. New cure strategies will need to limit the size of the reservoir at early time points.
One mechanism that could predispose HIV-1 towards active replication or transcriptional repression and latency is signaling through the T cell receptor (TCR). Engagement of the TCR and costimulatory CD28 molecule result in a multitude of cellular outcomes that influence HIV-1 replication including cytoskeleton reorganization, the activation of transcription factors, enhanced RNA polymerase II (RNAP II) processivity, and chromatin remodeling [14–16]. We hypothesized that the magnitude of T cell signaling during HIV-1 infection will dictate the course of the infection. In order to manipulate signal strength received by a T cell at the time of HIV-1 infection, we utilized chimeric antigen receptors (CARs) that recapitulate T cell receptor and CD28 signaling. By modulating the affinity with which these CARs bind to their ligand, we can differentially deliver signals to target cells.
Using these CARs, we demonstrate that stronger T cell signaling at the time of HIV-1 infection increases subsequent HIV-1 transcription and replication. Robust signals also facilitated the formation of latently infected cells that were readily inducible upon secondary stimulation. Minimal signaling through CARs, although sufficient for HIV-1 integration, failed to support viral replication and generated a deep-seated latent infection. Transcriptional elongation of HIV-1 provirus was limited by RNAPII pausing in the absence of CAR signaling; however, strong CAR signaling correlated with decreased negative elongation factor (NELF) binding and enhanced RNAPII processivity. Our results suggest a model in which signaling strength influences HIV-1 transcription and establishment of latency at the time of initial infection of CD4+ T cells.
To examine how signaling cascades downstream from the T cell receptor regulate HIV-1 transcription we utilized CARs (Fig 1A). Intracellular signaling domains for the CARs include CD3ζ with its immunoreceptor tyrosine-based activation motifs (ITAMs) and the CD28 costimulatory domain with its four critical tyrosine residues [17]. Furthermore, a mCherry tag provides a marker for positive selection of CAR+ cells. The extracellular ligand-binding domains of the CARs consist of a single chain variable fragment (scFv) that recognizes receptor tyrosine-protein kinase erbB-2 (Her2) [18,19]. By using different scFvs, a library of CARs with binding affinities for Her2 ligand spanning three logs was generated (Fig 1B). CARs were transduced into Jurkat T cells and primary CD4+ T cells. By enriching for mCherry, we obtained CAR+ populations that were >90% pure (Fig 1C).
We confirmed that signaling through CARs mimicked aspects of TCR signaling and supported differential changes in both downstream gene expression and T cell phenotypes. CD69, a transmembrane lectin and a marker for CD4+ T cell activation, was monitored by flow cytometry before and after receptor activation with Her2 ligand (Fig 2A). Primary CD4+ T cells transduced with either the low affinity or the high affinity receptors were stimulated with plate-bound Her2 ligand for 24 h. In the absence of ligand, less than 7% of the cells were positive for CD69, verifying that there is no ectopic CAR signaling. Activating cells with Her2 induced CD69 expression in the low affinity and high affinity receptors relative to their affinity for ligand. In addition, we analyzed the ability of the receptors to generate T cell subsets one week post stimulation by determining the expression of CCR7, a lymph node homing receptor, and CD45RA, a marker that is downregulated on memory subsets following activation through endogenous TCR. Similar expansion of CCR7+ CD45RA- T cell populations was observed in cells stimulated through the TCR and both the high and low affinity CARs (Fig 2B).
Finally, microarray analysis was performed to determine if CAR and TCR signaling resulted in similar global gene expression changes. A heatmap of genes whose expression was significantly altered compared to baseline upon activation through either the high and low affinity CARs or treatment with anti-CD3 and anti-CD28 is shown in S1 Fig. A spectrum of changes in gene expression downstream from the T cell receptor was observed that included subsets of genes that were induced or repressed to similar extents by the TCR and CARs as well as more graded differential responses in which the high and low affinity CARs induced intermediate changes as compared to the T cell receptor. We did observe donor-to-donor variation, which was expected. This may reflect intrinsic donor variation as well as sexual dimorphism because of the inclusion of 1 male and 2 female donors. Together with the CD69 analysis and memory cell markers, these data demonstrate that signaling through the chimeric antigen receptors facilitate a range of cellular effects associated with T cell activation and that the CARs can be used as tools to modulate T cell signals.
To determine whether T cell signaling influences viral infection, Jurkat T cells expressing low affinity or high affinity CARs were plated on Her2-coated wells and simultaneously infected with VSV-G pseudotyped NL4-3.Luc, a single-cycle HIV-1 clone which contains a luciferase reporter in place of Nef. VSV-G allowed us to bypass potentially confounding effects from receptor/chemokine receptor signaling due to gp120 binding and focus specifically on CAR-associated signaling cascades. To assess whether signaling influenced the establishment of infection, we measured levels of HIV-1 proviral DNA using a previously described nested Alu-PCR approach [20]. We modified the assay by designing primers to luciferase to estimate the relative frequency of HIV-1 integration without confounding signals from the lentiviral vectors used to express the CARs (see Materials and Methods). CAR-associated signaling did not affect the infection of Jurkat cells since we detected comparable levels of provirus regardless of the presence or absence of CAR ligand (Fig 3A). When HIV-1 expression was measured by luciferase activity, Jurkat cells infected in the context of strong T cell signaling expressed greater than 10-fold more HIV-1 compared to untreated controls (Fig 3B). In contrast, engagement of the low affinity receptor led to a modest 3-fold expression compared to unstimulated cells despite a similar proviral load as the high affinity CAR-expressing cells. These data indicate that strong T cell signaling at the time of infection facilitates HIV-1 expression without enhancing provirus integration.
We confirmed that these differences were due to downstream signaling emanating from the CARs by using the Src kinase inhibitor PP2. In the presence of PP2, the increase in HIV-1 expression upon cellular stimulation was attenuated, consistent with T cell signaling as a regulator of HIV-1 expression (S2 Fig). The pharmacologically inactive version of this inhibitor, PP3, had no effect on the ability of CARs to influence HIV-1 expression.
We validated these results using primary CD4+ T cells that were transduced with either the low affinity or high affinity CAR. Following transduction, cells were allowed to return to a resting state as monitored by low CD69 expression before infection with HIV-1 in the absence or presence of the ligand Her2. Consistent with the data from Jurkat cells, similar levels of proviral DNA were detected in primary T cells regardless of CAR signaling (Fig 3C). Depending on the donor, cells that received robust stimulation at the time of infection expressed 2.5- to 11-fold more HIV-1 compared to untreated controls. Stimulating through the low-affinity CARs led to more modest luciferase expression when compared to untreated cells (Fig 3D).
To gain insight into whether there is a threshold or minimal T cell signal required for HIV-1 infection and replication, we transduced Jurkat T cells with CARs that spanned a range of binding affinities (Fig 1B). These cells were infected with NL4-3.Luc as described above in the absence or presence of Her2. Although the high affinity condition supported HIV-1 infection and transcription, the intermediate and low affinity receptors did not support HIV-1 expression (Fig 4A). This was despite similar levels of infection as determined by measuring proviral DNA (Fig 4B) and a greater than 10-fold increase in binding affinity for the Her2 ligand above that of the low affinity CARs. These data suggest that T cell signaling controls HIV-1 expression by a digital on/off mechanism since viral expression does not linearly correlate with signal strength.
We hypothesized that differential T cell signaling during infection alters the size of the inducible latent reservoir. To examine this, we infected CAR-expressing primary CD4+ T cells with VSV-G pseudotyped BRU-dENV-GFP in the presence of Her2 ligand. One week post infection, cells were sorted for both mCherry expression as a marker for the CAR and lack of GFP expression in order to enrich for latently infected cells. CARpos/GFPneg cells were reactivated with PMA plus ionomycin or left unstimulated to control for spontaneous HIV-1 reactivation (Fig 5A). PMA plus ionomycin significantly reactivated HIV-1 expression within cells that had been initially infected in the context of high-affinity CAR, resulting in a 3- to 9-fold increase in the percentage of GFP positive cells (Fig 5B) and a 1000-fold induction of HIV-1 mRNA measured by qRT-PCR (Fig 5C, S3 Fig). However, the observed reactivation of HIV-1 was modest in cells infected at the time of stimulation through the low affinity CAR. Less than a 2-fold change was observed in the percentage of GFP+ cells, and only a 200-fold induction of HIV-1 mRNA was detected in reactivated latently infected cells expressing low affinity CARs. Thus, there was a ~5-fold increase in HIV-1 mRNA induction for cells expressing high affinity CAR upon reactivation compared to reactivation in cells expressing low affinity CAR (Fig 5C, S3 Fig).
A panel of latency reversing agents were tested for their abilities to reactivate the latently infected cells generated by the different CARs. Cells were reactivated with antibodies to CD3 and CD28, the HDAC inhibitor SBHA, and the PKC agonist Bryostatin (Fig 5B). Secondary stimulation through the endogenous T cell receptor with anti-CD3+CD28 reactivated latent HIV-1 in cells that had been infected and stimulated through the high-affinity receptor, resulting in a 3- to 7-fold increase in the percentage of GFP positive cells. However, anti-CD3+CD28 treatment resulted in either no reactivation or a modest 1.8-fold reactivation in cells stimulated through the low affinity CAR at the time of HIV-1 infection. SBHA did not strongly induce HIV-1 expression but there was a trend of greater virus reactivation in cells that had received stronger TCR signaling at the time of HIV-1 infection. Treatment with Bryostatin did not lead to robust activation for either low affinity or high affinity CAR-expressing cells. In general, there appeared to be different reservoir sensitivities to latency reversal agents between cells that had received strong or weak signaling during infection, especially when comparing cells derived from the same donor. Our data suggest that despite comparable amounts of integrated HIV-1 proviruses, robust signaling at the time of infection was not only necessary for active proviral transcription but also supported the generation of a population of latently infected cells that could be readily induced to express HIV-1. The population of latent cells generated in response to weaker CAR signaling was more resistant to latency reversal suggesting that HIV-1 in these cells was strongly repressed.
We were interested in mechanisms that governed HIV-1 repression following integration in the absence of sufficient T cell signaling; therefore, we examined the binding of transcriptional regulators on the HIV-1 LTR by chromatin immunoprecipitation (ChIP). Jurkat T cells expressing low or high affinity CARs were infected with NL4-3.Luc in the absence or presence of Her2 ligand. One day post-infection, cells were fixed and chromatin was prepared for ChIP.
Since HIV-1 proviral latency correlates with a positioned nucleosome that is downstream of the transcriptional start site, we explored whether the LTR was associated with post-translationally modified histones as an indicator of chromatin organization. ChIPs for acetylated histone H3 showed no significant difference in binding of the HIV-1 LTR between cells infected in the absence or presence of T cell signaling (Fig 6A). Therefore, chromatin accessibility does not appear to be limiting HIV-1 proviral transcription following infection.
We then examined RNAP II processivity by measuring RNAP II occupancy at multiple points, including the transcriptional start site and downstream in the HIV-1 tat gene. RNAP II was detected at the HIV-1 transcriptional start site whether cells were activated through a CAR or were unstimulated (Fig 6B). However, signaling through the high affinity receptor resulted in an increase in downstream RNAP II by greater than 4-fold, whereas only modest levels of RNAP II were found downstream in the absence of signals or following weak signaling (Fig 6C).
Since these data indicated a role for transcriptional pausing, we examined if CAR signaling altered the presence of the pausing factor negative elongation factor (NELF) at the HIV-1 transcriptional start site. Using ChIPs, we determined that signaling through the high affinity receptor diminished binding of NELF at the HIV-1 LTR by greater than 85% (Fig 6D). These data support a model in which a lack of robust T cell signaling limits HIV-1 transcription by establishing a paused polymerase complex.
Previous studies suggest that cell signaling may be a key regulator of HIV-1 expression and latency. The latent reservoir is enriched for antigen specific T cells, including those that respond to CMV, HSV, tuberculosis, and HIV [21–25]. Furthermore, the use of superantigens during viral entry increases HIV-1 replication [26]. Partial activation, cellular polarization, cell-to-cell contact, and/or infection of resting quiescent cells through perturbation have also been suggested to bias infections towards latency [11,27–31]. Therefore, the extent of cell activation is a key determinant in regulating the course of HIV-1 infection including the formation of the reservoir.
We have shown that differential signaling through CARs, which mimic TCR signaling, influences HIV-1 transcription and latency. In the lymph node, a primary site for both HIV-1 replication and the persistent latent reservoir [32–34], T cells will sample lymph node resident cells in search for antigen. Some of these interactions, facilitated by the presentation of the T cell cognate antigen, will result in robust T cell activation, clonal expansion, and changes in gene expression. However, most MHC complexes will lack cognate antigen and initiate weak signaling [35,36]. Using multiple CARs whose affinities for the Her2 ligand span several logs, we can deliver a range of signaling inputs to model the spectrum of T cell receptor signaling events. Our data indicates that stronger T cell activation at the time of infection, which would be more similar to antigen specific responses, correlates with robust HIV-1 expression as well as the establishment of inducible latently infected cells. We validated these findings in both Jurkat cell lines and primary CD4+ T cells derived from multiple donors although the magnitude of responses from primary cells was more variable as would be predicted. Additional factors may compensate for suboptimal T cell receptor signaling including cytokine-induced stimulation and interactions with antigen presenting cells that would engage both costimulatory molecules and inhibitory receptors. The contribution of the T cell receptor pathway and how this is integrated with other signaling events to influence HIV-1 infection and latency is a critical question that needs to be addressed.
Having a library of CARs with a range of binding affinities allowed us to determine if HIV-1 responds to signaling in an analog fashion correlating with signal input or is digitally regulated by specific thresholds resulting in all-or-none responses [37]. Signaling through the CARs with affinities that were intermediate did not support active transcription despite a greater than 10-fold increase in binding affinity compared to our low affinity receptor. These results would suggest that TCR signaling provides more of an on/off switch in regulating HIV-1 transcription and that there exist signaling thresholds that must be overcome to assure efficient HIV-1 transcription and replication.
Signal transduction and gene expression are inherently noisy processes, and stochastic events are hypothesized to drive HIV-1 latency. That latency and HIV-1 replication are driven by episodic bursts of proviral transcription and Tat levels has been supported by mathematical modeling and experiments using engineered virus models [38–40]. Even if latency is driven by random fluctuations of provirus transcription, T cell associated signals are strong modulators of noise, and targeting these pathways could enhance treatments directed at HIV-1 reactivation [41]. Weak signaling, such as those induced by the low affinity chimeric antigen receptors, may be inadequate to alter the inherent noise within the system, whereas robust TCR signals through the high affinity CAR increase the probability of stochastic events. It is important to appreciate that although signaling and transcription are subject to stochastic variation, these are coordinated and combinatorial processes that lead to defined patterns of gene expression and phenotypic outcomes [42].
Regulated aspects of transcription include assembly of multi-subunit complexes such as RNAP II and associated cofactors, chromatin, and transcription factors at the LTR. Our data suggest that the association of NELF with RNAP II is regulated by TCR signaling. Multiple positive and negative signals are known to converge on NELF-driven transcriptional pausing. P-TEFb relieves NELF repression through phosphorylation [43] and is itself regulated by cellular stress and signals [44–46]. Furthermore, we have shown that NELF interacts with co-repressors including NCoR1-GPS2-HDAC3 at the HIV-1 promoter [47] which may reinforce HIV-1 latency, especially during chronic infection, by facilitating post-translational modifications of histones and chromatin organization.
We propose that strength of signal at the time of infection acts as a bifurcating event leading to either robust transcription and the establishment of an inducible latent reservoir or minimal transcription and deep-seated latency. Our observations are consistent with the previous characterization of patient reservoirs that identified three subsets of latently infected cells: a small population of cells carrying inducible provirus, a larger population of cells with intact proviruses that are difficult to reactivate, and many defective proviruses [48]. Successful purging of the latent reservoir may require the use of a cocktail of latency reversing agents or the development of novel strategies to block reactivation [49–51].
Jurkat CD4+ T cells (E6-1) and human embryonic kidney 293T cells were obtained from American Type Culture Collection (ATCC). Jurkat cells were cultured in RPMI 1640, 5% FBS (Corning, Inc.), 100 units/mL penicillin (Invitrogen), 100 μg/mL streptomycin (Invitrogen), and 2mM L-glutamine (Invitrogen). HEK293T cells were cultured in Dulbecco’s Modified Eagle Medium, 10% FBS, 100 units/mL penicillin, 100 μg/mL streptomycin, and 2mM L-glutamine. Cells were grown at 37° C with 5% CO2.
Primary CD4+ T cells were derived from de-identified healthy blood leukapheresis packs purchased from NY Biologic. Mononuclear cells were enriched from leukopaks by centrifugating through Histopaque gradient (Sigma-Aldrich). CD4+ T cells were isolated by negative selection using EasySep Human CD4+ T Cell Enrichment Kits from STEMCELL Technologies. CD4+ cells were maintained in RPMI 1640, 10% FBS, 100 units/mL penicillin, 100 μg/mL streptomycin, and 2mM L-glutamine at 37° C with 5% CO2. Prior to transduction with CARs, primary cells were supplemented with 10 units/mL IL-2 and 10 ng/mL IL-7. Following transduction, IL-2 was removed from most culture conditions. All cells and cell lines were split every 2–3 days.
CARs were driven by a SFFV promoter in the lentiviral vector pHR [18,19]. pNL4-3.Luc.R-E- was obtained from NIH AIDS Reagent Program. BRU-ΔEnv-GFP has been described before [52]. Lentiviruses were made by transfection of vectors, VSV-G, Rev, Tat, and Gag-Pol constructs into HEK293T cells with 45μL polyethylenimine (1 mg/mL) per 6x106 cells. Supernatants were collected, filtered with 0.45μm syringe filter (Corning), concentrated by centrifuging through a 20% sucrose gradient, and titered with CEM cells [53]. We used a range of multiplicity of infections, but most viruses and lentiviruses within this paper were concentrated to approximately 1x106 IU/mL. HIV-1 viruses were made similarly but only required the viral plasmid and VSV-G.
For transductions with CAR vectors, a minimum of 1x106 primary and Jurkat cells were stimulated for 5–6 h with 10 μg/mL PHA, washed in PBS, and spinoculated with lentivirus and 5 μg/mL polybrene (Millipore) at 1200g for 90 min. Cells were then supplemented with fresh RPMI and IL-7, cultured overnight, and washed in PBS 18 h later. Cells were rested for one week to return to a resting state as confirmed by low CD69 expression prior to HIV-1 infection.
Non-tissue culture treated plates were coated overnight at 37°C with 1 μg/mL Her2 (Recombinant Human ErbB2/Her2 Fc Chimera Protein from R&D Systems, 1129-ER). Her2 solution was removed from wells, plates were washed 3 times in PBS, and wells were blocked for 1 h with a 5% FBS-PBS solution.
Jurkat or primary CD4+ T cells were infected and simultaneously plated in Her2-treated wells. For experiments in which latently infected cells were generated, cells were spinoculated in the Her2-treated wells at 1200xg for 90 min and then supplemented with fresh RPMI and IL-7. Following overnight infection, cells were washed and either lysed or maintained in fresh media in the absence of Her2.
For reactivation of latent cells, mCherry (CAR) positive and GFP (HIV) negative cells were sorted at 6 or 7 days post HIV-1 infection. Cells were cultured with the following concentrations of LRAs: 5 ng/mL PMA (Fisher Scientific) and either 10 or 100uM ionomycin (Sigma-Aldrich) for 2.5 h, Dynabeads human T-activator CD3/CD28 beads at a ratio to cells of 1:1 for 24 h, 50 μM SBHA (Sigma-Aldrich) for 24 h, and 25nM Bryostatin (Sigma-Aldrich) for 24 h. Cells reactivated with PMA were washed in PBS and re-plated in media. All reactivated cells were incubated with 10 ng/mL IL-7. Cells were cultured overnight prior to fixation for flow analysis.
For some experiments, cells were treated with 10 μM PP2 or PP3 (Calbiochem—Millipore Sigma) at the time of infection.
Flow data were collected on an LSRII from BD Biosciences. Zombie UV Fixable Viability Kit (BioLegend) was used as live/dead stain for reactivation experiments. We had a range of 250 to 6000 events per data point and a minimum cut-off of 250 events in Fig 5B. The mean number of all events was 1364 and the median number was 678. All cells were washed and fixed in a final concentration of 2% paraformaldehyde prior to analysis. Cell sorting was performed on a MoFlo Astrios from Beckman Coulter. All flow experiments performed at Boston University School of Medicine Flow Cytometry Core Facility.
Cell activation and phenotypes were determined by CD69 expression (Brilliant Violet 421 anti-human CD69 antibody; Clone FN50, BioLegend) and CCR7 and CD45RA expression (Pe/Cy7 anti-human CCR7 antibody; Clone G043H7, BioLegend and PerCP/Cy5.5 anti-human CD45RA antibody; Clone HI100, BioLegend). We had a minimum of 1000 events to be included as a data point in Fig 2.
Primary CD4+ T cells were transduced with CARs and allowed to return to a resting state for 1 week prior to cell sort. Cells were then stimulated overnight with plate-bound Her2 as described above. Untransduced CD4+ T cells were left unstimulated or were plated in a solution of 1 μg/mL CD28 (Mouse Anti-Human CD28, #555725, BD Biosciences) on previously coated wells of 1 μg/mL CD3 (Mouse Anti-Human CD3, #555329, BD Biosciences). We used a minimum input of 2.5x104 primary cells per experimental condition. Cells were then washed in PBS and lysed for RNA extraction using Qiagen miRNeasy Mini Kit (#217004). Microarrays and statistical support were provided by BU Microarray and Sequence Resource Core Facility. cDNA was made and samples were run on a Human Clariom S Array. Human Clariom S CEL files were normalized to unstimulated cells to produce gene-level expression values using the implementation of the Robust Multiarray Average (RMA) [54] in the affy package (version 1.36.1) [55] included in the Bioconductor software suite (version 2.11) [56] and an Entrez Gene–specific probe set mapping (21.0.0) from the Molecular and Behavioral Neuroscience Institute (Brainarray) at the University of Michigan [57, 58]. Array quality was assessed by computing Relative Log Expression (RLE) and Normalized Unscaled Error (NUSE) using the affyPLM package (version 1.34.0). Analyses of variance were performed using the f.pvalue function in the sva package (version 3.4.0). Differential expression was assessed by performing Student's t test on the coefficients of linear models created using the lmFit function in the limma package (version 3.14.4). In this way, a one-way ANOVA p value was obtained using a linear mixed effects modeling approach to account for differences between donors.
Correction for multiple hypothesis testing was accomplished using the Benjamini-Hochberg false discovery rate (FDR) [59]. All microarray analyses were performed using the R environment for statistical computing (version 2.15.1). All genes with FDR q values below 0.01 were plotted on a heatmap and arbitrarily separated into 5 clusters based on expression profiles.
Determination of gender was based upon log2 expression of the Y-linked genes DDX3Y, KDM5D, RPS4Y1, USP9Y, and UTY.
4x105 Jurkat cells were washed and lysed for luciferase analysis 24 h post infection, while 4x105 primary T cells were measured at 4 days post infection. Luciferin (Promega) was added and luciferase activity was measured via BioTek Synergy HT Microplate Reader.
A minimum input of 8x105 cells per experimental condition was lysed in Tris-EDTA buffer prior to Alu-PCR. Nested PCR strategy was adapted from Agosto et al., 2007 [20]. Briefly, integrated HIV-1 DNA was amplified using forward primers for the luciferase sequence and reverse primers for human Alu (see S1 File). The first reaction was performed on a TProfessional Thermocycler from Biometra according to the following conditions: 4 m at 95° followed by 20 cycles of 15 s at 93°C, 15 s at 50°C, and 2.5 m at 70°C. A second round of amplification was then performed using a forward primer, a reverse primer, and a probe for real time PCR within the HIV-1 3’ R / U5 region (see S1 File). The amount of amplified copies of HIV-1 was determined based on an NL4-3 plasmid copy standard. The second reaction was performed on an Applied Biosystems QuantStudio 3 Real-Time PCR system with heating for 4 m at 95° and real-time PCR conditions of denaturation for 15 s at 95°C, annealing for 30 s at 60°C, and extension for 1 m at 72°C.
We lysed 1.5x105 cells per experimental condition in TRIzol Reagent (Invitrogen). RT-PCR for HIV-1 mRNA was performed using forward primers and reverse primers for unspliced HIV-1 tat, and all values were normalized against beta-actin as a housekeeping gene (see S1 File). The reaction was performed on an Applied Biosystems QuantStudio 3 Real-Time PCR system with heating for 15 m at 94°C and real-time PCR conditions of denaturation for 15 s at 94°C, annealing for 30 s at 60°C, and extension for 30 s at 72°C.
5x106 Jurkat cells were infected with NL4-3.Luc. ChIP was performed 24 h later according to Natarajan et al., 2013 [47] with the addition of a nuclei isolation step using Farnham Lysis Buffer prior to sonication with a Bioruptor Pico. Specific details are listed in S1 File. Antibodies used included anti-NELF-d (Antibody TH1L from Proteintech Group), anti-RNA Polymerase II antibody (Clone N20 from Santa Cruz Biotechnology), anti-histone H3 antibody (Product 06–599 from Millipore Sigma), and Normal Rabbit IgG (Product 12–370 from Millipore Sigma).
Primers used for the transcriptional start site include the forward primer at +30 and the reverse primer at +239. Primers used for transcriptional elongation include the forward and reverse primers within the tat gene (see S1 File).
Except for microarray analysis detailed above, all statistical analysis performed using unpaired Student’s t test with significance thresholds of *p<0.01, **p<0.001, and ***p<0.0001. Because our experiments were performed on a sample population under the same conditions, we assumed that our data would be normally distributed. In agreement with this assumption, all experimental data points were less than 2 standard deviations from the mean. Where appropriate, normality was tested with a Shapiro-Wilk test.
|
10.1371/journal.pntd.0004677 | The wMel Strain of Wolbachia Reduces Transmission of Chikungunya Virus in Aedes aegypti | New approaches to preventing chikungunya virus (CHIKV) are needed because current methods are limited to controlling mosquito populations, and they have not prevented the invasion of this virus into new locales, nor have they been sufficient to control the virus upon arrival. A promising candidate for arbovirus control and prevention relies on the introduction of the intracellular bacterium Wolbachia into Aedes aegypti mosquitoes. This primarily has been proposed as a tool to control dengue virus (DENV) transmission; however, evidence suggests Wolbachia infections confer protection for Ae. aegypti against CHIKV. Although this approach holds much promise for limiting virus transmission, at present our understanding of the ability of CHIKV to infect, disseminate, and be transmitted by wMel-infected Ae. aegypti currently being used at Wolbachia release sites is limited.
Using Ae. aegypti infected with the wMel strain of Wolbachia that are being released in Medellin, Colombia, we report that these mosquitoes have reduced vector competence for CHIKV, even with extremely high viral titers in the bloodmeal. In addition, we examined the dynamics of CHIKV infection over the course of four to seven days post feeding. Wolbachia-infected mosquitoes remained non-infective over the duration of seven days, i.e., no infectious virus was detected in the saliva when exposed to bloodmeals of moderate viremia, but CHIKV-exposed, wild type mosquitoes did have viral loads in the saliva consistent with what has been reported elsewhere. Finally, the presence of wMel infection had no impact on the lifespan of mosquitoes as compared to wild type mosquitoes following CHIKV infection.
These results could have an impact on vector control strategies in areas where Ae. aegypti are transmitting both DENV and CHIKV; i.e., they argue for further exploration, both in the laboratory and the field, on the feasibility of expanding this technology beyond DENV.
| New approaches to preventing chikungunya virus (CHIKV) infection are needed because the endemic range of this virus is expanding and because current methods are limited to controlling mosquito populations, and this approach has not effectively controlled this virus. A promising candidate for arbovirus control and prevention relies on the introduction of the intracellular bacterium Wolbachia into Aedes aegypti mosquitoes. Wolbachia biocontrol has advanced from laboratory experiments demonstrating that Wolbachia reduces virus replication to small-scale field trials demonstrating that Wolbachia are capable of spreading through wild Ae. aegypti populations. This primarily has been proposed as a tool to control dengue virus (DENV) transmission; however, Wolbachia infections confer protection for their insect hosts against a range of pathogens including CHIKV in Ae. aegypti. Medium-scale Wolbachia deployments are imminent or in certain instances have commenced. Therefore, assessing whether or not Wolbachia-infected Ae. aegypti are effective against CHIKV will help inform the viability of Wolbachia biocontrol for CHIKV control. Our study provides valuable evidence that could justify expanding this type of control program to other Ae. aegypti-transmitted arboviruses, primarily CHIKV.
| Chikungunya virus (CHIKV; Togaviridae, Alphavirus) has recently re-emerged out of Africa and caused explosive outbreaks of arthritic disease in Southeast Asia, India, Europe and currently the Americas [1–4]. The current outbreak in the Americas is cause for great concern because CHIKV is spreading nearly uncontrolled with at least 44 countries experiencing autochthonous spread [5]. Infection with CHIKV results in a severe febrile illness, called chikungunya fever. Clinically, it resembles dengue fever and several other arboviral diseases [6], but it is more associated with joint pain, which in some patients can progress to chronic arthralgia that lasts for months to years [7]. CHIKV disease can be highly debilitating and has a pronounced economic impact on both the affected individual and the countries which experience the outbreaks, resulting in great losses in productivity [8–10]. CHIKV is transmitted to humans by the mosquitoes Aedes aegypti and Aedes albopictus. The distribution of these mosquitoes explains the recent global spread of the virus and invasion of the Americas [4,5,11]. Both mosquito species have demonstrated the capacity to sustain CHIKV transmission cycles and both have been associated with CHIKV outbreaks [1]; however, the etiologic strain of CHIKV, a member of the old Asian lineage [12], causing the current outbreak does not efficiently infect Ae. albopictus, suggesting that most CHIKV transmission in the Americas will occur via Ae. aegypti [5].
Despite the continued spread of the virus, there remains no effective antiviral therapy or licensed vaccines. Therefore, new approaches to preventing CHIKV are needed because the endemic range of this virus is expanding and because current methods are limited to controlling mosquito populations. To date, mosquito control has not prevented invasion of this virus into new locales or controlled the virus when it arrives [13]. A promising candidate for arbovirus control and prevention relies on the introduction of the intracellular bacterium Wolbachia into Ae. aegypti mosquitoes. Wolbachia biocontrol has advanced from laboratory experiments demonstrating that certain strains of Wolbachia shorten the lifespan of the mosquito [14] while simultaneously reducing virus replication [15] to small-scale field trials demonstrating that Wolbachia are capable of spreading through wild Ae. aegypti populations [16–18]. This primarily has been proposed as a tool to control dengue virus (DENV) transmission [19–21]; however, Wolbachia infections confer protection for their insect hosts against a range of pathogens including for Ae. aegypti against CHIKV [22,23] and for Ae. albopictus against CHIKV [24]. As a result, this technology currently is being evaluated in five countries around the globe (Australia, Brazil, Colombia, Indonesia, and Vietnam) for its potential to control DENV transmission.
The approach is well-established that Wolbachia infection confers protection against DENV transmission by Ae. aegypti. In contrast, the ability of CHIKV to infect, disseminate, and be transmitted by wMel-infected Ae. aegypti is far less established [23]. For example, van den Hurk et al. (2012) tested the wMel strain of Wolbachia, but they only assayed Ae. aegypti vector competence for CHIKV at a single time point (12 days post feeding) with a single bloodmeal titer, and only could detect virus in the saliva via qRT-PCR [23]. And Moreira et al. (2009) tested the wMelPop strain of Wolbachia against CHIKV [22], which no longer is being utilized by the Eliminate Dengue Program (EDP) because mosquitoes infected with this strain of Wolbachia displayed reduced fitness in small-scale field releases [18]. Therefore, we assessed vector competence for CHIKV in wMel-infected and wMel-free Ae. aegypti from Medellin, Colombia, because at present our understanding of the ability of CHIKV to infect, disseminate, and be transmitted by wMel-infected Ae. aegypti currently being used at Wolbachia release sites is limited. This becomes particularly important if one considers that vector competence of Ae. aegypti for certain viruses likely is governed to a large extent by vector genotype x virus genotype (G x G) interactions in genetically diverse, natural Ae. aegypti populations [25]. This challenges the general relevance of conclusions from laboratory systems that consist of a single combination of mosquito and virus genotypes [25,26]. These Wolbachia-infected mosquitoes were created as part of a collaboration with the EDP in Colombia and in the spring of last year (2015), medium-scale deployments of these mosquitoes began in the DENV metropolitan area of Medellin [see www.eliminatedengue.com/colombia]. Our results suggest that Wolbachia effectively blocks the transmission potential of Colombian Ae. aegypti for CHIKV and wMel infection has no impact on the lifespan of mosquitoes as compared to wild type mosquitoes following CHIKV infection. To our knowledge, this is the first description of the effects of naturally acquired CHIKV infection (i.e., exposure to virus was accomplished by feeding on a viremic host) on Wolbachia-infected mosquito vector competence. All previous studies (including those mentioned for CHIKV, as well as the numerous studies described with DENV) have relied on animal blood spiked with cultured virus or have relied on viremic human blood from a membrane feeder. These data argue for the expansion of this technology to CHIKV in South America and are useful and germane in the broader context of CHIKV-mosquito interactions. Additionally, knowledge about factors shaping vectorial capacity (e.g., probability of daily survival) will be informative for a more accurate appraisal of CHIKV transmission and the likelihood of establishing Wolbachia infection in natural mosquito populations.
This study was carried out in strict accordance with recommendations set forth in the National Institutes of Health Guide for the Care and Use of Laboratory Animals. All animals and animal facilities were under the control of the School of Veterinary Medicine with oversight from the University of Wisconsin Research Animal Resource Center. The protocol was approved by the University of Wisconsin Animal Care and Use Committee (Approval #V01380).
African Green Monkey kidney cells (Vero; ATCC #CCL-81) were grown in Dulbecco’s modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum (FBS; Hyclone, Logan, UT), 2 mM L-glutamine, 1.5 g/l sodium bicarbonate, 100 U/ml of penicillin, 100 μg/ml of streptomycin, and incubated at 37°C in 5% CO2. Aedes albopictus mosquito cells, (C6/36; ATCC #CRL-1660) were maintained in MEM supplemented with 10% FBS, 2 mM L-glutamine, 1.5 g/l sodium bicarbonate, 0.1 mM non-essential amino acids, 100 U/ml of penicillin, 100μg/ml of streptomycin, and incubated at 28°C in 5% CO2. CHIKV isolate 99659 (GenBank:KJ451624), originally isolated from a 33 year old male in the British Virgin Islands with a single round of amplification on Vero cells, was obtained from Brandy Russell (Centers for Disease Control and Prevention, Ft. Collins, CO, USA). Virus stocks were prepared by inoculation onto a confluent monolayer of C6/36 mosquito cells. This CHIKV strain is related phylogenetically to strains recently identified in Asia with most of them sharing a specific four amino-acid deletion in the nsP3 gene [3], and is representative of CHIKV strains circulating in Colombia [27].
Ae. aegypti used in this study were maintained at the University of Wisconsin-Madison as previously described [26]. Two populations of mosquitoes were used in this study. Wild type (WT) mosquitoes (not infected with Wolbachia) were established from eggs collected from ovitraps placed around the municipality of Bello, a northwest suburb of Medellin, Colombia. The Wolbachia-infected (wMelCOL; infected with the wMel strain of Wolbachia pipientis) mosquito line was created by crossing uninfected field strains with a wMel-infected laboratory strain of Ae. aegypti essentially as described in [27]. The wMel-infected laboratory population of Ae. aegypti originated from eggs provided by Scott O’Neill (Monash University, Victoria Australia). Wolbachia infection status was routinely verified using PCR with primers specific to the IS5 repeat element [19].
Mosquitoes were exposed to CHIKV by feeding on isoflurane anesthetized CHIKV-infected Ifnar-/- mice. These mice have abrogated type I interferon signaling and as a result develop lethal infection, with muscle, joint, and skin serving as primary sites of replication [28,29]; as well, as developing high viremia. Ifnar-/- mice on the C57BL/6 background were obtained from Eva Harris (University California-Berkeley, Berkeley, CA) and were bred in the pathogen-free animal facilities of the University of Wisconsin-Madison School of Veterinary Medicine. Groups of three and six week-old mixed sex mice were used for mosquito exposures because viremia varied with age. Mice were infected in the left, hind foot pad with either 103 plaque forming units (PFU) of CHIKV in 50 μl of animal diluent (AD: 1% heat-inactivated FBS in Dulbecco’s PBS) for three week-old mice or 104.5 PFU of CHIKV in 50 μl of AD for six week-old mice. Uninfected mosquitoes (both WT and wMelCOL) were allowed to feed on mice two days post infection at which time sub-mandibular blood draws were performed and serum was collected to verify viremia. Three week-old mice yielded an infectious bloodmeal concentration of 9.51 log10 PFU/ml ± 0.09 (mean ± standard deviation; n = 6) and six week old mice yielded an infectious bloodmeal concentration of 6.90 log10 PFU/ml ± 0.14. These bloodmeal titers were consistent with human viremias observed in the field [30–32].
Infection, dissemination, and transmission rates were determined using long established procedures [33,34]. Briefly, three- to six-day-old female mosquitoes were sucrose starved for 14 to 16 hours prior to exposure to mice. Mosquitoes that fed to repletion were separated into cartons and maintained on 0.3 M sucrose in an environmental chamber at 26.5°C ± 1°C, 75% ± 5% relative humidity, and with a 12 hour photoperiod within the Department of Pathobiological Sciences BSL3 Insectary facility at the University of Wisconsin-Madison. All samples were screened by plaque assay on Vero cells. Dissemination was indicated by virus-positive legs. Transmission was defined as release of infectious virus with salivary secretions, i.e., the potential to infect another host, and was indicated by virus positive-salivary secretions.
All CHIKV screens and titrations for virus quantification were completed by plaque assay on Vero cell cultures. Duplicate wells were infected with 0.1 ml aliquots from serial 10-fold dilutions in growth media and virus was adsorbed for one hour. Following incubation, the inoculum was removed, and monolayers were overlaid with 3 ml containing a 1:1 mixture of 1.2% oxoid agar and 2X DMEM (Gibco, Carlsbad, CA) with 10% (vol/vol) FBS and 2% (vol/vol) penicillin/streptomycin. Cells were incubated at 37°C in 5% CO2 for two days for plaque development. Cell monolayers then were stained with 3 ml of overlay containing a 1:1 mixture of 1.2% oxoid agar and 2X DMEM with 2% (vol/vol) FBS, 2% (vol/vol) penicillin/streptomycin, and 0.33% neutral red (Gibco). Cells were incubated overnight at 37°C and plaques were counted.
Infection, dissemination, and transmission rates were analyzed using an Exact unconditional test [35]. Saliva CHIKV titers were analyzed using a Bootstrap t-test and survival data were analyzed using Kaplan-Meir analysis and log-rank statistics.
In Colombia, all four DENV serotypes actively circulate in many parts of the country and there has been a significant increase in the number of severe dengue cases since re-emergence [36]. The rise in cases coincided with an increase in Ae. aegypti populations that also have expanded into new geographic locales. Similar to the country as a whole, Medellin, the second largest city in the country, also had a significant increase in dengue cases, despite the presence of a national integrated vector control strategy. This provided the impetus for new approaches to preventing DENV transmission. In fact, deployment of Wolbachia-infected Ae. aegypti began in Medellin early last year (2015) to assess the efficacy of this technology in reducing DENV transmission in endemic populations. Not surprisingly, CHIKV reached Colombia in August 2014 [25], and since its introduction, there have been over 300,000 cases of CHIKV detected. Again, current vector control measures were insufficient in preventing invasion of this virus into the country or controlling it after invasion. Although primarily designed as a biocontrol tool for DENV, evidence suggests that Wolbachia can limit infection in Ae. aegypti with CHIKV [23]; therefore, Wolbachia-infected Ae. aegypti could potentially be used to simultaneously control DENV and CHIKV. As a result, we evaluated whether Colombian mosquitoes infected with the wMel strain of Wolbachia reduced CHIKV transmission potential.
Here, we verified that the phenotype of reduced vector competence existed in Wolbachia-infected laboratory colonies of Colombian Ae. aegypti for CHIKV. Adult, female, mosquitoes were exposed to infectious bloodmeals containing CHIKV and mosquitoes that ingested blood containing virus were assayed for infection, dissemination, and transmission potential at 7 and 14 days (d) post feeding (PF). As expected, infection, dissemination, and transmission rates were high for WT exposed to blood containing CHIKV at a concentration of 9.51 log10 PFU/ml (Table 1). Although viral titer in the bloodmeal was high, CHIKV viremia in humans can vary drastically (ranging from 101−109 PFU/ml), and therefore was consistent with observations in the field [30–32]. Furthermore, evidence suggested that infection and dissemination rates were dose-dependent and rates increase with the titer of the ingested bloodmeal (see [37] for review). Our first goal then was to determine if there was a threshold in which a high viral infectious dose could overwhelm the system and negate the protection conferred by Wolbachia. Interestingly, there was a significant reduction (Exact Unconditional Test) in infection, dissemination, and transmission rates for wMelCOL exposed to blood containing CHIKV; i.e., Wolbachia infection in Colombian Ae. aegypti completely blocked CHIKV transmission at 7d PF and significantly reduced infection and dissemination rates at 14d PF (Table 1). These data were consistent with other strains of wMel-infected Ae. aegypti when exposed to CHIKV [23] or DENV [21,38]; i.e., Wolbachia infection does not completely ablate transmission of virus, but rather delays the extrinsic incubation period (EIP) of the virus and reduces the transmission potential of CHIKV-infected mosquitoes.
Recently, Ye et al. (2015) demonstrated that Wolbachia-infected mosquitoes exhibited fewer infective days compared to WT mosquitoes, and their data suggested that Wolbachia-infected mosquitoes were infective at earlier timepoints [38]; however, they relied on qRT-PCR to detect and quantify virus, which does not differentiate infectious from non-infectious virus [39]. The plaque assays used here quantified infectious particles. Furthermore, it also has been demonstrated that this strain of CHIKV could be detected in the saliva of Ae. aegypti as early as 3d PF, albeit at very low levels [40]. To ascertain if wMelCOL had the potential to transmit CHIKV at earlier time points, we assessed the dynamics of infection in WT and wMelCOL over time following an infectious bloodmeal more in agreement with viremias detected in Colombian patients (6.90 log10 PFU/ml) [25] versus a high viremic infectious bloodmeal (>9.0 log10 PFU/ml). After a CHIKV-infected bloodmeal of moderate viremia, WT mosquitoes quickly became infective (Fig 1A–1C) and peaked at 53% infective (10/19) at 5d PF (Fig 1C). In contrast, wMelCOL remained non-infective over the duration of seven days (Fig 1C), but a large proportion (39%-70%) of wMelCOL had established infections (Fig 1A) and a moderate number (11%-29%) also disseminated virus (Fig 1B). Likewise, after a CHIKV-infected bloodmeal of high viremia, WT mosquitoes quickly became infective (Fig 2A–2C) and maintained infectivity (Fig 2C). In contrast, wMelCOL remained non-infective over the duration of seven days (Fig 2C), with the exception of a single mosquito with CHIKV-positive saliva on day six. A large proportion (up to 95% at 4d PF) of wMelCOL had established infections (Fig 2A) and a moderate number (21–70%) also disseminated virus (Fig 2B). Infectious virus also was detected in the saliva of wMelCOL on day 14 PF (Table 1). WT mosquitoes exposed to a bloodmeal of high viremia had viral titers in the saliva consistent with WT exposed to a bloodmeal of moderate viremia (Fig 3).
We then investigated whether CHIKV had a negative effect on mosquito survival, because probability of daily survival is an important parameter in estimating vectorial capacity. It is critically important to understand how virus infection impacts vector survival if accurate predictions of transmission dynamics are to be made, because low mosquito survival will reduce the likelihood of onward transmission of the infecting virus to a new host. There has been inconsistency among reports of the effects of arboviruses on mosquito survival, and to our knowledge no reports on the impact of CHIKV infection on mosquito survival. A recent meta-analysis involving various vector-virus combinations found that, overall, arboviruses do reduce the survival of their mosquito vectors [41]. And, others have suggested that the presence of wMel infection can lengthen the lifespan of mosquitoes as compared to WT following DENV infection, suggesting that DENV infection is costly to mosquitoes and that Wolbachia is conferring some protection to the host [38]. Here, the presence of wMel infection had no impact on the lifespan of mosquitoes as compared to WT following CHIKV infection (p = 0.369 and p = 0.429; Fig 4A and 4B, respectively), nor was there any indication that CHIKV infection was overly costly to WT mosquitoes (Fig 4B). Certainly, mosquitoes survived the relatively short EIP of CHIKV (Figs 1 and 2). It also is important to note that we explored the effects of naturally acquired CHIKV infection (i.e., exposure to virus was accomplished by feeding on a viremic host) on mosquito survival; whereas, most previous studies have relied on animal blood spiked with cultured virus, which may or may not have influenced the magnitude of the observed effect. Furthermore, recent studies suggested that viral titer in the bloodmeal might impact mosquito survival; i.e., high viral titers in the blood lead to increased mosquito mortality [42]. Here, unusually high mortality was not observed in mosquitoes exposed to blood containing CHIKV at a concentration of >9.0 log10 PFU/ml, i.e., a very high viral titer in the bloodmeal (Fig 4). These data are in concordance with a recent study by Carrington et al. (2015) that demonstrated that DENV infection adds minimal cost to Ae. aegypti when mosquitoes were exposed to DENV by feeding on infected humans, and there was no relationship between survival and human plasma viremia levels [43]. Although a direct comparison cannot be made, our data suggest that the relationship between CHIKV and Ae. aegypti is also relatively benign; but, we cannot rule out that CHIKV and/or Wolbachia infection may impart additional costs not measured here, e.g., reduced fecundity [44]. Finally, Wolbachia biocontrol depends on Wolbachia infections being maintained stably at high levels within natural mosquito populations as well as continuing to exhibit virus interference. Wolbachia may not stably persist if there are changes in maternal transmission, cytoplasmic incompatibility, and/or fitness effects to the mosquito as a result of Wolbachia infection. Wolbachia infection did not shorten the lifespan of infected mosquitoes (Fig 4B), which bodes well for the success of this strategy, but work still is needed to assess the long-term stability of infection and changes in host fitness effects following invasion in Colombia.
In sum, Wolbachia biocontrol has been proposed primarily as a tool to control DENV transmission [19], but Wolbachia infections also confer protection for Ae. aegypti against CHIKV and to some extent yellow fever virus (YFV) [23] as well. And, as a result of the explosive outbreak of CHIKV and now Zika virus in the Western hemisphere [12,45–47], all four of these viruses co-circulate in many parts of the tropics. The possibility exists that Wolbachia biocontrol could be used as a multivalent strategy for all of these Ae. aegypti-transmitted arboviruses. At the very least, these results warrant further exploration, both in the laboratory and the field, on the feasibility of expanding this technology beyond DENV and informing whether Wolbachia biocontrol can be used to supplement or replace existing vector control strategies.
|
10.1371/journal.pntd.0003636 | Validation of a Microsphere Immunoassay for Serological Leptospirosis Diagnosis in Human Serum by Comparison to the Current Gold Standard | A microsphere immunoassay (MIA) utilising Luminex xMap technology that is capable of determining leptospirosis IgG and IgM independently was developed. The MIA was validated using 200 human samples submitted for routine leptospirosis serology testing. The traditional microscopic agglutination (MAT) method (now 100 years old) suffers from a significant range of technical problems including a dependence on antisera which is difficult to source and produce, false positive reactions due to auto-agglutination and an inability to differentiate between IgG and IgM antibodies. A comparative validation method of the MIA against the MAT was performed and used to determine the ability of the MIA to detect leptospiral antibodies when compared with the MAT. The assay was able to determine samples in the reactive, equivocal and non-reactive ranges when compared to the MAT and was able to differentiate leptospiral IgG antibodies from leptospiral IgM antibodies. The MIA is more sensitive than the MAT and in true infections was able to detect low levels of antibody in the later stages of the acute phase as well as detect higher levels of IgM antibody earlier in the immune phase of the infection. The relatively low cost, high throughput platform and significantly reduced dependency on large volumes of rabbit antisera make this assay worthy of consideration for any microbiological assay that currently uses agglutination assays.
| Leptospirosis is a zoonotic disease caused by spirochaetes of the genus Leptospira and affects millions of people, worldwide, each year. Laboratory diagnosis of leptospirosis currently relies on methods that are flawed in many areas. Current methods are outdated, time consuming and expensive. They rely on a continuous supply of animal products (rabbit anti-sera) and require specialist expertise and equipment. The current gold standard diagnostic assay for leptospirosis (MAT) cannot determine IgG from IgM antibodies and relies on live cultures, which presents problems in the way of maintenance and attenuation. Development of a new diagnostic assay for serological diagnosis of leptospirosis that is specific, sensitive and able to discriminate between IgG and IgM classes of antibodies—as well as being more cost effective—will significantly improve the capabilities for detecting leptospirosis infections. It will provide medical professionals with more valuable diagnostic information and public health professionals with improved epidemiological information.
| Leptospirosis is considered to be the most widespread zoonotic disease in the world [1] with clinical diagnosis proving challenging due to the non-specific nature of symptoms associated with the disease. There are some 300 leptospiral serovars belonging to a number of different serogroups. Currently there are 24 sero-groups of pathogenic leptospires based on their antigenic relatedness [2]. Leptospirosis was first reported in Australia in 1933 in the state of Queensland and has since been isolated Australia wide [3] with Queensland reporting the majority of these cases (57.6%) [4]. In 2011 the reported incidence of leptospirosis in Queensland was 3.4 cases per 100,000 people and overall in Australia the incidence was 0.84 cases per 100,000 people [5]. At present, 24 serovars of Leptospira spp are recognised in Australia and in recent years a dramatic increase in the incidence of leptospirosis cases in Australia (particularly Queensland) has been noted with environmental factors believed to be the main influence on this increase [6].
Diagnosis of leptospirosis occurs at two stages—during the acute phase the live organism can be detected by two methods. Polymerase chain reaction (PCR) testing is a useful molecular detection tool for rapid qualitative diagnosis of leptospirosis in its earliest stage [7]. Serum or blood samples provided for PCR testing must be collected within a precise timeframe (0–8 days post onset) to enable diagnosis. Blood culture isolation can also be utilised in the early stages of leptospiral infection (0–10 days post onset), however this method is time consuming, requires specialised media and equipment and can take months for a serovar specific result [8]. The immune phase of a leptospiral infection is characterised by the presence of leptospiral antibodies and diagnosis is based on serological methods with the microscopic agglutination test (MAT) considered the current gold standard [9]. If the stage of the disease is unknown, both acute and immune phase tests are performed. Other serological test methods have previously been developed including flow cytometry [10], complement fixation testing [11], indirect hemagglutination assay [12] an IgM dipstick assay [13] and an IgM enzyme-linked immunosorbent assay (ELISA) in a number of formats [14,15]. Each of these assays has its advantages and disadvantages [16] and the type of assay used for diagnosis is generally dependant on the facilities available.
Serological diagnosis of leptospirosis in humans in Queensland, Australia is currently performed by screening with a commercially available leptospirosis IgM ELISA followed by the MAT as a reference and confirmatory test. The MAT method has many disadvantages as it requires specialist expertise, fresh leptospirosis cultures, is labour intensive, costly and is capable of determining total antibody only. The current endemic routine panel for MAT testing in Queensland, Australia consists of 16 serovars, with representatives from a number of different serogroups. Each sample submitted for MAT is screened against this panel and any reactive samples are then serially diluted and retested to determine an end point. Results are reported as a titre with the end point being the final dilution of serum at which 50% or more of the leptospires are agglutinated. This assay permits the testing of up to 20 samples per day on a routine basis. The MIA has the ability to simultaneously test large numbers of samples against large numbers of serovars as well as determine individual IgG and IgM titres. These factors alone would be enormously beneficial in the laboratory diagnostics and epidemiological studies of leptospirosis.
Bead based suspension array technology (xMap, Luminex) has the capacity to multiplex up to 500 individual analytes in a single well and has been shown to be a successful diagnostic tool for serology in many applications [17,18,19]. This assay platform is based on magnetic coated polystyrene beads filled with two coloured fluorescent dyes in differing ratios resulting in 500 distinct bead sets. Each bead set can be coated with a different antigen and mixed to allow the simultaneous measurement of antibody response to up to 500 different antigens. This high-throughput screening system allows processing of high numbers of patient samples per day. Its speed, sensitivity, and accuracy of multiple binding events measured in the same small volume have the potential to replace many clinical diagnostic and research methods and deliver data on hundreds of analytes simultaneously [20]. The microsphere immunoassay (MIA) that has been validated in this study was adapted from the method described by Luminex Corp (2000) and can be utilised as a routine serology testing protocol for leptospirosis.
The development and validation of a high quality, reliable serological assay is pertinent to the ability of a laboratory to sero-diagnose diseases in humans. Assay development begins with the identification of a need for improved diagnostic capabilities and the benefits that can be obtained from such an assay. A Luminex microsphere immunoassay (MIA) for leptospirosis antibody detection has the potential to function both as a high sensitivity, high throughput screening assay as well as a high specificity assay for determination of serovar level antibodies. This paper assesses the leptospirosis MIA in human samples as a screening assay to determine reactive, equivocal and non-reactive samples. Validation is performed by comparison to the leptospirosis IgM ELISA and the current gold standard, the microscopic agglutination test (MAT) as the basis for defining the performance characteristics of the MIA.
The study protocol was approved by the Public and Environmental Health Research Committee and the Humans Ethics Committee, Queensland Health Forensic and Scientific Services. All human samples utilised in this study were de-identified and allocated a generic number.
Sixteen Australian endemic pure leptospiral cultures, Table 1, were grown for 5–7 days in 3mL EMJH broth at 30°C. These antigens were then quantitated using a Petroff-Hausser grid and centrifuged at 4°C for 25 mins. The supernatant was removed and the pellet resuspended in 500μL phosphate buffered saline (pH 7.5). All cultures were then diluted to obtain a concentration of 1.8 x 109 per mL. These diluted antigens were used to coat 16 individual Bio-Plex Pro Magnetic COOH Bead-sets. Coupled beads were then checked for sensitivity and specificity using rabbit anti-sera of known serovar and titre, obtained from MAT results (See method below).
This study utilised 200 serum samples which were selected from human serum samples submitted for routine leptospirosis serology to the WHO/FAO/OIE Collaborating Centre for Leptospirosis Reference and Research during 2012 and 2013. These samples were submitted from Queensland hospitals and private laboratories. One hundred and eighty of these samples had leptospirosis IgM ELISA reactive serology, 12 had non-reactive leptospirosis IgM ELISA serology and the remaining 8 samples were not tested previously using leptospirosis IgM ELISA. All leptospirosis IgM ELISA testing was performed at a Queensland hospital or private laboratory prior to the samples being received at the WHO/FAO/OIE Collaborating Centre for Leptospirosis Reference and Research. Routine diagnostic MAT was performed on all samples at the WHO/FAO/OIE Collaborating Centre for Leptospirosis Reference and Research and results recorded against 16 routinely used, endemic serovars.
Forty-eight additional samples with reactive serology for Dengue Virus (24), Barmah Forest Virus (8), Ross River Virus (8) or Rabies Virus (8) antibodies were obtained from the Queensland Health Public and Environmental Health Virology Laboratory. These samples had previously been tested by ELISA IgM (Dengue virus), ELISA IgG (Rabies virus) or Alphavirus Hemagglutination Inhibition total antibody (HAI) (Ross River virus and Barmah Forest virus) and were used to assess whether cross reactions exist in the leptospirosis MIA.
In addition to the 200 samples used for the validation, 20 sets of paired samples with a non-reactive leptospirosis acute sample and reactive leptospirosis convalescent sample on the MAT were also obtained and analysed using the MIA to determine a timeline for the detection of leptospiral antibody. The results for these twenty additional samples are shown separately.
Leptospiral antigens were covalently coupled to individual Bio-Plex Pro Magnetic COOH bead-sets (Table 2) using the Bio-Rad Amine Coupling kit and methods from Luminex Corp. Coupling is achieved via carbodiimide reactions involving the primary amino groups on the protein and the carboxyl functional groups on the bead surface. The bead yield per coupling reaction is approximately 2,500 beads per well (in a 96-well microtitre plate). For optimum results in the MIA, the coupled beads were diluted 1:4 in Triton-X detergent and 100 beads in 100μL buffer were used for the immunoassay. Each individual coupled bead-set was diluted in phosphate buffered saline (PBS) to give a reading of approximately 100 beads per bead-set per well. The working dilution and specificity of each bead-set was validated prior to use in a diagnostic capacity by utilising serovar-specific rabbit antisera and the IgG method as described below, substituting the secondary antibody with an anti-rabbit IgG (RPE). Bead-sets were considered to be valid for use if the targeted serovar produced an antibody response to that specific bead-set.
Two microsphere immunoassays (IgG and IgM) were performed on 200 serum samples taken from the routine MAT submissions which included samples with MAT titres (serial dilutions) ranging from < 1:50 (non-reactive) to 1:6400. Samples with an MAT titre between 1:50 and 1:200 were considered equivocal and samples with a titre 1:400 or above were considered reactive. Pooled convalescent serum from patients with recent leptospirosis infections, confirmed by PCR (on acute sample) and MAT, was used as the positive control serum in each microsphere immunoassay. Negative patient serum, confirmed by negative PCR and serology) was pooled and used as negative control serum. These controls were monitored each run to ensure the assay was consistent.
A 96-well filter plate was pre-wetted with 150μL PBS per well and vacuum applied. One-hundred μL of the diluted coupled beads were then added to each required well of the pre-wetted 96-well microtitre filter plate and vacuum applied. Serum samples for the IgG immunoassay were diluted 1:400 in PBS in 1mL micronic tubes. One-hundred μL of the diluted samples were added to the plate which was then incubated for 45 minutes on a shaker (750rpm) at room temperature. The plate was then vacuum-washed three times with 150μL PBS per well. 100μL of a diluted secondary antibody (anti-human IgG) with a fluorescent tag (RPE) was added to each well followed by a second 45 minute incubation and vacuum wash as per previous step. Finally, 150μL PBS was added to each well and the plate placed back on a shaker at room temperature for at least 10 minutes prior to analysis.
Serum samples for the IgM immunoassay were treated with Siemens Rheumatoid factor (RF) absorbent (at a dilution of 1:2) and diluted to a final concentration of 1:800 in PBS. The plate was prepared as per the IgG immunoassay. The secondary antibody—anti-human IgM with a fluorescent RPE tag—was used in this assay for conjugation.
All plate wells were then analysed using Luminex xMap technology on a BioPlex 200 Platform. The MIA results were reported as mean fluorescent intensity (MFI) and were deemed congruent or incongruent relative to the standard of comparison (MAT). Cut-off values for reactive samples were determined using five reactive sera for each MAT titre ranging from 1:100 to 1:6400 (Table 3), and developing a standard curve (R-Biopharm, 2012) using the titres obtained from MAT testing and comparing them with the mean fluorescent intensities from the MIA titrations. Fig. 1 shows the reactive sera MAT titres plotted against the MFI’s and the standard curve that resulted. From this curve, cut-off points were determined (Table 4). Positive/negative ratios were used to determine the cut-off point for non-reactive samples. During the validation and determination of cut-off points the results reactive high and reactive low were used to ensure that the MAT and the MIA results were comparable. All patient results were reported as reactive, non-reactive or equivocal.
Sensitivity (the ability of the MIA to correctly determine the presence of leptospiral antibody) was determined by running known reactive (true positive) samples on the MIA and calculating the proportion of reactive samples detected. True positive samples are samples known to be reactive by paired sample testing with the Gold standard, the microscopic agglutination test.
Assay specificity was assessed using two methods. The first involves running known non-reactive (true negative) samples and calculating the proportion of non-reactive samples detected by the MIA. True negatives are defined as non-reactive samples known to be non-reactive by paired sample testing with the Gold standard assay, microscopic agglutination test. False positives are reactive samples determined by the test assay (MIA) that are non-reactive by the gold standard.
The second test of specificity ensured that samples that have been shown to have reactive serology for other pathogens are not cross reacting with the leptospirosis MIA.
Within-run repeatability was determined by running four samples 20 consecutive times on one assay run for both IgG and IgM assays. Two of these samples had an equivocal result for at least one serovar on both assays, one sample had a reactive result for at least one serovar on both assays and the remaining sample was non-reactive for all 16 serovars for both IgG and IgM assays.
Repeatability is also monitored continuously as a quality control measure by monitoring positive (reactive) and negative (non-reactive) controls with expected and accepted MFI ranges for each control serum in every assay. If the control serum results were outside of these ranges, the run was deemed to have failed and was repeated.
Of the 200 samples tested, 180 samples were reactive for leptospirosis IgM by ELISA (Table 5). Twelve samples were IgM ELISA non-reactive and eight samples did not have previous IgM ELISA results; comparisons could only be made with the MAT and MIA for these eight samples. The MAT confirmed 27 of the leptospirosis IgM ELISA reactive samples had evidence of leptospiral total antibody and suggested that the remaining 153 IgM ELISA reactive samples were non-reactive (titre of < 1:50). These results suggest a substantial gap in the diagnostic performance of the ELISA and the MAT.
The MIA results (in mean fluorescent intensity—MFI) for the 27 MAT reactive samples also indicated reactive serology (MFI > 1200). Of the 173 non-reactive MAT samples, 126 were non-reactive on the MIA and the remaining 47 had low reactivity on the MIA, suggesting better sensitivity in the MIA. The results for five of these 47 samples, which have been confirmed as true leptospiral infections by PCR or blood culture, are shown in Table 6.
The MIA detected leptospiral antibody in 74 (41%) of the 180 ELISA IgM reactive samples. The remaining 106 ELISA IgM reactive samples were non-reactive on the MIA and the non-reactive IgM ELISA samples were also non-reactive on the MIA. The 8 samples that were not previously tested by ELISA were non-reactive on the MIA also (Table 5).
Of the 20 sets of additional paired samples with an MAT non-reactive acute sample and MAT reactive convalescent sample, 12 of these pairs demonstrated equivocal or reactive IgM MFI results for the acute samples with a significant rise in MFI in the convalescent samples on the IgM MIA. The results for the remaining eight pairs of samples were consistent between the MAT and the MIA. Table 7 shows the results for the paired samples comparing the MAT titre and the MIA IgM and IgG results. These samples were included in this study to show that IgM can be detected earlier or, at least at the same time, by the MIA when compared with the MAT in true leptospiral infections, as determined by a four-fold rise in serology.
Of the 48 reactive viral serology samples only one showed reactive IgG and IgM serology for leptospirosis (this sample was previously reactive for Dengue virus serology) and the remaining 47 samples were non-reactive for both leptospirosis IgG and IgM.
The four samples used to test within-run repeatability showed comparable results in each well across each of the 16 serovars. Table 8 shows the mean fluorescent intensity and standard deviation values for each of the four samples used in the repeatability testing for one of the serovars in the IgM immunoassay. Samples 1 and 2 were non-reactive. Sample 3 was reactive and sample 4 was in the equivocal range. The expected values were derived from comparison of the MIA mean fluorescent intensity with MAT titres. Repeatability was assessed across one run with one operator as, at the time of testing, only one operator was available to perform this testing.
The aim of diagnostic serology is to determine reactive and non-reactive samples for a particular infectious agent. By definition, a validated assay consistently provides test results that identify samples as being reactive or non-reactive for a selected analyte, and, by inference, accurately predicts the disease status of patients with a predetermined degree of statistical certainty [21]. The aim of this study was to validate a microsphere immunoassay (MIA) using Luminex xMap technology for diagnostic leptospirosis serology screening. The validation process was performed using a comparative method—that is comparing the new assay with the current gold standard assay. Sixteen leptospiral antigens have been coupled to 16 individual magnetic bead-sets and validated as a panel for routine diagnostic leptospirosis serology. This assay gives a qualitative result—Reactive, Equivocal or Non-Reactive and has the ability to determine recent from past infection by differentiating between IgM and IgG antibodies—something that is more difficult to achieve with microscopic agglutination testing (MAT) as this test can only determine total antibody. The class of antibody detected by the MIA can be used to determine the stage of the infection which is valuable for clinicians as it can determine treatment regimens for patients or in the case of a past infection, can suggest that something other than leptospirosis is causing symptoms. Information regarding new infections is also vital from a public health perspective as it can provide information on what serovars of leptospirosis are currently circulating and indicate the areas where these infections are occurring.
All leptospirosis serology reactive samples by MAT were detected by MIA suggesting that congruence is 100% when compared to the MAT. Results from the non-reactive samples, as well as the paired samples suggest, however, that the MIA is more sensitive than the MAT. In true infections (as demonstrated by paired sample serology testing with a minimum four fold rise in titre) the MIA was able to detect low level antibody in the later stages of the acute phase as well as pick up higher levels of IgM antibody earlier in the immune phase of the infection. The MAT results indicated that these samples were non-reactive in the acute/early immune phase. The MAT generally becomes positive between day 8 and day 10 of infection [22] however, results from this validation suggest that the MIA could detect antibody in the earlier stages of infection development and increase the likelihood of the clinician submitting a convalescent sample for confirmation of infection status.
The leptospirosis IgM ELISA has previously been shown to have poor specificity, as low as 41%, when used according to the manufacturer’s instructions [22]. All leptospirosis IgM ELISA reactive samples tested in Queensland pathology laboratories are sent to the WHO/FAO/OIE Collaborating Centre for Leptospirosis Reference and Research for confirmation testing. In this study it was found that of 180 leptospirosis IgM ELISA reactive samples only 15% (27/180) of these showed reactive results on the MAT. This could be due to a lower level of antibody which is not detected by the MAT at a dilution of 1 in 50 or a non-specific antibody reaction. In this study, 41% (74/180) of the leptospirosis IgM ELISA reactive samples had reactive IgG and/or IgM serology on the MIA, again suggesting the level of antibody in these particular samples may be too low for the MAT to detect. Also, this again shows that there may be some non-specific reactions occurring in the IgM ELISA, which are not seen on the MIA. The MIA is therefore advantageous as a screening test as it reduces the large numbers of samples that are unnecessarily sent for confirmation testing by MAT. It has also been suggested that false positivity can also occur in the leptospirosis IgM ELISA due to the presence of persistent IgM from past infections [23]. The MIA screening test eliminates these results by looking at the levels of the individual IgG and IgM antibodies across paired specimens. A low level or non-reactive IgM result and a plateaued reactive IgG would be suggestive of a past infection—something not currently visible on the leptospirosis IgM ELISA or the MAT.
The MIA results suggest that the beads coated with leptospiral antigen are specific for leptospiral antibodies and show no cross-reactivity with other viral agents. The one case in this study where a Dengue Virus reactive serology sample also showed leptospiral antibodies is likely to be a true leptospirosis infection occurring simultaneously with a Dengue Virus infection. Leptospirosis and Dengue Virus infections are both common in northern parts of Queensland (where this sample was from) as they are both associated with tropical and sub-tropical regions where extreme weather events occur [24]. In many cases samples are submitted for both arbovirus testing (including Dengue virus) and leptospirosis testing at the same time.
The results from the MIA show that reproducibility is possible and accurate when compared to the MAT. A major disadvantage of the MAT is attenuation of the live leptospiral cultures. It has been shown that over time, leptospiral cultures lose their antigenicity and therefore become less effective [25]. Also, day to day, the cultures can be different—more or less dense or contaminated—which makes reproducing results accurately a difficult task on the MAT. This issue is overcome with the MIA as the antigens (leptospiral cultures) are wild type cultures with a known passage number and are all diluted to a known concentration (1.8 x 109) prior to the bead coupling process. This ensures that there are equal amounts of each antigen available in every test. Another major advantage of the MIA over the MAT is that there is no need to maintain stocks of live leptospiral cultures for daily use. Pure cultures are only used in the MIA as antigens for bead coupling and these antigens can be centrifuged, diluted and frozen at -20°C for up to six months [26]. Currently, performing the MAT on a routine basis requires sub-culturing more than 200 tubes per week, maintaining four stocks of cultures.
When comparing the MAT and the MIA the advantages of the latter are obvious. Firstly, the MIA is less time consuming—a full plate of 88 samples can be run in around three hours. To run the same number of samples on the MAT, it would take twice the time for a full panel of 16 serovars excluding analysis. The MIA is also less labour intensive as it does not require adding 16 individual cultures to each well on a 96 well plate for each individual patient. These savings combined as well as the reagent costs suggest that the MIA is also less costly than the MAT. An analysis of laboratory and assay costs shows that the current diagnostic serology method (MAT) is performed at a cost of $AUD6.95 (excluding labour) to the leptospirosis reference laboratory per sample per 16 serovars [27]. In comparison, the MIA costs $AUD4.95 per sample (excluding labour) per 16 serovars. Secondly, the MIA uses a total of 7μL of serum (2μL for the IgG assay dilutions and 5μL for the IgM assay dilutions) compared with 50μL of serum used in the MAT. Thirdly, the MIA has the ability to detect and differentiate both IgG and IgM antibodies whereas the MAT can only detect total antibody and cannot give an accurate indication of the stage of infection in a single sample. The MIA can potentially include up to 500 analytes in the one assay, therefore, there is potential to be able to include all known leptospirosis serovars (~250) in one test at one time. Given the number of bead-sets available for microsphere immunoassays other applications could potentially involve the inclusion of a number of different viral and bacterial agents in one assay. For example, leptospirosis antibody detection and Dengue Virus antibody detection could be combined into one routine diagnostic test.
In conclusion, the results from this validation suggest that the leptospirosis MIA is a beneficial diagnostic screening tool for leptospirosis serology testing. This assay is able to determine reactive, equivocal and non-reactive samples when compared to the MAT. It is able to differentiate leptospiral IgG antibodies from leptospiral IgM antibodies which will provide vital diagnostic information as well as provide a better epidemiological picture. Further investigations will include validation of each individual serovar to enable serovar specific results to be reported and validation of a microsphere immunoassay for detection of leptospiral antibodies in animal samples will also be looked at in the future.
|
10.1371/journal.pntd.0002180 | False Positivity of Non-Targeted Infections in Malaria Rapid Diagnostic Tests: The Case of Human African Trypanosomiasis | In endemic settings, diagnosis of malaria increasingly relies on the use of rapid diagnostic tests (RDTs). False positivity of such RDTs is poorly documented, although it is especially relevant in those infections that resemble malaria, such as human African trypanosomiasis (HAT). We therefore examined specificity of malaria RDT products among patients infected with Trypanosoma brucei gambiense.
Blood samples of 117 HAT patients and 117 matched non-HAT controls were prospectively collected in the Democratic Republic of the Congo. Reference malaria diagnosis was based on real-time PCR. Ten commonly used malaria RDT products were assessed including three two-band and seven three-band products, targeting HRP-2, Pf-pLDH and/or pan-pLDH antigens. Rheumatoid factor was determined in PCR negative subjects. Specificity of the 10 malaria RDT products varied between 79.5 and 100% in HAT-negative controls and between 11.3 and 98.8% in HAT patients. For seven RDT products, specificity was significantly lower in HAT patients compared to controls. False positive reactions in HAT were mainly observed for pan-pLDH test lines (specificities between 13.8 and 97.5%), but also occurred frequently for the HRP-2 test line (specificities between 67.9 and 98.8%). The Pf-pLDH test line was not affected by false-positive lines in HAT patients (specificities between 97.5 and 100%). False positivity was not associated to rheumatoid factor, detected in 7.6% of controls and 1.2% of HAT patients.
Specificity of some malaria RDT products in HAT was surprisingly low, and constitutes a risk for misdiagnosis of a fatal but treatable infection. Our results show the importance to assess RDT specificity in non-targeted infections when evaluating diagnostic tests.
| Rapid diagnostic tests (RDT) for malaria are currently rolled-out as the backbone of parasite-based diagnosis, and their diagnostic accuracy is sufficiently high to substitute microscopy. One decade ago, attention has been given to occurrence of limited false positivity in a number of malaria RDTs, but false positivity of RDTs has remained poorly documented since then. In the last years, the number of available RDT products has dramatically increased and test performance has improved. False positivity may therefore not be perceived as a problem anymore. In this manuscript, we demonstrate that specificities of malaria rapid diagnostic tests detecting parasite antigens are seriously affected by human African trypanosomiasis (sleeping sickness), with values down to 11%. Malaria constitutes the main differential diagnosis of human African trypanosomiasis, and the false-positive results for malaria RDTs increase the risk of misdiagnosis or delayed diagnosis of human African trypanosomiasis which is a fatal but treatable infection.
| Traditional diagnosis of malaria relies on microscopic detection of Plasmodium in thick blood films, which is labour-intensive, time-consuming and requires technical skills. Malaria rapid diagnostic tests (RDT) offer an attractive alternative for microscopy. They detect parasite antigens in blood through an antibody-antigen reaction, made visible by a red line on a nitrocellulose strip. Test formats include two- and three-band tests. Two band RDT products consist of a control line and a test line, three-band tests have a control line and 2 test lines, of which at least one usually targets a Plasmodium falciparum (Pf) specific antigen. The main target antigens are histidine rich protein 2 (HRP-2) specific for Pf and Plasmodium lactate dehydrogenase (pLDH), either detecting all human infective species (pan-pLDH) or specific for Pf (Pf-pLDH). Accuracies of RDTs are now such that they can substitute microscopy: average sensitivity and specificity have been estimated 95% for HRP-2 test lines, and respectively 93 and 99% for pLDH detection tests [1]. Of note, the HRP-2 antigen can persist for up to ≥6 weeks after (effectively treated) Pf infection [2], [3]. Besides this, false positivity has been attributed to the rheumatoid-factor [4]–[6], but has also been observed for anti-nuclear antibody, anti-mouse antibody and rapid plasma reagin positive samples [7], and has been attributed mainly to the HRP-2 test line. Limited false positivity has been noted in hepatitis C, schistosomiasis, toxoplasmosis, dengue, leishmaniasis and Chagas disease [7]–[9].
Occurrence of false positivity in RDTs may be particularly relevant for infections prevailing in malaria endemic regions, for which the differential diagnosis includes malaria. An example of such an infection is human African trypanosomiasis (HAT) [10]. Human African trypanosomiasis (HAT), or sleeping sickness, is a fatal but treatable infection caused by the parasites Trypanosoma brucei (T.b.) gambiense and rhodesiense, which are transmitted by tsetse flies. Association of HAT with a strong polyclonal B-cell activation and with rheumatoid factor-like anti-immunoglobulin antibodies has been reported [11]. Furthermore, T.b. gambiense infection has been shown to decrease the specificity of antibody detection tests for HIV diagnosis [12]. Based on these reports we put forward that HAT may be associated with false positivity in malaria RDTs. The objective of this study was therefore to examine the specificity of some commonly used malaria RDT products in HAT.
Before enrolment into the study, participants gave written informed consent. Parents or guardians provided consent on behalf of all child participants. Ethical clearance for the study was obtained from the institutional review board of ITM and the ethical committees of the University Hospital in Antwerp, Belgium (study registration number B30020108363) and of the Ministry of Health of the Democratic Republic of the Congo (DR Congo).
HAT patients and paired non-HAT endemic controls were prospectively included in DR Congo, Bandundu Province between July and December 2010. Malaria transmission in most parts of DR Congo, including in Bandundu Province, is high (>1 case per 1000 population) and DR Congo is considered as suffering from the highest malaria burden in Africa [13], [14]. Moreover, almost 80% of all notified HAT cases originate from DR Congo, and almost half of them are detected in Bandundu where the HAT prevalence is estimated at 0.36% [15]. Participants were identified during HAT screening activities of the dedicated HAT mobile team of Masi-Manimba, or included at the HAT treatment centres of Masi-Manimba and Bonga-Yasa. Inclusion criteria for HAT patients were the presence of trypanosomes in blood, lymph and/or cerebrospinal fluid (irrespective of disease stage), and being 12 years or older. Exclusion criteria were pregnancy, being previously treated for HAT and being moribund. To each HAT patient, a control was matched, fulfilling the following criteria: same gender and age and being permanent resident from the same or a neighbouring village. Inclusion criteria for controls were absence of clinical evidence for HAT (no swollen lymph nodes or neurological symptoms), absence of trypanosome specific antibodies in whole blood detected by card agglutination test for trypanosomiasis [16]; no trypanosomes in blood detected by the mini anion exchange centrifugation technique [17] and being 12 years or older. Exclusion criteria were identical as for HAT patients. From the participants, clinical and epidemiological parameters in conjunction to HAT were collected, including the use of anti-malaria medication in the month prior to inclusion. Blood was drawn on EDTA and on heparin. For conservation of DNA, 0.5 ml of EDTA blood was mixed with an equal volume of GE buffer (6 M guanidium, 0.2 M EDTA, pH 8.0) and stored at ambient temperature until DNA extraction. Aliquots of blood taken on EDTA and of plasma prepared from the blood taken on heparin were snap frozen in liquid nitrogen and shipped to ITM where specimens were stored at −70°C until use. Two thick and thin blood films were prepared and Giemsa stained; one set remained in Kinshasa, the double was shipped to ITM.
DNA was extracted from EDTA blood mixed with GE buffer using the Maxwell 16 DNA Purification Kit (Promega, Madison, Wisconsin, USA). A four primer real-time PCR was performed for detecting single and mixed infections of Plasmodium falciparum (Pf), Plasmodium vivax (Pv), Plasmodium malariae (Pm) and Plasmodium ovale (Po) as previously described [18], [19]. A negative and positive control for each Plasmodium species was included for each test run. If Pf-PCR, Pv-PCR, Pm-PCR or Po-PCR was positive, malaria-PCR was considered positive, if negative for all species, malaria-PCR was considered negative. Standard microscopy was performed on thick blood films by an expert microscopist. Parasite density was assessed by counting asexual parasites against 200 white blood cells (WBC) in thick blood films and converted into parasites/µl using the standard of 8000 WBC/µl [20].
In total, ten rapid diagnostic tests for malaria were evaluated. Malaria RDTs in a cassette format were selected based on (i) demonstrated diagnostic accuracy [7], [21] and/or (ii) use by the national malaria control programme and non-governmental organizations in trypanosomiasis endemic countries. In addition, RDTs detecting Pf-pLDH were added. The following two-band tests were used: 1° Paracheck Pf (Orchid Biomedical Systems, Goa, India); 2° ICT Malaria Pf Cassette Test (ICT Diagnostics, Cape Town, Republic of South Africa); 3° Advantage Pan Malaria Card (J. Mitra & Co Pvt. Ltd., New Delhi, India). The three-band tests were: 4° Malaria Antigen Pf (HRP-2/pLDH) (Standard Diagnostics Inc., Hagal-Dong, Korea); 5° SD malaria Ag Pf/Pan (Standard Diagnostics Inc); 6° SD Malaria Antigen Pf (Standard Diagnostics Inc.); 7° ICT Malaria Combo Cassette Test (ICT Diagnostics); 8° Carestart Malaria HRP2/pLDH (Pf/Pan) Combo Test (Acces Bio, New Jersey, USA); 9° Carestart Malaria pLDH (Pf/pan) (Acces Bio); 10° First Response Malaria Ag (pLDH/HRP2) Combo Rapid Diagnostic Test (Premier Medical Corporation Ltd., Daman, India). Tests were performed according to the instructions of the manufacturers, using thawed EDTA blood. Three experienced readers who were blinded to each other's and to other results, except for the HAT status, read the RDT results, at daylight and within the prescribed delay. After reading, RDT results were immediately photographed. Test line intensities were scored as negative (no line visible), faint (barely visible), weak (paler than the control line), medium (equal to the control line), strong (stronger than the control line). Invalid tests (no control line) were repeated. Based on the three scores, a consensus line intensity score was established. If two out of three readers had scored the line intensity the same, this was taken as the consensus. If the three readers had scored the intensity differently, the photograph was used for taking the final decision. For analysis of test line results, faint, weak, medium and strong line intensities were considered positive, absence of test lines was recorded as negative. Interpretation of the RDT result was as prescribed by the manufacturer. An RDT was considered negative for malaria when all test lines were negative, and positive when at least 1 test line was positive.
The presence and concentration of rheumatoid factor in plasma of all Pf-PCR negatives was measured using the VITROS chemistry Products RF reagent on the VITROS 5600 Integrated system (Ortho-Clinical Diagnostics, Buckinhamshire, UK). The procedure consists of an antigen-antibody reaction occurring between rheumatoid factor from the test sample and denatured human IgG adsorbed to latex particles from the test reagent, resulting in agglutination. The agglutination is detected as an absorbance change, with the magnitude of the change being proportional to the quantity of rheumatoid factor. The detection limit was 9 IU/ml, as suggested in the instructions, values ≥12 IU/ml were considered abnormal.
For calculation of specificity and sensitivity, PCR was used as the reference method (Supporting information file S1). Specificity of HRP-2 or Pf-pLDH lines was assessed respectively as the number of HRP-2 or Pf-pLDH line negatives/number of Pf-PCR negatives; sensitivity was defined as the numbers of HRP-2 or Pf-pLDH line positives/number of Pf-PCR positives. Likewise, specificity of pan-pLDH lines was calculated as the number of pan-pLDH negatives/number of malaria PCR negatives and the sensitivity as the number of pan-pLDH positives/number of malaria PCR positives. Specificity of three-band RDTs for diagnosis of malaria was calculated as the number of RDT negatives/number of malaria PCR negatives or number of RDT negatives/number of Pf-PCR negatives if the RDT consisted only Pf specific test lines. Proportions were calculated with 95% exact confidence intervals (CI). In a primary analysis, differences between proportions were tested for significance using the Fisher exact probability test (STATA 10.0), in order to avoid data loss and as the effect of matching, which was done for another study component, was expected to be small on RDT performance. For comparison of specificity, this primary analysis was followed by a secondary analysis using McNemar chi square, after matching controls and HAT again taking into account the PCR result. For sensitivity, the low number of PCR-matched samples precluded meaningful comparison with the McNemar test. Inter-reader reliabilities were assessed for the test results expressed as positive and negative line readings and kappa values for each pair of readers were calculated.
An overview of the study population and reference test results is shown in table 1. In total, 117 HAT patients and 117 controls were included in the study. Median age was 28 years and 45% of the study participants were male. Significantly more HAT patients than controls had taken anti-malaria drugs in the month prior to inclusion (p = 0.01). Twenty-two (9.9%, CI 5.9–13.8) thick blood films were positive for Plasmodium. Due to low Plasmodium parasite densities, the Plasmodium species could be defined only in 16 thin blood films, and was Pf in 15, Po in one. In total, 79/233 participants were malaria-PCR positive (1/234 sample of blood on GE buffer missing), including all 22 that were thick blood film positive. Species identification as defined by PCR were 87% Pf (69/79), followed by Po (26.6%, 21/79), Pm (21.5%, 17/79) and Pv (1.3%, 1/79).
Specificity of Pf-pLDH test lines in three malaria RDT products was 97.5–100% (table 2). There was no significant difference in specificity between controls and HAT patients. False positive test lines were faint (1–3/163) or weak (1/163, figure 1). Sensitivities of Pf-pLDH test were respectively 21.1–28.9% in controls (n = 38) and 9.7% in HAT patients (n = 31). True positive reactions consisted of faint (4–8/69) and weak lines (0–5/69, figure 1). For none of the RDT products, a significant difference in sensitivity between controls and HAT patients was observed. On the complete series of samples tested (n = 233), there was almost perfect agreement between at least two out of three readers assessing the Pf-pLDH lines (maximal kappa values between two readers of 0.89–0.96).
Specificity of HRP-2 test lines in seven malaria RDT products for controls and HAT patients are summarized in table 2. In controls, HRP-2 test line specificity was between 87.3 and 100% while in HAT, it was between 67.9 and 98.8%. Significantly lower specificity of the HRP-2 test line in HAT compared to controls was observed for five out of seven RDT products. After matching, the difference in specificity was lost for the Carestart Malaria HRP2/pLDH (Pf/Pan) Combo Test. In participants not having taken anti-malaria drugs prior to inclusion, specificities in controls (n = 63) and HAT (n = 53) ranged respectively 88.5–100% and 67.9–100%, and specificity of HRP-2 test lines remained significantly lower in HAT compared to negative controls for two RDT products (ICT Malaria Pf Cassette Test, ICT Malaria Combo Cassette Test). In Malaria Antigen Pf (HRP-2/pLDH), SD malaria Ag Pf/pan and Carestart Malaria HRP2/pLDH (Pf/Pan) Combo Test, a similar tendency towards lower specificity of HRP-2 in HAT was observed, but the difference lost significance. Line intensity scores are shown in figure 1. False positivity was mainly due to presence of faint HRP-2 test lines (1–28/163) and to a lesser extent due to weak HRP-2 test lines (0–9/163). No medium or strong false positive HRP-2 test lines were observed. With respectively 22/163 samples being HRP-2 false positive in ≥two RDT products and 13/163 in ≥three RDT products, false positivity appeared to be at random distributed over the samples.
Sensitivity of HRP-2 test lines in Pf-PCR positive controls (n = 38) and HAT patients (n = 31) ranged between 47.4–71.1% and 25.8–74.2% respectively. No significant difference in sensitivity of HRP-2 was observed between controls and HAT patients. The majority of positive lines scored weak (8–23/69), medium (1–4/69) or strong (0–8/69, Figure 1), with the exception of the Carestart Malaria HRP2/pLDH (Pf/Pan) Combo Test where mainly faint test lines (25/69) were seen (Figure 1).
On the complete series of samples tested (n = 233), there was substantial to almost perfect agreement between at least two readers in HRP-2 line intensity (maximal kappa value between two readers of the seven RDTs were 0.75–0.95).
For pan-pLDH tests as well, a significant lower specificity in HAT patients compared to controls was observed with five out of seven RDT products (Table 2). These differences were confirmed after taking into account matching. Pan-pLDH test line specificities in controls ranged between 86.3–100% versus between 13.8–97.5% in HAT patients. False positive pan-pLDH test lines were mainly of faint or weak line intensity (respectively 1–38 and 1–36/153, figure 1), but false positive medium (1–11/153) and strong (14/153) intensity pan-pLDH test lines were observed as well, especially with ICT Malaria Combo Cassette Test. With respectively 62/153 samples being pan-pLDH false positive in ≥two RDTs and 29/153 in ≥three RDTs, it appeared that the same samples were responsible for most false positives in the different RDTs.
Sensitivity of pan-pLDH test lines in malaria-PCR positive controls (n = 44) was 0–63.6%, in HAT patients (n = 35) it was 11.4–77.1%. Sensitivity in controls was significantly lower than in HAT patients for two out of seven RDTs: ICT Malaria Combo Cassette Test (respectively 25.0% (13.2–40.3) versus 77.1% (58.9–89.6), p<0.001) and Carestart Malaria pLDH (Pf/pan) (respectively 0.0% (0–8.0) versus 11.4% (3.2–26.7), p<0.04). The majority of positive line intensities scored only faint (4–29/79) or weak (0–22/79), but some medium (0–2/79) or strong tests line intensities (0–7/79) were observed as well. On all 233 samples tested, there was substantial to almost perfect agreement between at least two readers (maximal kappa values between 2 readers of the 7 RDTs of 0.79–1).
Specificity of the combined test result in the seven three-band RDTs in controls and HAT patients are summarized in table 2. For five out of seven RDT products, combining HRP-2 with either Pf-pLDH or pan-pLDH, specificity was significantly lower in HAT patients compared to controls. These differences were confirmed after taking into account matching. In both tests combining Pf-pLDH with pan-pLDH test lines (SD Malaria Antigen and Carestart Malaria pLDH (Pf/pan)), no difference between HAT patients and controls was observed, and specificities were 96.3 to 100%.
After subtraction of patients who were on anti-malaria drugs prior to sampling, specificities in controls (n = 55) and HAT (n = 50) for the different RDTs ranged between 78.2–100% and 12.0–100% respectively, and remained significantly lower in HAT patients compared to controls for three out of five three-band RDT products incorporating HRP-2 test lines (SD Malaria Antigen, ICT Malaria Combo Cassette Test, Carestart Malaria HRP2/pLDH (Pf/Pan) Combo Test). In Malaria Antigen Pf (HRP-2/pLDH) and First Response Malaria Ag (pLDH/HRP2) Combo Rapid Diagnostic Test, a similar tendency towards lower specificity in HAT was observed, but the difference lost significance.
Proportions of rheumatoid factor positive samples were 7.6% (6/79) among controls (maximum concentration 47 IU/ml) and 1.2% among HAT patients (1/84, maximum 45 IU/ml) respectively. This difference was not statistically different (p = 0.06). There was no association between false positivity of any of the HRP-2, pan-pLDH, or Pf-pLDH test lines and the presence and concentration of rheumatoid factor (p values ranging 0.3 to 1).
We demonstrated that specificity of seven out of ten malaria RDT products was significantly lower in HAT patients compared to controls. RDT products generated false positive test results in various proportions. The problem was most pronounced with pan-pLDH test lines and occurred to a lesser extent with HRP-2 test lines. The Pf-pLDH test lines were not affected. Rheumatoid factor was not associated with false positivity.
Our study has some limitations. It was set up to test the specificity of the RDTs and not their sensitivity to detect clinical malaria. Furthermore, our data only apply to chronic T.b. gambiense HAT, from which we collected specimens in a single HAT focus, and not to the more acute disease form caused by T.b. rhodesiense, and did not address other diseases that might influence malaria RDT specificity. Further, to assess the influence of the persistence of HRP-2, we did a posteriori calculations for specificity excluding patients with history of use of anti-malaria medication one month prior to inclusion. However, given reported persistence of HRP-2 for up to 6 weeks or even longer, this one-month period may have been too short to rule out all interference by HRP-2 persistence [2], [3]. This exclusion was further based on self-reported treatment history but not on parasitologically demonstrated infection in the month prior to sampling. Although the effect on test specificities was limited, the reduced numbers in each group resulted in loss of significance for some tests. Finally, the exact nature of the confounding factor in HAT was not identified.
Despite these limitations, this is the first study carrying out an evaluation of specificity of malaria RDT products in HAT patients. Problems with specificity of malaria RDT products for other non-malarial infectious pathogens have been observed for dengue, schistosomiasis, leishmaniasis and Chagas disease but on limited sample numbers only [7]. Other strengths are the high number of HAT patients tested, and the fact that the study was organised prospectively. Furthermore, RDTs were run side-by- side in a reference setting, limiting sources of variation. PCR was preferred as reference test over microscopy due to its higher sensitivity to detect Plasmodium infection [19], enabling us to eliminate microscopy false negatives from the specificity calculations. Taking microscopy as a more conservative reference test gives a relative underestimate of the specificity of the RDTs (Supporting information file S2). Finally, matching of controls and HAT patients based on the PCR results again confirmed differences in specificity observed with the Fisher exact test, despite data loss.
Although the number of RDT products including Pf-pLDH test lines was limited to three, we confirm that the Pf-pLDH test line seems to be less prone to false positivity, but our findings are in contrast with the high specificity of the pan-pLDH test line previously described [6], [8], [9].
Previous reports have mainly focused on the presence of rheumatoid factor as a reason for false positive HRP-2 test line reactions [4]–[6]. Rheumatoid factor was present in 1.2% of HAT patients only, excluding it as the major cause of false positive malaria RDTs in this patient group. Our results contradict previous findings of rheumatoid factor-like anti-immunoglobulin antibodies in 74% of HAT patients [11]. It is not clear if differences in reagent and methodology, rabbit immunoglobulin in a solid phase anti globulin assay previously [22], versus human IgG in a latex agglutination assay in our measurements, account for these different results. It also seems unlikely that the high IgG concentrations, known to be present in T.b. gambiense HAT sera [23] and known to possibly interfere with quantitation of rheumatoid factor, would account completely for the absence of measurable rheumatoid factor. Although we presently did not assess anti-nuclear antibody, anti-mouse antibody or rapid plasma reagin to be co-responsible for false positivity [7], the abnormally high immunoglobulin concentrations in blood of HAT patients [23] may account on their own for the variability of false positive reactions of HAT specimens in malaria RDTs, to which some brands of tests may be more susceptible than others, depending on the reaction conditions and antibodies used for antigen detection.
Although it cannot be entirely ruled out, it seems improbable that problems of quality among RDT test lots would have accounted for the low malaria RDT specificity in HAT. All RDT products were stored according to the instructions of the manufacturers and were used before their expiry date. Lot to lot variation was assessed for ICT Malaria Combo Cassette Test and obtained specificities were similarly low for the second as for the first lot tested (data not shown). Cross-reactions of specific trypanosome antigens present in blood of HAT patients with the antibodies used in the malaria RDTs also seems unlikely, because of the large variation of proportions of false positive results between the different RDT products.
Although malaria constitutes the main differential diagnosis of HAT [10], this is the first study evaluating accuracy of malaria antigen detection RDTs in HAT. Although in areas with high malaria endemicity such as Bandundu, people of 12 years or older have acquired sufficient immunity against malaria parasites, so that if they have fever or feel unwell the reason should not be automatically ascribed to malaria even if the RDT, or any other diagnostic test, is positive, mistaking HAT for malaria is probably a frequent event. This is indirectly supported by our finding that HAT patients had taken significantly more anti-malaria treatment than controls. In DR Congo, RDTs for malaria are deployed since 2010, with Paracheck Pf being reported to be most often used in 2010 [24], and SD malaria Ag Pf/Pan being actually recommended by the national control program for malaria. Without doubt, other malaria RDTs circulate. Problems with specificity of some malaria RDTs may increase the risk of misdiagnosis or delayed diagnosis of HAT, which is already an important problem [25]. Our findings therefore show the interest of constituting a HAT specimen test panel for evaluation of false positive reactions and including HAT specimens in RDT product testing rounds [7]. This should not only be done for malaria, but also for HIV antibody detection RDTs tests, for which low specificity during HAT infection has been reported earlier [12]. Furthermore, in HAT endemic areas, awareness among health personnel of the possibility of HAT infection, even if tests for other diseases are positive, should be increased. Rapid tests for screening on HAT specific antibodies are now available, and allow identification of HAT suspects even at the level of the health centre [26].
We withhold from proposing a particular malaria RDT to be used in HAT and malaria co-endemic regions since actually we cannot present relevant data on sensitivity of the RDTs in clinical malaria. Parasite densities in immune patients exhibiting symptoms range from 2,500 in infants to 30,000/µl in adults [2]. The highest parasite densities observed in our participants were 978 and 694 parasites/µl, and corresponding blood samples tested positive in all malaria RDTs (data not shown). Although we confirm higher test sensitivity for HRP-2 compared to pLDH test lines [1], the low sensitivity of malaria RDTs in the actual samples can be ascribed to low parasite densities and is of less clinical importance for malaria immune patients. Sensitivity of malaria RDT therefore remains to be tested in a clinically relevant target population [27] before an appropriate malaria RDT for a HAT endemic area can be proposed.
Our results emphasize the importance, in general, to explore the impact of non-targeted infections when evaluating diagnostic tests.
|
10.1371/journal.pbio.2004867 | Dependence of innate lymphoid cell 1 development on NKp46 | NKp46, a natural killer (NK) cell–activating receptor, is involved in NK cell cytotoxicity against virus-infected cells or tumor cells. However, the role of NKp46 in other NKp46+ non-NK innate lymphoid cell (ILC) populations has not yet been characterized. Here, an NKp46 deficiency model of natural cytotoxicity receptor 1 (Ncr1)gfp/gfp and Ncr1gfp/+ mice, i.e., homozygous and heterozygous knockout (KO), was used to explore the role of NKp46 in regulating the development of the NKp46+ ILCs. Surprisingly, our studies demonstrated that homozygous NKp46 deficiency resulted in a nearly complete depletion of the ILC1 subset (ILC1) of group 1 ILCs, and heterozygote KO decreased the number of cells in the ILC1 subset. Moreover, transplantation studies confirmed that ILC1 development depends on NKp46 and that the dependency is cell intrinsic. Interestingly, however, the cell depletion specifically occurred in the ILC1 subset but not in the other ILCs, including ILC2s, ILC3s, and NK cells. Thus, our studies reveal that NKp46 selectively participates in the regulation of ILC1 development.
| Group 1 innate lymphoid cells (ILCs) comprise two subsets: natural killer (NK) cells and ILC1s. Although NK cells and ILC1s are functionally distinct, a factor that regulates one subset but not the other has not been identified. In the current study, we discovered that NKp46, a marker expressed by both NK cells and ILC1s, is critical for the development of ILC1s but not NK cells. In mice lacking NKp46, and in wild-type (WT) mice depleted of immune cells by irradiation and then transplanted with bone marrow (BM) cells lacking NKp46, ILC1s that express cell surface receptor tumor necrosis factor (TNF)-related apoptosis-inducing ligand (TRAIL) were almost completely absent in all organs or tissues examined, including liver, spleen, BM, and small intestine (SI). In contrast, the cell number and signature cytokine expression of all other ILC subsets—namely NK cells, ILC2s, and ILC3s—were not significantly affected. Collectively, our findings provide new evidence supporting an essential role for NKp46 in the development of ILC1s.
| Natural cytotoxicity receptor NKp46, encoded by the Ncr1 gene, is a natural killer (NK) cell–activating receptor that plays roles in regulating the NK cell’s clearance of virus and rejection of tumor [1]. Following binding to its putative ligands, the receptor activates intracellular signaling through immune-receptor tyrosine-based activating motifs (ITAMs) [2]. Some non-NK innate lymphoid cell (ILC) populations also express NKp46, including the ILC1 subset (Lin─NKp46+NK1.1+CD49b─ CD49a+) [3] of group 1 ILCs and the ILC3 subset (Lin─CD127+RORγt+) of group 3 ILCs [4]. However, the role of NKp46 in these non-NK ILCs is still poorly understood. We previously reported that NKp46 defines a subset of NKT cells susceptible to malignant transformation in the presence of interleukin 15 (IL-15) and has a role in the NK cell clearance of herpes simplex virus 1 [5,6]. In the current study, we aimed to unravel the role of NKp46 in regulating the development and function of NKp46+ ILCs, especially ILC1s, using a genetic approach.
An NKp46 knockout (KO) mouse model—in which Ncr1, the gene encoding NKp46, was replaced with green fluorescent protein (gfp) (Ncr1gfp/gfp) [5,7]—was used in this study. The development of ILC populations was assessed in different organs and tissues comparing wild-type (WT) (Ncr1+/+), heterozygous (Ncr1gfp/+), and KO (Ncr1gfp/gfp) mice. The gating strategy for ILC1 flow cytometric analysis is shown in Fig 1A. A clear NK1.1+NKp46─ population (more than 50%) was observed within the NK1.1+ population in the liver (S1A Fig), consistent with previous studies [8,9]. The intensity of CD49a surface expression was higher in the NK1.1+NKp46+CD49b─ CD49a+ ILC1 population than that in the NK1.1+NKp46─ CD49b─ CD49a+ population, and the latter population displayed a phenotype of CD3─/dim (S1 and S2 Figs). Dadi and colleagues recently defined an ILC1-like population, named T cell receptor (TCR) lineage type 1 innate-like T cells (ILTC1), with a phenotype of NK1.1+NKp46─CD49a+ [10]. Moreover, NKp46 is considered a reliable marker for ILC1s and NK cells [9–11]. Thus, in the current study, the ILC1 subset was defined as NK1.1+NKp46+(or GFP+ for KO mice)CD49b─CD49a+ (Fig 1A), which does not include the NK1.1+NKp46─CD49b─CD49a+ population. It was surprising that, when compared to Ncr1+/+ littermates, proportionally the ILC1 subset (CD49b─CD49a+) was nearly completely absent in the liver of Ncr1gfp/gfp mice and was significantly decreased in the liver of Ncr1gfp/+ mice (Fig 1A, bottom panel; Fig 1B, upper panel; S2 Fig, upper panel), while the percentage of the NK1.1+NKp46─(or GFP─ for KO mice)CD49b─CD49a+ population seems to be no different among the Ncr1gfp/gfp, Ncr1+/gfp, and Ncr1+/+ groups (S2 Fig, bottom panel). Furthermore, quantification of the ILC1 subset showed that the absolute cell quantity was also drastically reduced in the liver of the Ncr1gfp/gfp mice and moderately reduced in the liver of Ncr1gfp/+ mice compared to their Ncr1+/+ littermates (Fig 1B, lower panel). We also used markers including CD62L, Eomesodermin (Eomes), and T-box expressed in T cells (T-bet) to confirm our observation regarding dependency on NKp46 for ILC1 development by comparing Ncr1gfp/gfp mice and/or Ncr1+/gfp mice to Ncr1+/+ mice (S3 and S4 Figs). The development of ILC populations was also assessed in other organs and tissues using the liver as our point of reference and comparing results in WT (Ncr1+/+) mice to Ncr1gfp/gfp mice (Fig 1C and 1D). Cell quantification by flow cytometry indicated that ILC1s (CD49b─CD49a+) were nearly completely absent in the bone marrow (BM), spleen, and small intestine (SI) of Ncr1gfp/gfp mice (Fig 1E). In contrast, the absolute number of NK cells, which are the main population that expresses NKp46 in tested organs or tissues, did not significantly change in Ncr1gfp/gfp mice compared to their Ncr1+/+ littermates, consistent with a previous report [7] (Fig 1E, lower panel).
TNF-related apoptosis-inducing ligand (TRAIL) is a functional protein, selectively expressed on the ILC1 subset, that plays an essential role in mediating the cytotoxicity of this population against target cells through triggering the death receptor-transduced signaling pathway [12]. The expression of TRAIL and lack of CD49b expression can also be used to distinguish ILC1s from NK cells [13] because resting NK cells in WT mice do not express TRAIL, while ILC1s do (Fig 2A, left). Consistent with the data shown in Fig 1 regarding the near-complete lack of CD49a+CD49b─ ILC1s, the TRAIL+CD49b─ population of NKp46+(or GFP+)NK1.1+ type I ILCs was barely detectable in the liver (Fig 2A and 2B) and other organs (Fig 2C) of Ncr1gfp/gfp mice compared to their Ncr1+/+ littermate controls.
Due to our observation that NKp46 deficiency restricted the development of ILC1s, we next set out to test whether this effect also occurred in other ILC subsets. However, using the gating strategy as described in Fig 3A, we did not observe a significant difference in the quantities or frequencies of Lin─CD127+Gata3+ ILC2s and Lin─CD127+RORγt+ ILC3s when Ncr1gfp/gfp mice were compared to their Ncr1+/+ littermate controls (Fig 3B and 3C). There was an insufficient quantity of ILC1 cells from Ncr1gfp/gfp mice to study how this NKp46 deficiency affects ILC1 function(s); however, we did observe that the interferon γ (IFN-γ) production by NK cells in response to co-stimulation with IL-12 and IL-18 was unaltered between cells isolated from Ncr1gfp/gfp mice versus those from Ncr1+/+ littermate controls (Fig 3D). Likewise, IL-22 production by ILC3 cells isolated from Ncr1gfp/gfp mice versus those from Ncr1+/+ littermate controls was not significantly different (Fig 3E). Consistent with our results, Satoh-Takayama and colleagues previously demonstrated that NKp46 is not required for IL-22-mediated intestinal innate immune cells in the gut to defend against Citrobacter rodentium [14]. Together, these results suggest that NKp46 does not control homeostasis or signature ILC cytokine production of ILC2s, NK cells, or ILC3s, but does selectively participate in the regulation of ILC1 development.
Although NK cells and ILC1s are closely related, ILC1s do not develop through Lin─CD122+NK1.1─DX5─ NK cell precursors (NKPs) but can develop through Lin─c-kitlowα4β7+CD127+CD25─ Flt3─ common helper innate lymphoid precursors (CHILPs). Both NKPs and CHILPs are derived from a common lymphoid progenitor (CLP) [15]. The proportion of both types of precursors are similar in the BM of Ncr1gfp/gfp and Ncr1+/+ mice (S5 Fig). To validate that NKp46 deficiency results in the near-complete absence of ILC1s, and to test whether this phenomenon is cell intrinsic or extrinsic, BM transplantation was undertaken. For this purpose, irradiated WT CD45.1 recipients were engrafted with Ncr1gfp/gfp (KO) or Ncr1+/+ (WT) CD45.2 donor BM cells (Fig 4A) or a 1:1 mixture of KO and WT (S6 Fig) via a tail-vein injection. Two weeks later, quantification of various ILC subsets was undertaken by flow cytometric analysis. Donor cells and host cells were distinguished by staining cells with an anti-CD45.2 antibody. We found that there was a nearly complete absence of ILC1s in the liver, spleen, and BM of WT recipients when Ncr1gfp/gfp BM cells were used as donor cells. However, the ILC1 population was present in significantly larger quantities when Ncr1+/+ BM cells were used as donor cells (Fig 4B and 4C and S6 Fig). In contrast, the quantity of ILC2 cells and ILC3 cells in the SI was not affected in recipient mice engrafted with Ncr1gfp/gfp or Ncr1+/+ BM donor cells (Fig 4D and 4E). The moderate increase in the proportion of NK cells could be occurring due to the complete lack of ILC1s among NKp46+NK1.1+ type I ILCs (Fig 4B top right; Fig 4C bottom right). Collectively, these results further confirm that ILC1 development depends on NKp46, and this dependency is cell-intrinsic (Fig 1).
In conclusion, our findings provide novel evidence that NKp46 plays a critical role in ILC1 development. Previous studies in this area focused on transcriptional control of ILC development. It is known that T-bet and Eomes regulate NK cell development, Gata3 controls ILC2 development, and RORγt defines the ILC3 lineage [15]. Several transcription factors—such as nuclear factor interleukin 3 regulated (Nfil3), runt related transcription factor 3 (Runx3), and T-bet—have been found to control ILC1 development; however, these factors do not play a selective role in determining ILC1 development [15]. That is, these transcription factors have some overlapping roles in several types of ILCs and thus individually cannot determine the fate of ILC1 development. For example, T-bet has been shown to play a role not only in ILC1 development but also in NK cell and ILC3 development [15]. Here, we identified a receptor that selectively determines the developmental fate of ILC1s. Our study also supports the notion that ILC1s and NK cells belong to different lineages, although both belong to group 1 ILCs, and both can produce IFN-γ when activated (e.g., by cytokines). Our study also shows that Ncr1gfp/gfp mice may serve as a useful animal model for investigating the physiological or pathological functions of ILC1s, given their near-complete absence in various organs, while the development and function of other ILCs are kept nearly intact, with the exception of functions related to NKp46’s role in NK cells [7,16].
All animal experiments were performed according to the protocol (# 2012A00000090), which has been approved by The Ohio State University Institutional Animal Care and Use Committee (IACUC). No human subjects were involved in this study.
NKp46 KO C57BL/6 mice with Ncr1 replaced with gfp (Ncr1gfp/gfp) [5,7] and their heterozygous littermates (Ncr1gfp/+) were used. NKp46 KO C57BL/6 mice were previously described [7]. Congenic CD45.1+ C57BL/6 mice for transplant experiments were purchased from Jackson Laboratory. All mice used in these studies were 10–12 weeks of age.
Cells were labelled with flow antibodies for 15 minutes in the dark in PBS containing 2% BSA. Labelled cells were washed twice and resuspended in PBS containing 2% BSA. The prepared samples were analyzed using an LSR-II flow cytometer (BD Bioscience) or sorted using an Aria II cell sorter (BD Bioscience). Anti-CD19-PeCy7 (561739), anti-CD3-PeCy7 (552774), anti-NK1.1-BV421 (562921), anti-NKp46-FITC (560756), anti-NKp46-AF647 (560755), anti-CD49b-PE (553858), anti-CD49a-PerCP-Cy5.5 (564862), anti-CD62L-APC (553152), anti-T-bet-APC (561264), anti-CD45.2-AF700 (560693), anti-CD45.2-FITC (553772), anti-Gata3-AF647 (560068), anti-RORγt-PE (562607), anti-CD127-V450 (561205), anti-CD117-PE (553869), anti-LPAM-1-APC (562376), anti-Flt3-BV421 (566292), anti-Ly-6A/E-APC (565355), and anti-CD122-PE (553362) were purchased from BD Bioscience. Anti-IFN-γ-AF-700 (505823) and anti-CD25-Pacific Blue (102022) were purchased from Biolegend. Anti-CD253-APC (17-5951-82), anti-Eomes-PE (12-4875-82), and anti-CD127-PerCP-Cy5.5 (45-1271-80) were purchased from eBioscience.
For BM cell transplantation, CD45.2+ donor BM cells collected from Ncr1gfp/gfp mice (5×106), Ncr1+/+ mice (5×106), or a mixture of these 2 types of cells at a ratio of 1:1 (10×106 in total) were injected IV into the CD45.1+ congenic recipient mice, which were lethally irradiated (4 cGy twice on the same day) using an X-ray irradiator. The detailed protocol was described in our previous studies [17,18].
1 × 105 NK cells were seeded into a 96-well plate and cultured with or without IL-12 plus IL-18 for 24 h. Supernatants were then harvested to detect IFN-γ production, which was assessed by the Mouse IFN-gamma Uncoated ELISA Kit (Catalog #88-7314-88, Invitrogen)
Student's t test or paired t test was used to analyze two independent or paired groups, respectively. A p value less than 0.05 was considered statistically significant.
|
10.1371/journal.pgen.1005686 | Comparative Genomic Analyses of the Human NPHP1 Locus Reveal Complex Genomic Architecture and Its Regional Evolution in Primates | Many loci in the human genome harbor complex genomic structures that can result in susceptibility to genomic rearrangements leading to various genomic disorders. Nephronophthisis 1 (NPHP1, MIM# 256100) is an autosomal recessive disorder that can be caused by defects of NPHP1; the gene maps within the human 2q13 region where low copy repeats (LCRs) are abundant. Loss of function of NPHP1 is responsible for approximately 85% of the NPHP1 cases—about 80% of such individuals carry a large recurrent homozygous NPHP1 deletion that occurs via nonallelic homologous recombination (NAHR) between two flanking directly oriented ~45 kb LCRs. Published data revealed a non-pathogenic inversion polymorphism involving the NPHP1 gene flanked by two inverted ~358 kb LCRs. Using optical mapping and array-comparative genomic hybridization, we identified three potential novel structural variant (SV) haplotypes at the NPHP1 locus that may protect a haploid genome from the NPHP1 deletion. Inter-species comparative genomic analyses among primate genomes revealed massive genomic changes during evolution. The aggregated data suggest that dynamic genomic rearrangements occurred historically within the NPHP1 locus and generated SV haplotypes observed in the human population today, which may confer differential susceptibility to genomic instability and the NPHP1 deletion within a personal genome. Our study documents diverse SV haplotypes at a complex LCR-laden human genomic region. Comparative analyses provide a model for how this complex region arose during primate evolution, and studies among humans suggest that intra-species polymorphism may potentially modulate an individual’s susceptibility to acquiring disease-associated alleles.
| Genomic instability due to the intrinsic sequence architecture of the genome, such as low copy repeats (LCRs), is a major contributor to de novo mutations that can occur in the process of human genome evolution. LCRs can mediate genomic rearrangements associated with genomic disorders by acting as substrates for nonallelic homologous recombination. Juvenile-onset nephronophthisis 1 is the most frequent genetic cause of renal failure in children. An LCR-mediated, homozygous common recurrent deletion encompassing NPHP1 is found in the majority of affected subjects, while heterozygous deletion representing the nephronophthisis 1 recessive carrier state is frequently observed amongst world populations. Interestingly, the human NPHP1 locus is located proximal to the head-to-head fusion site of two ancestral chromosomes that occurred in the great apes, which resulted in a reduction of chromosome number from 48 in nonhuman primates to the current 46 in humans. In this study, we characterized and provided evidence for the diverse genomic architecture at the NPHP1 locus and potential structural variant haplotypes in the human population. Furthermore, our analyses of primate genomes shed light on the massive changes of genomic architecture at the human NPHP1 locus and delineated a model for the emergence of the LCRs during primate evolution.
| Genomic instability is a major contributor to de novo mutations that can occur in the process of human genome evolution [1–3]. Genomic rearrangements can be mediated by various mechanisms, including nonallelic homologous recombination (NAHR), nonhomologous end joining, mobile element insertion (e.g. long interspersed element (LINE)-mediated retrotransposition) and replication based mechanisms [4]. Low copy repeat (LCR) mediated NAHR plays a significant role in genomic instability resulting in rearrangements associated with genomic disorders [5]. LCRs, also known as segmental duplications, are two or more repeated sequences that usually span 10–400 kilobases (kb) each and share >95% DNA sequence identity [6,7]. LCRs are highly homologous, and constitute ~5–10% of the human and great ape genomes [6,8,9]. LCRs provide substrates for NAHR-mediated crossing-over that results in structural variants (SVs) including copy number variants (CNVs) such as duplications and deletions of large genomic segments [5] or copy number neutral events such as inversions [10–12].
Numerous NAHR-mediated rearrangements are associated with genomic disorders by affecting dosage sensitive genes. For example, Potocki-Lupski syndrome (PTLS, MIM #610883) or Smith-Magenis syndrome (SMS, MIM #182290) are frequently caused by an ~3.7 megabases (Mb) NAHR-mediated common recurrent duplication or deletion, respectively. These recurrent rearrangements of 17p11.2 utilize directly oriented proximal and distal SMS-REPs as substrates for NAHR [13–18].
LCRs originated from genomic evolutionary processes and can facilitate responses to selective pressure by creating new genes that may contribute to lineage-specific phenotypes. LCRs can also configure local genomic structure in a manner that contributes significantly to disease susceptibility [19–24]. Because of their repetitive nature and structural complexity, LCRs can confound the accuracy of human and nonhuman mammalian genome assemblies. Discerning long stretches of paralogous, highly identical sequences can be difficult; this problem becomes particularly challenging when there are more than two copies in a haploid genome [6,25,26], and consequently LCRs are likely under-represented in draft genome assemblies for many species. Mappability of the short sequencing reads from next generation sequencing techniques can be reduced within LCRs, and as a result multiple experimental molecular and computational approaches are often required to characterize SVs relative to the human haploid reference in a given personal genome. Several efforts have demonstrated the value of thoroughly scrutinizing complex genomic regions to better understand the human genome and discern variation that may be important to health, evolution, and susceptibility to diseases [27–33].
The human chromosomal region 2q13-2q14.1 represents the product of head-to-head fusion of two ancestral chromosomes forming human chromosome 2 [34]. This evolutionary fusion event is unique to the human genome, and is responsible for the chromosome number difference (46 versus 48) between human and the great apes including chimpanzee (Pan troglodytes), gorilla (Gorilla gorilla) and orangutan (Pongo abelii). The fusion of two subtelomeric regions from two ancestral chromosomes (analogous chromosomes 2A and 2B in the great apes) introduced substantial complexity to this region. A common recurrent 290 kb deletion encompassing Nephrocystin-1 (NPHP1, MIM *607100), a gene that maps to the centromeric portion of the human 2q13 region, is associated with several diseases. Juvenile-onset nephronophthisis 1 (NPHP1, MIM #256100) is an autosomal recessive cystic kidney disorder causing chronic renal failure in children. Homozygous NPHP1 deletion is found in ~80% of patients born to consanguineous parents and in ~60% of sporadic cases [35]. In addition to nephronophthisis 1, the same NPHP1 deletion has also been identified in patients with Senior-Loken syndrome-1 (SLSN1, MIM# 266900) and Joubert syndrome 4 (JBTS4, MIM# 609583) with distinct phenotypes [35–37]. Moreover, a recent study demonstrates that heterozygous NPHP1 deletion CNV in combination with NPHP1 point mutations (SNVs) can lead to Bardet-Biedl syndrome (BBS, MIM# 209900) [38]. The NPHP1 deletion is recurrent, and results from NAHR-mediated unequal crossing-over involving the directly oriented flanking LCRs [39]; the frequency of heterozygous NPHP1 deletion is estimated to be approximately 1/400 in normal individuals from northern European descent [38]. Dittwald et al explored a clinical database containing chromosomal microarray (CMA) data from 25,144 patients, of which NPHP1 duplications (N = 233) and deletions (N = 118) were found to be the most commonly observed copy number aberrations (combined ~1.4%) compared to CNVs from other loci [5].
The complex genomic architecture of the human 2q13 region, especially the NPHP1 locus, provides the foundation for different SVs that may be observed in personal genomes among human populations. The polymorphic nature of this locus was previously demonstrated, and its evolution, prevalence and potential impact to disease susceptibility warrant further investigation. In this study, we utilized a combination of genomic technologies, including array-comparative genomic hybridization (aCGH) and optical mapping (OM), to identify novel SV haplotypes at the NPHP1 locus and clarify the relative frequencies of specific haplotypes in the human population. We further utilized comparative sequence alignments of primate genome sequences and aCGH to construct a model for the evolution of the genomic architecture at the NPHP1 locus in nonhuman primates and human genomes. Unexpectedly, we found that this region displays evidence for incomplete lineage sorting, such that the structure of this region in humans is more similar to that of gorillas than to the orthologous region in chimpanzees or orangutans. The results also confirmed a dramatic genomic expansion of the NPHP1 locus during primate evolution and revealed a pattern of LCR evolution that may be explained by a model of multi-step, serial segmental duplication [32].
The exceedingly complex and polymorphic genomic architecture of the NPHP1 locus presented difficulty during the assembly of the haploid human genome reference. This becomes readily apparent by the comparison between two recent updates of the human reference assembly, GRCh37/hg19 and NCBI36/hg18, at the NPHP1 locus. The 800 kb sequences flanking each end of NPHP1 in hg18 and hg19 were compared by Miropeats [40]. Two (Gaps II and III) of the three major gaps (Gaps I, II and III) in hg18 are closed in hg19; Gap II corresponds to a region encompassing an ~45 kb LCR distal to NPHP1 that is only present in hg19 (Fig 1A).
We delineated the LCR structure of the NPHP1 locus (hg19, chr2: 110,080,914–111,762,639) using data from the UCSC Genome Browser “Segmental Dups” track (http://genome.ucsc.edu/index.html). Two LCRs of approximately 358 kb in length, which we termed 358PROX (centromeric) and 358DIST (telomeric), flanked NPHP1 in an inverted orientation–this can also be manifested by the Miropeats (Fig 1A). Similarly, two LCRs of approximately 45 kb in length, termed 45PROX (centromeric) and 45DIST (telomeric), were embedded within 358PROX and 358DIST, and thus had an inverted orientation. One additional LCR paralogous to 45PROX and 45DIST, termed 45MID, was revealed in build hg19 after the closure of Gap II from hg18, resulting in a total of three copies of the 45 kb LCRs in the haploid reference. The 45MID and 45PROX, flanking NPHP1 in direct orientation, are presumably the LCR pairs responsible for the NAHR-mediated NPHP1 deletion (Fig 1A). Self-alignments of the DNA sequences (chr2: 110,080,914–111,762,639) using NCBI BLAST tool (http://blast.ncbi.nlm.nih.gov/Blast.cgi) confirmed the genomic structure of the NPHP1 locus in the reference by revealing the two major groups of LCRs and their relative orientations (S1A Fig).
We computationally characterized the LCRs at the NPHP1 locus using the human reference sequences. Pairwise alignments of both groups of 45 kb LCRs and 358 kb LCRs against their individual consensus sequences revealed a high percentage of sequence identities for both the 358 kb LCRs and 45 kb LCRs (Table 1). PR domain-containing protein 9 (PRDM9) recognizes a degenerative 13-mer motif (5’-CCNCCNTNNCCNC-3’) that is critical to recruit recombination machinery required for crossovers in at least 40% of all human homologous recombination hot spots [4,41–44]. NAHR crossover studies suggest that the frequency of the PRDM9 hotspot motifs within LCR regions is one of the parameters correlated with the rate of NAHR mediated genomic rearrangement [5,17]. Characterization of the PRDM9 hot spot motif in the LCRs of the NPHP1 locus may elucidate potential crossover sites. At the NPHP1 locus, 12 motifs were found in each of the three 45 kb LCRs, while 161 and 154 hotspot motifs were found in 358PROX and 358DIST, respectively (Table 1). LCRs contribute to the complex genomic architecture of this region, and could incite genomic instability. The high degree of sequence identity between LCR pairs and the density of PRDM9 hotspot motifs (Table 1, S1B Fig) may additionally contribute to the instability and increase the recurrent rearrangement frequency at the NPHP1 locus. The high similarity (>99.6%) between the corresponding paralogous LCRs in humans also indicates that gene conversion may occur frequently at the NPHP1 region [45].
NAHR events between directly oriented LCRs generate deletions or duplications; while NAHR events between inverted-oriented LCRs lead to inversions–such copy number neutral SVs may impose weaker selection forces than deletions and duplications do, and are thus more likely to be found as population polymorphisms [27–29]. In fact, the SV haplotype at the NPHP1 locus identified in the reference genome, arbitrarily designated as the H1 SV haplotype, is not the only SV haplotype in the human population. Experimental evidence suggested the presence of at least three alternative SV haplotypes (H2, H3 and H4, Fig 1) before the first draft of the human genome assembly [39]. These alternative SV haplotypes share an inversion of the NPHP1 region, encompassing NPHP1 and 45MID, between 358PROX and 358DIST (NPHP1 inversion). Besides the NPHP1 inversion, H2 and H4 appear to have several other SVs unique to themselves – 45PROX was lost in H2, while both 45PROX and 45DIST were lost in H4 (Fig 1B).
We hypothesized that the majority of the structural polymorphism at the NPHP1 locus can derive from (1) the NPHP1 inversion, (2) the copy number loss of one or more of the 45 kb LCRs, or (3) a combination of these two events. To obtain further evidence to support these SV haplotypes and their prevalence across human populations, we examined the fosmid libraries from the Human Genome Structural Variation project (HGSV, http://humanparalogy.gs.washington.edu/structuralvariation/) to search for individual discordant fosmids representing SVs [12]. We identified 78 discordant fosmids representing losses of either 45PROX or 45DIST and 31 discordant clones representing NPHP1 inversions between 358PROX and 358DIST in a total number of 17 individuals (Fig 2A). Each of the 17 individuals had at least one discordant fosmid indicating loss of the 45 kb LCR, while 13/17 had at least one discordant fosmid indicating the NPHP1 inversion, suggesting that the current human reference genome actually presents a minor SV allele. We further examined the copy number distribution of the 45 kb LCRs utilizing the dataset published by Conrad et al [2]. CNVs in 450 individuals from different ethnicity groups were genotyped. These individuals include 180 CEU (Utah residents with ancestry from northern and western Europe), 180 YRI (Yoruba in Ibadan, Nigeria), 45 JPT (Japanese in Tokyo, Japan) and 45 CHB (Han Chinese in Beijing, China). Various copy numbers of the 45 kb LCRs, ranging from two to six, were observed at different frequencies in each population (Fig 2B). The distributions of copy numbers across different populations were not significantly different (Kruskal-Wallis rank sum test, p-value = 0.6766, Fig 2C). Aggregating all the populations, 1%, 13%, 56%, 25% and 5% of the entire examined population has two, three, four, five and six copies of the 45 kb LCRs, respectively (Fig 2B). The frequencies of copy number derived from Conrad et al also correlated well with those derived from PFGE experimental data by Saunier et al [39], in which 13%, 21% and 1.3% of 152 control individuals from an undefined ethnicity were found to harbor three, five and six copies of 45 kb LCRs, respectively (Fig 2B). It is likely that the most common copy number of the 45 kb LCRs in a diploid genome is four, which deviates from the copy number of six that would exist in an individual with homozygous H1, as in the haploid reference genome. Thus the six-copy state may be a minor genotype that was represented in only 1.3%-5% of the general population.
The observation of polymorphic structural variants, including copy number polymorphisms of the 45 kb LCRs and NPHP1 inversion, prompted us to search for novel SV haplotypes. OM constructs ordered restriction maps (Rmaps) from single-molecules of DNA, which are assembled into genome-wide contigs that can be compared to an in silico restriction map from the human reference in order to discern SVs [46]. The OM can be used as an independent validation method for SVs revealed by other methods, such as fosmid sequencing data, which suggested SVs including the NPHP1 inversion and loss of the 45 kb LCRs in HapMap individual NA15510 (Fig 3A). The SwaI OM contig assembly of NA15510 and its alignment to an in silico human reference (hg19) further validated a homozygous H2 SV haplotype at the interrogated locus (chr2: 109,943,987–111,547,676), with four copies of the 45 kb LCRs and homozygous NPHP1 inversion (Fig 3B). Additionally, NA10860 and NA18994 yielded OM results supporting genotypes identical to NA15510 (Fig 3B).
Fluorescence in situ hybridization (FISH) was performed in an attempt to delineate the organization of NPHP1 and the 45 kb LCRs at the NPHP1 locus. FISH experiments were designed with fosmid probes independently targeting NPHP1 (fluoresces green, G) and the 45 kb LCRs (fluoresces red, R). A fosmid probe targeting ~1.3 Mb proximal to NPHP1 was used as an “anchor probe” (fluoresces blue, B, S2A Fig). In the majority of interphase cells from the lymphoblastoid cell line of NA15510 (42/50), we observed resolved signals of four red and two green (R4G2), which represented four copies of the 45 kb LCR and two copies of NPHP1 in a diploid genome (S2B Fig). This result was consistent with the copy number data from OM. Additionally, a signal pattern of red-green-red, representing an SV haplotype also consistent with the OM data, was observed in the interphase cells (S2B Fig); however, such organization of signals was confounded by a signal pattern of yellow-red, the yellow of which was likely to represent an overlapping signal between red and green due to a two dimensional representation of a three dimensional reality and the close physical proximity, or overlapping in the z–plane, of a red and green signal. Moreover, using the blue “anchor probe” as a third color, we observed a signal pattern of blue-red-green-red in 25/50 interphase cells examined. This experimental result was also consistent with the OM data observed for this sample (S2C Fig). However, such a pattern was not uniformly and consistently found (S2C Fig). In aggregate, these results may be explained by the close proximity of the components being targeted, the three dimensional relative spatial positioning, and the less organized structure of chromosomes in interphase cells.
We then performed custom-designed aCGH (S3A Fig) to validate the copy number estimations from OM. We used the DNA sample from NA10851, a genome that has four copies of the 45 kb LCRs, as the universal reference for aCGH experiments. As a proof of principle, aCGH comparing six DNA samples (NA18517, NA15510, NA18994, NA10860, NA18555 and NA12878) with NA10851 confirmed the copy number of the 45 kb LCRs estimated by the Conrad et al study (Fig 3C, S3B Fig). These samples included the three samples (NA15510, NA18994, NA10860) interrogated by OM. The consistency of copy numbers predicted by Conrad et al, aCGH, OM and the corroboration of the independent experimental assays including FISH further substantiated the accuracy of OM analysis for discerning CNV/SVs.
A total number of eight DNA samples were analyzed by OM, and the contig assemblies at the NPHP1 locus revealed different SV haplotypes (Table 2). In addition to NA10860/NA18994/NA15510, MM52 (a multiple myeloma primary tumor sample) [47] and HF087 (an oligodendroglioma primary sample) [48] were also found to be homozygous for the H2 SV haplotype by OM (Table 2). In the DNA from CHM (complete hydatidiform mole, CHM1h-TERT) [46], the OM consensus map showed an allele with the loss of the 45PROX but without the NPHP1 inversion, presenting a potential novel SV haplotype (termed H5, Fig 3E and 3F, Table 2). CHM is derived from fertilization of an enucleated egg with a single sperm [29], and the haploid nature of the CHM genome facilitates accurate mapping and assembly by eliminating allelic variations from the diploid genomes. The OM analysis of H1-ES-P208 (human embryonic stem cell line, passage 208) revealed alleles with three different SV haplotypes, including one H5 SV haplotype and two other novel SV haplotypes, one with the loss of the 45DIST and the NPHP1 inversion (termed H6) and the other one with losses of both 45PROX and 45DIST without the NPHP1 inversion (termed H7) (Fig 3G and 3H, Table 2). It is interesting that three different SV haplotypes in H1-ES-P208 were identified by OM, suggesting potential mosaicism. An NAHR-mediated inversion could potentially occur between the inverted 358 kb LCRs to convert H5 to H6 or vice versa. The highly identical paralogous LCRs may facilitate this rearrangement during the additional mitoses from the 208 cell culture passages of H1-ES-P208 [49,50]. aCGH analysis of H1-ES-P208 supported the copy number of three in its diploid genome as inferred by the OM analysis. Thus it led to the possibility that an admixture of cells with H5/H7 or H6/H7 combinations, both of which represent three copies of the 45 kb LCRs, could be present in the H1-ES-P208 cell line (Table 2). This finding should be further validated using orthogonal approaches that may delineate SV haplotypes. Unfortunately, the H1-ES-P208 cell line is no longer available. The DNA sample from HCC1937 (a lymphoblastoid cell line from primary ductal carcinoma) revealed two different SV haplotypes, H2 and H5. These results emphasize the structural complexity of the NPHP1 locus and indicate that two (or potentially sometimes more reflecting mosaic states) SV haplotypes may be observed in the genome of one individual. However, although novel SV haplotypes are identified based on data from the aforementioned cell lines, it remains to be examined how representative they are of the different human populations worldwide.
The 45 kb LCRs are responsible for the NAHR-mediated recurrent NPHP1 deletion. We have shown above that the copy number of this LCR is highly dynamic in the human population. To better understand the origin of this complexity and assess homologous genomic regions in closely related species, we investigated this region in great apes (chimpanzees, gorillas and orangutans) and Old World monkeys (rhesus macaques [Macaca mulatta] and baboons [Papio anubis]). Analysis of the evolutionary history of this interval may represent a unique opportunity to characterize the emergence of a repeat sequence that causes susceptibility to a specific disease in humans. We were able to trace the human 45 kb LCR locus back to its ancestral origin by comparing several nonhuman primate genomes with the SV haplotypes observed in humans.
Based on the OM data, the deletion of the 45 kb LCR resulted in an ~40 kb loss in the human subjects compared to the human reference, with an ~6 kb mismatching sequence remaining at the place of the 45 kb LCR loss (Fig 3B). Alignments of human discordant fosmid clones with fully sequenced inserts, mapped to either the proximal or the distal side of human NPHP1 (hg19), revealed a shared “deletion/insertion” haplotype with the 45 kb LCR deletion and a 5936 bp insertion (5936Ins) at the deletion breakpoint junction (Figs 4A and 3B, S4 Fig). The 5936 bp stuffer sequence could not be uniquely mapped to any position of the human genome reference hg19 using BLAT (http://genome.ucsc.edu/cgi-bin/hgBlat). We further investigated the origin of the 5936Ins using BLAST (http://blast.ncbi.nlm.nih.gov/Blast.cgi) with the Nucleotide collection (nr/nt) database. Interestingly, in addition to the five fully sequenced clones previously found in the human fosmid libraries, we found two chimpanzee BAC clones, CH251-328D3 and CH251-71L9, which encompassed a sequence highly identical (98%) to the human 5936Ins. The two clones mapped to a region in the chimpanzee chromosome 2A –a position syntenic to the human NPHP1 locus. This finding indicated that the 5936Ins that was present at some but not all human NPHP1 loci was also present in at least some chimpanzee genomes. We subsequently used BLAT to search for this sequence in the genomes of baboon, rhesus macaque, orangutan, gorilla and chimpanzee. This 5936Ins could be aligned to the reference genomes of all these species with increasing sequence identities, meaning it may exist as an ortholog in Old World monkeys and great apes (Table 3).
From ~85 million years ago (Mya) to the present, primate genomes have undergone substantial sequence change during evolution. Genomic segmental duplications have been one significant aspect of that process [21,51,52]. The human 5936Ins that is conserved among baboons, rhesus macaques and great apes motivated us to use comparative genomic analyses to understand the dynamic structural changes that occurred at the NPHP1 locus during the evolution of the human genome.
DNA sequences of 800 kb flanking each side of NPHP1 were downloaded from UCSC Genome Browser (http://genome.ucsc.edu/index.html) for baboon, rhesus macaque, orangutan, gorilla, chimpanzee and human. According to our previous analysis, a minor allele is presented at the NPHP1 locus in the hg19. As a result, we modified the human reference sequence so that it reflected the more common H2 SV haplotype (Table 2). We manually constructed the H2 SV haplotype sequences by: (1) deleting the sequences of 45PROX, (2) inserting the 5936Ins and (3) inverting the sequences between 358PROX and 358DIST (Fig 5A–5F). Miropeats [40] was subsequently used to perform local alignments between the human H2 and each individual primate reference genome. Alignment between human H1 and H2 revealed the gain of the 45PROX in H1, the overall sequence similarity, and the inverted orientation of the paralogous 358 kb LCRs (Fig 5A). Old World monkeys, including baboons and rhesus macaques, diverged from the ancestors of humans about 25–33 million years ago (Mya) and share 94.9% sequence identity with the human genome [53]. Due to lower sequence identity between baboon/macaque and human, the Miropeats threshold was lowered to display the alignments between more distant species; this change in the threshold correlated with more noise in the alignments. The diagram of alignments suggested that sequences homologous to the human 45 kb LCRs do not exist in the current baboon/macaque reference genomes, as the only traces between these references and human H2 were sparse, and likely indicated noisy alignments (Fig 5E and 5F). A region on the left side of both baboon and rhesus macaque NPHP1 could be aligned to a portion of both human 358 kb LCRs, indicating that a smaller region orthologous to part of the human 358 kb LCRs exists in both of the baboon and macaque references, but it appears as only a single copy per haploid genome (Fig 5E and 5F). In both baboon and rhesus macaque, the intra-species Miropeats alignments also revealed a lack of the pattern of paralogous LCRs, which was observed in the human reference genome (S5E and S5F Fig). The great ape lineages including orangutans, gorillas and chimpanzees diverged from the human evolutionary lineage about 12–16 Mya, 6–8 Mya and 4.5–6 Mya, respectively, with increasing sequence identity to the human genome (S1 Table) [53]. Miropeats revealed similar patterns of alignment of orangutan and chimpanzee’s references versus human H2: two paralogous 45 kb LCRs in the human reference were either directly or invertedly aligned to a single genomic region in the references of orangutan and chimpanzee; while the two 358 kb LCRs in the human reference were partially aligned to a single genomic region on the left side of NPHP1 in the orangutan and chimpanzee references, with chimpanzee appearing to have a larger partial alignment (Fig 5B and 5D). Although the 45 kb and 358 kb LCRs in human could be aligned to a region in orangutan and chimpanzee references with longer sequence homology, these genomic segments were lacking paralogous LCR partners. Consistently, intra-species alignments by Miropeats did not reveal any pattern of paralogous LCRs in their reference genomes (S5B and S5D Fig).
Gorilla is a more distant species from human than chimpanzee according to the standard evolutionary phylogeny; however, it had a Miropeats pattern more consistent with that of a human-human alignment. Each of the two human 45 kb LCRs in H2 could be apparently aligned to three genomic regions surrounding NPHP1 in the gorilla reference, suggesting an SV haplotype in gorilla that is orthologous but with lower sequence homology (Fig 5C). Moreover, each of the two human 358 kb LCRs could also be aligned to the flanking sequences of gorilla NPHP1 at the paralogous positions, suggestive of a potential orthologous locus in gorilla (Fig 5C). Intra-species alignment of the gorilla reference genome to itself revealed paralogous alignment of long sequences flanking NPHP1, suggesting that a genomic architecture similar to the one observed in humans at the NPHP1 locus at least in the gorilla reference (S5C Fig).
Miropeats also revealed the conservation of the human 5936Ins in the nonhuman primate genomes. As shown by Miropeats, the 5936Ins observed in the human H2 was also identified in the reference genomes of baboon, rhesus macaque, orangutan and chimpanzee (Fig 5B, 5D–5F). Interestingly, the gorilla reference contained two regions similar to the human 5936Ins that were adjacent to two gorilla 45 kb LCR orthologs (Fig 5C). The relative transitions of the human 45 kb LCRs and 5936Ins in different nonhuman primates derived from Miropeats were confirmed by BLAT using the sequences of the human 45MID and the 5936Ins as templates (Table 3). Since Miropeats was applied to reference genomes to reveal similarities and differences between humans and nonhuman primates, the comparison results may not represent the general populations of queried species. Orthogonal experimental approaches, including aCGH and copy number analysis of genomic sequencing data generated from multiple primate individuals, were used to further investigate the copy number changes in the nonhuman primates comparing to humans.
Interphase FISH experiments were performed on lymphoblastoid cell lines of one chimpanzee (CRL-1868) and one gorilla (CRL-1854) using the human fosmid probes described above to explore the genomic architecture at the NPHP1 locus in each species. The experiment illustrated the 45 kb LCR orthologs in both species tested. The majority of the scored gorilla interphase cells (46/50) and all of the scored chimpanzee interphase cells (50/50) showed R2G2, suggesting one copy of the 45 kb LCR ortholog and one copy of NPHP1 on the haploid genome of each individual (S2D and S2E Fig). Interestingly, resolved signals of R4G2 was observed in a minority population (4/50) of gorilla interphase cells, indicating a potential two-copy configuration of the 45 kb LCR ortholog in the gorilla haploid genome.
Inter-species aCGH were performed to further validate the copy number alterations indicated by sequence alignments (S6A Fig). Genomic DNAs from baboon (N = 1), rhesus macaque (N = 2), orangutan (N = 1), gorilla (N = 3) and chimpanzee (N = 7) were used to compare with human genomic DNA (NA10851) on the previously described aCGH. The quality of the hybridization positively correlated with the sequence identities between different primates and human (S1 Table). Comparing to the human genome, a large portion of genomic sequences flanking NPHP1 appeared to be nonexistent or have lower copy number in nonhuman primates, and the degree of similarity to human varied from baboon to chimpanzee (S6 Fig). After careful examination of the aCGH data of the 358 kb LCR locus, we achieved an estimation of variation of the genomic content between nonhuman primates and human. Sequences totaling sizes of 146 kb (40.85%), 156 kb (43.6%), 88 kb (24.6%), 24 kb (6.7%) and 72 kb (20.1%) appeared to be nonexistent (aCGH log2 ratio was lower than -1) in the genomes of baboon, rhesus macaque, orangutan, gorilla and chimpanzee, respectively, while approximately 123 kb (34.4%), 93 kb (26.0%), 115 kb (32.1%), 156 kb (43.6%) and 197 kb (55.0%) appeared to exist in the genomes of these primates, albeit at lower copy number (aCGH log2 ratio was between 0 and -1). The latter observation indicated that these genomic regions might constitute the ancestral nonduplicated segments orthologous to the human LCRs. The indication of copy number variants is derived from the log2 ratio of aCGH probes, which largely relies on the degree of hybridization based on sequence similarity. Thus, these data can reflect genome differences and phylogenic distance between the two species being compared. The sizes of genomic segments with copy number changes described above are estimates, and may be refined by testing a large cohort of nonhuman primates.
Since the 45 kb LCRs in human are directly related to the recurrent NPHP1 deletion, and a haplotype consisting of a 45 kb LCR loss accompanied with a 5936Ins is frequently observed, we were interested in understanding the evolution of this haplotype. As previously shown, the copy number of the 45 kb LCRs ranged from two to six in humans without NPHP1 deletion. Moreover, aCGH of eight DNA samples from patients with homozygous recurrent NPHP1 deletion revealed two copies of the 45 kb LCRs in seven individuals and four copies in one individual (Table 4). NAHR-mediated NPHP1 deletion between a pair of directly-oriented 45 kb LCRs reduces one copy, creating a recombinant copy from two substrate copies, of the 45 kb LCRs after the deletion event. Thus the majority (14/16) of haplotypes examined prior to NPHP1 deletion would have two copies of the 45 kb LCRs–this is consistent with the observation that copy number of two is the most frequently observed copy number in a haploid human genome.
Copy numbers of the 45 kb LCR orthologs in nonhuman primates were also estimated by aCGH using DNA from NA10851 as reference. In the tested baboon (N = 1) and rhesus macaques (N = 2), the species average log2 ratios targeting the human 45 kb LCRs were consistent with a complete absence of the 45 kb LCR ortholog (Fig 6B). In the tested orangutan (N = 1) and chimpanzees (N = 7), the species average log2 ratios indicated a reduced copy number of the 45 kb LCR orthologs in comparison to the reference human DNA used in aCGH (Fig 6C). In the tested gorillas (N = 3), the species average log2 ratio was close to zero, indicating that the copy number of the human 45 kb LCRs equaled the copy number of the gorilla orthologs (Fig 6C). Mean log2 ratios of the 45 kb LCRs or orthologous region for each individuals tested, including the hybridizations between human DNA samples without NPHP1 deletion (N = 10), are shown in Table 4.
Similar copy number of LCRs at the human and gorilla NPHP1 locus could be independently validated using existing and larger datasets. Dumas et al performed cDNA aCGH to survey genome-wide gene CNVs across a number of primate lineages [21]. We investigated CNVs at the NPHP1 locus using their published dataset, and correlated them with the CNV findings from our comparative analysis. Examination of two data points of human cDNAs (AA937147 and AI820499) located at the human 45 kb LCR locus revealed the relative copy numbers examined in each individual comparing to the human reference. Consistently, the gorilla genomes showed roughly comparable signal intensity compared to the human genomes, while the chimpanzee and orangutan genomes presented lower signal intensity for these two data points (S7 Fig). Although both AA937147 and AI820499 in rhesus macaque and baboon presented lower signal intensity than human, large variations were observed between these two data points (S7 Fig). These latter findings in rhesus macaque and baboon suggest that AA937147 and AI820499, although located at the same LCR locus, might have different copy numbers inside each species. Alternatively, the large variation could also be due to the essential absence of the 45 kb LCR ortholog in the macaque and baboon genomes. Moreover, Sudmant et al analyzed read-depth profiles from whole genome sequencing (WGS with a median coverage of ~25×) data of 10 humans, 32 gorillas, 23 chimpanzees and 17 orangutans. Copy number analysis based on sequencing read-depth revealed a similar copy number of the 45 kb LCRs and their orthologous regions in the tested humans and gorillas (p-value = 0.1705, Welch Two Sample t-test). Moreover, the copy number of the 45 kb LCRs in human is higher when comparing to the orthologous region in the tested chimpanzees (p-value < 0.0003, Welch Two Sample t-test) and orangutans (p-value < 0.0003, Welch Two Sample t-test) (Fig 6D and S7 Fig). These aggregated data provided evidence to partially support the observation of the similar genomic architectures between humans and gorillas.
Combined OM and aCGH approaches appear to be a versatile route for delineating SV haplotypes in a structurally complex locus like NPHP1. Array CGH can detect CNVs as small as a few hundred base pairs in size in test samples when compared to a reference. However, balanced SVs, e.g. inversions, cannot be detected by aCGH. Unlike aCGH, OM creates large datasets of ordered Rmaps from individual genomic DNA molecules, which through analysis reveal genome structures. Alignments between an optical map from test samples and the in silico generated reference maps reveal both CNVs and copy number neutral inversions and translocations. Errors associated with enzymatic cleavage or fragment sizing are inevitable, but can be modeled and dealt with by algorithms and software designed to work with large Rmap datasets for the construction of accurate maps [55–57]. Importantly, the CNVs were called consistently by both OM and aCGH in the samples tested in this study (Table 2). These data suggest that OM and aCGH, two orthogonal genomic approaches for SV characterization, can complement each other to provide a comprehensive and accurate SV haplotype. Furthermore, sequencing technologies may greatly facilitate the delineation of SV haplotypes in a region with complex genomic structures. The highly complex and repeating nature of genomic regions enriched with LCRs can challenge short read sequencing approaches and introduce mapping artifacts, some of these limitations may be potentially overcome by single-molecule long-read sequencing technologies, such as single-molecule real-time (SMRT) sequencing [58].
The evolutionary history of structural changes in a complex region such as 2q13 can be reconstructed by appropriate comparisons among related populations and species. The formation of human chromosome 2 through a telomeric fusion of chromosomes 2A and 2B was originally documented using high-resolution G-banding technique [34]. Later, using cosmid sequencing, two inverted arrays of telomeric repeats (5’(TTAGGG)n3’) in a head-to-head orientation (5’(TTAGGG)n- (CCCTAA)m3’) were found at the 2q fusion breakpoint (2qFus) [59]. In our study, we analyzed the origin of the SV haplotypes of the NPHP1 locus, which is about 3 Mb proximal to 2qFus. The comparative genomic analyses among nonhuman primates and humans suggest a trend of genomic expansion at the NPHP1 locus during primate evolution. Aggregating the sequence alignments and copy number analyses using aCGH and WGS data of baboon, rhesus macaque, orangutan, gorilla, chimpanzee and human, we propose the following model.
First, the present-day human 45 kb and 358 kb LCRs may be formed by gradual expansion and propagation of primate orthologous sequences into paralogous regions early in the evolution of the apes, after they diverged from the Old World monkeys. The nonhuman primate orthologs of the human 45 kb LCRs, not found in the current baboon and rhesus macaque reference sequences, emerge prior to the divergence of orangutan and human as they are found in both lineages (45MID, Fig 7). This 45kb sequence expanded in the lineage leading to chimpanzees and exhibit increasing sequence identity to the human 45 kb LCRs. These sequences then propagated into paralogous regions and eventually formed the 45 kb LCRs now observed in the human genome. A similar unknown mechanism may be suitable for explaining the expansion and paralogous propagation of the nonhuman primate orthologs of the human 358 kb LCRs (Fig 7). In the human lineage, the order of appearance of the 45 kb LCRs may be inferred by molecular clock analysis based on reference DNA sequence comparisons excluding insertion/deletion (indel) events. Comparative analysis suggests that 45MID in the human haploid reference is the ancestral copy among the three copies. The sequence identity between 45MID and 45PROX or 45DIST is slightly lower than that between 45PROX and 45DIST using the Smith-Waterman local alignment algorithm (Table 1). This suggests that 45DIST and 45PROX are paralogous propagation products of 45MID. The observation of the 45MID ortholog location adjacent to NPHP1 in both chimpanzee and orangutan supports the contention that the 45MID in human is the ancestral copy. However, the sequence-based molecular clock analysis is based on a priori hypothesis that the paralogous LCRs evolved independently. Therefore, it may be confounded by the potential interactions between paralogous sequences, such as homologous recombination leading to gene conversion, considering the fact that the 45PROX and 45MID are imbedded within the two highly homologous and inverted 358 kb LCRs.
Secondly, the human 45 kb LCRs imbedded in the 358 kb LCRs may be formed by paralogous insertion of the duplicated segments at the site where the 5936Ins were lost. This model is supported by the identification of the human 5936Ins orthologous sequences in all the primate reference genomes with increasing identities. The 5936Ins orthologs are present in all lineages studied, whereas the 45 kb LCR orthologs are only present in the orangutan, gorilla and chimpanzee genomes (Fig 7). Thus, the haplotype in human with the 5936Ins embedded in the 358 kb LCRs is the likely ancestral haplotype, whereas the haplotype possessing the imbedded 45 kb LCRs in the 358 kb LCRs may be more recent and could be formed by the deletion of the 5936Ins followed by a 45 kb LCR insertion. Therefore, the haplotype without 45PROX and 45DIST may be the most ancestral haplotype in the human lineage.
The structural complexity at the NPHP1 locus may also exist at a population level within the human species. We delineated various SV haplotypes in seven diploid human genomes and one haploid human genome using OM. These include five diploid genomes (NA10860, NA18994, NA15510, MM52 and HF087) that are homozygous for H2 (H2/H2), one (H1-ES-P208) with a possible admixture of cells with H5/H7 and H6/H7 combinations, one (HCC1937) with H2/H5, and one haploid genome (CHM) with only H5. Within15 chromosomes/SV haplotypes, NPHP1 inversions are found in at least 11, which account for 73.3% of the total alleles. Moreover, cross-correlation between aCGH (testing copy number of 45 kb LCRs) and OM (testing NPHP1 inversion) shows that none of the genomes investigated above harbor the haploid reference allele (Table 2). On a population basis, 11% of YRI, 3% of CHB/JPT and 1% of CEU harbor 6 copies of the 45 kb LCRs, thus possess at most two reference alleles (homozygous); 31% of YRI, 31% of CHB/JPT and 16% of CEU harbor 5 copies of the 45 kb LCRs, thus possess at most one reference allele (heterozygous) in each examined individual. The aggregate experimental evidence above strongly suggests that the SV haplotype at the NPHP1 locus is highly variable, as has been observed for the complex LCRs at the iso17q susceptibility locus [33]. Furthermore, the allele represented in the human reference genome may actually be a minor allele, suggesting a potential necessity for improvement of the reference genome and again reiterating the limitations of the haploid human reference and the lack of representation of CNV/SV variant alleles. In the current study, we used OM to identify various SV haplotypes in seven diploid human genomes and one haploid human genome. However, haplotype analysis of the NPHP1 locus was not performed in a large general population or across ethnicities in either humans or nonhuman primates studied. Thus the frequency of each SV haplotype in the general human population cannot be estimated. Perhaps large-scale assays, with methods such as single-molecule and long-read sequencing, will benefit the further investigation of the complex NPHP1 locus.
Highly identical LCRs at the NPHP1 locus lead to local genomic instability, which subsequently results in variable SV haplotypes. A new SV haplotype may be generated from an existing one by a simple inversion between two invertedly oriented LCRs, e.g. H3 from H1 (Fig 1B). In some cases, depending on the breakpoint, an inversion may generate a novel SV haplotype that may change the CNV susceptibility of the disease-associated genes, and the resultant SV haplotype may have potential clinical significance [28,29]. At the NPHP1 locus, the two directly oriented 45 kb LCRs flanking NPHP1 are the substrates for the NAHR event resulting in the common recurrent deletion of the gene. However, intra-chromosomal NAHR-mediated deletions would be inhibited if a chromosome lacks the flanking directly oriented LCRs. Interestingly, in our study, we identified SV haplotypes that appear to be resistant to NAHR-mediated NPHP1 deletion as a result of the loss of either one or both of the directly-oriented 45 kb LCRs flanking NPHP1. For example, OM analysis of the haploid genome of CHM reveals an H5 SV haplotype losing the 45 kb LCR on the centromeric side of NPHP1 which is utilized as a flanking substrate for unequal crossing-over (Fig 3). Therefore, these findings suggest that there may exist protective alleles that potentially inhibit intra-chromosomal NAHR, while the inter-chromosomal event may also be prevented if the protective allele exists in a homozygous state in an individual. These SV haplotypes could potentially reduce the frequency of the NPHP1 common recurrent deletion in the human population. This study works in concert with the previous studies regarding the correlation between local genomic structure and individual’s susceptibility to acquiring disease-associated alleles [28,29]. Both protective and susceptible SV haplotypes likely exist at other disease-associated loci with similar structural complexity [22]. The SV haplotypes (H5, H6 and H7) elucidated by OM were identified in human cell lines. Thus, they may potentially reflect tissue culture events generated in the cell lines tested. However, these results are parsimoniously explained by the underlying genomic architecture and mechanistic first principles, and thus the observed results in cell lines likely represent the organismal genome structure, reflect the genomic instability at this locus and indicate the potential existence of H5, H6, and H7 in the personal genomes of individuals in human populations. Nevertheless, it remains to be examined how representative these SV haplotypes are for human populations at large.
The NPHP1 locus in the gorilla reference genome is an interesting example of evolutionary complexity at two levels: complexity of sequence structure and complexity of population-level evolutionary genetics. Although whole genome comparisons indicate that gorillas diverged from human ancestors before chimpanzees did [60], the gorilla reference genome has a configuration of SV haplotype more similar to human than the chimpanzee does at the NPHP1 locus (Fig 7). Distinctly, instead of completely deleting the gorilla 5936Ins ortholog as observed in the human SV haplotypes, the gorilla 45 kb LCR orthologs appears to be imbedded inside of the gorilla 5936Ins ortholog. The coexistence of these orthologs indicates that the SV haplotype found in the gorilla reference may be an intermediate state different from the SV haplotypes in human. Alternatively, the observation of the human pattern of SV haplotype in gorilla could potentially reflect an assembly error due to the local complexity in the gorilla reference genome; this complexity involves both genomic gaps and the presence of the partial 45 kb LCR and 5936Ins orthologs within the 358 kb LCR orthologs. Analyses using genome-wide cDNA arrays [21], the WGS read-depth analysis, and the genomic aCGH (performed in this study) confirm a comparable copy number of the 45 kb LCR and its ortholog between the human H2 SV haplotype and the gorilla genomes tested (Fig 6, S7 Fig, S8 Fig, Table 4). Thus, the aCGH and sequence alignment data are suggestive of an analogous structural haplotype at the NPHP1 locus between gorilla and human that differs from that of chimpanzee (S6 Fig). Copy number analyses of DNA samples from a large number of gorillas will provide additional evidence supporting this hypothesis.
It is intriguing that the gorilla genomic structure for this region appears to be more similar to human than chimpanzee. The estimated divergence time of human-gorilla lineages is approximately 6–8 Mya, which is earlier than estimated human-chimpanzee divergence (approximately 4.5–6 Mya). Moreover, the sequence identity of human versus chimpanzee (99%) is slightly higher than that of human versus gorilla (98.4%) [53,60]. The genome-wide evidence reflects the most commonly accepted evolutionary phylogeny that has humans more closely related to chimpanzees than to gorillas, i.e. ((H-C)-G). However, a recent study of the gorilla genome shed light on the complex evolutionary phylogeny by providing compelling evidence that incomplete lineage sorting affects 15% of the human-chimpanzee-gorilla genomes [60]. That is, whole genome analyses demonstrate that 70% of the gorilla genome follows the ((H-C)-G) standard phylogeny, while 15% of the gorilla genome segments exhibit ((H-G)-C) whereas another 15% exhibit ((C-G)-H). The structural variation results using aCGH and WGS read-depth analysis suggest that the region including NPHP1 and its adjacent LCRs fall in one of the segments with the alternative ((H-G)-C) pattern. It is plausible that the last common ancestor of humans, chimpanzees and gorillas was polymorphic for NPHP1 haplotypes, and segregating for haplotypes that resemble the human/gorilla structure and the chimpanzee structure. If this were true, the pattern of variation across the three species can be explained by the retention of one ancestral haplotype in humans and gorillas, and the loss of that haplotype with retention of the more ancestral form in chimpanzees [52,61]. Furthermore, these results lead to the prediction that gorillas will be more susceptible than other nonhuman primates to mutations that delete NPHP1 and thus cause a disease similar to NPHP1. Our findings also suggest that the complex structure of the human NPHP1 region was established prior to the fusion of the two ancestral chromosomes that formed the present human chromosome 2, as neither gorillas nor chimpanzees exhibit this fusion.
In summary, we computationally and experimentally characterized the genomic architecture and identified novel SV haplotypes of the NPHP1 locus in the human 2q13 region. The more commonly observed alternative SV haplotypes suggest the current human genome reference represents a minor allele. For such complex loci enriched with LCRs, the accuracy of assembly may be compromised. Thus detailed exploration using various comparative genomic analytical methods is needed to document the human genome structure and the stages of its evolution in a more comprehensive way. NAHR-mediated NPHP1 deletion occurs between the two flanking directly oriented LCRs. Here, we found that potential “protective alleles” lacking directly oriented LCR flanking NPHP1 also exist in the examined human genomes, and such structure may protect those alleles from NPHP1 deletion mediated by intra-chromosomal NAHR, or inter-chromosomal events in the homozygous state. Such potential “protective alleles” may also exist in other “NAHR-susceptible” loci, with the drawback of such alleles being a decrease in genome plasticity that facilitates evolution. A large number of genomic disorders are associated with loci with complex genomic architectures that introduce risk of genomic rearrangements (e.g. 17p11.2 and SMS/PTLS). The assessment of the disease risk of these disorders can be facilitated by accurate determination of the alternative SV haplotypes via single molecule analysis, including OM or long-molecule/long-read sequencing. Moreover, we assessed the origins of the complexity of the NPHP1 locus using inter-species comparative genomic analysis, and we found evidence supporting the genomic expansion and propagation of LCRs during primate evolution. The generation of LCRs may occur in a multi-step manner, and the higher order of genomic complexity constituted by LCRs may render the genome susceptible to instability and DNA rearrangements.
Sequences of LCRs were downloaded from UCSC Genome Browser (http://genome.ucsc.edu/cgi-bin/hgTables). The coordinates used for sequence downloading are: chr2:110688766–110733137 (45PROX), chr2:110983705–111031088 (45MID), chr2:111153517–111197896 (45DIST), chr2:110494432–110852754 (358PROX) and chr2:111033788–111392192 (358DIST).
Pairwise alignments of the LCRs in the same group were performed using “pairwiseAlignment” R package [62]. It was performed as a type of local alignment that considers the penalty from end gaps. Sequence identity was calculated after the alignment.
We aligned the LCR sequences using the Clustal W2 algorithm [63] and determined the percentage of identical positions over a 100 base pair window along the length of the LCR. We created a position weight matrix (PWM) based on a previously reported recombination hotspot motif [41] and subsequently assessed each LCR for matches to the motif’s PWM and its reverse complement using the Biostrings package implemented in the R Statistical Programing Language (http://www.r-project.org/, http://www.bioconductor.org/packages/release/bioc/html/Biostrings.html). We indicate the positions of strong (>85% of the maximum possible score) matches along the edge of each LCR with a triangle.
Fosmid libraries of 17 individuals (ABC7, ABC8, ABC9, ABC10, ABC11, ABC12, ABC13, ABC14, ABC16, ABC18, ABC21, ABC22, ABC23, ABC24, ABC27, WIBR2, JVI) were downloaded from Human Genome Structural Variation Project (HGSV, http://humanparalogy.gs.washington.edu/structuralvariation/). All end sequence pairs (ESPs) mapped to hg19 build were manually filtered according to the mapping quality and chromosomal location. Discordant fosmids were selected based on the annotations judging the distance and orientation between the ESPs. A UCSC Genome Browser custom track was created for the discordance fosmids identified in NPHP1 locus based on genomic coordinates and relative orientations of the ESPs (http://genome.ucsc.edu/cgi-bin/hgCustom).
Miropeats program was used to descriptively illustrate the genomic architecture by plotting the inter-/intra-species alignments of the reference genome. ICAass (v 2.5) algorithm was used to perform DNA sequence comparisons, and Miropeats (v 2.01) was then applied for converting the comparisons into graphical display based on the position and matching quality (a threshold set up by users) [40]. According to the time of divergences and overall sequence similarities upon evolution, different thresholds gauging the length of DNA sequence homology (“seed”) were chosen in order to show the feature of alignments between different primates. In our study, a threshold of 500 were set for baboon/human, macaque/human and orangutan/human pairs, while 1000 were used for gorilla/human, chimpanzee/human and all the intra-species alignments. Miropeats were performed between sequences from genomic intervals +/-800 kb of NPHP1 of each genome build used.
Optical mapping [46,48,55–57,65–71] is a single-molecule, whole-genome analysis system for the comprehensive discovery and characterization of SVs. Large genomic DNA molecules (from 300 kb to multi-Mbs) were extracted, stretched and immobilized on positively charged glass surfaces via capillary flow within microfluidic devices fabricated using soft lithography [65]. Hydrodynamic forces generated by capillary flow combine with DNA/surface electrostatic interactions to stretch and immobilize very long molecules. DNA molecules, thus presented, were restriction digested (SwaI or BamHI, New England Biolabs), stained with YOYO-1 (an intercalating fluorochrome; Invitrogen) and imaged using a custom-designed, fully-automated, epifluorescence microscopy imaging system [65]. Restriction endonuclease sites undergo double-stranded breakage followed by DNA relaxation at the cut ends, which present as micron-sized gaps along stretched DNA molecules. Acquired images were then automatically analyzed using custom machine vision software [65,66], which yielded large datasets of single molecule ordered Rmaps. Using an iterative assembly process that leverages Bayesian inference approaches and cluster computing [46,55–57,67,68], the Rmaps datasets were then assembled into multimegabase map contigs that were later joined to span entire chromosomes. The iterative assembler clusters single-molecule maps using pairwise alignments to a reference genome, and then assembles these map clusters using a maximum-likelihood Bayesian assembler to generate contigs and consensus restriction maps. The assembled genomes were then viewed within a custom genome visualization environment (Genspect) that allows detailed inspection of the primary data underlying called SVs. Lastly, the assembly/analysis pipeline automatically tabulates a list of SVs that were inspected and manually curated in another custom visualization software (GnomSpace) to characterize all SVs in the analyzed genomes.
Great ape CNV data derived from whole genome sequencing (WGS) were downloaded as bigBed files from http://eichlerlab.gs.washington.edu/greatape-cnv/tracks/, and bigBedToBed was used to extract annotations within the region of interest. The human genome hg18 coordinates of the CNV annotations were converted to hg19 coordinates using liftOver with default parameters. A UCSC track hub was generated in order to visualize the CNVs lifted over to hg19. The average number of copies for each sample (great apes or human) was calculated for each genomic window of 500 bp in the region of interest.
Fosmids clones (G248P81805G1, G248P88963C6, and G248P88660A3) were obtained from BACPAC Resources Center (BPRC) as “stab-cultures”. Clones were cultured in LB medium containing 12.5μg/ml chloramphenicol). Fosmids were extracted from a suspension culture with QIAGEN Plasmid Midi Kit. G248P81805G1 is from NPHP1; G248P88963C6 is from the 45 kb LCRs; and G248P88660A3 is from a conserved region ~1.3 Mb proximal to NPHP1.
Cultured lymphoblastoid cell lines from the individuals NA15510 (human), CRL-1854 (gorilla), and CRL-1868 (chimpanzee) were harvested using 10ug/ml Colcemid (Roche) for 30 minutes followed by 0.075M (hypotonic) treatment for 10 minutes at 37°C. The cells were then fixed using Carnoy’s fixative (3 methanol: 1glacial acetic acid). The cell pellet obtained from the harvesting was used to prepare the “dropped slides” for FISH.
Directly labeled custom “home-brewed” probes were produced from fosmid clones mentioned above. The home brewing process was performed using Nick Translation Kit (Abbott molecular). Green dUTP, Orange dUTP (herein referred to as red) and Aqua dUTP (Abbott Molecular) were used to label G248P81805G1, G248P88963C6 and G248P88660A3, respectively. Signal validation was also verified by observing the metaphases on an inverse DAPI function.
The slides were aged using 2X SSC at room temperature for 2 minutes followed by sequential dehydration using 70%, 80% and 100% ethanol for 2 minutes each. Metaphases on slides were marked and 5ul of the probe mixture was added. Target DNA and the probes were co-denatured at 75°C for 5 minutes followed by hybridization at 37°C for 16 hours. The slides underwent post wash with 2X SSC at 37C for 2 minutes. The slides were left to air dry after post wash, and 10ul of DAPI-II counterstain (Dako, Agilent Technologies) was added to the slides. Metaphases were viewed using Olympus Florescence microscope (Olympus America) and analyzed using Cytovision software v3.6.
|
10.1371/journal.pgen.1007914 | RPGRIP1L is required for stabilizing epidermal keratinocyte adhesion through regulating desmoglein endocytosis | Cilia-related proteins are believed to be involved in a broad range of cellular processes. Retinitis pigmentosa GTPase regulator interacting protein 1-like (RPGRIP1L) is a ciliary protein required for ciliogenesis in many cell types, including epidermal keratinocytes. Here we report that RPGRIP1L is also involved in the maintenance of desmosomal junctions between keratinocytes. Genetically disrupting the Rpgrip1l gene in mice caused intraepidermal blistering, primarily between basal and suprabasal keratinocytes. This blistering phenotype was associated with aberrant expression patterns of desmosomal proteins, impaired desmosome ultrastructure, and compromised cell-cell adhesion in vivo and in vitro. We found that disrupting the RPGRIP1L gene in HaCaT cells, which do not form primary cilia, resulted in mislocalization of desmosomal proteins to the cytoplasm, suggesting a cilia-independent function of RPGRIP1L. Mechanistically, we found that RPGRIP1L regulates the endocytosis of desmogleins such that RPGRIP1L-knockdown not only induced spontaneous desmoglein endocytosis, as determined by AK23 labeling and biotinylation assays, but also exacerbated EGTA- or pemphigus vulgaris IgG-induced desmoglein endocytosis. Accordingly, inhibiting endocytosis with dynasore or sucrose rescued these desmosomal phenotypes. Biotinylation assays on cell surface proteins not only reinforced the role of RPGRIP1L in desmoglein endocytosis, but also suggested that RPGRIP1L may be more broadly involved in endocytosis. Thus, data obtained from this study advanced our understanding of the biological functions of RPGRIP1L by identifying its role in the cellular endocytic pathway.
| The desmosome is a type of intercellular junction, essential for cells to adhere to one another. Abnormalities in desmosomes can cause disorders in the hair, skin, and heart, some of which are severe or even fatal. Here, we discovered that RPGRIP1L, a protein known to regulate cilia formation and function, is essential for stabilizing desmosomes of skin keratinocytes. Specifically, suppressing the Rpgrip1l gene in mice or in keratinocytes disrupted the ultrastructure of desmosomes, and compromised cell-cell adhesion in vivo and in vitro. We found that knocking down RPGRIP1L in keratinocytes aberrantly accelerated the internalization of cell membrane desmogleins, key desmosomal cadherins. Inhibiting endocytosis effectively rescued these phenotypes. Biotinylation assays confirmed that desmogleins are likely the primary targets of RPGRIP1L. Interestingly, membrane proteins that are not directly associated with the desmosomes were also found to be internalized in RPGRIP1L-knockdown cells, raising the possibility that RPGRIP1L might regulate endocytosis more broadly. Findings from this study not only identified RPGRIP1L as a regulator of the desmosomes, but also expanded our understanding of cilia-related proteins in the formation of the desmosomal junctions.
| Retinitis pigmentosa GTPase regulator interacting protein 1-like (RPGRIP1L, also known as NPHP8, MKS5, KIAA1005, or Ftm in mouse) is a transition zone protein that has important roles in regulating cilia formation and function [1–5]. Mutations in the RPGRIP1L gene cause Joubert syndrome (JBTS) and Meckel syndrome (MKS) [6,7], two severe ciliopathies that are characterized by central nervous system malformation, cystic kidneys, polydactyly, retinal degeneration, and retinal dystrophy [8]. RPGRIP1L participates in the assembly of the ciliary transition zone, autophagy, and activation of the ciliary proteasome [9], whereas mutant RPGRIP1L interferes with ciliary functions, leading to dysplasia of affected organs [6,7,10].
In the skin, RPGRIP1L is essential for hair follicle morphogenesis by regulating primary cilia formation and hedgehog signaling [11]. Interestingly, RPGRIP1L is also expressed in interfollicular epidermal keratinocytes, many of which are not ciliated [12], suggesting that RPGRIP1L may exert cilia-independent functions in the skin.
Desmosomes are anchoring junctions that are essential for functionalities of tissues that are subjected to constant mechanical stress, such as the skin and the heart. Desmosomal junctions are composed of transmembrane cadherins, desmogleins and desmocollins, and cytoplasmic proteins, including junction plakoglobin (JUP), plakophilins, and desmoplakin (DSP) [13,14]. The adhesion function of desmosomal junctions is dependent on the intercellular anchorage of desmogleins and desmocollins.
The assembly and disassembly of the desmosomes is highly dynamic, and intercalates with cellular events associated with the regulation of the cytoskeleton, intracellular trafficking, ubiquitination, and molecular signaling [13]. Forward and reverse genetic studies continue to uncover new players involved in the formation of the desmosomes, which collectively contribute to the establishment of a comprehensive regulatory network of desmosome assembly and homeostasis.
Mutations in genes encoding desmosomal proteins can cause a range of heritable disorders that affect the skin, hair, and heart, such as monilethrix, woolly hair, palmoplantar keratoderma, and arrhythmogenic right ventricular cardiomyopathy [15–19]. Moreover, disruption of desmosomal junctions by autoantibodies can cause pemphigus, a family of devastating autoimmune disorders characterized by severe intraepithelial blistering in the skin or mucous membranes [20,21]. Loss of desmosomal proteins has, at least in some cases, been linked to cancer development or progression [20,22]. Understanding the cellular and molecular mechanisms underlying the assembly and disassembly of desmosomal junctions is important for the understanding of the pathogenesis of desmosome-related disorders.
In this study, we uncovered a previously unknown function of RPGRIP1L in the formation of the desmosomal junctions. We found that disrupting the Rpgrip1l gene in mice or keratinocyte cell lines resulted in desmosomal abnormalities that are associated with aberrant internalization of desmogleins. These findings revealed RPGRIP1L as a regulator of desmosome formation and function, and suggested a broader role of RPGRIP1L in the assembly of cellular organelles, including the ciliary transitional zone and the desmosome.
Rpgrip1l is ubiquitously expressed in the skin, including the epidermis, dermis, and hair follicles [11]. In mouse epidermis, the Rpgrip1l transcript, as determined by in situ hybridization, is consistently expressed in basal epidermal keratinocytes and, to a lesser extent, in spinous and granular cells (Fig 1A). The RPGRIP1L protein is enriched between the basal body (marked by gamma-tubulin, γ-Tub) and ciliary axoneme (marked by acetylated α-tubulin, α-Tub) of ciliated basal keratinocytes (S1A Fig), or enriched at the centrioles of unciliated keratinocytes (S1E Fig), but below detection in Rpgrip1l knockout (Rpgrip1l–/–) epidermis (S1I Fig). Since differentiated epidermal keratinocytes are rarely ciliated [12,23,24], the widespread expression pattern of Rpgrip1l in the epidermis raised the possibility that Rpgrip1l performs functions beyond regulating ciliogenesis and ciliary functions.
Indeed, skins of 50% of E18.5 Rpgrip1l–/–embryos (n = 16) exhibited focal intraepidermal blistering, predominantly between basal and suprabasal keratinocytes, marked by KRT14 and KRT1, respectively (Fig 1B, middle panels). In severe cases, cell-cell detachments could also be observed in the spinous and granular layers (Fig 1B, right panels). To further explore this relatively sporadic blistering phenotype, which was not sufficiently characterized in a previous study [11], and to circumvent perinatal lethality associated with severe developmental abnormalities in Rpgrip1l–/–mutants, including exencephaly and ventricular septal defects [2,25], we cultured skins isolated from E18.5 embryos. Organ-cultured Rpgrip1l–/–skins exhibited widespread blistering between the basal and suprabasal layers (Fig 1D and 1E), suggesting that blistering is progressive as the skin becomes mature. Histologically, this intraepidermal blistering phenotype was not associated with discernable cytolysis of keratinocytes, detachment of the basement membrane (Fig 1C), or apoptosis (S2 Fig). Thus, these findings suggest that the blistering phenotype observed in Rpgrip1l–/–skin may be associated with abnormalities in keratinocyte adhesion.
The desmosomal junctions play essential roles in epidermal adhesion and were, therefore, examined in Rpgrip1l–/–mutants. Immunofluorescence labeling revealed reduced expression of desmoglein 1 (DSG1), desmoglein 3 (DSG3), JUP, and desmocollin 1 (DSC1) in Rpgrip1l–/–skin (Fig 2 and S3 Fig). Specifically, these proteins were diffusely localized to the cytoplasm of keratinocytes of Rpgrip1l–/–skin, in contrast to the predominant plasma membrane localization in control skins (Fig 2 and S3 Fig). The expression patterns of DSP and desmocollin 2 and 3 (DSC2/3) appeared slightly perturbed in Rpgrip1l–/–skin, whereas the expression of plakophilin 1 (PKP1) did not seem to change, as judged by immunofluorescence microscopy (Fig 2 and S3 Fig).
Transmission electron microscopy (TEM) revealed that desmosomes between basal and suprabasal keratinocytes were significantly shorter in Rpgrip1l–/–mutants (Fig 3A and 3B). Moreover, in Rpgrip1l–/–skin, the electron dense midline of desmosomes was less prominent or invisible, the keratin filament attachment was reduced, and the outer electron dense plaque appeared less dense or disorganized (Fig 3A). Similar defects were occasionally observed between spinous keratinocytes (Fig 3A). These findings demonstrated that the blistering phenotype in Rpgrip1l–/–skin is correlated with abnormalities in the desmosomal junctions. Interestingly, adherens junctions, as assessed by immunofluorescence labeling of E-cadherin (CDH1), α-catenin (CTNNA1), and β-catenin (CTNNB1), exhibited only subtle perturbations in Rpgrip1l–/–skin (S3 and S4 Figs). Taken together, these data suggest that RPGRIP1L may be required for keratinocyte adhesion primarily through regulating desmosomal junction formation in vivo.
RPGRIP1L is expressed in many cell types, and is enriched at the base of cilia for its cilia-related functions [1,2,6,7,10,26]. To evaluate the roles of RPGRIP1L in desmosome formation, we first examined the expression pattern of RPGRIP1L in HaCaT cells and normal human epidermal keratinocytes (NHEKs). HaCaT cells do not form primary cilia, and NHEKs rarely form primary cilia (3.4 ± 2.6%) after serum starvation, in comparison to mouse embryonic fibroblasts (MEFs) which do (66.7 ± 12.5%) (Fig 4A). In HaCaT cells and NHEKs, RPGRIP1L is enriched at the centrosomes (marked by γ-TUB) and diffusely distributed in the cytoplasm (Fig 4A). These findings suggest that the potential role of RPGRIP1L in desmosome formation is independent of its role in ciliogenesis in keratinocytes.
We subsequently knocked down the endogenous RPGRIP1L gene in HaCaT cells by siRNAs (Fig 4B and 4C). Knockdown cells were then treated with high calcium to allow desmosomal junctions to form, then subjected to dispase dissociation assay as illustrated in Fig 4D. RPGRIP1L-knockdown markedly compromised the integrity of the epidermal sheet, resulting in significantly increased fragmentation (Fig 4E and 4F). This experiment indicated that RPGRIP1L is functionally required for cell-cell adhesion of keratinocytes in vitro.
To further confirm in vivo findings, desmosomal junctions were evaluated in RPGRIP1L-knockdown HaCaT cells and NHEKs. RPGRIP1L-knockdown did not significantly affect cell viability (S5 Fig), but resulted in marked reduction of desmoglein 1 and 2 (DSG1/2) and desmoglein 3 (DSG3) proteins as determined by western blotting (Fig 5A and S6 Fig), but not mRNA (S7 Fig). The protein levels of DSP, PKP1, plakophilin 2 (PKP2), and JUP were unaffected, whereas those of DSC2/3 increased in RPGRIP1L-knockdown cells (Fig 5A and S6 Fig). Immunofluorescence labeling demonstrated that the membrane localization of many desmosomal proteins, including DSG1/2 and DSG3, were significantly reduced in RPGRIP1L-knockdown cells (Fig 5B). These findings suggest that disrupting RPGRIP1L expression in keratinocytes impairs the stability and membrane localization of desmosomal proteins.
At the ultrastructural level, RPGRIP1L-knockdown HaCaT cells exhibited desmosomal abnormalities that are similar to those observed in vivo, including disrupted midline and reduced keratin attachment (Fig 5C). Taken together, these in vitro results further substantiated the role of RPGRIP1L in maintaining structural integrity of the desmosomes.
In contrast, RPGRIP1L-knockdown did not result in discernable changes in the expression pattern of intermediate filaments in HaCaT cells as demonstrated by KRT14 immunostaining (S8 Fig).
Because disrupting Rpgrip1l resulted in consistent changes in the desmogleins under both in vivo and in vitro conditions, and the blistering phenotype observed in Rpgrip1l–/–skins is similar to what is seen in pemphigus, a severe blistering disorder caused primarily by the disruption of the desmogleins, we focused our investigation on the desmogleins.
The formation of desmosomes is highly dynamic and can be arbitrarily divided into the assembly and disassembly phases. In HaCaT cells, desmosomes start to assemble when the cells are exposed to high calcium. We found that knocking down RPGRIP1L during desmosome assembly (0.5, 1, and 3 hours after shifting to high calcium, as illustrated in Fig 6A) did not impair the accumulation of DSG1/2 to the plasma membrane, as determined by immunofluorescence labeling (Fig 6B and quantification in 6C), suggesting that RPGRIP1L might be dispensable for desmosome assembly. In contrast, in the disassembly assay (as illustrated in Fig 6D), where DSG1/2 and DSG3 were examined 24 hours after calcium switch, in conjunction with 1-hour EGTA treatment to further induce desmosome disassembly, the plasma membrane localization of DSG1/2 or DSG3 was significantly decreased in RPGRIP1L-knockdown cells such that DSG1/2 or DSG3 appeared discontinuous along, or in some cases absent from the plasma membrane (Fig 6E and 6G, respectively). Quantifications of membrane and cytoplasmic signal intensity showed that the membrane/cytoplasmic ratio of DSG1/2 or DSG3 was significantly reduced in knockdown cells, a phenotype that was further exacerbated upon EGTA treatment (Fig 6F and 6H, respectively). These findings suggest that loss of RPGRIP1L may cause increased internalization of cell surface desmogleins. This result is consistent with a well-established model in which increased desmoglein endocytosis leads to a decrease in both cell surface and steady-state levels of desmogleins, as seen in Figs 2 and 5 and S6B Fig [27,28].
Desmoglein internalization is mediated by multiple endocytic mechanisms and remains a subject of further investigation [29–37]. Nevertheless, blocking endocytosis could prevent the internalization of desmogleins that are present on the cell surface. Here, we utilize two well-established approaches to blocking endocytosis to determine whether aberrantly accelerated internalization of desmogleins in RPGRIP1L-difficient cells is functionally responsible for the loss of membrane desmogleins. One approach was to use dynasore, a specific inhibitor of dynamin GTPase activity [38,39], to suppress endocytosis. Dynasore had been shown capable of stabilizing desmosomal junctions through blocking endocytosis [38]. The other approach was to use hyperosmotic sucrose to suppress endocytosis [40]. Dynasore or sucrose was added 24 hours after shifting to high calcium and 2 hour prior to fixation, as illustrated in Fig 7A.
Pretreating RPGRIP1L-knockdown cells with 50 μM dynasore for two hours was sufficient to rescue the increased internalization of DSG1/2, as determined by immunofluorescence labeling (Fig 7B). Specifically, the membrane localization of DSG1/2 in RPGRIP1L-knockdown cells was restored to a level comparable to that of control knockdown (Fig 7B, upper panels, and quantifications in c). Furthermore, dynasore treatment also overcame the additive effects of both knockdown- and EGTA-induced DSG1/2 internalization (Fig 7B, lower panels, and quantifications in 7C). Similarly, hyperosmotic sucrose effectively rescued RPGRIP1L knockdown-induced DSG1/2 internalization in HaCaT cells, even in the presence of EGTA, as demonstrated by immunofluorescence labeling and quantification (Fig 7B, right panels, and 7C). Collectively, these rescue experiments suggest that elevated endocytosis is functionally responsible for RPGRIP1L knockdown-induced internalization of the desmogleins.
To directly monitor the endocytosis of desmogleins, we performed an internalization assay, in which the internalized DSG3 can be quantified by labeling DSG3 with AK23, a monoclonal antibody against the extracellular domain of DSG3 [41], in live cells as previously demonstrated [29]. Specifically, as outlined in Fig 7D, AK23 was used to label cell surface DSG3 on ice, one hour prior to shifting to 37°C to trigger internalization (in the presence of control or PV IgG). One hour later, AK23-labeled but not internalized DSG3 was stripped off cell surface through acid wash. Cells were then fixed, and the internalized/AK23-labled DSG3 was detected by immunofluorescence.
In control siRNA-treated cells, control IgG treatment resulted in low levels of internalization of DSG3 (Fig 7E, upper left panel). As expected, PV IgG induced marked internalization of DSG3 (Fig 7E, upper right panel, and quantifications in 7F). Remarkably, RPGRIP1L-knockdown also resulted in marked internalization of DSG3 (Fig 7E, lower left panel), an effect as robust as PV IgG, as quantified in Fig 7F, reinforcing prior findings that the loss of RPGRIP1L can be pathogenic. More interestingly, DSG3 internalization was further increased by dual PV IgG and RPGRIP1L siRNA treatments (Fig 7E, lower right panel, and quantifications in 7F). This additive effect suggested that loss-of-RPGRIP1L and PV IgG may promote DSG3 endocytosis through distinct mechanisms.
The above findings established a role of RPGRIP1L in regulating the endocytosis of cell surface desmogleins in epidermal keratinocytes. To determine whether RPGRIP1L might be more broadly involved in endocytosis, we evaluated the steady state-rate of endocytosis of cell-surface proteins by biotinylation assays [42]. First, the level of DSG3 decreased in whole cell lysate of RPGRIP1L-knockdown cells (Fig 7G), whereas biotinylated (endocytosed) DSG3 markedly increased in RPGRIP1L-knockdown cells (Fig 7G), a finding that is consistent with the above data (Figs 5B, 6G and 7E). The total protein level of DSC3 increased, whereas biotinylated (endocytosed) DSC3 increased marginally in this biotinylation assay (Fig 7G), also consistent with previously obtained data (Fig 5A and 5B).
Next, we examined cell membrane-associated proteins other than desmogleins, specifically epidermal growth factor receptor (EGFR) and CDH1. Despite the comparable total levels of EGFR and CDH1 in control and RPGRIP1L-knockdown cells (Fig 7G, left panels), RPGRIP1L-knockdown correlated with increased levels of biotinylated (endocytosed) EGFR and CDH1 (Fig 7G, right panels). Although the desmogleins cross-regulate with many cell surface proteins, including EGFR [43,44] and CDH1 [45–47], increased endocytosis of EGFR and CDH1 in association with RPGRIP1L-knockdown nevertheless suggests that RPGRIP1L may regulate endocytosis more broadly, which is worthy of future investigation.
Collectively, data obtained from this study suggest that increased endocytosis of desmogleins is primarily responsible for the keratinocyte adhesion defects associated with RPGRIP1L deficiency, thereby establishing a role of RPGRIP1L in stabilizing the desmosomes in skin.
Increasing evidence suggests that cilia-related proteins perform important cellular functions beyond regulating cilia formation or function [48]. In this study, we demonstrated that a ciliary protein, RPGRIP1L, is required for the maintenance of desmosomal junctions through regulating endocytosis of desmogleins in epidermal keratinocytes, thereby extending the functions of cilia-related proteins to cell-cell adhesion. Intriguingly, JBTS and MKS patients, who harbor loss-of-function mutations in the RPGRIP1L gene, do not exhibit blistering phenotypes. It is possible that blistering in these patients is underdiagnosed, or that abnormalities in desmosomal junctions exist but are subclinical. It is also plausible that the mutant RPGRIP1L gene products in patients retain a certain level of functionality, whereas genetically disrupting the Rpgrip1l locus leads to more catastrophic phenotypes in mice, by which blistering was observed. Further understanding the molecular mechanisms through which RPGRIP1L participates in desmosomal junction formation will shed light on how RPGRIP1L performs diverse functions in ciliogenesis and desmosome formation.
RPGRIP1L is highly enriched at the transition zone of cilia, but is also distributed in the cytoplasm and at the plasma membrane [6,49]. Thus, our finding that RPGRIP1L performs functions beyond the cilia is not entirely surprising. Data obtained from this study support a role of RPGRIP1L in stabilizing desmogleins at the plasma membrane, thus qualifying RPGRIP1L as a regulator of desmoglein internalization.
The precise molecular mechanism through which RPGRIP1L regulates desmoglein internalization remains to be uncovered. Without a strong presence at the plasma membrane, it is unlikely that RPGRIP1L interacts with desmosomal proteins at the cell membrane as previously described for Lis1, adducin, and flotillins [50–52]. Rather, in keratinocytes, RPGRIP1L is enriched at the base of the cilium as well as the centrosomes, both hubs for intracellular trafficking. RPGRIP1L may modulate desmoglein internalization by interacting with cytoplasmic regulators of the desmosomes, or through signaling, such as PKCα [37,38,53–56] or p38/MAPK [30,57–59]. Because keratinocyte proliferation, differentiation, and apoptosis are not markedly impaired in RPGRIP1L-deficient cells, the mechanism through which RPGRIP1L regulates desmoglein internalization is likely to be specific to the desmosomal regulatory network.
It is well established that RPGRIP1L physically interacts with NPHP4 during cilia formation or function [7,49,60]. NPHP4 is not only required for proper cilia formation, but also implicated in the formation of tight junctions [61]. In the current study, we provided evidence that RPGRIP1L is required for the proper formation and function of the desmosomal junctions, primarily through regulating endocytosis of desmogleins. Interestingly, we also observed, albeit inconsistently, changes in components of adherens junctions, including CDH1, CTNNA1, and CTNNB1. Considering the well-documented cross-talk between these intercellular anchoring junctions [62,63], we postulate that the desmosomes are the primary targets of RPGRIP1L in keratinocytes. This evidence nevertheless raised the possibility that RPGRIP1L may be more broadly involved in cell-cell junctions through the RPGRIP1L-NPHP4 protein complex [7,49]. The functional requirement of NPHP4 or the RPGRIP1L-NPHP4 protein complex in desmosome formation remains to be determined.
The desmogleins were consistently the most markedly endocytosed proteins. In unbiased biotinylation assays, elevated internalization of DSG3 was also correlated with increased endocytosis of other cell surface proteins in RPGRIP1L-knockdown cells. It remains to be determined whether this is a mere correlation, or whether RPGRIP1L is functionally associated with endocytosis of other cell surface molecules. Generalization of the potential role of RPGRIP1L in internalization of membrane molecules may further our understanding of endocytosis and recycling in both ciliated and unciliated cells. Given that the current knowledge of the molecular functions of RPGRIP1L is limited to its role in proteasome activity and autophagy at the base of the cilium [26,64], it is plausible that RPGRIP1L also participates in endocytosis through regulating protein degradation, a potential mechanism that needs to be further dissected.
The current study focused on the desmogleins, whose expression levels and membrane localizations were found to be consistently compromised in vivo and in vitro, mimicking key pathological features observed in pemphigus. We found that components of the desmosome were not equally affected in Rpgrip1l–/–skin and RPGRIP1L-knockdown cells. The levels of most other desmosomal proteins remained unchanged with the exception of DSC2/3 increasing, whereas PKP1 did not exhibit increased internalization in RPGRIP1L-knockdown cells. It is plausible that the increased DSC2/3 might have helped PKP1 to associate with the cell membrane in RPGRIP1L-deficient cells. In light of these findings, we postulate that desmogleins may be the primary targets of RPGRIP1L, whereas changes in other desmosomal components are secondary or compensatory. Indeed, several prior studies demonstrated the protective role of plakophilin in pemphigus or skin fragility models [34,35,65,66].
In conclusion, data obtained from this study implicate RPGRIP1L in the maintenance of desmosomal junctions through restricting desmoglein endocytosis, thereby revealing a cilia-independent function of RPGRIP1L in epidermal keratinocytes.
All procedures related to mice were performed in accordance with the European Directive 2010/63/EU and the French application decree 2013–118 on the protection of animals used for scientific purposes, and were approved by the local ethical committee "Comité d'éthique Charles Darwin" (approval number 2015052909185846), or in accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health of the United States, and approved by the Institutional Animal Care and Use Committee of Stony Brook University (approval number 2012-1974-R2-9.14.18-MI).
The Rpgrip1l mouse model was described previously [2,4,5,25]. HaCaT cells were transfected with RPGRIP1L siRNAs (HSC.RNAI.N015272.12.1 and 2, Integrated DNA Technologies, Coralville, IA). Non-targeting (Negative Control) siRNA (NC-1, Integrated DNA Technologies) was used as control siRNA. Twenty-four hours after transfection, cells were switched to high calcium (1.5 mM CaCl2) for designated durations. EGTA was used at 2 mM. Dynasore (Sigma-Aldrich, Saint Louis, MO) and sucrose were used at 50 μM and 400 mM, respectively. PV IgG were purified in Dr. Payne’s laboratory and used at 400 μg/ml. Normal human IgG (Sigma-Aldrich, Saint Louis, MO) was used as control IgG. Method details are provided in the Supplementary Materials and Methods (S1 Text) online.
Skins excised from E18.5 embryos were cultured as previously described [67]. Briefly, dorsal skins harvested from E18.5 embryos were cultured at the air-liquid interface in DMEM and 10% fetal bovine serum at 37° C and 5% CO2. Cultured skins were then fixed in 10% buffered formalin and processed for routine histology analysis.
Freshly isolated tissues were fixed immediately in buffered formalin, embedded in paraffin, and processed for routine hematoxylin and eosin (H&E) staining or other examinations. Most immunofluorescence labeling of tissue specimens and cells was performed on formalin-fixed paraffin-embedded tissue sections as described previously [68,69]. RPGRIP1L, cilia, DSG3, and CTNNA1 were detected on frozen sections of the skin (Supplementary Methods). TUNEL staining was performed with DeadEnd Fluorometric TUNEL System (Promega, Fitchburg, WI). Primary antibodies are listed in S1 Table. AlexaFluor-conjugated secondary antibodies (1:200) were obtained from Life Technologies (Carlsbad, CA). Sections were sealed in mounting medium with or without DAPI (Vector Laboratories, Burlingame, CA). Images were acquired by a Nikon 80i fluorescence microscope, fitted with a Nikon (Melville, NY) DS-Qi1Mc camera, or by a Leica (Wetzlar, Germany) SP5C Spectral confocal laser-scanning microscope, and processed with Photoshop 5.5 CS (Adobe System Incorporated, San Jose, CA).
To quantify fluorescence intensity of skin tissues, the mean intensity of randomly selected epidermal regions (2 regions per specimen, n ≥ 3 mice per group, excluding the cornified layer), was measured with the NIS-Element analysis software, as described elsewhere [50].
To obtain the ratio of membrane/cytoplasmic fluorescence intensity, the mean peak pixel values at the two edges of a cell (representing the plasma membrane) and the mean pixel value between the peaks (representing the cytoplasm) were obtained by the Plot Profile tool of the ImageJ software (1.43u, National Institute of Health, Bethesda, MD), and presented as a membrane/cytoplasmic ratio.
The dispase dissociation assay was performed as described previously [70]. Briefly, 24 hours after transfection, confluent cells were incubated in high calcium medium (1.5 mM CaCl2) for 24 hours. Subsequently, cells were washed with DPBS and incubated with dispase II (2.4 U/ml in EMEM with 10% FBS and 1.5 mM CaCl2, Roche, Indianapolis, IN), for 20 min at 37°C. Detached monolayers were subjected to mechanical challenge by inverting 50 times in 4 ml PBS in a 15-ml Falcon tube. Cell fragments were imaged and counted under a dissecting microscope (Stemi 2000-C, Carl Zeiss, Obserkochen, Germany).
The IgG internalization assay was performed to detect internalized DSG3 as previously described [27,29]. Briefly, HaCaT cells were incubated with a monoclonal antibody (AK23) against the extracellular domain of DSG3 [41], in media containing 1.5 mM calcium for 30 minutes on ice. Cells were then washed and incubated with PV IgG (400 μg/ml) or normal human IgG (400 μg/ml) at 37° C for one hour to induce DSG3 internalization. Subsequently, cells were treated with acid wash solution (100 mM glycine, 20 mM magnesium acetate, 50 mM potassium chloride, pH 2.2) to remove surface-bound DSG3 antibody before fixation. For knockdown studies, cells were transfected with siRNA prior to calcium switch. Images were acquired by Leica SP5C Spectral confocal laser scanning microscope under the same color intensity threshold and analyzed using ImageJ. Quantification was done by counting green fluorescent puncta in randomly sampled microscopic fields with Analyze Particle, which was then normalized by the number of cells so that the net result reflects the average number of puncta (internalized DSG3) within one cell.
Cell surface protein biotinylation and endocytosis assays were used to measure internalization of cell surface proteins, as described previously [42]. Briefly, cells were incubated with freshly prepared 2 mg/ml Sulfo-NHS-SS-biotin (EZ-Link™Sulfo-NHS-SS-Biotin; Thermo Fisher Scientific, Waltham, MA) for 30 min at 4°C after two washes in ice-cold PBS2+ (PBS with 1.5 mM CaCl2 and 1.5 mM MgCl2) for biotinylation to occur. Cells were then washed and incubated in three changes of quenching solution (100 mM glycine in PBS2+), 10 minutes each, on ice. After a PBS2+ wash, cells were incubated in the pre-warmed high calcium growth media containing 2 mM EGTA for 30 min to trigger internalization. Stripping control cells were kept at 4°C to block internalization. Subsequently, non-internalized biotin was stripped by washing with cold NT buffer (150 mM NaCl, 1.0 mM EDTA, 0.2% BSA, 20 mM Tris, pH 8.6, and 50 mM Tris(2-Carboxyethyl) phosphine Hydrochloride), and cell lysates were collected in RIPA buffer containing protease inhibitor. Biotinylated proteins were pulled down with streptavidin magnetic beads (Thermo Fisher Scientific) at 4°C overnight, eluted in Laemmli buffer at 95°C, and analyzed by immunoblotting. Rabbit anti-DSG3 (Bio-Rad AbD Serotec, Raleigh, NC) was used to detect biotinylated DSG3. Target proteins were examined in a minimum of three experiments. Results from one representative experiment are shown.
All quantifications are presented as mean ± SD. Student’s t-test was used unless otherwise stated. One-way ANOVA and two-way ANOVA were conducted using the GraphPad software. P < 0.05 was considered statistically significant.
Additional Materials and Methods information is provided in the Supplementary Materials and Methods (S1 Text) online.
|
10.1371/journal.pntd.0001311 | Preexisting Japanese Encephalitis Virus Neutralizing Antibodies and Increased Symptomatic Dengue Illness in a School-Based Cohort in Thailand | Dengue viruses (DENVs) and Japanese encephalitis virus (JEV) have significant cross-reactivity in serological assays; the clinical implications of this remain undefined. An improved understanding of whether and how JEV immunity modulates the clinical outcome of DENV infection is important as large-scale DENV vaccine trials will commence in areas where JEV is co-endemic and/or JEV immunization is routine.
The association between preexisting JEV neutralizing antibodies (NAbs) and the clinical severity of DENV infection was evaluated in a prospective school-based cohort in Thailand that captured asymptomatic, non-hospitalized, and hospitalized DENV infections. Covariates considered included age, baseline DENV antibody status, school of attendance, epidemic year, and infecting DENV serotype. 942 children experienced at least one DENV infection between 1998 and 2002, out of 3,687 children who were enrolled for at least one full year. In crude analysis, the presence of JEV NAbs was associated with an increased occurrence of symptomatic versus asymptomatic infection (odds ratio [OR] = 1.55, 95% CI: 1.08–2.23) but not hospitalized illness or dengue hemorrhagic fever (DHF). The association was strongest in children with negative DENV serology (DENV-naive) (OR = 2.75, 95% CI: 1.12–6.72), for whom the presence of JEV NAbs was also associated with a symptomatic illness of longer duration (5.4 days for JEV NAb+ versus 2.6 days for JEV NAb-, p = 0.048). JEV NAbs were associated with increased DHF in younger children with multitypic DENV NAb profiles (OR = 4.05, 95% CI: 1.18 to 13.87). Among those with JEV NAbs, the association with symptomatic illness did not vary by antibody titer.
The prior existence of JEV NAbs was associated with an increased probability of symptomatic as compared to asymptomatic DENV illness. These findings are in contrast to previous studies suggesting an attenuating effect of heterologous flavivirus immunity on DENV disease severity.
| Dengue viruses (DENVs) and Japanese encephalitis virus (JEV) have significant cross-reactivity in serological assays, but the possible clinical implications of this remain poorly understood. Interactions between these flaviviruses are potentially important for public health because wild-type JEV continues to co-circulate with DENV in Southeast Asia, the area with the highest burden of DENV illness, and JEV vaccination coverage in this region is high. In this study, we examined how preexisting JEV neutralizing antibodies (NAbs) influenced the clinical severity of subsequent DENV infection using data from a prospective school-based cohort study in Thailand that captured a wide range of clinical severities, including asymptomatic, non-hospitalized, and hospitalized DENV infections. We found that the prior existence of JEV NAbs was associated with an increased occurrence of symptomatic versus asymptomatic DENV infection. This association was most notable in DENV-naives, in whom the presence of JEV NAbs was also associated with an illness of longer duration. These findings suggest that the issue of heterologous flavivirus immunity and DENV infection merits renewed attention and interest and that DENV vaccine developers might incorporate detailed assessments of preexisting immunity to non-DENV flaviviruses and histories of vaccination against non-DENV flaviviruses in evaluating DENV vaccine safety and efficacy.
| The dengue viruses (DENV) and Japanese encephalitis virus (JEV) are closely-related members of the virus family Flaviviridae. DENV and JEV co-circulate in the Indian subcontinent and in Southeast Asia, where they are important causes of human disease and mortality. The co-occurrence of JEV and DENV has been documented in Thailand since 1969, when severe epidemics of each were observed in the Chiang Mai valley region [1]. There is no licensed DENV vaccine and vector control efforts have been largely ineffective in containing transmission. While inactivated and live-attenuated JEV vaccines are licensed for use in humans, vaccination does not interrupt the primary JEV transmission cycle involving pigs, waterfowl, and Culicine mosquitoes [2]. Despite reported high levels of JEV vaccination (estimated to be 84% in 1998 and 98% in 2008), infections continue to be detected in Thailand each year [3].
JEV and DENV exhibit significant serological cross-reactivity, which can complicate assessment of the relative burdens of each in co-endemic areas and their possible interactions [4], [5]. There exists limited, inconclusive evidence regarding the clinical implications of prior JEV exposure or JEV vaccination and the severity of subsequent DENV infection. Using observed interactions between DENV serotypes as an analogy, JEV/DENV cross-reactive immunity may possibly be protective [6], detrimental [7], or inconsequential. Hoke et al. reported that recipients of inactivated JEV vaccine (JEVAX) experienced a non-significant decrease in the occurrence of DHF relative to placebo during the first two years after vaccination and that, among DHF patients, vaccinees experienced milder disease [8]. There was no evidence of an association between JEV vaccination and the occurrence of DHF in a study of hospitalized DENV patients in Bangkok [9]. There has been no reported increase in adverse events following live-attenuated DENV vaccination of JEV-immune volunteers [10], [11]. However, DENV vaccine recipients have demonstrated heightened and broadened DENV antibody responses and antibody responses of longer duration in the setting of preexisting heterologous flavivirus immunity [12]–[14].
Two human studies have provided evidence of a possible protective effect of the opposite sequence; i.e., DENV exposure followed by JEV infection. Burke et al found decreased clinical severity in JEV-infected hospitalized patients with higher levels of flavivirus-cross-reactive IgG in Thailand, presumed to be attributable to prior DENV infection [15]. Hammon observed that following the eradication of DENV from Guam in 1945, a subsequent large JEV epidemic in 1947 caused illness in those who were less likely to have been exposed to DENV previously, namely young children and adult expatriates[16], [17].
Animal and in vitro studies of cross-reactivity between various combinations of flaviviruses suggest that the nature of the interactions need not be bidirectional and that the influence of a given virus may vary by serotype and even strain [18]. There exists ample evidence that DENV infection may be enhanced in vitro with heterotypic DENV antibodies [19] and also with antibodies to non-DENV flaviviruses, including JEV [20]. However, a study using sera from JEV-immune Thai individuals found no evidence of enhancement of DENV-2 infection in vitro [21]. Animal studies have suggested a protective role of DENV immunity upon JEV challenge in mice [22] and a protective role of WNV immunity upon subsequent challenge with another member of the JEV antigen complex in a variety of animal models [23]–[26]. In summary, there has been evidence of both protective and detrimental interactions between heterologous flaviviruses; the mechanisms and epidemiological implications of these associations remain unclear.
Potential interactions between flaviviruses are important for public health because wild-type JEV continues to co-circulate with DENV in Southeast Asia, the area with the highest burden of DENV illness, and JEV vaccination coverage in this region is high. As DENV vaccines advance toward licensure and implementation in co-endemic regions, an improved understanding of what constitutes protective immunity with DENV exposure is necessary. Given ambiguous findings from prior studies, we examined how preexisting JEV immunity influenced the clinical severity of subsequent DENV infection using data from a prospective school-based cohort study in Thailand [27], [28]. The availability of information on asymptomatic DENV infections as well as outpatient and hospitalized dengue cases provided a unique opportunity to assess these interactions.
The prospective cohort study that generated these data was co-administered by the University of Massachusetts and the Armed Forces Research Institute of Medical Sciences (AFRIMS) for the years 1998–2002. TE, AR, DL, and AN were involved in the design and conduct of the cohort study, for which all subjects provided written informed consent. ST and RG are among the lead scientists presently managing the Department of Virology at AFRIMS, where specimens and data from this study are maintained. The lead author, KA, used only preexisting demographic and laboratory data for the purposes of this retrospective analysis.
The cohort study was approved by the Human Use Review and Regulatory Agency of the Office of the Army Surgeon General, the Institutional Review Board of the University of the Massachusetts Medical School, and the Ethical Review Board of the Ministry of Public Health, Thailand. Secondary data analysis for the purpose of this publication was approved by the Institutional Review Board at Emory University.
Data were collected during a five-year, school-based, prospective cohort study for DENV infections in children in Northern Thailand. The study design and methods have been described previously [27], [28]. Briefly, the study was conducted in Kamphaeng Phet Province from 1998–2002. In January 1998, 2,214 children were recruited from grades 1 through 5 at twelve primary schools. New participants were enrolled from the 1st grade class in January of each year and eligible participants were re-enrolled. The numbers of children enrolled at the start of the active surveillance period each year were 2,044 in 1998; 1,915 in 1999; 2,203 in 2000; 2,011 in 2001; and 1,759 in 2002. A total of 3,687 children were enrolled for at least one year during the study period, with an average of 2.9 years of follow-up per child. At enrollment, the children or their parents were asked about prior JEV vaccination and age of vaccination.
Serum samples were collected for dengue serology four times each year (January, June, August, and November). All serological and virological testing was performed at AFRIMS in Bangkok, Thailand. Active case surveillance of the participants was conducted from June 1 to November 1, with potential illnesses identified based on absence from school, visit to a school nurse or public health clinic, or admission to the hospital. Absent students were visited by village health workers and evaluated with a symptom questionnaire and an oral temperature. Acute blood samples were obtained for students with a history of fever within 7 days of fever onset, as were 14-day convalescent samples.
Acute and convalescent blood specimens from incident febrile illnesses were tested using immunoglobulin M (IgM) and G (IgG) enzyme immunoassays for DENV and JEV. Acute DENV infections were defined serologically as a DENV-specific IgM level ≥ 40 units and with DENV-IgM > JEV-IgM. The infecting DENV serotype was identified from acute blood specimens using serotype-specific reverse-transcriptase polymerase chain reaction (RT-PCR) or virus isolation. Symptomatic infections were defined as a documented history of febrile illness with virologic or serologic evidence of acute DENV infection. Charts of hospitalized children were independently reviewed and classified as DF or DHF and assigned a severity grade following WHO criteria [29]. If a child experienced a febrile DENV illness but did not meet the criteria for DHF, they were characterized as having DF.
The duration of illness was derived from home visit data and hospitalization records. For non-hospitalized illnesses, the duration of illness was the length of time from the date of school absence or clinic visit to the date of the last home visit at which the child exhibited symptoms consistent with DENV infection. For hospitalized illnesses, the duration of illness was the pre-hospitalization time plus the time in the hospital.
Routine specimens were tested for the presence of hemagglutination inhibiting (HI) antibodies against DENV-1 – DENV-4 and JEV using standard methods [30]. The reference virus strains were DENV-1 (Hawaii), DENV-2 (New Guinea ‘C’), DENV-3 (H87), DENV-4 (H241), and JEV (Nakayama). Asymptomatic DENV infections were defined as a four-fold or greater rise in hemagglutination inhibition (HI) titers for any of the four DENV serotypes between two consecutive routine serum samples and without a concurrent four-fold rise in JEV HI titers, or with a concurrent four-fold rise in JEV HI titers but with higher HI titers for any DENV serotype than for JEV.
Plaque reduction neutralization titers (PRNTs) were obtained for pre-infection samples using standard methods [31]. Briefly, LLC-MK2 cell monolayers were infected with DENV1 – DENV4 and JEV in the presence of serial dilutions of heat-inactivated patient plasma. The reference virus strains were: DENV-1 (16007), DENV-2 (16681), DENV-3 (16562), DENV-4 (1036), and JEV (Nakayama), for which the sources and passage histories have been previously described [32]. The lowest dilution of serum tested was 1∶10 (corresponding to a final dilution of 1∶20 when combined with an equal volume of DENV); the dilutions and titers reported herein refer to the initial serum dilution. The concentration of patient plasma that resulted in a 50% reduction in plaque formation was calculated using log probit regression. The reciprocal titer of this dilution was defined as the PRNT50. A PRNT50 <10 was defined as undetectable or ‘negative’ and a titer ≥10 as ‘positive.’ PRNT assays were performed for asymptomatic infections only if the child had not missed school during the observation period. This was done in an attempt to exclude missed symptomatic DENV illnesses that could have been misclassified as asymptomatic seroconversions.
Children were grouped into three categories of DENV immunity based upon their pre-infection DENV antibody profiles: DENV-naïve (pre-infection PRNT50<10 for all DENV serotypes), DENV-monotypic (pre-infection PRNT50 ≥10 for a single DENV serotype), and DENV-multitypic (pre-infection PRNT50 ≥10 for two or more DENV serotypes). To attempt to discern cross-reactive immunity from serotype-specific immunity, we further stratified DENV-multitypics according to mean age of infection, with the logic that older children would have had more time to experience multiple DENV infections and generate multiple serotype-specific responses, while younger children with multitypic profiles would, on average, reflect more cross-reactivity. Preexisting JEV immunity was dichotomized as JEV-positive (pre-infection PRNT50≥10) and JEV-negative (pre-infection PRNT50<10).
By definition, acute phase blood samples were not available from children with asymptomatic DENV infections for direct detection of the infecting virus. Approximately one-quarter of symptomatically infected individuals were RT-PCR negative, likely because the acute specimen was drawn when the child was no longer viremic. Based upon evidence that DENV transmission in this community is highly clustered, we assumed that the serotypes detected among RT-PCR positive cases attending a given school during a given epidemic year were representative of the serotypes causing asymptomatic and RT-PCR negative symptomatic infections at that school [33], [34]. To minimize the possible biases inherent to this assumption, imputation of an individual's unknown infecting serotype using community-level data was performed only for children residing in communities that had a single serotype detected during that epidemic year. If more than one serotype was detected at a given school during a given year, the infecting serotype was left as missing.
The JEV vaccine used in Thailand and throughout Southeast Asia, JEVAX, is a formalin-inactivated, mouse brain-derived JEV vaccine. The first dose of JEVAX in Thailand is administered concurrently with the diphtheria/tetanus/pertussis vaccine at 18 months of age. A second dose is administered 1 to 4 weeks later and a third (booster) dose is administered at 30 to 36 months of age. The seroconversion rates with JEVAX in Thai children have been estimated to be 50% following primary vaccination, 64–80% following secondary vaccination, and 100% following receipt of the third dose [35]. Based upon these observations of suboptimal seroconversion rates with the two dose regimen, the third dose was added to the Expanded Program for Immunization (EPI) regimen in Thailand in 2000.
A Phase III trial of JEVAX was conducted in 1984 in Kamphaeng Phet, Thailand, which demonstrated favorable efficacy [8]. JEVAX began to be slowly incorporated into the EPI in Thailand in 1988 and became part of Kamphaeng Phet's EPI in 1992. Vaccination histories for the cohort study were obtained by self-report and subsequent verification with clinic or family documents was not possible. However, given the ages of the children enrolled (five to fourteen years of age) and the timing of the study (1998 to 2002), it is likely that the majority of children enrolled in the cohort study would have received two doses of JEV vaccine. Younger children would likely have been vaccinated as infants, as part of the EPI program, and older children would likely have been vaccinated during the large “catch-up" vaccination programs taking place in the province in the 1990s. The JEVAX vaccine strain in use during the 1990s in Thailand was the Nakayama strain, the country has since switched to the Beijing strain.
Bivariate analyses were performed using chi-square testing for categorical variables and ANOVA or nonparametric testing for continuous variables. Logistic regression models for symptomatic versus asymptomatic infection were constructed using SAS' GENMOD procedure, accounting for the clustering of observations by school. The best model was then chosen by backward regression. Analyses were performed using SAS software, version 8 (SAS Institute, Cary, NC), SPSS for Windows version 10.0 (SPSS Inc., Chicago, IL), and R version 2.10.1 (R Foundation for Statistical Computing, Vienna, Austria).
A total of 942 children experienced at least one DENV infection between 1998 and 2002, out of 3,687 children who were enrolled for at least one full year. While some experienced multiple DENV infections during their enrollment, analyses are restricted to the first detected infection for each child. Further characterization of infections as symptomatic or asymptomatic was limited to those 569 cases that occurred within the active surveillance period (June 1 – November 1). Three-hundred-sixteen (56%) of these infections were asymptomatic and 253 (44%) were symptomatic. Of the symptomatic cases, 217 (86%) were DF and 36 (14%) were DHF (Table 1). No deaths were attributed to DENV infection.
Children with JEV NAbs were more likely to experience symptomatic infection than children without JEV NAbs (57% versus 46%, p = 0.021 by χ2 testing, Table 1). Of the 569 first detected DENV infections, 479 had pre-infection DENV and JEV neutralizing antibody titer information available for analysis (84.2%). 90% of the missing NAb data were associated with asymptomatic infections, for reasons described above.
There were no differences by age, gender, or school in the proportions of infections that were symptomatic. The proportion of infections that were symptomatic varied by year, from 26% in 2000 to 51% in 2001 (p = 0.032, by χ2 testing across all strata of epidemic year). There were no differences in the proportion symptomatic by pre-infection DENV antibody status. There was no difference in the proportion symptomatic by reported JEV vaccination history and, among those reporting a history of vaccination, no difference in the time between vaccination and the first detected infection (mean±SD 5.01±2.22 years for asymptomatics, 4.93±2.11 for symptomatics).
Children developing DF were most likely to be JEV NAb positive, those experiencing asymptomatic infection the least (50% versus 39%, p = 0.017, by χ2 testing across all strata of infection severities) (Table 2). There were no differences in the proportions of children that were JEV NAb positive by age or gender. The proportion JEV NAb positive varied by school and year. DENV-naives were most likely to be JEV NAb positive (54%), then DENV-multitypics (47%), then DENV-monotypics (26%) (p = 0.001 by χ2 testing). Those reporting a history of JEV vaccination were more likely to be JEV NAb positive (46% versus 31% p = 0.012 by χ2 testing). Among those reporting a history of vaccination, there was no difference in time since vaccination (mean±SD 5.01±2.27 years for JEV NAb positives, 4.96±2.19 for JEV NAb negatives).
JEV NAb positivity was associated with an increase in the odds of symptomatic infection in unadjusted analysis (OR = 1.55, 95% CI: 1.08 to 2.23) (Table 3). There were no significant differences in the occurrence of hospitalized illness or the occurrence of DHF.
Individuals with JEV NAbs were more likely to experience symptomatic DENV infection for all strata of preexisting DENV immunity, though this association was strongest and significant only for DENV-naives (OR = 2.75, 95% CI: 1.12 to 6.72) (Figure 1a). The association was non-significant and weakest for children with DENV-monotypic immunity prior to infection. Younger and older children with DENV-multitypic immunity had approximately the same increased odds of symptomatic infection with JEV NAbs; neither association was significant.
The directions of the associations between JEV NAbs and hospitalized illness (Figure 1b) and JEV NAbs and DHF (Figure 1c) were not consistent across strata of preexisting DENV immunity and the associations were largely non-significant. The association between JEV NAbs and DHF was significant only for younger DENV-multitypics, who had increased odds of DHF with JEV NAbs (OR = 4.05, 95% CI: 1.18 to 13.87).
The presence of JEV NAbs was associated with an increased duration of DENV illness in DENV-naives (5.70 versus 2.69 days, p = 0.045) (Table 4). For DENV-monotypics and multitypics, no difference in the duration of illness was observed.
Among those with JEV NAbs prior to infection, there was no difference in the geometric mean titer between asymptomatically and symptomatically infected individuals (p = 0.45 by Mann-Whitney U test, Figure 2). The distributions of the titers were similar.
The greatest number of RT-PCR-positive illnesses in the cohort was associated with DENV-2; the highest proportion of hospitalized infections was observed with DENV-3 (Figure 3a). The infecting DENV serotype was unknown for all asymptomatic infections and for 25.2% of symptomatic infections. 40.0% of children missing serotype data resided in communities with a single serotype in circulation the year of their infection and were therefore eligible for imputation. The probability of symptomatic infection was increased with JEV NAbs for all DENV serotypes, but the association was not significant for any serotype in subgroup analysis (Figure 3b). Limiting the comparisons to symptomatic, RT-PCR positive infections (with no imputation), the direction of the association between JEV NAbs and hospitalized illness was highly variable (Figure 3c). DENV-3 infection in the setting of preexisting JEV NAbs was associated with decreased hospitalized illness (15.0% hospitalized with JEV NAbs versus 46.7% hospitalized without, p = 0.004 by χ2 test).
The positive, significant association between the presence of JEV NAbs and symptomatic infection remained after controlling for age, pre-infection DENV immunity, and epidemic year and accounting for clustering of observations by school (data not shown). The adjusted odds ratio was 1.70 (95% CI 1.15 to 2.51).
The serological cross-reactivity between DENV serotypes is well-documented and has been linked to both cross-protection and enhanced disease. In contrast, the clinical implications of serological cross-reactivity observed between DENV and other non-DENV flaviviruses remain unclear. In this study, we characterized the association between preexisting JEV antibodies and the clinical severity of subsequent DENV infection in a prospective study of school-children in Thailand. We report the novel finding that JEV NAbs were associated with the increased occurrence of symptomatic DENV infection.
The increased occurrence of symptomatic illness with preexisting JEV NAbs was most pronounced in children who were DENV-naïve (i.e., presumably experiencing their first DENV infection), for whom JEV NAbs were also associated with a longer duration of illness. It is important to note that given the sensitive method of identifying febrile illnesses in this cohort, ‘symptomatic DENV illnesses’ ranged from a single day of fever to a prolonged and debilitating disease course. It is therefore notable that we observed an increase in the duration of illness in those with JEV NAbs, suggesting that their influence was to increase the occurrence and severity of clinically-meaningful DENV illness. The significant findings in DENV-naïve children are noteworthy because in this group, JEV NAbs are more likely to reflect a ‘true’ prior exposure to JEV or JEV vaccine in this group and less likely to have arisen as a cross reactive response to a prior DENV infection. The association between JEV NAbs and symptomatic illness was not as strong for DENV-monotypics and -multitypics, which could be due to cross-protection with increasing DENV immunity and/or a confounding effect with the inclusion of JEV NAbs as a result of cross-reactivity.
Given the absence of confirmatory data from other human cohorts or animal models, we must place these findings in the context of prior in vitro investigations. This study may be most analogous and in accordance with Putvatana et al, who found no evidence of antibody-dependent enhancement of DENV-2 infection in vitro using sera of JEV-immune Thai individuals [21]. We report an increase in non-hospitalized DENV illnesses with preexisting JEV NAbs, but not DHF (in unadjusted analysis and for nearly all subgroups), which may suggest that enhancement is less likely to be the mechanism of increased DENV illness. However, the number of DHF cases was low in this cohort study and therefore the power to detect an association between JEV antibodies and severe DENV illness was limited. It is also possible that studies of this association, both in vivo and in vitro, could have different conclusions based upon the serotypes and strains under consideration [18]. Indeed, while only limited serotype-specific analyses were possible with these data, the nature of the association between JEV NAbs and hospitalized illness did not appear to be consistent across serotypes. Finally, one important caveat with this in vivo study is that while an association was indeed detected between JEV NAbs and DENV illness, it is possible that JEV NAbs are in fact a marker of another underlying biological function that was not considered in this analysis, such as cell-mediated immunity.
Notably, there was not a strong independent association in this study between preexisting DENV immunity and the clinical severity of a subsequent DENV infection; this may at first glance appear to be in contrast to prior studies linking secondary DENV infection with an increased occurrence of hospitalized illness or DHF [36], [37]. However, the present study uniquely focused on predictors of symptomatic (primarily non-hospitalized) infection, for which the influence of prior DENV exposures and DENV immunity may be different. Second, while DENV-naives and DENV-monotypics may be somewhat reliably characterized as having no prior DENV infections and one prior DENV infection, respectively, the DENV-multitypic group likely comprises a mixture of children with multiple prior DENV infections as well as children with a single prior DENV infection and a persistent cross-reactive response. This ‘mixing of effects’ within the DENV-multitypic group may have clouded the association between DENV immunity and the clinical severity of DENV infection.
It is possible that some children were misclassified with respect to pre-infection JEV serostatus in this study. The prevalence of JEV NAbs in children experiencing DENV infection in the cohort was 45%, remarkably low given that vaccine coverage was estimated to exceed 80% at the time of the study. This low seropositivity is likely due to waning of the antibody response, which is a well-documented phenomenon with JEV inactivated vaccines and particularly with the two-dose regimen [35], [38]. 60% of children who lacked JEV NAbs ‘seroconverted’ to become JEV NAb positive in the post-season sample following a DENV infection, which may reflect an anamnestic response to prior JEV vaccination. Additionally, 76% of DENV-naïve/JEV-negatives exhibited a secondary-type response by ELISA during acute infection, further suggesting that negative JEV and DENV antibody titers do not preclude the possibility of a prior flavivirus infection or JEV vaccination.
It is conceivable that different sources of JEV NAbs, which could not be discerned in this study, may modulate the severity of DENV infection in different ways. Given the cross-reactivity between JEV and DENV in serological assays, it is possible that JEV NAbs in some cases were present as a cross-reactive response to a prior DENV infection. Further, while JEV vaccination was widespread during the study period, wild-type JEV continued to circulate and cause human infections. In summary, the JEV antibodies detected in the cohort may have arisen as a result of JEV vaccination, JEV infection, cross-reactivity from DENV infection, or combinations of these. Future studies should seek to distinguish between vaccine-derived JEV immunity and immunity derived from natural exposure, perhaps by analysis of NS1-specific antibodies [39].
The identification of asymptomatic DENV seroconversions using sequential HI antibody data is a unique strength of this study. However, it should be noted that the sensitivity of this method to detect post-primary DENV infections, or DENV infections in JEV immunes, has not yet been validated against a gold standard. It is therefore possible that some DENV infections were missed or misclassified as JEV infections (i.e., false negatives), or that assay and/or biological variability caused an elevation of HI titers where no new infection had in fact occurred (i.e., false positives).
This first report of an association between preexisting JEV NAbs and DENV illness warrants further study. Because this study was conducted on a relatively limited temporal, spatial, and demographic scale, similar analyses should be repeated in other cohorts. It would be of particular interest to compare DENV-endemic regions where live-attenuated and inactivated JEV vaccines are in use. Possible associations between other flaviviruses and DENV illness should be evaluated, and possible effect-modifying effects of the infecting DENV serotype, should be explored. It would be of great interest to investigate the association between JEV cell-mediated immunity induced by JEV vaccination and wild-type infection and the occurrence of DENV illness.
The biological mechanism of this association remains to be elucidated. However, an intriguing mechanistic possibility may be found in the literature surrounding other inactivated virus vaccines. Some inactivated vaccines have been linked with increased disease, perhaps due to the generation of an epitope-restricted, lower titer immune response. Early efforts to develop inactivated vaccines against measles and respiratory syncytial virus were abandoned after they were linked with the occurrence of atypical, occasionally severe disease [40], [41] There have been no reports to date of immuno-pathological responses following inactivated flavivirus vaccination in humans, though a recent study in mice reported that low doses of JEVAX were associated with increased viral load and death following subsequent Murray Valley encephalitis virus challenge, relative to placebo, high doses of JEVAX, and the live Chimerivax-JEV vaccine [42].
In summary, we report that the prior existence of JEV NAbs was associated with an increased probability of symptomatic DENV illness in a cohort of school-children in Thailand. These findings have public health importance in that DENVs co-circulate with other flaviviruses in much of their geographic range (e.g., JEV in Asia, yellow fever virus in Africa and South America, and West Nile virus in various locations in both hemispheres) and JEV vaccination is common throughout South and Southeast Asia. We suggest that the issue of heterologous flavivirus immunity and DENV, usually considered to be inconsequential or perhaps protective, merits renewed interest and investigation. In particular, the findings indicate that DENV vaccine developers should include preexisting flavivirus immunity and vaccination histories in assessments of vaccine safety and efficacy. The results of these studies may be important for shaping DENV vaccine implementation strategies.
|
10.1371/journal.pbio.1000595 | Genome-Wide and Phase-Specific DNA-Binding Rhythms of BMAL1 Control Circadian Output Functions in Mouse Liver | The mammalian circadian clock uses interlocked negative feedback loops in which the heterodimeric basic helix-loop-helix transcription factor BMAL1/CLOCK is a master regulator. While there is prominent control of liver functions by the circadian clock, the detailed links between circadian regulators and downstream targets are poorly known. Using chromatin immunoprecipitation combined with deep sequencing we obtained a time-resolved and genome-wide map of BMAL1 binding in mouse liver, which allowed us to identify over 2,000 binding sites, with peak binding narrowly centered around Zeitgeber time 6. Annotation of BMAL1 targets confirms carbohydrate and lipid metabolism as the major output of the circadian clock in mouse liver. Moreover, transcription regulators are largely overrepresented, several of which also exhibit circadian activity. Genes of the core circadian oscillator stand out as strongly bound, often at promoter and distal sites. Genomic sequence analysis of the sites identified E-boxes and tandem E1-E2 consensus elements. Electromobility shift assays showed that E1-E2 sites are bound by a dimer of BMAL1/CLOCK heterodimers with a spacing-dependent cooperative interaction, a finding that was further validated in transactivation assays. BMAL1 target genes showed cyclic mRNA expression profiles with a phase distribution centered at Zeitgeber time 10. Importantly, sites with E1-E2 elements showed tighter phases both in binding and mRNA accumulation. Finally, analyzing the temporal profiles of BMAL1 binding, precursor mRNA and mature mRNA levels showed how transcriptional and post-transcriptional regulation contribute differentially to circadian expression phase. Together, our analysis of a dynamic protein-DNA interactome uncovered how genes of the core circadian oscillator crosstalk and drive phase-specific circadian output programs in a complex tissue.
| The circadian clock is a timing system that allows organisms to keep behavioral, physiological, and cellular rhythms in resonance with daily environmental cycles. In mammals, such clocks use transcriptional regulatory loops in which the heterodimeric transcription factor BMAL1/CLOCK plays a central role. While defects in circadian clock function have been associated with diabetes, obesity, and cancer, the molecular links between the circadian clock and such output pathways are poorly characterized. Here, we mapped DNA-binding sites of BMAL1 in mouse liver during one circadian cycle. Our temporal analysis revealed widespread daily rhythms in DNA binding, with maximum levels peaking at midday. In the list of candidates, core circadian genes stood out as the most strongly bound, often showing multiple binding sites. Interestingly, BMAL1 targets were highly enriched in genes involved in carbohydrate and lipid metabolism, and also in transcription factors, in particular nuclear receptors. Our results suggest that the mammalian clock uses BMAL1 to control transcriptional output programs both directly and indirectly. Additionally, the DNA specificity of BMAL1 binding revealed the importance of tandem E-box elements, which may favor strong binding and precise timing of daily gene expression. Taken together, our work confirms BMAL1's primary function as a master regulator of the core circadian oscillator, while demonstrating that it also contributes in a more distributed fashion to a variety of output programs.
| Circadian clocks provide higher organisms with cell-autonomous and organ-based metronomes that control temporally gated and tissue-specific gene expression or metabolic programs [1]–[4]. In the liver, such programs have been implicated in detoxification [5], glucose homeostasis [6],[7], cholesterol biosynthesis [8],[9], and gating of the cell cycle [10],[11]. The mammalian clock depends on a cell-autonomous [11] core oscillator that is built around interlocked transcriptional feedback loops. These use a variety of transcriptional regulators: the basic helix-loop-helix (bHLH) PAS domain proteins CLOCK, NPAS2, and BMAL1 [12],[13], orphan nuclear receptors of the REV-ERB [14] and ROR families [15], and the DEC bHLH repressors [16]. In addition, important co-regulators such as PER and CRY proteins mediate negative feedback by repressing their own transcriptional activators, BMAL1/CLOCK [17]–[20]. Among all these regulators, the Bmal1 gene is the only single gene in the circadian network whose knockout results in arrhythmicity [21],[22]. BMAL1 functions as a heterodimeric complex, BMAL1/CLOCK, that activates transcription of its targets via E-boxes [12],[23],[24]. The DNA-binding activity of BMAL1/CLOCK is thought to cycle because of circadian changes in post-translational modifications [25],[26]. The core oscillator exerts its function by controlling temporally gated outputs, notably metabolic functions [5],[7],[27]. Transcriptional regulation of circadian output is known to occur both directly via the core clock transcription factors and indirectly, as, for example, via the PAR-bZip regulators DBP/TEF/HLF, which are themselves controlled by BMAL1/CLOCK [28]. Thus, circadian output function is controlled via a hierarchical network of transcription regulators that drives vast programs of tissue-specific gene expression both in the suprachiasmatic nucleus [29] and in peripheral tissues [29]–[34] in the mouse. Notably, these transcript rhythms cover the full range of expression phases, which thus begs the question about the mechanism behind phase-specific circadian gene expression. It has been proposed that virtually any peak expression phase can be achieved by suitably tuned regulatory sequences that integrate a small number of phase-specific core regulators [35]. Here we investigate the degree to which BMAL1 recruitment to the genomic DNA is itself rhythmic and to what extent peak binding carries phase information for downstream circadian mRNA expression.
To address these questions and further dissect the hierarchical structure of circadian clock networks, we perform a time series chromatin immunoprecipitation (ChIP) analysis for the master clock regulator BMAL1 in mouse liver. This allows us to identify a comprehensive set of direct BMAL1 targets in a circadianly controlled tissue, to model the DNA-binding specificity of BMAL1 in vivo, and to determine how tightly the phase of mRNA output follows rhythmic protein-DNA interactions. Our results reveal the pervasiveness of circadian protein-DNA interactions in a mammalian tissue by showing widespread rhythmic and phase-specific binding of BMAL1 to coding and non-coding genes. This enables us to characterize the cooperative interactions of BMAL1/CLOCK complexes at tandem E-box elements (E1-E2), and to emphasize the complexity of circadian phase control that involves transcriptional and post-transcriptional mechanisms.
To obtain a time-resolved and genome-wide map of BMAL1 target sites, we performed ChIP in mouse liver at 4-h time intervals during one light-dark cycle. Following initial testing of ChIP efficiency by quantitative PCR (qPCR) (Figure S1), two independent BMAL1 ChIP time courses were subjected to ultra-high-throughput sequencing to yield about 20 million tags per time point (Table S1) and were analyzed via a bioinformatics pipeline that combines existing and novel methods. Briefly, we used the MACS software [36] to detect regions with enriched BMAL1 binding compared to an input chromatin sample (see Materials and Methods). To efficiently reject spurious signals and accurately estimate the location of binding sites, we developed a model-based deconvolution method for ChIP combined with deep sequencing (ChIP-Seq) data (see Text S1). We identified 2,049 bona fide BMAL1 binding sites in mouse liver. Among the top 200 sites, more than 90% are significantly rhythmic (Fisher test, p<0.05; see Materials and Methods), while the proportion drops to 60% for all sites (1,319 sites) (Figure 1A). Consistent with previously published results [24],[37], the binding phases are sharply distributed around Zeitgeber time (ZT) 4 to ZT8 (Figure 1A and 1B), which confirms BMAL1 as a highly phase-specific circadian transcription factor. At peak time, the binding signal (measured in number of unique tags in a site) spans over one order of magnitude, and sites near reference clock genes (RCGs) clearly stand out as the most strongly bound sites (Figure 1C; Text S3), i.e., 26 out of the 41 RCG sites are among the top 5% binding sites. In addition, RCGs often have multiple robustly rhythmic binding sites. For example, the Dbp gene has three sites: at the promoter and in the first and second introns (Figure 1D), with peak-to-trough amplitudes greater than 10-fold, similar to those measured with qPCR (Figure S1), with some residual binding at ZT18 compared to input chromatin. The three sites clearly overlap with DNase I hypersensitive sites mapped in [28] and also with evolutionarily conserved regions in the genome, suggesting that these sites are under purifying selection. Similarly, Rev-Erbα shows three strongly rhythmic sites, two near the promoter and one 8 kb upstream (Figure 1E), which could be involved in DNA looping with the promoter sites. A vast majority of RCGs, including the Per1/2, Cry1/2, Dec1/2, Rev-Erbβ, Rorγ, E4bp4, and Hlf/Tef genes, show similarly strong signals (Figure S2). Moreover, we also find binding sites at recently identified targets like Gys2 [38], Nampt [39],[40], and Wee1 [10].
To study the location of BMAL1 binding sites relative to genes, we annotated each site with the nearest Ensembl transcript, including coding and non-coding genes. Positioning of BMAL1 sites with respect to the Ensembl annotation shows that 40% of the sites are within 1 kb, and 60% within 10 kb, of an annotated transcription start site (TSS) (Figure 2A) (random expectation is 15%, p<10−16, binomial test). Viewed on a finer scale, the 40% of sites within 1 kb of TSSs cluster slightly upstream of TSSs (50–100 bp upstream), while no similar correlation is observed for the 3′ ends of transcripts (Figure 2B). Compared to genomic frequencies, BMAL1 sites are strongly enriched in promoter regions (±2 kb around TSS) of coding genes and depleted inside genes (Figure 2C). To assess whether BMAL1 might also control non-coding genes, we considered all transcripts with a binding site within 10 kb and found that the majority of sites are close to coding genes (more than 50%), while few are found near RNA genes or microRNAs (Figure 2D; Text S2). Moreover, we found that BMAL1 binds in accessible and transcriptionally active chromatin regions, as 83% of the sites are located near genes that are expressed (defined as expressed above the median in RNA-Seq liver data [41]; Figure S3A; see Materials and Methods), which represents a highly significant fraction (p<10−15, rank test). Phylogenetic analysis showed that the conservation of BMAL1 sites increases with the strength of binding (Figure 2E), with the first 100 sites showing very high conservation (median PhastCons conservation scores near 1). Importantly, this is not simply a consequence of strongly bound sites tending to fall near TSSs (Figure S3B), as all Ensembl TSSs show lower conservation (Figure 2E). We further assessed conservation levels in both proximal sites (within 1 kb of an annotated Ensembl TSS) and distal sites, and found that both categories of sites were significantly more conserved than control regions (taken 500 bp downstream of each site), with distal sites showing on average less conservation than proximal sites (Figure 2F). On the same scale, sites close to RCGs showed strong conservation among mammalian species.
Functional annotation analysis with DAVID [42],[43] identified enriched annotation clusters, the most prominent ones relating to carbohydrate and lipid metabolism, as well as transcriptional regulation in general (Table S2). This supports the finding that glucose metabolism is a major hepatic function directly controlled by BMAL1 [6],[7],[29],[38]. For example, glycolytic enzymes and transporters that were previously implicated in the circadian control of glucose homeostasis, e.g., Pck1 and Glut2 [7], as well as G6Pase [44], are identified as putative targets. As the mRNAs of these genes cycle with a phase that is expected for BMAL1/CLOCK targets, our data argue these key nodes are direct BMAL1 targets. Supporting this scenario, loss of function mutants have shown that BMAL1 and CLOCK are involved in glucose homeostasis [6],[45]. Similarly, lipid synthesis, notably sterol and triglyceride metabolism, is significantly enriched among BMAL1 targets, which substantiates the action of the core clock in these pathways. Interestingly, the most enriched functional cluster is transcriptional regulation: in total, 82 DNA transcription factors show BMAL1 binding, including 18 nuclear receptors, all expressed in liver (Table S3; [27]), 15 basic-leucine zipper proteins, 6 bHLH factors, and 10 zinc fingers (Table S3), indicating a hierarchic organization of circadian output programs. Notice, though, that only a minority of theses sites show binding strengths comparable to those of canonical clock genes. Unexpectedly, the Bmal1 promoter itself shows a weak BMAL1 site, the significance of which is unclear at this point. More than 30% of these factors show rhythmic mRNA abundance on expression arrays (Table S3). To assess whether these factors are also circadianly active, we applied a bioinformatics analysis that combines known transcription factor consensus sites with mRNA measurements to infer active transcription factors [46],[47]. This method predicts a transcription factor as circadianly active when its putative targets, identified as those genes showing a conserved consensus binding site in their promoter, show phase coherent circadian expression (see Materials and Methods). Out of 22 factors with represented consensus sites, this analysis predicted circadian activity for those binding the DBP/HLF/TEF/E4BP4, REV-ERB/ROR, HIF1A, PPARα, and BACH1 consensus motifs (Figure S4), thus supporting a functional role for cyclic BMAL1 binding to the promoters of these regulators. Finally, enrichment of Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways found cancer pathways as highly enriched in BMAL1 targets (DAVID, p<0.001; Table S4), notably in components of the cell cycle and in transforming growth factor beta (TGFβ) signaling (Figure S5). Specifically, we identify previously described [10],[11],[48]–[50] and novel links between the circadian clock and the cell cycle. For example, the G2-M-transition inhibitor Wee1 is a putative target. Likewise, several cyclins of the G1-S transition (Ccne1, Ccne2, Ccnd1, and Ccnd3) and their partner, cyclin-dependent kinase 4 (Cdk4), are also bound by BMAL1. Notably, several of these genes (e.g., Ccnd2 and Ccne2) show circadian mRNA expression (Figure S5A). Other important pathways at the threshold of significance that have been previously linked to circadian function include the insulin [6],[45],[51],[52] and Pparα [53]–[55] signaling pathways (Figure S5).
Having discussed genomic positioning and functional annotations of BMAL1 sites, we aimed at refining current models for the DNA-binding specificity of BMAL1/CLOCK in vivo. To this end, we performed de novo motif searches and applied hidden Markov models (HMMs) to the genomic sequences surrounding the 2,049 binding sites. As expected, a MEME [56] analysis in short windows of ±50 bp around the predicted binding location (see Materials and Methods) clearly identified E-box signals as the strongest cis-element (Figure 3A). We also found an Sp1 motif, which is consistent with 40% of sites being located near TSSs [57] (Figure 2B). In the window considered, we did not identify other sequences that could indicate the involvement of further co-factors. On the other hand, a positional analysis of the E-box sequences indicates that these frequently occur in tandem with a spacer constraint of six or seven nucleotides (Figure 3B), reminiscent of the E1-E2 element [58],[59]. This configuration prompted us to train a nucleotide profile using a HMM that considers both single and variably spaced tandem elements (Figure S6B), similar to our previous model [58]. As the binding signal spans more than a decade (Figure 1D), sites bound by BMAL1 were weighted using the number of tags at peak binding for the training of the HMM. The sequence-specific profile converges toward two E-box elements, with inferred stringencies (cutoffs) that tolerate about one (E1) and three (E1-E2) mismatches (Figure 3C; Table S5). The genomic positions of the consensus sequences co-localize tightly with the predicted centers of the ChIP signals, i.e., they are mostly within ±25 bp (Figure 3D), which is largely because of the accuracy of the deconvolution method in localizing the binding sites. Overall, 13% of all BMAL1 sites had E1-E2 elements with spacers of 6 bp (7%) or 7 bp (6%), while in RCGs this fraction represented 29% of the sites, covering 53% of genes with at least one E1-E2 site. To investigate the influence of single and tandem E-boxes for BMAL1 binding, we divided the BMAL1 sites into three classes: sites with no E-box (Ø), sites with a single E-box (E1), and sites with E1-E2 elements (E1-E2). We found that E1-E2 sites have significantly more BMAL1 tags and more rhythmic binding profiles than E1 alone or empty sites (Figure 3E). Moreover, both strongly and weakly bound BMAL1 sites harbor significantly more E1-E2 elements than control regions taken 500 bp downstream of each site (Student's t test, p<2.2×10−16; Figure 3F). In summary, our sequence analysis showed that E1-E2 tandem repeats are overrepresented in BMAL1 sites and that the presence of such regulatory sites favors strong binding.
The identified sequence elements prompted us to further characterize how BMAL1 complexes interact with DNA at these sites. We thus performed electromobility shift assays (EMSAs) with nuclear extracts from mouse livers. Using oligonucleotide probes from ChIP-Seq sites with E1-E2 sequences in the Dbp promoter, the Dbp intron 2, and the Per2 promoter, we observed three main protein-DNA complexes, present in all probes (Figure 4A). Supershift assays with BMAL1 and CLOCK antibodies indicate that the two slowest migrating complexes, hereafter termed 2BC and BC, contain BMAL1 and CLOCK (Figure 4B). The supershift assay results also exclude the possibility of other DNA-binding complexes involving either one but not both. The third and fastest migrating complex most likely represents other E-box binding bHLH proteins expressed in liver, such as the abundant protein USF1, as discussed in [28]. Of the two BMAL1/CLOCK-containing complexes, 2BC shows stronger binding, which decreases when the spacing between the E1 and E2 sites increases (Figure 4C). In contrast, BC does not seem to be affected. This argues for a cooperative interaction between two BMAL1/CLOCK heterodimers at the E1-E2 sites that is reduced and eventually lost when the spacing increases. This is reflected in the pattern for the 9-bp spacer (sp9), which is comparable to that of a probe with an intact E1 site and a mutated E2 site (E1-mE2 probe). Finally, cross-linking protein-DNA complexes in combination with two-dimensional EMSA confirms that the BC and 2BC complexes have the same composition, i.e., they both contain CLOCK and BMAL1 but no other DNA-binding proteins (Figure 4D). Taken together, these data indicate a cooperative binding of two BMAL1/CLOCK heterodimers at E1-E2 elements.
The data presented so far suggest that E1-E2 sites favor strong binding in vivo, which could result from cooperative binding of two BMAL1/CLOCK heterodimers at these elements. To substantiate the hypothesis that E1-E2 sites function as strong transcriptional enhancers, we performed transactivation assays by expressing BMAL1/CLOCK heterodimers in 293T cells and measured luciferase reporter constructs driven by wild-type E1-E2 sites from the Dbp intron 2 and Per2 promoter sites, or by mutated sites with only one E1 site. In both cases, the constructs with only one E1 site (termed E1-mE2) show significantly reduced activity compared to the constructs with intact E1-E2 sites, namely about 50% for the Dbp and 70% for the Per2 site (Figure 5A). Consistent with the EMSA results, reporter constructs with only the E2 site (mE1-E2 and E2-E2; Figure 5B) show transactivation levels comparable to background, underlining the importance of the E1 moiety. However, the E1-E1 construct had an activity similar to that of E1-E2, indicating that cooperativity can compensate for weaker binding affinity. When the spacing is increased from seven to ten nucleotides (sp10; Figure 5A), the activity is reduced to levels similar to those in E2 mutants (E1-mE2), suggesting that the interaction between the two BMAL1/CLOCK heterodimers is reduced when the phasing of the two binding sites is altered. According to this interpretation we observed that the transactivation increased again for spacers corresponding to one additional full helical turn of the DNA, i.e., spacers of 16–17 bp (Figure 5C). Notably, the intronic BMAL1 sites in Rev-Erbα and Rev-Erbβ harbor a 16-bp E1-E2 element. These results thus argue that tandem E1-E2 sites play a role in determining the magnitude of BMAL1-dependent transactivation, which parallels our finding that such elements favor strong BMAL1 binding in the liver (Figure 3E).
Our positional analysis of BMAL1 sites showed that more than 60% are located less than 10 kb from a TSS, which was emphasized by the strong enrichment of sites in promoter regions (Figure 2B). To assess whether BMAL1 binding near coding genes, i.e., located less than 10 kb from a TSS, is predictive of a circadian mRNA expression pattern and to determine a possible functional role for the E1-E2 element, we compared the putative targets with mRNA expression profiles in liver sampled around the clock [31]. The set of BMAL1 targets was highly enriched (p<2×10−16, two-sample Wilcoxon test) in circadian mRNA profiles (Figure 6A), also when we restricted our analysis to liver-specific genes (see Materials and Methods), excluding the possibility that this would merely reflect the numerous circadianly expressed transcripts in liver. Stratifying the analysis according to binding strength, we found that strong binding is highly predictive of rhythmic mRNA expression. Namely, for all BMAL1 sites with a TSS within 10 kb, 100% of targets robustly cycled among the top 10, 85% among the top 20, over 50% among the top 100, and 29% in total (Figure 6B). Consistent with the maximal binding of BMAL1 around ZT6 (Figure 1A and 1B), the expression phase of the rhythmic targets peaked around ZT10 (Figure 6C). Interestingly, the distribution of expression phases in targets with or without E1-E2 elements differed significantly: although targets harboring E1-E2 elements showed a similar mean phase compared to targets without or with single E-boxes, these genes showed a tighter mRNA phase dispersion (Rao homogeneity test for circular data [60], p<0.01), suggesting a role for the E1-E2 element in controlling the precision of the circadian expression phase (Figure 6D). However, we did not observe a significant difference in the mean nor in the dispersion between the 6-bp and 7-bp spacer variants of E1-E2. Therefore, these results suggest that a fair fraction of the BMAL1 sites induce rhythmic transcription, and that E1-E2 elements play a role in the precise temporal expression of BMAL1 targets.
To further study the temporal relationship between rhythmic BMAL1 binding, transcription, and mRNA accumulation, we quantified temporal profiles of pre-mRNA and mRNA for canonical clock genes showing strongly rhythmic BMAL1 sites. Comparing the profiles of BMAL1 binding and pre-mRNA accumulation identified several outcomes: (i) for early genes, Rev-Erbα, Rev-Erbβ, Dbp, Tef, and Dec2, the pre-mRNA closely follows binding without significant delay, suggesting that transcription largely depends on BMAL1/CLOCK (Figure 7A); (ii) genes such as Per1, Per2, and Cry2 show pre-mRNA accumulation levels that are delayed by less than 4 h compared to BMAL1 binding, suggesting that other regulators contribute to transcription (Figure 7B); (iii) finally, Cry1, Rorγ, and E4bp4 show pre-mRNA accumulation profiles that are delayed by about 12 h, indicating that other regulators are dominant in determining the phase of transcription (Figure 7C). The mRNA profiles followed pre-mRNA accumulation with short delays of maximally 4 h (see Dbp, Tef, and Dec2). We expected that longer lived mRNA transcripts would show delayed phase and reduced amplitude compared to pre-mRNA profiles, which was supported by the Gys2, March8, and Qdpr genes (Figure S7A); proxies for mRNA half-lives from mouse embryonic stem cells [61] and fibroblasts [62] showed consistency in these cell types (Figure S7B). To test the prediction that transcription in early targets depends largely on BMAL1, while additional regulators contribute to the other cases, we compared mRNA accumulation for several genes in wild-type and Bmal1−/− animals at both peak (ZT6) and trough (ZT18) BMAL1 activity time points (Figure 7D). We found that the expression of the early genes, Rev-Erbα and Dbp, was strongly suppressed in Bmal1−/− mice. Moreover, genes of the intermediate or late types showed similar (e.g., Per2) or higher (e.g., Per1, Cry1, Cry2, Rorγ, and E4bp4) levels of expression in Bmal1−/− compared to wild-type, indicating that for these categories, Bmal1 can act as a repressor, either directly [63] or indirectly [15]. Our Bmal1−/− mRNA data are consistent with measurements obtained at different time points in light-dark time courses [64], and for dark-dark time courses [15]. For Tef, the former data indicate a regulation in between early and intermediate types. Taken together, these results show that the phase of BMAL1 binding explains temporal accumulation of the early circadian transcripts. In addition, genes with delayed pre-mRNA profiles indicate that other circadian regulators contribute to transcription. Therefore, additional data for circadian activators and repressors will be key to further dissecting the transcriptional logic by which the binding amplitudes and phases of such regulators are integrated at circadian promoters.
Circadian gene expression relies on rhythmic transcription mediated by transcription factors, among which is the master regulator BMAL1/CLOCK in mammals. In this study, we identify more than 2,000 sites, of which 60% are rhythmically bound by BMAL1 in mouse liver under physiological light-dark conditions (Figure 1B). As liver tissue is mainly entrained through circadian signals from the suprachiasmatic nucleus or from feeding cues, we expect little differences with dark-dark conditions. Nevertheless, future studies in dark-dark conditions will allow estimating the changes in BMAL1 binding that are strictly dependent on the core clock. Our results substantiate at the genome-wide level the model [24],[65],[66] that rhythmic protein-DNA interactions in mammals underlie phase-specific circadian gene expression, which is reminiscent of widespread circadian binding found for dCLK in Drosophila [67], or the circadian WHITE COLLAR COMPLEX (WCC) in Neurospora [68]. Importantly, we found that peak BMAL1 binding is fairly narrowly centered around ZT6, indicating that it does not contribute much to flexibility in specifying phase at this regulatory level. As BMAL1 can form functional bHLH heterodimers with CLOCK and NPAS2 [12],[13], our data do not distinguish between targets specific for either partner. In liver, NPAS2 protein is weakly expressed [69]; however, our EMSA analysis (Figure 4B) with liver extracts did not indicate that putative BMAL1/NPAS2 complexes bind E-boxes or tandem E1-E2 elements. Similarly the BMAL1 paralog BMAL2, which is very weakly expressed in liver at the mRNA level [41],[70], can form functional BMAL2/CLOCK dimers [71]–[73] but those are not recognized by our antibody, which is highly specific to BMAL1 (Figure S8). Interestingly, we find that strongly bound BMAL1 sites exhibit high phylogenetic conservation among placental mammals, which is even more pronounced in RCGs. As recent studies showed that CEBPA and HNF4A binding in the liver could be highly species-specific [74], it would be interesting to compare our results with BMAL1 ChIP data from livers in other mammalian species.
Surprisingly, the distribution of binding strengths showed relatively few (<50) sites with binding strengths comparable in magnitude to those of core circadian genes. This indicates that BMAL1 plays a major transcriptional role in the core oscillator, while the many weaker sites suggest that it controls diverse output programs in a more distributed fashion. Among the strongest targets, known circadian genes are indeed largely overrepresented, and we found that many bona fide regulatory elements for BMAL1/CLOCK, e.g., those in Dbp [24],[62], Per2 [71], and Per1/2/3 [59],[72], were strongly bound by BMAL1. Several of those elements, e.g., in Dbp or Per2, contain previously identified E1-E2 elements [58]. However, this selectivity cannot be explained by sequence-specific binding alone. Although strongly bound sites are enriched in E1-E2 consensus sites, we also find sites with such elements that are bound more weakly (Figure 3E). As the measured ChIP signal is determined by a combination of sequence-specific binding, cooperative interaction with co-regulators, and chromatin accessibility, it is difficult to determine what distinguishes strong from weaker sites. We have just argued that sequence specificity is only partially informative, and differences in accessibility are also unlikely, as we showed that 83% of the sites fall near expressed genes. Thus, it may be that yet uncharacterized cooperative interactions with co-regulators, or cooperative interactions between multiple BMAL1 sites, are primarily responsible for the strong binding at core circadian genes (Figure 8). One candidate co-regulator could be the SP1 protein, which was suggested to bind DNA circadianly [37], and also found as an enriched cis-element (Figure 3A). Supporting the scenario of multiple interacting BMAL1 sites, we found that circadian genes often contain multiple BMAL1 binding sites (Figure S2), which could be involved in long-range DNA interactions, as proposed for the estrogen receptor [75].
Previous bioinformatics analyses, including our own, identified evolutionarily conserved E-boxes and E1-E2 sites as putative BMAL1/CLOCK consensus sites in vertebrates [58],[59],[76], both of which were shown to drive rhythmic transcription in luciferase reporter assays [59],[76]. Here we established in vivo that both simple and tandem E-boxes are characteristic of BMAL1 target genes. While the sites comprising E1-E2 elements are overall in the minority, these sites contain a number of distinguishing features: (i) more than half of the RCGs have such sites bound in vivo; (ii) E1-E2 sites are associated with strong and rhythmic binding (Figure 3E); (iii) finally, the comparison with microarray data indicates that E1-E2 sites show comparably tighter mRNA expression phases (Figure 6D). Our in vitro experiments show that BMAL1/CLOCK binding to E1-E2 elements involves a cooperative and spacing-dependent interaction between the tandem sites, consistent with the constraint in the spacer length that was identified computationally [58],[59]. Together, our data argue that single E-boxes in the genomic context are sufficient to recruit BMAL1/CLOCK heterodimers rhythmically, while E1-E2 elements may play a role in the core clock to ensure precise ticking of the circadian clock.
A central question was to study the relationship between circadian DNA binding and mRNA expression. Although the nature of ChIP experiments does not imply that circadian oscillations in DNA binding necessarily lead to a circadian modulation of the transcription rates, the body of experiments and analyses shown here indicate that a large fraction of the BMAL1 sites lead to circadian modulation in transcription. For instance, a significant fraction of BMAL1 targets show robustly circadian mRNA expression, with a peak phase that is delayed by a few hours compared to peak BMAL1 binding. Indeed, the analysis of binding profiles shows that BMAL1 binding is mainly restricted to ZT4–ZT8, while the phases of mRNA expression are centered at ZT10–ZT12, with a distribution that is broader than that of binding (Figure 6C). Analysis of pre-mRNA and mRNA levels of core clock genes in wild-type and mutant Bmal1−/− animals indicated that transcription of genes with early phases (in phase with BMAL1 binding) depended predominantly on BMAL1, while that of delayed genes involved further regulators. Other regulators that have been implicated in the tuning of circadian expression phase include the DEC [77] and CRY [66] repressors. The finding that delayed genes tended to be upregulated in the knockout condition suggests that BMAL1 could act as a repressor either via direct [63] or indirect mechanisms [15], as has been previously proposed. While the genetic data [15] indicate that the delays reflect a primary regulation by the Rev-Erb/Ror repressor/activator pair, we showed that these genes nevertheless do have rhythmically bound BMAL1 binding sites. Moreover, the timing of mRNA expression can also be influenced by post-transcriptional mechanisms that regulate the stability of the transcripts, such as those mediated by microRNA. In fact, transcript stability affects not only the phase but also the amplitude of the mRNA accumulation. If the amplitude of the pre-mRNA is weak already, a long mRNA half-life can cause the mRNA accumulation to be practically constitutive, as exemplified by March8 mRNA levels (Figure S7A). For this reason, the fraction of cyclic mRNA transcripts among BMAL1 targets probably underestimates the fraction of functional sites, i.e., those that drive rhythmic transcription.
The large number of transcriptional regulators among putative BMAL1 targets emphasizes the pervasiveness of the circadian oscillator in liver function and shows the hierarchical control of circadian output function. Accordingly, circadian transcriptional regulators controlled by BMAL1/CLOCK can transmit their phase information to downstream targets, a model that is supported by regression analyses that predict circadian activity for several of those targets (Figure S4). These findings substantiate regulatory links that were proposed in previous computational studies aimed at reconstructing the circadian transcription regulatory network [78]–[80]. In our ontology analysis, nuclear receptors appeared as the most overrepresented annotation cluster, which may reflect their role in serving as a relay between the circadian clock and metabolic processes [27],[81], as well as in orchestrating tissue-specific circadian gene expression [30]. BMAL1 also appears to directly control specific pathways such as glucose metabolism (Gys2, Glut2, and Pck1) and triglyceride metabolism (Insig1/2 and Pnpla2). This dual, direct and indirect, regulation of circadian output function is emphasized by the presence of feed-forward loops (FFLs) [82] among targets, and might be implicated in the control of circadian expression phase. For instance, BMAL1 binds P450 oxydoreductase (Por), which was previously identified as a DBP/HLF/TEF target [5] with robust cyclic mRNA expression [7]. Similarly, BMAL1 binds both Hif1α and its known target Vegfa (Figure S5A). Interestingly, HIF1α, which we also predicted to be circadianly active (Figure S4), has been previously linked to the circadian clock as a CLOCK interacting protein [83] and in large-scale small interfering RNA perturbation experiments [47]. A number of studies have suggested that transcriptional regulation of cell cycle components by the circadian clock would lead to temporal gating of cell division [10],[11],[48]–[50]. Our data provide a number of additional links between these processes, in particular for regulators of the G1-S transition. Therefore, the circadian clock appears to not only interact with the cell cycle at G2-M [10] but might also influence entry into S phase.
In conclusion, our circadian time course ChIP analysis showed that BMAL1 binds over 2,000 sites in the mouse genome. In addition, we found highly phase-specific binding patterns, peaking at ZT6. The distribution of binding strength rapidly decays, i.e., we find at most a few dozen sites with magnitudes in the range of those found at core oscillator genes or PAR-bZip transcription factors. This strengthens the idea of BMAL1's primary function as master regulator of the circadian clock, with weaker contributions to a variety of output programs. At the genomic sequence level, strong sites also more frequently harbor highly conserved tandem E1-E2 sites, and the latter are bound cooperatively by dimers of BMAL1/CLOCK heterodimers. Genes with such elements also showed more tightly distributed phases in their mRNA expression. However, while some genes are principally regulated by BMAL1/CLOCK, other targets exhibit more complex temporal patterns in their precursor and mature RNA, hinting at contributions from further regulators. The large number of transcription factors among BMAL1 targets is reminiscent of the hierarchic organization of circadian output pathways in mouse liver. This network structure may provide flexibility in the control of tissue-specific output programs by peripheral oscillators.
Animals were housed under a 12-h light/12-h dark regimen with food and water available ad libitum. ZT0 is defined as the time when the lights are turned on. Animals were housed for 3 wk under the indicated photoperiods. The age of the animals was between 3 and 4 mo. All animal care and handling was performed according to the State of Geneva's law for animal protection.
For each time point, livers from two mice were pooled to prepare chromatin as in [24]. For the BMAL1 ChIP, a polyclonal anti-rabbit antibody to a C-terminal peptide was raised and purified using standard techniques. The specificity of the antibody for BMAL1 was ascertained using SDS-PAGE with nuclear extract from wild-type and Bmal1−/− animals (Figure S8); extracts were provided by Frédéric Gachon (University of Lausanne). Sepharose-protein A beads (GE Healthcare) were prepared according to manufacturer indications and resuspended in RIPA buffer (50 mM Tris-HCl [pH 8], 150 mM NaCl, 2 mM EDTA [pH 8], 1% Triton X-100, 0.5% sodium deoxycholate, 0.1% SDS) supplemented with Roche Complete Protease Inhibitor Cocktail. Chromatin (250 µl) was pre-cleared by incubating with 60 µl of bead suspension for 1.5 h at 4°C on the rotating wheel. Pre-cleared chromatin was then incubated with 4 µl of BMAL1 antibody for 5 h at 4°C on the rotating wheel. Bead suspension (35 µl) was added to each reaction, and incubation was continued for 3 h at 4°C on the rotating wheel. Beads were then washed three times with wash buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA [pH 8], 150 mM NaCl, 20 mM Tris-HCl [pH 8]) and once with final wash buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA [pH 8], 500 mM NaCl, 20 mM Tris-HCl [pH 8]). Co-immunoprecipitated DNA fragments were eluted from the beads in 120 µl of 1% SDS and 100 mM NaHCO3 for 15 min at 30°C and then treated with 1 µl of RNase A for 1 h at 37°C. Co-immunoprecipitated DNA fragments were incubated overnight at 65°C with Proteinase K and then purified using Qiaquick PCR Purification Kit (Qiagen). For real-time PCR quantification, the equivalent of 5 µl of chromatin of each reaction was used in a 20-µl reaction using the primers and TaqMan probes listed in Table S6, using an ABI 7900HT PCR machine (Applied Biosystems). For Illumina sequencing, two sets of libraries were prepared with independent BMAL1 ChIP time courses (library A: one time course; library B: pool of three time courses), and a total of 16 lanes were sequenced on an Illumina Genome Analyzer 2 machine. To prepare the input library, samples from the six time points were pooled at equal amounts, and one lane was sequenced.
At each time point, sequenced DNA reads from both libraries were pooled and mapped to the mouse genome (Mus musculus National Center for Biotechnology Information m37 genome assembly [mm9; July 2007]) using Bowtie [84] with three mismatches and only one hit allowed on the genome. If several reads coming from the same library mapped at the same genomic position and on the same strand (redundant tags), we considered this as a PCR duplicate and only one read was kept for the rest of the analysis. The numbers of reads per time point are shown in Table S1. To normalize for differences in sequencing depth among the time points, the number of tags per position in each BMAL1 ChIP-Seq library was rescaled by the total number of mapped tags in this library, and then for each time point, the numbers of tags in each library were summed up. The number of tags in a binding site is expressed as the number of non-redundant tags per 107 aligned tags, with the best sites in the range of 200. The list of all sites with their annotations is given in Text S2.
At each time point separately, BMAL1-bound regions were detected by MACS [36] with the following parameters: shift = 75, bandwidth = 150, genome size = 2.4 Gb, and the input chromatin sample as control data; overlapping binding regions were merged. In each region, a refined estimate of the binding location was obtained using a deconvolution algorithm that models the expected distribution of tags on the positive and negative strands (see Text S1). This was done on a single track in which all tags from all time points were merged. Local maxima in the deconvolved signal were used to call binding site positions for the rest of the analysis. The deconvolution methods also allowed us to efficiently reject spurious sites, leaving us with a total of 2,049 trustable binding sites. For each binding site, the signal in windows of ±250 bp were quantified for each time point and subjected to rythmicity analysis.
Each binding site was annotated with the Ensembl transcript having the closest TSS using the R package biomaRt. The Ensembl transcript ID was then used to retrieve further annotations such as Mouse Genome Informatics symbol, Entrez Gene ID, and Affymetrix Mouse 430 probe ID. Mouse liver RNA-Seq data from [41] were used to define the liver transcriptome (threshold was set at 1.35 reads per kilobase per million mapped reads, corresponding to the 50th percentile; see Figure S3A).
DNA sequences and placental mammals PhastCons conservation scores [85] in windows of ±50 bp around the center of each binding site were retrieved from Ensembl and the UCSC Genome Browser database, respectively. To analyze correlations in the positions of E-boxes, the sequences for all BMAL1 binding sites were scanned with a weight matrix (Figure S6A), and the resulting likelihood scores were converted to occupancies using a sigmoid transformation with threshold corresponding to one mismatch. The correlation signal was then computed on the occupancies. The HMM was trained using the sequences under all BMAL1 binding sites, weighted proportionally to the number of BMAL1 tags at peak binding, using the model architecture shown in Figure S6B. To compute the position of E1-E2 instances with respect to binding sites, we extracted weight matrices from the trained HMM with spacing ranging from 6 to 7 bp and scanned windows of ±250 bp around each binding site.
Time series expression data were from [31] using the plus-doxycyclin condition, which mimics wild-type light-dark conditions. Liver-expressed genes for these data were defined as having mean log2 (expression) over the 12 time points greater than 3.5 (Figure S3A). The 24-h Fourier component (F24) and phase were computed using established methods [86], and the p-value associated with 24-h rhythmic expression (also for cyclic binding) was computed using a Fisher test for one specific period [87] for a time series at 4-h intervals of even length N:(1)
Times series data as above [31] were combined with position-specific weight matrices (PSWMs) from the SwissRegulon database [47] to predict transcription factor activities using a regression model similar to that in [47]. Briefly, we fitted the following multi-linear model:(2)where Egt is the mean-centered expression level of the gene g at time t, Ngm is the number of predicted conserved sites for motif m, and Amt is the activity of the motif m at time t. It denotes an intercept. To compute the Ngm matrix, windows of ±2,500 bp around each Ensembl transcript annotated with an Affymetrix Mouse 430 probe ID were scored with the corresponding PSWM, and the likelihoods along the sequence were summed up [46] and weighted by a factor C0.05, where C stands for the product of the PhastCons conservation scores in that sequence. Given expression data Egt and the occupancies Ngm, the unknown activities Amt are then inferred using standard least squares regression.
To quantify pre-mRNA and mRNA levels with real-time RT-PCR, whole cell RNA was isolated according to [88]. For each time point, the extracted RNA from four livers was pooled (in each case two of the four livers were from the animals used for the chromatin preparation). For the Bmal1−/− samples at ZT6 and ZT18 (provided by Frédéric Gachon), total RNA from two livers was pooled. Pooled RNA (0.5 µg) was reverse-transcribed using random hexamers and Superscript reverse transcriptase (Invitrogen). The cDNA equivalent to 20 ng of total RNA was PCR-amplified in an ABI 7900HT PCR machine using the primer and TaqMan probes listed in Table S7. The relative levels of each RNA were calculated on the basis of 2−CT and normalized to the corresponding levels of Gapdh RNA. Each mRNA time course was normalized by its mean value, and the data shown represent the mean±standard deviation of three independent time courses.
EMSA and preparation of nuclear extracts were performed as in [37] with the following modifications. EMSA probes were prepared by dissolving forward and reverse oligonucleotides (listed in Table S8) in 100 mM NaCl, annealing them by warming them to 95°C and letting them cool down to 25°C over the course of several hours. Annealing oligonucleotides (30 µl, 25 ng/µl) were incubated with 4 µl of Klenow fill-in buffer, 2 µl of 5 mM dATP/dGTP/dTTP, 2 µl of 3,000 Ci/mmol 32-dCTP, and 2 µl of 5 U/µl Klenow fragment for 15 min at room temperature. Radiolabeled probes were then purified using Qiaquick Nucleotide Removal Kit (Qiagen) and resuspended in 15 µl of H2O. For supershift experiments, 1 µl of purified antibody was added immediately before the addition of the radioactive probe. The antibodies used were anti-BMAL1 and anti-CLOCK from [28]. Two-dimensional EMSA was performed as in [28] with the following modification: the protein-DNA complexes were separated on a 4% acrylamide gel by electrophoresis (first dimension).
293T cells were cultured in Dulbecco's Modified Eagle Medium supplemented with 10% fetal bovine serum (Invitrogen) and 1.5% streptomycin/penicillin antibiotics (Cellgro) under 5% CO2 at 37°C. Twenty-four hours after seeding at 1.5×105 cells/ml, cells were transfected using LipofectAMIN 2000 (Invitrogen). At 28 h after transfection, cells were harvested, and the luciferase activity was determined by using Dual Luciferase Reporter Assay (Promega) on a luminometer (EnVision 2104 MultiLabel Reader, PerkinElmer). Transactivation assays were performed using 1,200 ng of total DNA per well (300 ng of pDEST26-BMAL1, 300 ng of pDEST26-CLOCK, 50 ng of different pGL3-Promoter constructs [firefly luciferase], phRL-SV40 [renilla luciferase]) and a total of 1,200 ng of pDEST26-LACZ plasmids. Different E-box motifs were inserted upstream of the SV40 promoter of pGL3-Promoter vector (Promega) by using annealed primers (Table S9) and ligated into KpnI-XhoI sites.
Illumina sequencing data for the BMAL1 ChIP are available at Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/), accession number GSE26602. Processed BigWig files that can be visualized on the UCSC Genome Browser as a custom track to generate graphs such as Figure 1D and 1E are available at http://circaclock.epfl.ch. The fully annotated (including binding strength) 2,049 sites are provided in Text S2.
|
10.1371/journal.pntd.0001652 | Multispacer Sequence Typing Relapsing Fever Borreliae in Africa | In Africa, relapsing fevers are neglected arthropod-borne infections caused by closely related Borrelia species. They cause mild to deadly undifferentiated fever particularly severe in pregnant women. Lack of a tool to genotype these Borrelia organisms limits knowledge regarding their reservoirs and their epidemiology.
Genome sequence analysis of Borrelia crocidurae, Borrelia duttonii and Borrelia recurrentis yielded 5 intergenic spacers scattered between 10 chromosomal genes that were incorporated into a multispacer sequence typing (MST) approach. Sequencing these spacers directly from human blood specimens previously found to be infected by B. recurrentis (30 specimens), B. duttonii (17 specimens) and B. crocidurae (13 specimens) resolved these 60 strains and the 3 type strains into 13 species-specific spacer types in the presence of negative controls. B. crocidurae comprised of 8 spacer types, B. duttonii of 3 spacer types and B. recurrentis of 2 spacer types.
Phylogenetic analyses of MST data suggested that B. duttonii, B. crocidurae and B. recurrentis are variants of a unique ancestral Borrelia species. MST proved to be a suitable approach for identifying and genotyping relapsing fever borreliae in Africa. It could be applied to both vectors and clinical specimens.
| In Africa, relapsing fevers are caused by four cultured species: Borrelia crocidurae, Borrelia duttonii, Borrelia hispanica and Borrelia recurrentis. These borreliae are transmitted by the bite of Ornithodoros soft ticks except for B. recurrentis which is transmitted by louse Pediculus humanus. They cause potentially undifferentiated fever infection and co-infection with malaria could also occur. The exact prevalence of each Borrelia is unknown and overlaps between B. duttonii and B. crocidurae have been reported. The lack of tools for genotyping these borreliae limits knowledge concerning their epidemiology. We developed multispacer sequence typing (MST) and applied it to blood specimens infected by B. recurrentis (30 specimens), B. duttonii (18 specimens) and B. crocidurae (13 specimens), delineating these 60 strains and the 3 type strains into 13 species-specific spacer types. B. crocidurae strains were classified into 8 spacer types, B. duttonii into 3 spacer types and B. recurrentis into 2 spacer types. These findings provide the proof-of-concept that that MST is a reliable tool for identification and genotyping relapsing fever borreliae in Africa.
| In Africa, relapsing fevers (RF) are arthropod-borne diseases caused by four cultured species Borrelia crocidurae, Borrelia duttonii, Borrelia hispanica and Borrelia recurrentis [1]. Transmission is by the bite of Ornithodoros soft ticks for the first three species whereas Pediculus humanus louse feces transmit B. recurrentis [2], [3]. In Tanzania, molecular investigations of human and tick specimens further provided evidences for two additional, yet uncultured Borrelia species [1], [4]. Each one of the four cultured Borrelia species is more prevalent in one geographical area of Africa with B. hispanica being reported in Morocco [5], B. crocidurae in Senegal [6], B. duttonii in Tanzania [7] and B. recurrentis in Ethiopia [8]. However, the precise area of distribution of each Borrelia is unknown and may overlap as both B. duttonii and B. crocidurae have been reported in Togo and Tanzania [1], [9].
In these regions of Africa, RF was reported to be the most prevalent bacterial disease, accounting for 8.8% of febrile patients in Togo [9]. In Senegal, average incidence is 11 per 100 person-years [10]. The main clinical symptom of infection is recurrent undifferentiated fever associated with high bacteremia; RF are therefore often diagnosed as malaria and cases of malaria co-infection with have been reported [9], [11], [12]. RF are treatable by antibiotics. Severity ranges from asymptomatic to fatal, particularly if left untreated and can be associated with significant pregnancy loss or peri-natal mortality [13], [14], [15].
The African RF Borrelia are very closely related species as illustrated by 16S rRNA gene sequence variability ≤1% [2]. Accordingly, a previous comparison of B. duttonii and B. recurrentis genomes indicated that the two organisms formed a unique bacterial species [16]. Such a close genetic and genomic proximity challenged the development of laboratory tools for the accurate discrimination between the African RF Borrelia and genotyping [16]. Sequencing the 16S rRNA and the flagellin genes is unsatisfactory since African RF Borrelia differ by only one base in the flagllin gene sequence and have 16S rRNA gene sequence similarity above 99% [17]. Analysis of the intrergenic spacer (IGS) located between the 16S and 23S rRNA genes only explored the variability between B. duttonii and B. recurrentis [1]. Moreover, IGS sequence overlapped between one B. duttonii phylogenetic group and one B. recurrentis group [1] with a second overlap disclosed with subsequent analyses of further material [7].
We previously observed that multispacer sequence typing (MST), a PCR-sequencing-based method for bacteria genotyping, was efficient in typing otherwise homogenous bacterial species such as the plague agent Yersinia pestis [18] and the typhus agent Rickettsia prowazekii [19]. Ongoing study of the B. crocidurae genome in our laboratory gave us the opportunity to develop MST for African RF Borrelia and to deliver the proof-of-concept that MST is a suitable method for both the species identification and genotyping of RF Borrelia in Africa.
B. crocidurae Achema strain, B. recurrentis A1 strain and B. duttonii Ly strain were grown in BSK-H medium (Sigma, Saint Quentin Fallavier, France) supplemented with heat-inactivated 10% rabbit serum (Eurobio, Courtaboeuf, France). B. recurrentis DNA was extracted from 21 blood specimens collected in 1994 in Addis Ababa, Ethiopia Dr. S. J. Cutler (School of Health, Sports and Bioscience, University of East London, London UK). Likewise, B. recurrentis DNA extracted from 9 blood specimens collected in 2011 in Bahir Dah, Highlands of Ethiopia were provided by SC Barker (Parasitology section, School of Chemistry and Molecular Bioscience, University of Queensland, Brisbane, Australia) and KD Bilcha and J Ali (University of Gondar, Ethiopia). In addition, B. duttonii DNA extracted from 17 blood specimens collected in Mvumi, Tanzania were also provided by Dr. S. J. Cutler. B. crocidurae DNA was extracted from 13 blood specimens collected in 2010 in Senegal by C. Sokhna (URMITE, Dakar, Senegal) including 11 specimens from Dielmo and 2 specimens from Ndiop. DNA was extracted from these specimens using QIAamp DNA Blood mini kits (QIAGEN, Hilden, Germany) according to the manufacturer's instructions.
The B. crocidurae genome (Genbank accession number CP003426–CP003465) has been sequenced and annotated in our laboratory using pyrosequencing technology on a Roche 454 GS FLX sequencer. The draft genome is comprising of one closed chromosome and scaffolds representing the plasmids. Spacer sequences extracted from B. crocidurae strain Achema, B. recurrentis strain A1 (Genbank accession number CP000993) and B. duttonii strain Ly (Genbank accession number CP000976) genomes using perl script software were compared using ssaha2 software [20]. Spacers were pre-selected for a 300 to 800-bp length. Pre-selected spacers were further analyzed for sequence similarity in order to exclude spacers with <0.1% interspecies sequence similarity. PCR primers were then designed using primer3 software (http://fokker.wi.mit.edu) in order to amplify the entire sequence of each of the selected spacers.
Five microliters of Borrelia DNA and 10 pmol of each primer (Eurogentec, Seraing, Belgium) were added to the PCR mixture, containing 0.4 U Phusion DNA Polymerase (Finnzymes, Espoo, Finland), 4 µl of 5× Phusion HF Buffer (Finnzymes) and 0.4 µl of 10 mM dNTPs. The volume was adjusted to 24 µL by adding distilled water. Thermal cycling was performed on a 2720 DNA thermal cycler (Applied Biosystems, Courtaboeuf, France) with an initial 30-sec cycle at 98°C followed by 35 cycles consisting of 10 seconds at 98°C, 30 seconds at 58°C and 1 minute at 72°C, followed by a 10-min final extension step at 72°C. To rule out amplicon carry-over, nucleotide-free water negative control was used throughout the steps of the protocol. PCR products were purified prior to sequencing by using the Nucleo-Fast 96 PCR Kit (Macherey-Nagel, Hoerdt, France). Three microliters of the resulting DNA were added to each primer mixture comprised of 10 pmol of each primer, 4 µL water and 3 µL BigDye Terminator reaction mix (Applied Biosystems). Sequencing thermal cycling was performed on a Applied Biosystems DNA thermal cycler with an initial 5-min cycle at 96°C followed by 25 cycles consisting of 30 seconds at 96°C, 20 seconds at 55°C, and 4 minutes at 60°C, followed by a 7-min final extension step at 15°C. Sequencing products were purified using sephadex plates (Sigma-Aldrich, Saint Quentin Fallavier, France) and sequencing electrophoresis was performed on a 3130 Genetic Analyzer (Applied Biosystems).
The nucleotide sequences were edited using ChromasPro software (www.technelysium.com.au/chromas.html). Similarities between spacers were determined after multiple alignments using the MULTALIN software [21]. MST discrimination power was calculated using the Hunter-Gaston Index [22]:where D is the numerical index of discrimination, N is the total number of isolates in the sample population, s is the total number of different types, and nj is the number of isolates belonging to the jth type.
The five spacer sequences analyzed herein were concatenated and neighbor-joining phylogenetic tree was reconstructed using the maximum likelihood method in PhyML 3.0 [23]. Each particular sequence of a given spacer was assigned to a spacer type (ST) number.
This study was approved by the IFR48 Ethic Committee. All patients provided informed written consent.
Chromosome sequence alignment of the three Borrelia reference genomes studied herein revealed that 23 intergenic spacers that were common to all three species. Of these, five spacers fulfilled our selection criteria and were named MST2, MST3, MST5, MST6 and MST7. Use of the PCR primers listed in Table 1 to amplify each of the five spacers produced amplicons ranged from 333-bp to 738-bp and sequence reads ranging from 246-bp to 543-bp (Table 1).
Pairwise comparison of the five spacers (Table 2 and figure 1) revealed they had species-specific sequence with interspecies sequence differences relying on single nucleotide polymorphism in 36 (90%) cases, deletion in 3 (7.5%) cases and insertion in 1 (2.5%) case. Comparing B. duttonii MST7 with B. crocidurae and B. recurrentis MST7 yielded 93% and 97% similarity, respectively, whilst comparing B. crocidurae MST7 with B. recurrentis MST7 showed 93% similarity. The other four spacers yielded pairwise sequence similarity of 97–99% (Table 2). Sequences for each allele of each spacer have been deposited in GenBank under accession number (JQ398815: JQ398841) as well as in our local data base (http://www.ifr48.com).
While the concatenation of the five spacers yielded a discrimination index of 0.825,1, this index was of 0.7814 for MST2, 0.6896 for MST6, 0.6749 for MST5, 0.6623 for MST7, and 0.6579 for MST3. Concatenation of the five spacers yielded 8 STs named ST6–ST13 for the 13 B. crocidurae samples and the B. crocidurae Achema type strain (Table 3; Figure 2). 3 STs named ST1–ST3 for the 18 B. duttonii samples and the B. duttonii Ly type strain and 2 STs named ST4–ST5 for the 30 B. recurrentis samples and the B. recurrentis A1 type strain. MST2 sequencing classified latter samples into ST-4 (11 samples) and ST-5 (19 samples) due to the insertion of a G at position 190. The genotype ST-5 represented 47.6% (10 out 21 samples) detected in 1994 and all the nine samples detected in 2011.
The phylogenetic tree constructed after concatenation of the five intergenic spacer sequences separated the RF Borrelia into three clades, each clade containing only one Borrelia species (figure 2). A first clade comprised of all the 30 B. recurrentis isolates; a second clade comprised of three groups representing the three B. duttonii spacer types and a last clade comprised of 7 B. crocidurae spacer types.
PCR-derived data reported herein were interpreted as authentic as the negative controls used in every PCR-based experiment remained negative, all the PCR products were sequenced and experiments yielded reproducible sequences. We therefore established the proof-of-concept that MST could be used for species identification and genotyping of 3 out of 4 cultured RF borreliae (B. hispanica was not available for this study) in Africa. MST combines the sensitivity of PCR with unambiguous, portable data yielded by sequencing. Indeed, all the sequences determined are freely available in GenBank and in our local database website at ifr48.com. Therefore, any laboratory with a capacity in PCR-sequencing could easily confirm and compare their data with that reported herein to further increase the knowledge of RF Borrelia species and genotypes circulating in African countries.
In the present study, five intergenic spacers were selected from the alignment of B. crocidurae, B. duttonii and B. recurrentis reference genomes, representing approximately ∼0.2% of the total genome length. The spacers were scattered across the chromosome thus representative of the whole genome. Such a multi-target approach offers distinct advantages over the one single locus methods previously used, such as the 16S–23S IGS for typing that may be less representative of the whole genome. Based on this spacer sequencing, a total of 61 RF strains could be separated into 12 STs. Interestingly, we observed that isolates grouped into three clades corresponding to the three Borrelia organisms under study. Indeed, MST yielded no overlap between B. duttonii and B. recurrentis organisms contrary to that observed when using IGS typing [1], [7]. We observed that sequencing MST7 spacer alone accurately discriminated between B. duttonii and B. recurrentis with 3% sequence divergence, a result not previously achieved. Therefore, sequencing MST7 spacer alone could be used for the molecular identification of RF Borrelia in Africa at the species level, but not for genotyping which requires sequencing the four other spacers in addition to MST7.
Further analysis indicated that each one of the three Borrelia species under study was comprised of several spacer-types. B. recurrentis was the least diverse Borrelia comprising of only two very closely related groups. This finding supports the previous genomic analysis that concluded that B. recurrentis was a subset of B. duttonii [16]. In our study also, there was an inverse correlation between the RF Borrelia MST diversity and the reported mortality rate for these RF Borrelia [8], [15].
Despite the fact that we tested a small set of B. crocidurae, nevertheless we found a high diversity index in this species since 13 B. crocidurae samples collected in Senegal yielded 7 MST types and the B. crocidurae Achema type strain collected in Mauritania yielded an additional MST type. This first genotyping method for B. crocidurae is therefore very promising to probe its geographic repartition as well as potential association of B. crocidurae genotypes with vectors. Indeed, four genogroups could be identified in O. sonrai ticks collected in Senegal and Mauritania [24]. In this study, B. crocidurae flagellin sequence was found identical among the four O. sonrai tick groups but the B. crocidurae infection rate significantly differed among the four tick groups; MST may help studying such discrepancy and may reveal previously unknown relationships between B. crocidurae genotypes and O. sonrai genotypes. Moreover, a recent study indicated that B. crocidurae may be transmitted by soft tick Ornithodoros erraticus in Tunisia, challenging O. sonrai as the only B. crocidurae vector in West Africa [25]. MST is new laboratory tool to question whether the unexpected higher diversity in B. crocidurae than in B. duttonii and B. recurrentis is linked to a more complex cycle involving several mammals and ticks species.
Present data indicate that MST is offering a new sequencing-based technique for further exploring the identification and genotypes of RF Borrelia in vectors and clinical specimens collected in Africa.
|
10.1371/journal.pntd.0007322 | Geographic shifts in Aedes aegypti habitat suitability in Ecuador using larval surveillance data and ecological niche modeling: Implications of climate change for public health vector control | Arboviral disease transmission by Aedes mosquitoes poses a major challenge to public health systems in Ecuador, where constraints on health services and resource allocation call for spatially informed management decisions. Employing a unique dataset of larval occurrence records provided by the Ecuadorian Ministry of Health, we used ecological niche models (ENMs) to estimate the current geographic distribution of Aedes aegypti in Ecuador, using mosquito presence as a proxy for risk of disease transmission. ENMs built with the Genetic Algorithm for Rule-Set Production (GARP) algorithm and a suite of environmental variables were assessed for agreement and accuracy. The top model of larval mosquito presence was projected to the year 2050 under various combinations of greenhouse gas emissions scenarios and models of climate change. Under current climatic conditions, larval mosquitoes were not predicted in areas of high elevation in Ecuador, such as the Andes mountain range, as well as the eastern portion of the Amazon basin. However, all models projected to scenarios of future climate change demonstrated potential shifts in mosquito distribution, wherein range contractions were seen throughout most of eastern Ecuador, and areas of transitional elevation became suitable for mosquito presence. Encroachment of Ae. aegypti into mountainous terrain was estimated to affect up to 4,215 km2 under the most extreme scenario of climate change, an area which would put over 12,000 people currently living in transitional areas at risk. This distributional shift into communities at higher elevations indicates an area of concern for public health agencies, as targeted interventions may be needed to protect vulnerable populations with limited prior exposure to mosquito-borne diseases. Ultimately, the results of this study serve as a tool for informing public health policy and mosquito abatement strategies in Ecuador.
| The yellow fever mosquito (Aedes aegypti) is a medically important vector of arboviral diseases in Ecuador, such as dengue fever and chikungunya. Managing Ae. aegypti is a challenge to public health agencies in Latin America, where the use of limited resources must be planned in an efficient, targeted manner. The spatial distribution of Ae. aegypti can be used as a proxy for risk of disease exposure, guiding policy formation and decision-making. We used ecological niche models in this study to predict the range of Ae. aegypti in Ecuador, based on agency larval mosquito surveillance records and layers of environmental predictors (e.g. climate, elevation, and human population). The best models of current range were then projected to the year 2050 under a variety of greenhouse gas emissions scenarios and models of climate change. All modeled future scenarios predicted shifts in the range of Ae. aegypti, allowing us to assess human populations that may be at risk of becoming exposed to Aedes vectored diseases. As climate changes, we predict that communities living in areas of transitional elevation along the Andes mountain range are vulnerable to the expansion of Ae. aegypti.
| Mosquito-borne disease transmission poses an ongoing challenge to global public health. This is especially true in much of Latin America, where arboviral disease management is complicated by the proliferation of mosquito vectors in tropical conditions, frequently coupled with limited resources for medical care and comprehensive vector control services [1]. In Ecuador, the yellow fever mosquito (Aedes aegypti) is of particular medical importance as it is a competent vector for several established and emerging viral diseases, including all four serotypes of dengue virus (DENV), chikungunya (CHKV), Zika virus (ZKV), and yellow fever virus (YFV) [2–5]. The Ae. albopictus mosquito, also a competent vector of arboviruses, was recently reported for the first time in the city of Guayaquil, Ecuador [6]. Mosquito-borne diseases caused by arboviruses transmitted by Aedes spp. have no treatment beyond palliative care, and with the exception of yellow fever, there are no clinically established vaccines [7–9]. As a result, mosquito surveillance and control remain the best tools available for preventing and managing outbreaks of arboviral disease.
In Ecuador, the Ministry of Health, or Ministerio de Salud Pública (MSP), oversees public health vector control services in the country, including mosquito surveillance, indoor residual spraying, larvicide application, and ultra-low volume (ULV) fogging. The MSP conducts larval index (LI) surveys at the household level, wherein containers of water are sampled for mosquito larvae. Larval indices (e.g. household, container, and Breteau) are among the most common indicators used by public health agencies to establish mosquito presence and quantify abundance, which are key considerations for understanding localized transmission potential and planning larval source reduction [10]. Although cost effective relative to the delivery of clinical services, mosquito abatement and surveillance activities are nevertheless limited by financial constraints, necessitating informed strategies for focusing resources and personnel [11,12]. This becomes a critical factor when developing surveillance and control programs on very large scales, such as an entire country, where misdirection of program activities can rapidly deplete program funding. Advancing the understanding of where vectors of interest can occur on the landscape would provide valuable guidance in communicating risk of exposure and avoiding the pitfalls associated with indiscriminately rolling out interventions.
Like many mosquito species, the presence of Aedes spp. on the landscape is closely tied to environmental conditions [13–15]. Adult survival and larval development are largely driven and restricted by temperature, while successful oviposition and larval emergence rely on the persistence of standing water in the environment [16–21]. In contrast with other medically important mosquito species in the region, such as Anopheline vectors of malaria, Ae. aegypti typically does very well in heavily urbanized environments, largely due to their reproductive strategy of exploiting small volumes of water in manmade containers around the home as larval habitats [22]. In landscapes with heterogeneous topography, elevation also serves as a limiting factor for mosquito distributions, as temperature and precipitation change with elevation [23,24]. Situated in northwestern South America, Ecuador exemplifies high landscape diversity, with hot, humid areas of low elevation along the Pacific coast in the west and interior Amazon basin in the east, and the cool, arid Andes mountain range in the central portion of the country (Fig 1). Historically, the western coastal and interior regions experience a much higher incidence in mosquito-borne diseases than mountainous areas, where sharp increases in elevation and decreases in temperature limit the geographic distribution and vectorial capacity of the mosquito vector.
The present-day distribution of Ae. aegypti is broadly defined by regional temperature and precipitation trends, but global climate change has the potential to significantly alter the future geographic range of mosquito vectors [3]. The Intergovernmental Panel on Climate Change has established four representative concentration pathways (RCP), or different scenarios for future greenhouse gas emissions, which are the basis for modeling future climates. Even under the most conservative of these scenarios (RCP 2.6), mean global temperatures are projected to increase [25]. As temperature trends increase globally, it has been estimated that observed patterns in the distribution of mosquito vectors will shift accordingly; previous studies have projected that Aedes mosquitoes will increase their global range as temperature and rainfall patterns become more suitable under various climate change scenarios [18,26–28]. Modeling and visualizing changes in mosquito distributions at the national level will provide a useful tool for managing disease and planning the delivery of health services, as public health resources can be better allocated in anticipation of disease emergence in naïve populations driven by mosquito range expansions.
Ecological niche models (ENMs) have been used to estimate current potential distributions in insect populations, including mosquitoes, as well as range expansions resulting from environmental and climate changes [29–32]. Ecological niche modeling methodologies have been applied to many systems, spanning regional to global scales, in an effort to estimate Ae. aegypti distribution and the associated risk of exposure to humans, often indicating that water availability and land cover factor heavily into overall mosquito habitat suitability [3,29,33,34]. In Ecuador and other areas of Latin America, elevation also becomes a limiting factor for Ae. aegypti presence, though it is suggested that climate change may allow for the encroachment of mosquitoes into higher elevations [32,35]. While many prior studies have utilized records of adult stages of mosquitoes for ENMs, this study leverages existing larval surveillance data collected in Ecuador as an indicator of species presence, providing a predictive tool about the source of mosquitoes in the environment. This complements predictive models for adult stages, particularly in considering potential for intervention, as it can target larvicidal approaches, rather than reactive adulticidal spraying methods. The Genetic Algorithm for Rule-Set Production (GARP) is a machine-learning algorithm that builds species ENMs using presence-only occurrence records and continuous environmental variables [36]. The genetic algorithm (GA) employed by GARP to build rule-sets for distribution models is stochastic in nature, resulting in a set of models from a single dataset of species occurrence records and allowing for the assessment of agreement between resulting models. This methodology offers a robust option for modeling the potential distribution of species on a landscape from presence-only records, as absence of a species is difficult to discern through historical records and field sampling (e.g. entomological surveys) [36,37]. GARP also provides a platform for projecting future climate scenarios onto the landscape with the natively generated rule-sets for species distribution prediction, allowing for the estimation of future geographic distributions [38].
Assessing current and future vector distributions in an ENM framework is useful for defining the spatial distribution and possible changes in risk exposure, using mosquito presence as a proxy for transmission risk. Previous work in Ecuador’s southern coast has focused on describing interannual variation in dengue transmission for a single region [39,40]. In this study, we contribute to the available climate-informed tools used by the public health sector in Ecuador to assist decision-making, examining potential geographic shifts in risk at broader spatial and temporal scales. We had three main objectives 1) use an ENM approach to estimate the current geographic range of Ae. aegypti in Ecuador using a unique set of larval survey data; 2) use projected climate data to model the future geographic range under a variety of climate change scenarios; and 3) compare current and future climate models to describe changes in Ae. aegypti range over time, where we hypothesized that larval Ae. aegypti distribution in Ecuador would expand into areas of higher elevation with projected increases in global temperature.
Presence only data on the occurrence of Ae. aegypti in Ecuador were made available for this study by the MSP. From 2000–2012 the MSP sampled aquatic larval mosquitoes from standing water in and around households in cities and towns throughout mainland Ecuador following standard protocols for entomological surveillance recommended by the World Health Organization [41]. These data were collected year-round by vector control technicians from the National Service for the Control of Vector-Borne Diseases (SNEM) of the MSP as part of routine vector surveillance activities. Upon entering households, technicians visually inspected all potential larval habitat sites inside and outside of the home. Live samples of juvenile mosquitoes from positive containers were collected and transported to local vector control offices, where laboratory technicians confirmed species identification. Although the possibility for confusion with Ae. albopictus exists, this species was only recently detected in Ecuador [6]. Originally used by the MSP as an indicator of mosquito abundance around households, positive LI records for Ae. aegypti were used in this study to indicate the presence of mosquitoes at a given location. These occurrence data were de-identified from households and aggregated to the administrative level of parroquia (township or parish) by the MSP for each year of the study. Figures were produced in ArcMap (ver. 10.4, ESRI, Redlands, CA) using shapefiles from the GADM database of Global Administrative Areas, ver. 2.8 (gadm.org), elevation data freely available from NASA’s Shuttle Radar Topography Mission (jpl.nasa.gov/srtm), georeferenced mosquito surveillance data provided by the MSP and edited by the authors for this project, and ENM output produced in the course of this study.
Parroquias (n = 991) represented in this data set range in size from roughly 2 km2 to over 8,000 km2. Therefore, we felt it prudent to reduce this high spatial variation prior to analyses. To correct for this extreme variation in the spatial resolution of aggregated presence data, the number of positive LI locations in a given parroquia were reassigned from the centroid of the administrative boundary to the centroids of cities, barrios (neighborhoods), and villages where MSP mosquito surveillance was conducted, ≤ 5km in urban extent. Human settlements were identified via a combination of OpenStreetMap (http://openstreetmap.org) and Google Earth (http://earth.google.com) satellite imagery in ArcMap. While satellite images were used to identify population dense areas, guiding disaggregation of LI data, this imagery was not used in mapping or creation of figures. Given the potential for uncertainty, a conservative approach to disaggregation was taken, where occurrence records were not included in the final dataset in cases of spatial ambiguity (e.g. cities larger than 5km in extent with a single occurrence record, multiple developments in an administrative unit exceeding the number of surveys conducted, etc). This method of informed disaggregation allowed for better spatial representation and improved model performance compared to ENMs built with aggregated data, without compromising de-identification (S1 Table).
Environmental coverage datasets for current climatic conditions, comprised of rasterized elevation and 19 derived biophysical variables (Bioclim), were compiled using publicly available interpolated weather station data (WorldClim ver. 1.4., http://worldclim.org) (Table 1) [42]. WorldClim provides long-term climate averages based on weather station records from 1950–2000, a period coinciding with the start of the MSP’s larval survey. Although more contemporary long-term averages of interpolated climate are available, these datasets have yet to incorporate models of future climate conditions into publicly available products. Because Ae. aegypti is primarily considered an urban vector in close association with human development, gridded human population density, adjusted to data from the United Nations World Population Prospects 2015 Revision, was also included as an environmental predictor for initial model building as a proxy for built land covers (Socioeconomic Data and Applications Center (SEDAC) Gridded Population of the World (GPW)) [43,44]. A resolution of 2.5 arc-minutes (i.e. 5km grid cells) was chosen for all raster layers to reflect variability in the resolution of geolocated data.
Environmental coverages for estimated future climatic conditions in the year 2050 were taken from forecasted Bioclim variables, allowing for direct comparison between current and future predicted ranges. We chose three general circulation models (GCMs) of physical climate processes commonly used in projecting shifts in species distributions, the Beijing Climate Center Climate System Model (BCC-CSM-1), National Center for Atmospheric Research Community Climate System Model (CCSM4), and the Hadley Centre Global Environment Model version 2, Earth-System configuration (HADGEM2-ES) under the four standard emissions scenarios (RCP 2.6, RCP 4.5, RCP 6.0, RCP 8.5) [25,45–49]. Gridded human population data available through SEDAC are only projected through the year 2020 [43]. To obtain human population for the year 2050, a simple linear extrapolation, wherein we assume a stable rate of growth, was performed on a pixel-by-pixel basis in ArcMap with available years of SEDAC data. Although a rudimentary means of estimating human population growth, the resulting trend mirrors more sophisticated cohort-based population estimates for Ecuador projected for the same time period [50,51].
Ecological niche models reflecting current and future climate conditions were built using DesktopGARP ver. 1.1.3 (DG) [37]. Although more contemporary methods of building ENMs are available, GARP was chosen for this study because of its demonstrated ability to produce models that are transferable to novel time periods [52]. Furthermore, while other methods of estimating species distributions are known to overfit geographic models to training data, an issue which could exacerbate any spurious errors in our disaggregated occurrence data, GARP has been shown in other studies to exclude a degree of outlier data from geographic predictions [53,54]. LI point records and environmental coverage datasets were prepared for modeling using the ‘GARPTools’ package (co-developed by C.G. Haase and J.K. Blackburn) in the program R (ver. 3.3.1). Spatially unique LI records (n = 478) were split into 75% training (n = 358) and 25% testing datasets (n = 119) for ten randomly selected iterations; training datasets were used in model building and testing datasets were used to compute model accuracy metrics [36,37,55,56]. Ten experiments were run in DG, each using one of the randomly selected LI training datasets and the full set of current environmental coverage variables (Table 2). Each experiment was run for 200 models, allowing for a maximum of 1,000 iterations with a convergence limit of 0.01. Occurrence training data were internally partitioned in DG into 75% training/25% testing for model building and subset selection, and top models were selected using the ‘Best Subsets’ option, specifying a 10% hard omission threshold and 50% commission threshold [57]. The top ten best subsets models from each GARP experiment were summated with the GARPTools package to assess model agreement and accuracy. Model accuracy metrics for each GARP experiment were calculated from the 25% testing dataset withheld from the model building process. Three standard measures of accuracy, calculated in GARPTools, were used to compare best subsets from each experiment: receiver operator characteristic (ROC) curve with area under the curve (AUC), commission (i.e. false positives), and omission (i.e. false negatives) [58]. The AUC is an indicator of a model’s ability to predict areas of species presence versus absence, with an AUC of 0.5 indicating a model that performs no better than random, and an AUC of 1.0 indicating a perfect model [58]. We additionally performed partial ROC (pROC) analyses for model accuracy, a method which addresses some of the limitations identified in the classic ROC approach [59]. Partial ROC analyses were performed with Niche Toolbox (ver. 0.2.5.4), specifying an omission threshold of E = 10 and 1000 bootstrap replicates, where resulting AUC ratios >1 indicate that model predictions are significantly better than random (p < 0.01) [59,60].
The model building process was then repeated in DG with the best performing training dataset (i.e. high AUC relative to low omission) to compare full model performance with more parsimonious sets of environmental variables. In addition to variable combinations selected based on previous literature, the GARPTools package was used to extract ruleset trends from the full model (e.g. prevalence and importance of given variables in the resulting model) to assemble additional candidate variable sets for model comparison. The subset of models with the highest AUC and lowest omission (i.e. best model) was chosen as the most probable estimate of current larval mosquito geographic distribution, and rulesets generated from the best model were then projected to the year 2050 for all combinations of GCMs and RCPs. To compare the relative changes in geographic predictions between current and future climate scenarios, the best subsets of current and projected future models for each RCP scenario were recoded as binary geographic distributions (i.e. presence and absence) in ArcMap, where cells with model agreement of ≥ 6 were considered present. Recoded distributions were combined using the ‘Raster Calculator’ tool in the Spatial Analyst extension of the program ArcMap, allowing for the visualization of range agreement across GCMs. The number of people at risk in areas of expanding mosquito distribution, where range expansion was predicted under at least one GCM, was estimated in ArcMap, using the Raster Calculator tool to extract information on GPW and extrapolated population for the year 2050.
The original dataset of LI occurrences in Ecuador, provided by the MSP, consisted of 3,655 collection events aggregated to 374 parroquia centroids, indicating the number of parroquias that had positive surveillance results for Ae. aegypti larvae during the study period. Disaggregation of these data yielded 478 spatially unique locations within these parroquias, corresponding with areas of human habitation regularly surveyed by the MSP. Incorporating prior knowledge regarding the agency’s collection of data in developed areas allowed for the adoption of a finer spatial scale for analysis without changing the overall distribution of larval mosquito presence in Ecuador (e.g. mosquitoes remained conspicuously absent in most high-elevation parroquias located in the Andes mountains).
Much of Ecuador was predicted to be suitable for the presence of Ae. aegypti larvae under current climatic conditions, with the notable exception of the eastern portion of the country associated with the Amazon basin and high elevation areas associated with the Andes mountain range, running north to south through the center of the country (Fig 2). This iteration of model subsets generated by GARP had the highest AUC, relative to low omission (AUC = 0.73, Avg. Commission = 63.47, Avg. Omission = 3.02), and was built with a reduced set of environmental variables including elevation, human population, maximum temperature of the warmest month, annual temperature range, mean temperature of the wettest month, mean temperature of the driest month, mean temperature of the warmest quarter, mean temperature of the coldest quarter, precipitation of the wettest month, precipitation seasonality, precipitation of the driest quarter, and precipitation of the coldest quarter (Table 3).
The projected geographic distribution of larval Ae. aegypti for the year 2050 (Figs 3B–3D, 3F–3H and S1 and S2), built with the best-performing selection of environmental coverages under four climate change scenarios, showed marked changes in pattern when compared with estimated mosquito presence under current conditions (Figs 3A and 3E and S1 and S2). Potential distributional shifts were generally consistent across GCMs, with slight range expansions into areas of higher elevation and large portions of the eastern Amazonian basin predicting mosquito absence (Figs 3 and S1 and S2). Combining the current and future model agreement rasters for best subset models by RCP revealed predicted areas of geographic stability in western Ecuador and the eastern foothills of the Andes, range contraction throughout much of Amazon basin in the east, and range expansions along transitional elevation boundaries over time (Fig 4). Range expansions and contractions were generally consistent across climate models, with the magnitude of distribution change increasing with more extreme climate change scenarios (Fig 4). Similarly, the human population with the potential to experience increased exposure to mosquito presence generally increases with RCP, with an additional 9,473 (RCP2.6), 11,155 (RCP4.5), 10,492 (RCP6.0), and 12,939 (RCP8.5) people currently living in areas of transitional elevation estimated at risk of becoming exposed under different climate change scenarios (Table 4).
The predicted current geographic distribution of Ae. aegypti suitability in Ecuador, under current climate conditions, largely reflects present-day risk maps for many of the mosquito-borne diseases currently circulating in the country, wherein populations living at high altitudes are not considered at-risk for transmission [61]. Predicted larval distributions are roughly continuous in the eastern and western portions of Ecuador, but are sharply restricted along increasing elevation gradients in the central portion of the country, the area corresponding with the location of the Andes mountain range (Fig 2) [9]. This conspicuous absence of mosquitoes in the Andes reflects the generally protective nature of high mountain elevations from mosquito presence, with all models predicting larval mosquito absence throughout central Ecuador (Figs 2–4 and S1 and S2). The predicted low habitat suitability for Ae. aegypti in the eastern portion of the Amazon basin is notable, as this is a region currently perceived as potentially higher risk for mosquito exposure by public health officials relative to mountainous regions, mostly owing to its low elevation, despite having generally low human population density (Fig 2). Although similar in elevation to regions of active disease transmission in the West, the hydrology and seasonal temperature patterns of the Amazon basin differ considerably from coastal areas. Previous work in this region suggests a great deal of spatial variability in the basin with regards to climate patterns, which drive differences in biodiversity [62–64]. Given that the mosquito life cycle depends heavily on the availability of water in the environment, spatial discrepancies in precipitation could account for the low model agreement of mosquito habitat suitability in the easternmost portion of the Amazon.
Range expansion of Ae. aegypti into higher elevations as a result of changing climate was supported across GCM models and emissions scenarios (Figs 3–4 and S1 and S2). All best model subsets suggest that areas of transitional elevation along the eastern and western peripheries of the Andes mountains may experience some level of increased exposure to the presence of mosquitoes, though much of the mountain range, including densely populated areas like the capital city, Quito, will remain unsuitable habitat. The intrusion of Ae. aegypti into areas of transitioning elevation represents a potential area of concern for public health managers, as communities in these areas are largely protected from mosquito exposure and associated diseases under current climatic conditions. Excluding travel-related cases, reporting of arboviral diseases in Ecuador’s mountain dwelling populations is quite low, although there are low-lying valleys near Quito that may be suitable for arbovirus transmission. Accordingly, the MSP primarily directs mosquito-borne disease outreach and intervention efforts to high-risk communities, particularly in large coastal cities with consistently high disease incidence, such as Guayaquil and Machala. As a result, communities situated in the foothills of the Andes will not necessarily have the same baseline risk perceptions and preventative behaviors as those communities burdened with historically high incidence of mosquito-borne diseases. This sets the stage for potential disparities in preventative knowledge and health services should Aedes mosquitoes expand into naïve populations [5,65].
Models projected to future climate scenarios predict the extirpation of Ae. aegypti in several areas of Ecuador, with a particularly large range contraction in the Amazon basin shown across scenarios. This finding is consistent with other studies on potential geographic shifts in arthropod vectors in response to climate change, which demonstrate that increasing temperatures do not necessarily lead to net increases in geographic disease risk, but rather shifts in distribution as high temperatures decrease habitat suitability [32,66]. While our models do not account for the possibility of vector populations adapting to changing climate, evidence suggests that ectotherms have a limited capacity for exceeding physiological thermal limits [67]. The potential loss of mosquito habitat in Ecuador has considerable implications for the public health sector. Localized extinctions would conserve valuable health resources by triggering allocation shifts as unsuitable areas no longer support active disease transmission.
Our findings are broadly consistent with a previous coarser scale ENM analysis of adult mosquitoes in Ecuador, which suggests that while Aedes mosquitoes may shift into highland areas under changing climate conditions, the total area of suitable habitat will ultimately decrease as localized climatic conditions favor extirpation [32]. However, models of Aedes distribution in the previous study were made through the year 2100, representing an extended time horizon for guiding agency decision making. While predicted ranges in 2100 are visually similar to results presented here, notable discrepancies exist between the spatial distributions predicted in our models and the previous study for 2050, where the previous model predicts widespread absence of mosquitoes in central Ecuador and presence throughout much of the eastern Amazon basin. In contrast to our methods, Escobar et al. [32] used a different niche modeling algorithm, a different model of climate change (A2), a coarser spatial resolution (20 km), and combined global species occurrence for two adult arbovirus vectors, Ae. aegypti and Ae. albopictus, to predict pooled arbovirus risk throughout Ecuador. Though Ae. aegpyti and Ae. albopictus are competent vectors of diseases that occur in Ecuador (e.g. dengue, chikungunya, Zika), these species differ significantly in their physiology, possibly driving observed discrepancies between the models of pooled adult Aedes spp. risk and larval Ae. aegypti range [68]. Reaching consensus across ENMs is a known area of conflict in ecology that requires more research, where various methodologies can lead to vastly different forecasts of geographic distributions and risk, making direct comparisons between models difficult [69]. Future studies combining multiple approaches and comparing the impact of input on models could help resolve this conundrum.
The scale of analysis used in this study presents a limitation in applying resulting ENMs for local management decisions. We chose a moderately low spatial resolution for this study (5km raster cells) to reflect the highest level of precision that could be assigned to larval mosquito occurrence (i.e. points could be matched to cities or clusters of villages, but not to individual households or neighborhoods). Arboviral disease transmission and larval mosquito presence, especially for Ae. aegypti, are typically managed at the household or neighborhood level, and although we can use these results to discuss regional changes in mosquito distribution throughout Ecuador, we cannot overstate the findings as a means to assess risk at the level of disease transmission [70]. Furthermore, the LI survey conducted by the MSP was limited in that focus was placed on sampling areas with perceived arbovirus transmission risk throughout Ecuador, especially households in densely populated urban centers and established communities where cases had been reported in the past. Low accessibility and human population density in Ecuador’s eastern basin region may have contributed to under sampling of mosquito presence in these areas, possibly accounting for low model agreement in this area. Ultimately, robust vector surveillance for Ae. aegypti in eastern Ecuador would be required to validate absence in this region, though such intensive ground-truthing would be wrought with logistical concerns, including diversion of scarce surveillance resources from high-demand management districts and the inherent difficulty of establishing “true” absence via surveys.
Aedes aegypti is a globally invasive species, owing much of its success to its close connection with human activity and urban environments. As such, predicted habitat suitability does not guarantee the introduction and establishment of a species in the future due to a myriad of factors, such as physical and geographical barriers to movement [71]. Patterns of human movement and land use also have the potential to influence mosquito expansion in ways that we cannot predict with ENMs. Additionally, microclimate can become a critical factor in determining true habitat suitability, and there are many examples of anthropogenic behaviors and structures providing a buffering effect, or refuge, against climatic conditions that would be otherwise physiologically limiting to insect vectors [5,72–75]. Dramatic shifts in species compositions in Ecuador, mediated by elevation, also occur on very fine spatial scales [76,77]. Moving forward, observed areas of range expansion on the edge of unsuitable habitat may be better modeled at finer resolutions, which would aid in making community-targeted management decisions based on estimated risk.
Based on the results of this study, we conclude that the geographic distribution of Ae. aegypti in Ecuador will be impacted by projected shifts in climate. Extensive changes in modeled vector distributions were observed even under the most conservative climate change scenario, and these changes, although consistent in pattern, became more evident with increasingly high greenhouse gas emissions scenarios. Although there is a continued need for surveillance activities, these findings enable us to anticipate transitioning risk of arboviral diseases in a spatial context throughout Ecuador, allowing for long-term planning of agency vector control strategies.
|
10.1371/journal.pbio.1002039 | Composition, Formation, and Regulation of the Cytosolic C-ring, a Dynamic Component of the Type III Secretion Injectisome | Many gram-negative pathogens employ a type III secretion injectisome to translocate effector proteins into eukaryotic host cells. While the structure of the distal “needle complex” is well documented, the composition and role of the functionally important cytosolic complex remain less well understood. Using functional fluorescent fusions, we found that the C-ring, an essential and conserved cytosolic component of the system, is composed of ~22 copies of SctQ (YscQ in Yersinia enterocolitica), which require the presence of YscQC, the product of an internal translation initiation site in yscQ, for their cooperative assembly. Photoactivated localization microscopy (PALM) reveals that in vivo, YscQ is present in both a free-moving cytosolic and a stable injectisome-bound state. Notably, fluorescence recovery after photobleaching (FRAP) shows that YscQ exchanges between the injectisome and the cytosol, with a t½ of 68 ± 8 seconds when injectisomes are secreting. In contrast, the secretin SctC (YscC) and the major export apparatus component SctV (YscV) display minimal exchange. Under non-secreting conditions, the exchange rate of YscQ is reduced to t½ = 134 ± 16 seconds, revealing a correlation between C-ring exchange and injectisome activity, which indicates a possible role for C-ring stability in regulation of type III secretion. The stabilization of the C-ring depends on the presence of the functional ATPase SctN (YscN). These data provide new insights into the formation and composition of the injectisome and present a novel aspect of type III secretion, the exchange of C-ring subunits, which is regulated with respect to secretion.
| The type III secretion system, also known as the injectisome, is a key virulence factor in many gram-negative bacteria, and is responsible for the transmission of bacterial proteins directly into host cells. While some elements of the system are well characterized, the cytosolic components involved in substrate recognition and handling are not well understood. One of the major questions is the role of the C-ring, an essential yet enigmatic cytosolic injectisome member. We used fluorescence microscopy to analyze the architecture and behavior of the C-ring in live Y. enterocolitica bacteria, a human pathogen. We found that in vivo, the C-ring assembles cooperatively with the help of additional copies of its own C-terminal fragment and has a highly dynamic structure, with C-ring subunits exchanging between the working injectisomes and a cytosolic pool. The rate of exchange is different between secreting and non-secreting injectisomes and depends on the function of the type III secretion ATPase, indicating that the stability of the complex is altered when functioning. This dynamic behaviour raises the possibility that the C-ring is a regulator of targeted protein delivery by the type III secretion system and makes the C-ring a viable target for the development of novel anti-virulence drugs.
| Bacteria that live in close contact with eukaryotic cells are frequently able to modulate host cell behavior. Various Gram-negative species employ a molecular syringe, the type III secretion system (T3SS) or injectisome [1, 2], to translocate effector proteins from the bacterial cytosol into the host cell. T3SSs are often involved in pathogenesis, and are crucial virulence factors for animal and plant pathogens [3–5], but are also employed to promote symbiosis [6, 7]. Thus, the translocated effectors vary greatly between species [8, 9]. In contrast, the export machinery itself is conserved across organisms [10, 11]. It is a large complex that comprises more than 15 different proteins present in one to more than a hundred copies each (S1 Fig). The distal part of the injectisome including the needle and the “base,” a series of membrane-spanning rings, is structurally well characterized [12, 13]. However, little is known about the structure and function of the proximal components, the “export apparatus” in the inner membrane (IM) and the cytosolic components that are essential for the function and regulation of the system. This report uses the unified Sct protein nomenclature [3] for general protein features, in combination with the Yersinia Ysc nomenclature (S1 Fig; S1 Table for names of homologous proteins).
Five conserved cytosolic proteins are essential for type III secretion, an ATPase (SctN; YscN in Yersinia), thought to detach chaperones and unfold export substrates [14], and its putative negative regulator (SctL; YscL) [15, 16]; a protein with homology to the central stalk of the FoF1-ATPase that stimulates ATPase activity (SctO; YscO) [17–19]; a homolog of the flagellar motor C-ring components FliM and FliN (SctQ; YscQ) [20]; and an additional accessory protein (SctK; YscK). All of these proteins interact with each other: yeast-two-hybrid and a yeast-three-hybrid experiment clearly suggest a sequence of interactions SctK-SctQ-SctL-SctN [20–23] while SctO has been shown to directly and functionally interact with SctN and SctL (S1 Fig) [19, 24]. The ATPase SctN and the C-ring component SctQ require each other and the two interacting proteins SctK and SctL to assemble at the cytosolic interface of the injectisome [25], suggesting the formation of one large cytosolic complex. The cytosolic components are required for various essential steps in type III secretion, including energy transduction, binding (or rejection) of substrates, and their preparation for export [14, 18, 23, 26, 27]. Despite this central role, surprisingly little is known about the molecular events and the organization of the proteins at the cytosolic interface, mostly because the components rarely co-purify with the rest of the machinery and are difficult to analyze in vitro in the absence of the other parts of the injectisome.
The T3SS is related to the bacterial flagellar motor with which it shares a common ancestor (see S1 Table for homologues) [28–31]. Recently, this homology was highlighted by the discovery that at least in some organisms, the cytosolic C-ring, which is formed by the two proteins FliM and FliN in the flagellum, also consists of two different proteins arising from a single gene in the injectisome [32, 33]: (i) the full length protein, SctQfull, which is similar in size and has some sequence homology to FliM in its N-terminal region, and (ii) a product from an internal translation start site, SctQC, which comprises about the C-terminal third of SctQ and is highly homologous to FliN. The flagellar C-ring is part of the switch complex [34–36], which can reverse the direction of rotation of the Escherichia coli flagellum upon binding of the activated response regulator CheY-P to FliM, allowing the bacterium to tumble and reorient [37, 38]. The C-ring is also required for the export of substrates through the flagellar T3SS, but this can be overcome by simple overexpression of the transcriptional regulators [39] or even the ATPase alone [40].
The injectisome is generally not thought to rotate. Remarkably however, the C-ring not only remains conserved among all known injectisomes, but is also absolutely essential for export of substrates, suggesting a divergent functional adaptation in the flagellum and the injectisome. This idea is supported by recent cryo-tomography studies [41, 42], where the C-ring, a prominent feature of the flagellum [42, 43], was absent in the injectisome, whereas both the major export apparatus component (FlhA in the flagellum/SctV in the injectisome) and the ATPase (FliI/SctN) formed distinct and comparable structures in both machines. Consequently, the structure and composition of the injectisome C-ring remain unclear. Similarly, its localization is ambiguous. The C-ring protein SctQ is present at the cytosolic interface of the injectisome [20] and co-localizes with other injectisome components [25]. However, complexes of the cytosolic components including SctQ only partially cofractionated with other injectisome components and were also found in the cytosolic fraction [21, 26], indicating that the C-ring is either not tightly bound or is present in two subpopulations. Insights into the role of the injectisome C-ring have been gained by interaction studies showing that the C-ring is involved in export cargo handling [23, 26]. In combination with SctK and SctL, the C-ring was shown to form a “sorting platform” that governs the order of substrate export [26]. However, the molecular mechanism of this function remains as elusive as structure and localization of the C-ring.
To study the composition and functional role of the injectisome C-ring, we generated and studied functional fluorescent fusions of various injectisome components in a Y. enterocolitica strain lacking the major virulence effectors [41]. Based on relative fluorescence intensities, we found that there are 22 ± 8 full-length C-ring subunits YscQ per injectisome. YscQC, the product of the internal translation start site, is not only required for substrate translocation, but also necessary for the localization of YscQfull and the correct assembly of the ATPase. YscQ exchanges between its docking site at the injectisome and a cytosolic pool in vivo. Its exchange rate under secreting conditions is significantly higher than under non-secreting conditions (t½ of 68.2 ± 7.9 s versus 134.3 ± 16.1 s), which links C-ring dynamics and effector export. This correlation depends on an active ATPase, suggesting a close functional relation between the two largest cytosolic components of the injectisome, C-ring and ATPase.
The composition of the C-ring is unknown. To estimate the number of YscQ proteins within the Yersinia T3SS, we compared the relative intensities of foci in bacteria expressing either EGFP-YscQ or EGFP-YscD. Members of the SctD family, to which YscD belongs, have been shown to form 24-mer ring structures [12, 44–48], which was recently corroborated for Yersinia by the crystal structure of YscD and the cryotomography structure of the injectisome [41]. We therefore devised a spot-detection algorithm and used the intensity of EGFP-YscD foci as a reference for EGFP-YscQ. Both proteins were expressed from the pYV virulence plasmid using their native promoters. For biosafety reasons, this and all subsequent experiments were performed in Y. enterocolitica strains that lack the major virulence effectors YopH, O, P, E, M, T and are auxotrophic for diaminopimelic acid [41]. While EGFP-YscQ was fully functional, EGFP-YscD was slightly less efficient in effector export. Importantly, both fusion proteins were stable, as detected by an immunoblot of total cellular proteins, using anti-GFP antibodies (S2 Fig). Based on the relative intensities of the foci, we calculated the number of YscQfull per injectisome to be 22.0 ± 8.3 under non-secreting conditions and 22.2 ± 6.8 under secreting conditions (Fig. 1). As the intensity of foci is determined at the microscopy plane corresponding to the center of the bacterium in the respective DIC image, some foci will be centered below or above this plane and will be detected with lower intensity, accounting for most of the observed variance. Supporting this, the intensity distribution curve for YscQ did not show any greater width than the curve for YscD (Fig. 1B and 1C). As the number of YscD per injectisome is thought to be fixed, this suggests that the stoichiometry of YscQ at the injectisome is similarly constant.
It was recently found that an internal translation initiation site in SctQ leads to the expression of a short C-terminal fragment, SctQC, which interacts with SctQfull [32, 33]. We deleted the internal start site of the Yersinia C-ring protein YscQ by mutating the corresponding Met218 to Ala and found that the resulting strain lacking YscQC did not export any effectors or translocators, confirming that the additional C-terminal fragment is essential for secretion in Yersinia [33]. This effect could be complemented in trans by YscQC, EGFP-YscQC, or mCherry-YscQC (Fig. 2A). EGFP-YscQC was stable (S2 Fig) and localized in foci at the bacterial membrane (Fig. 2B), like other fluorescent injectisome components [25, 49]. Surprisingly, when we analyzed the distribution of EGFP-YscQM218A, the full-length protein lacking the internal start site, it no longer localized to the injectisome, but was completely cytosolic. Upon in trans expression of YscQC, localization was recovered (Fig. 2B). Therefore, although YscQfull contains all parts of YscQC, it requires additional YscQC for proper localization. Similarly, EGFP-YscQC did not localize in foci in the absence of YscQfull (S3A Fig), showing that only the YscQfull:YscQC complex can correctly assemble. To test whether localization of the ATPase YscN also depends on YscQC, we imaged EGFP-YscN in YscQM218A. In the resulting absence of YscQC, EGFP-YscN remained completely cytosolic (S3B Fig), indicating that YscQC is required for the assembly of the complete cytosolic complex.
The expression of YscQC from plasmid revealed cooperative binding properties. To quantify this effect and to directly test the relative localization of YscQfull and YscQC, we expressed mCherry-YscQC in trans in an EGFP-YscQM218A background. Increasing levels of mCherry-YscQC led to a significant increase in the number of spots, rather than the intensity of spots, for both YscQ variants (Fig. 2C, 2D and 2F). Next, we analyzed the amount of mCherry-YscQC by immunoblot using polyclonal anti-YscQ antibodies (S4 Fig) and plotted the number of detected EGFP-YscQM218A spots against the amount of YscQC (Fig. 2E). The result shows a clear linear correlation (R2 > 0.98), suggesting that the formation of an observable focus is an all-or-nothing process and binding of full-size C-ring subunits to the injectisome is cooperative up to the maximal number of YscQfull per injectisome. EGFP-YscQM218A and mCherry-YscQC foci colocalized (Fig. 2C). While YscQfull required the presence of YscQC for its proper localization at the injectisome, overexpression of mCherry-YscQC did not remove EGFP-YscQM218A from its localization in foci at the injectisome (S5 Fig).
Protein complexes can be dynamic assemblies, adapting their structure and function to changing environmental conditions. Given the inconsistent data on the cytosolic versus injectisome-bound localization of the C-ring, we wanted to test whether the C-ring exchanges subunits in the assembled structure. To this end, we followed the ability of injectisome components to exchange with cellular pools. Besides the C-ring subunit YscQ, we analyzed the secretin YscC and the major export apparatus component YscV, as these proteins cover the major functional subcomplexes of the injectisome (cytosolic components, membrane rings, and export apparatus, respectively) (Fig. 3A). All fusion proteins were stable and fully functional for secretion (S2 and S6 Figs.). To ensure that the fluorescent tags have only minimal influence on the kinetics of effector export, we devised a sensitive assay based on the quantification of export of an engineered T3SS substrate, β-lactamase fused to the YopH secretion signal, YopH1–17-bla [50–52]. In this assay, the strains expressing YscC-mCherry, YscV-mCherry, and EGFP-YscQ showed export rates of over 75% of the wild-type (WT) strain (Fig. 3B). This shows that the fluorescent tags have very little influence on protein functionality, suggesting that the exchange rates of labeled proteins resemble those of unlabeled subunits.
To detect the exchange of YscC-mCherry, YscV-mCherry, and EGFP-YscQ, we analyzed fluorescence recovery after photobleaching (FRAP) in live cells expressing these fusion proteins from their native promoters on the pYV virulence plasmid (Fig. 3C and 3D). Bacteria were grown and the T3SS was induced under non-secreting conditions. Three hours after induction, cells were transferred to secretion-permissive medium. A single fluorescent spot was photobleached by a short pulse of a tightly focused laser beam and recovery of fluorescence was analyzed for up to 10 minutes. As expected for the stable secretin ring, YscC-mCherry spots showed very little recovery within this time (Fig. 3C and 3D; S1 Movie). Similarly, YscV-mCherry spots only minimally recovered fluorescence (Fig. 3C and 3D; S2 Movie). In contrast, EGFP-YscQ spots showed almost complete recovery (≥75%) within the first few minutes after photobleaching (Fig. 3C and 3D; S3 Movie). Analysis of 20 recovery curves of EGFP-YscQ spots showed an average recovery half-time t½ of 68.2 ± 7.9 s. Similar to the full-length protein EGFP-YscQ, subunits of the C-terminal fragment EGFP-YscQC also exchanged between the injectisome and the cytosol (S7 Fig). While this exchange could not be quantified because N-terminally labeled YscQC can only be expressed in trans, this observation is compatible with exchange of a stable YscQ:YscQC complex. Individual recovery curves of EGFP-YscQ and EGFP-YscQC did not show any discernible large steps, which suggests that small units, such as single YscQ proteins or small subcomplexes, exchange within functional injectisomes.
To assess C-ring dynamics on a single-molecule level, we analyzed YscQ under secreting conditions by photoactivated localization microscopy (PALM) (Fig. 3E and 3F) [53–56]. We detected two populations of PAmCherry-YscQ molecules: some only moved within a very small area over successive exposures (Fig. 3F, red), whereas others were mobile (Fig. 3F, blue). Interestingly, the immobile PAmCherry-YscQ molecules were also tightly concentrated in small foci at the cell membrane (Fig. 3E and 3F). Analysis of the diffusion coefficient of PAmCherry-YscQ and YscV-PAmCherry (Figs. 3G and S8) showed that more than 50% of the PAmCherry-YscQ molecules were mobile (single step diffusion coefficients > 0.15 μm2/s), while only a minority of YscV-PAmCherry molecules were, confirming the different mobility of injectisome components on a single-molecule level.
To assess whether the exchange of YscQ between its docked location at the injectisome and the cytosolic pool might be functionally significant, we compared the rates of protein exchange under non-secreting and secreting conditions. We found that EGFP-YscQ was significantly less dynamic under non-secreting conditions, with an average half-time of recovery t½ of 134.3 ± 16.1 s, compared to a t½ under secreting conditions of 68.2 ± 7.9 s (Fig. 4). This result indicates that subunit exchange is regulated and links the observed turnover of YscQ to the ultimate function of the T3SS, effector secretion. Assembly of the C-ring depends on the presence of other cytosolic components, including the ATPase YscN [25]. However, the functional relation between the two components is unclear. Therefore, we analyzed YscQ dynamics in a strain expressing EGFP-YscQ and a catalytically inactive version of the ATPase, YscNK175E [57]. Strains expressing YscNK175E are defective in effector secretion [58], but form the C-ring [25]. We observed that the mutation in the ATPase led to fast EGFP-YscQ exchange both under non-secreting and secreting conditions, 54.4 ± 13.7 s and 58.5 ± 10.2 s, respectively (Fig. 4). The average mobile fraction was approximately 80% for both tested strains and conditions. These results show that the C-ring is only stabilized under non-secreting conditions when the ATPase is functional.
The role of the cytosolic components is a long standing open question in the type III secretion field. Despite the fact that all conserved cytosolic constituents are essential for assembly and function of the injectisome, their working mechanism and function in secretion remain unclear. This is especially true for the C-ring. Its main role in the closely related flagellum, the switching of rotation direction, does not apply to the injectisome, and its composition and behavior remained elusive.
The C-ring component SctQ has an internal translation initiation site leading to the expression of a protein equivalent to the C-terminal third of SctQ, SctQC, in addition to the full-length protein, SctQfull [32, 33]. Low levels of effector secretion of Salmonella SPI-2 lacking SctQC led to the proposal that the short version acts as a chaperone for SctQfull [32]. The short translation product was however essential for secretion in Y. pseudotuberculosis [33]. In our present study, we found that the C-terminal fragment (YscQC in Yersinia) is required for the assembly of the cytosolic complex (Fig. 2). The requirement of YscQC for localization of YscQfull and the colocalization of both proteins (Fig. 2C) is compatible with the proposal that binding of the two versions triggers structural rearrangements in YscQfull [33]. Excess cytosolic YscQC did not titrate away YscQfull from the injectisome (S5 Fig), showing that the affinity of YscQfull to the complete localized cytosolic complex is higher than to YscQC alone. The loss of functionality in strains lacking the internal start site is therefore at least in part due to their inability to assemble the cytosolic components of the injectisome. This differs from the situation in the flagellum, where FliN (the homologue of YscQC) stabilizes FliM (YscQfull), but is not absolutely required for correct FliM localization [59–61] and indicates divergent adaptation of the C-ring to its respective functions in the two systems.
The stoichiometry and size of the injectisome C-ring were unknown. We therefore compared the relative intensity of foci in strains expressing EGFP-YscQ or EGFP-YscD and, based on the well-documented reference SctD (YscD), which is present in 24 subunits [12, 41, 44–46], determined the number of full-length YscQ subunits per injectisome to be 22 both in non-secreting and secreting conditions (with standard deviations of 8 and 7, respectively) (Fig. 1). These numbers are based on an average occupancy and therefore represent a lower boundary for the maximum number of binding sites for a dynamic component. However, we did not detect a broader intensity distribution for YscQ compared to YscD (Fig. 1B and 1C), which would be expected if the number of YscQ per injectisome would fluctuate. As the stoichiometry of YscD is assumed to be constant, this suggests that despite the dynamic exchange of YscQ, the number of YscQ at the injectisome at any given time remains relatively constant and our numbers are close to the maximal stoichiometry. The absence of distinctive larger steps in the recovery curves after photobleaching in combination with the mutual requirement of YscQfull and YscQC for assembly indicate that the (YscQfull):(YscQC)2 complex proposed by Bzymek and colleagues [33] could be the building block of the injectisome C-ring.
In the E. coli and Salmonella flagellum, the C-ring is composed of about 36 FliM:FliN4 subcomplexes [61–63], but the C-rings from different species have diameters between 34 and 57 nm, suggesting the copy number might be different between species [64]. If the arrangement of YscQfull:YscQC subcomplexes within the machinery is similar to the flagellum, the diameter of the C-ring will be determined by the size and arrangement of YscQfull. Assuming a globular shape of YscQfull and a circular arrangement of the subunits, the radius of a C-ring containing 22 YscQfull is 30.2 nm (Fig. 1D; see S1 Text for details). The injectisome C-ring would therefore be smaller and composed of fewer subunits than its flagellar counterpart. Unlike the flagellar C-ring, the injectisome C-ring has not been visualized in recent cryotomography studies [41, 42], despite strong evidence that it is present at the cytosolic interface of the injectisome [20, 22, 25, 26]. The reason for this is unclear, but our observation that the injectisome C-ring is not fixed might at least partially explain this phenomenon. In a structural comparison of the flagellum and the injectisome, Kawamoto and colleagues [42] detected a weak electron density connecting the edge of the SctV “export gate” to an unknown and structurally indiscernible part at the cytoplasmic surface of the membrane. In the flagellum, a similar density connects the “export gate” FlhA with the C-ring. In the injectisome, the diameter of the binding site of the electron density at the cytoplasmic surface was approximately 28 nm [42], which is larger than the IM rings, but close to our calculated C-ring diameter of 30.2 nm. If the injectisome C-ring indeed resides at this spot, it would be located very close to the IM, which might further impede its structural identification.
Interestingly, we found that in functioning T3SSs, YscQ is not a stable structural component, but exchanges between a docking site at the injectisome and a cytosolic pool. The recovery half-time under secreting conditions, equivalent to the average lifetime of a given protein in the complex, was 68.2 ± 7.9 s. In contrast, the secretin SctC (YscC) and the major export apparatus component SctV (YscV) showed less than 20% recovery over 600 s (Fig. 3). This was expected in the case of the secretin, which forms a very stable ring in the outer membrane [65]. The flagellar homologue of SctV, FlhA, was shown to display some mobility [59], but in a more recent study, the majority of FlhA foci did not show any detectable turnover in the 10 minutes after photobleaching [66]. Our data show that similarly, there is little or no SctV turnover in the injectisome, in agreement with the proposition that the IM export apparatus is anchored relatively stable within the MS-ring [67].
Protein complexes can be highly dynamic assemblies, with proteins exchanging and switching their conformation while the complex is actively functioning. Dynamic behavior has been shown for the flagellum, where the stator protein MotB [68] and the C-ring component FliM exchange [63, 69–71]. Interestingly, FliM exchange was strongly decreased when clockwise rotation was suppressed, both in the absence of CheY [69] and when the flagellar motor was locked in counterclockwise direction [71]. We therefore tested whether the exchange of C-ring subunits in the injectisome is linked to the function of the T3SS, translocation of substrates. Comparing C-ring dynamics under secreting and non-secreting conditions, we found that the YscQ exchange rate was decreased under non-secreting conditions in the WT strain (Fig. 4). Which factors might control or influence this change of dynamics? The easiest explanation would be that the C-ring shuttles effector-chaperone complexes from the cytosol to the injectisome. However, such a direct role, similar to the one proposed for the flagellar ATPase complex [72] appears at odds with the high observed rates of effector export, once the injectisome is activated for secretion [73, 74]. Schlumberger and colleagues reported export rates of about 20 effector proteins per second [74], which are much faster than the observed exchange rates of the C-ring, although our study was performed in a strain lacking the six main effectors YopH, O, P, E, M, T, and both the presence of these effectors and direct cell contact as opposed to secretion induced by Ca2+ depletion might influence the exchange rate of the C-ring. Likewise, dynamics are not simply controlled by extracellular Ca2+ concentrations nor solely by the secretion process itself, as subunit exchange was fast regardless of extracellular Ca2+ in an ATPase-inactive strain (Fig. 4). Regulation of C-ring dynamics therefore differs from the situation in the flagellar C-ring, where exchange correlates with the output, the rotation direction, regardless of how this output was caused [63]. Taken together, the results suggest that YscQ is actively stabilized in the non-secreting mode and that this stabilization requires the ATPase to be functional. In these conditions, the ATPase is probably bound to one or more chaperone-substrate complexes. Combined with earlier observations that the C-ring itself binds to cargo [23, 26], these results suggest that the C-ring might be stabilized by the chaperone-substrate complexes bound to the ATPase under non-secreting conditions. The resulting close functional connection of C-ring and ATPase is supported by previous data showing that both proteins require each other for assembly [25]. Whether the dynamic exchange extends to the ATPase and the two other cytosolic components, SctK and SctL, which are all required to localize the C-ring, remains to be determined. At the moment, the lack of sufficiently active fusion proteins prevents the detailed analysis of these components. Based on structural arguments, Kawamoto and colleagues [42] proposed that in the flagellum, a structural rearrangement occurs between the “on” and “off” state, possibly involving the regulator protein (FliH in the flagellum, SctL in the injectisome), which in the flagellum was shown to be required to connect the ATPase FliI both to the C-ring (FliN) [75] and the major export apparatus component (FlhA) [76]. Such a structural rearrangement might explain our observation that C-ring dynamics decrease under non-secreting (=“off”) conditions. Interaction studies under different conditions might reveal the structural and functional rearrangements upon activation of export, an important step towards understanding the molecular mechanism of type III secretion.
Many bacteria utilize the T3SS to establish permanent infections and therefore need to adapt the function of the injectisome to their current environment. Within the host, regulating long-term activity of the system is critical for the survival of the bacterium, however this activity and its functional regulation have not been studied in detail. The steps in type III secretion that are governed by the cytosolic components, selection and export of substrates, are obvious targets for regulation and most signaling pathways employ cytosolic response regulators. Therefore, it is conceivable that the cytosolic components are involved in regulating type III secretion. Dynamic exchange is a suitable target for regulation, as exemplified by the flagellum and the nuclear pore complex [77]. Exchange of C-ring subunits could therefore be used to regulate T3SS function, in line with the observed changes in dynamics under different conditions. To date, a small number of external conditions influencing type III secretion beyond activation has been described, notably oxygenation [78, 79] and pH [80]. The involved pathways are species specific, suggesting that the respective signal receivers at the injectisome are less conserved than other components. Intriguingly, parts of the C-ring protein, especially its N-terminal region, show a significantly higher sequence variation than most other injectisome components (S9 Fig) [81], which could be an evolutionary consequence of responding to cues from species-specific signaling pathways.
Over the last years, understanding of the composition, function, and regulation of the functional cytosolic components of the T3SS has lagged behind the determination of the structure of the membrane rings. Our discovery that the C-ring is composed of around 22 subunits and requires the presence of additional copies of its C-terminal fragment to assemble sheds light on the composition of this essential part of the T3SS. Furthermore, our finding that the C-ring is a dynamic component and that its exchange correlates both with the secretion status and the activity of the ATPase reveals a new aspect of how the injectisome works and responds to its environment, advancing our knowledge and appreciation of the molecular mechanisms and regulation of the complete T3SS.
Strains and constructs used in the experiments are listed in S2 Table. E. coli Top10 and BW19610 were used for cloning and E. coli SM10 λ pir+ for conjugation. E. coli strains were grown routinely on Luria–Bertani (LB) agar plates or in liquid LB medium at 37°C. Ampicillin and streptomycin were used at concentrations of 200 µg/ml and 100 µg/ml to select for expression vectors and suicide vectors. All Y. enterocolitica strains were based on strain IML421asd [41], which lacks the effectors YopH, YopO, YopP, YopE, YopM, and YopT and is auxotrophic for diaminopimelic acid due to an additional mutation in the aspartate-β-semialdehyde (asd) gene. Y. enterocolitica were routinely grown at 25°C in brain heart infusion (BHI) broth containing nalidixic acid (35 µg/ml) and diaminopimelic acid (80 µg/ml). Plasmids were generated using Phusion polymerase (Finnzymes). Mutators for modification or deletion of genes in the pYV plasmids were constructed as described earlier [49]. All constructs were confirmed by sequencing (Source BioScience).
Y. enterocolitica mutants were generated by allelic exchange, replacing the WT gene on the virulence plasmid by the mutated version. Completion of the allelic exchange was tested for by plating diploid bacteria on plates containing 5% sucrose [82].
Y. enterocolitica cultures for secretion assays and stoichiometry analysis were inoculated from stationary overnight cultures to an optical density at 600 nm (OD600) of 0.12 in BHI broth containing EDTA (5 mM) (BHI-EDTA, secreting conditions) or CaCl2 (5 mM) (BHI-Ca2+, non-secreting conditions) supplemented with nalidixic acid (35 µg/ml), diaminopimelic acid (80 µg/ml), glycerol (4 mg/ml), and MgCl2 (20 mM). After 1.5 h of growth at 25°C, the yop regulon was induced by shifting the culture to 37°C. Where indicated, expression of the pBAD constructs was induced by adding L-arabinose (0.2%, unless stated otherwise) to the culture just before the shift to 37°C. After 3 h of incubation at 37°C, cultures were used for further analysis.
At this point, bacteria for FRAP and PALM analysis were collected (2,400g, 4 min) and resuspended in HEPES-M22 complemented with diaminopimelic acid (80 µg/ml) and either 5 mM EDTA (secreting conditions) or 5 mM CaCl2 (non-secreting conditions).
Bacteria and supernatant (SN) fractions were separated by centrifugation at 20,800g for 10 min at 4°C. The cell pellet was taken as total cell (TC) fraction. Proteins in the SN were precipitated with trichloroacetic acid 10% (w/v) final for 1–8 hours at 4°C. Proteins were separated on Novex 4%–20% gradient SDS-PAGE gels (Life technologies). Unless mentioned otherwise, proteins secreted by 3 × 108 bacteria (SN) or produced by 2.5 × 108 bacteria (TC) were loaded per lane. Secreted proteins were stained using the Coomassie-based “Instant blue” staining solution (Expedeon). Immunoblotting was carried out using mouse polyclonal antibodies against GFP (Clontech 632459, 1:1,000) or mCherry (Clontech 632543, 1:1,000) or rabbit polyclonal antibodies against YscQ (MIPA235, 1:1,000). Detection was performed with corresponding rabbit anti-mouse or goat anti-rabbit secondary antibodies conjugated to horseradish peroxidase (Dako; 1:5,000), before development with Immobilon Western chemiluminescent substrate (Millipore). For the quantification of YscQC, the signal intensity of the region around the expected protein size in each lane of the immunoblot was measured using ImageJ and corrected for the background intensity in the lane lacking YscQC.
Y. enterocolitica strains were transformed with plasmids pAD372 or pAD374 (expressing GST-bla or YopH1–17-bla under an arabinose-inducible promoter) (see S2 Table). Cultures were inoculated and grown under non-secreting conditions as stated above. Expression was induced with 0.2% Arabinose (w/v) in parallel to the temperature switch to 37°C. Bacteria were collected by centrifugation (4 min, 2,400g, 37°C), and resuspended in prewarmed HEPES-M22 (M22 buffered with 20 mM HEPES instead of phosphate buffer) complemented with diaminopimelic acid (80 µg/ml), glycerol (4 mg/ml), and MgCl2 (20 mM) and 5 mM EDTA to induce secretion, followed by 30 min incubation at 37°C. In a 96 well Corning 3603 plate, 100 µl/well of each strain was added in triplicates and 20 µl/well of β-lactamase substrate solution (0.1 M Tris-HCl [pH 7.5], 20 µM Fluoricillin Green 495/525 [Life Technologies]) was added. Fluorescence was measured every 30 s for 15 min, using a BMG Labtech Fluostar photometer (Ex 490/10 nm, Em 530/12 nm) and the slope of the linear increasing region was determined. The results are averages of five independent experiments (four independent experiments for strain ΔYscN) with three technical replicates each.
For standard fluorescence imaging, determination of spot intensity, and deconvolution, 1.5 µl of resuspended bacterial culture was placed on a microscope slide layered with a pad of 2% agarose in HEPES-M22 (M22 buffered with 20 mM HEPES instead of phosphate) supplemented with diaminopimelic acid (80 µg/ml) and either EDTA (5 mM) (secreting conditions) or CaCl2 (5 mM) (non-secreting conditions). A Deltavision Spectris optical sectioning microscope (Applied Precision) equipped with a UPlanSApo 100 × 1.40 oil objective (Olympus) combined with 1.6× auxiliary magnification and an Evolve EMCCD camera (Photometrics) was used to take differential interference contrast (DIC) and fluorescence photomicrographs. For fluorophore visualization, either the GFP/hsGFP filter set (Ex 475/28 nm, Em 522/44 nm) or the mCherry/hsCherry filter set (Ex 575/25 nm, Em 634/63 nm) were used. DIC frames were taken with 0.05 s and fluorescence frames with 1.0 s exposure time. Per image, a Z-stack containing 7 to 15 frames per wavelength with a spacing of 150 nm was acquired. The stacks were deconvolved using softWoRx 5.5 with standard settings (Applied Precision). A representative DIC frame and the corresponding fluorescence frame were selected and further processed with the ImageJ software. Quantitative analysis was performed on the undeconvolved images.
For FRAP, a custom-built multi-color fluorescence microscope was used, which allows photobleaching of one fluorescent spot within the cell using slimfield while simultaneously monitoring the fluorescence emission using conventional widefield epifluorescence [68, 83]. 1.5 µl of resuspended bacterial culture were analyzed on a microscope slide layered with a pad of 2% agarose in the respective resuspension medium. Fluorescence was excited using a 473 nm or a 561 nm laser (for EGFP or mCherry, respectively). Fluorescence emission was imaged at 50 nm∕pixel in frame-transfer mode at 25 Hz by a 128 × 128-pixel, cooled, back-thinned electron-multiplying charge-coupled device camera (iXon DV860-BI; Andor Technology). The displayed images are unweighted averages of three consecutive frames. Analysis was performed on the raw data. Average fluorescence intensities of the spot (IS), the corresponding bacterium (IB), and the background (I0) were determined manually using ImageJ. The spot intensity ratio R = (IS − I0) / (IB − I0) was calculated for each frame. The pre-bleach ratio and post-bleach ratio were determined using ten frames immediately before and after photobleaching the spot with a 20 ms focused laser beam. Unless stated otherwise, the recovery curve was measured by 100 frames of 40 ms exposure over a period of 335–570 s. Values were normalized with the pre-bleach and post-bleach ratio and τ½ was calculated by fitting R(time) to an exponential curve using the formula R(time) = start + (start-end) * (Exp(−time/τ½) − 1) using the sequential quadratic programming algorithm in SPSS Statistics 21 (IBM), with the parameters set to start = [0.9; 1.1], end = [−0.1; 0.4]. Fits with an r2 value of less than 0.4 were excluded from further analysis. The recovery half-time (or half-life) t½ was calculated using he formula t½ = τ½* ln(2).
Photoactivated Localization Microscopy Y. enterocolitica cultures were grown and treated as given above for FRAP analysis. 15,000 consecutive frames were imaged with a frame rate of 15.26 ms on a custom-built single-molecule fluorescence microscope using a 405-nm laser (CNI) for photoactivation and a 561-nm laser (Oxxius) for PAmCherry excitation. Analysis was performed using the STORMTRACKER package for MATLAB (MathWorks) as described by Uphoff and colleagues [53]. The minimal track length was 5, the threshold used to discriminate between bound and unbound molecules was 0.15 μm2/s for the single-step diffusion coefficient.
The stoichiometry of YscQ was determined comparing the fluorescence intensities of fluorescent foci in strains expressing EGFP-YscQ or EGFP-YscD under secreting and non-secreting conditions. Both proteins were expressed from their native promoter on the pYV virulence plasmid. Images were corrected for background fluorescence only and were not processed otherwise to allow for direct comparison of fluorescence distribution and intensity. Automated spot counting in fluorescent 3D stacks was performed in MATLAB (MathWorks) using the electron tomography related Dynamo toolbox (http://dynamo-em.org/) [84]. The peaks were detected based on the fluorescent intensity being higher than the average background of the stack by 5 standard deviations of the pixel values. The minimal distance between two peaks was limited by 3 pixels (304 nm). For each strain and condition, 11 stacks from two independent experiments were imaged. In each stack, at least 100 spots were detected, the total number of analyzed spots was between 3,400 and 4,300 per condition. In parallel, ten stacks from two independent experiments were imaged for cells not expressing any fluorophore. Spot intensity was corrected for the background outside of cells and the average fluorescence background in cells not expressing a fluorophore. Each stack was treated as an independent value. Spot detection and quantification for EGFP-YscQM218A at different mCherry-YscQC levels was performed using MicrobeTracker 0.937 [85] and the automated SpotFinderZ algorithm using the following settings: Expand cells, 2 px; max width squared, 22.3 px2; min width squared, 1.11 px2; min height, 0.053 i.u.; max rel. sq. error, 97; max var/sq. height ratio, 1.23; min filtered/fitted ratio, 0.02187; and standard settings otherwise. More than 170 cells per condition were analyzed.
The numerical data used in all figures (Figs. 1–4) are included in S1 Data.
|
10.1371/journal.pntd.0005546 | Tiger on the prowl: Invasion history and spatio-temporal genetic structure of the Asian tiger mosquito Aedes albopictus (Skuse 1894) in the Indo-Pacific | Within the last century, increases in human movement and globalization of trade have facilitated the establishment of several highly invasive mosquito species in new geographic locations with concurrent major environmental, economic and health consequences. The Asian tiger mosquito, Aedes albopictus, is an extremely invasive and aggressive daytime-biting mosquito that is a major public health threat throughout its expanding range.
We used 13 nuclear microsatellite loci (on 911 individuals) and mitochondrial COI sequences to gain a better understanding of the historical and contemporary movements of Ae. albopictus in the Indo-Pacific region and to characterize its population structure. Approximate Bayesian computation (ABC) was employed to test competing historical routes of invasion of Ae. albopictus within the Southeast (SE) Asian/Australasian region. Our ABC results show that Ae. albopictus was most likely introduced to New Guinea via mainland Southeast Asia, before colonizing the Solomon Islands via either Papua New Guinea or SE Asia. The analysis also supported that the recent incursion into northern Australia’s Torres Strait Islands was seeded chiefly from Indonesia. For the first time documented in this invasive species, we provide evidence of a recently colonized population (the Torres Strait Islands) that has undergone rapid temporal changes in its genetic makeup, which could be the result of genetic drift or represent a secondary invasion from an unknown source.
There appears to be high spatial genetic structure and high gene flow between some geographically distant populations. The species' genetic structure in the region tends to favour a dispersal pattern driven mostly by human movements. Importantly, this study provides a more widespread sampling distribution of the species’ native range, revealing more spatial population structure than previously shown. Additionally, we present the most probable invasion history of this species in the Australasian region using ABC analysis.
| The Asian tiger mosquito, Aedes albopictus, is an aggressive mosquito that has expanded globally in the last century, chiefly due to the increase of human movements. It is a major public health concern due to its role in transmitting dengue, chikungunya and Zika viruses. Understanding how populations of Ae. albopictus are genetically related and how they have been introduced into new regions is important for controlling them and assessing their disease risk; few studies have explored this in the Indo-Pacific. In our study, we sampled a broader range of populations of Ae. albopictus in the Indo-Pacific to explore genetic patterns and to investigate the likely route of invasion through Australasia. We uncovered clear genetic groups throughout this region, but also found that some geographically distant populations are closely related, likely due to human-associated movements. We also found, that Ae. albopictus likely colonized New Guinea from mainland Southeast (SE) Asia before spreading to the Solomon Islands via either PNG or SE Asia. In contrast, the populations in Australia’s Torres Strait Islands were introduced from Indonesia. Interestingly, we found major genetic changes over time in some Torres Strait populations, less than a decade after its introduction, potentially reflective of a random reduction in genetic diversity (genetic drift) or a secondary invasion.
| Many species of mosquitoes are amongst the most invasive pests in the world. They have a long history of human-mediated introductions [1, 2] that have resulted in the spread of major epidemics (malaria, dengue, Zika, etc.) and the establishment of invading mosquitoes as a biting nuisance. The Asian tiger mosquito, Aedes albopictus (Skuse 1894) [3], is regarded as one of the most invasive mosquitoes in the world [4]. Native to tropical and subtropical Asia and multiple Western Pacific and Indian Ocean islands, Ae. albopictus now has a pan-global distribution [5–9]. Its initial movement from Southeast (SE) Asia toward the Indo-Malayan Peninsula and Indian Ocean islands may have resulted from the increase in human migration during the 17th and 18th centuries, with international trade (particularly the used-tire and ornamental bamboo trades) further facilitating its global spread in the 20th century [10]. It is among the primary vectors of several globally expanding and medically important arthropod borne viruses (arboviruses)–particularly dengue, chikungunya, yellow fever, and Zika—while also able to transmit at least 23 other arboviruses and canine heartworm [10]. Whilst at present, the yellow fever mosquito Aedes aegypti (Linnaeus 1762) [11] is responsible for most of the transmission of some of these important arboviruses, the increased cold tolerance of Ae. albopictus relative to Ae. aegypti suggests that it could extend the range of many of these diseases under the right circumstances [12, 13].
Genetic techniques can provide critical information to infer the routes and sources of invasive species as well as informing on the demographic history and genetic composition of founding populations [8, 14]. For mosquito vectors, this knowledge is not only useful for inferring invasion routes in order to focus biosecurity efforts, it can also inform us of the colonizing capacity, adaptability and behaviour of invading mosquito lineages [15, 16]. However, the high dispersibility of Ae. albopictus mediated by human activities can make it challenging to detect genetic variation between populations due to high gene flow facilitated by these activities [17]. Microsatellite markers have been shown to be useful for exploring Ae. albopictus genetic patterns as they evolve rapidly and can often detect subtle population structure [8, 18, 19]. Recently, approximate Bayesian computation (ABC) analysis has proven a powerful tool in testing the probability of competing invasion scenarios. This can provide us with crucial information regarding the timing and origin of mosquito introductions, which has been used recently for Aedes mosquitoes [20–22].
Overall, the population genetics of Ae. albopictus through the SE Asian-Indo-Pacific region requires further exploration and samples from this region (particularly Australasia) are often lacking from global population genetic studies, despite the importance of this region for vector research [23–25]. A study by Beebe, Ambrose [18] explored part of the Indo-Australasian invasion by Ae. albopictus and provided an interesting scenario where there is high human connectivity (largely maritime) spanning both oceanic barriers and complex geographic landscapes. The current study expands on this work to include SE Asian native populations (Myanmar, Thailand, Malaysia, Singapore, Indonesia), as well as several younger populations that appear to have been introduced within the last six decades (Papua, Papua New Guinea (PNG), Solomon Islands, Fiji, Christmas Is., Cocos (Keeling) Is., Nauru, Torres Strait Islands (Australia)). Additionally, we included populations from the United States (USA) and northern Asia (only for COI) to see how these populations fit into a broader geographic analysis. We use previously developed nuclear microsatellite markers [18] and mitochondrial cytochrome c oxidase subunit I (COI) sequences to investigate the population genetics of the Ae. albopictus within the Indo-Pacific region. Our primary aims were to uncover the most likely historical invasion route of Ae. albopictus into the Australasian region as well as to detail the genetic connectivity and population structure of Ae. albopictus throughout this broad geographic region that we refer to as the Indo-Pacific (the aforementioned populations). While there are some records of the progressive establishment of Ae. albopictus throughout this region, the origin/s and the precise timing of introductions require testing using genetic methods under a coalescent-based approach such as ABC analysis. Our secondary aim was to further investigate the 2005 colonization of the Torres Strait Islands, Australia [26]. Many of the Torres Strait Islands have undergone intense spraying efforts since the establishment of Ae. albopictus and the region experiences monsoon-dry seasons leading to regular population bottlenecks [27, 28]. We hypothesized that the genetic changes in neutral alleles may be detectable over time in these newly invaded and small island populations.
Mosquito collections involving HLC from the Solomon Islands (S1 Table) were approved by the Medical Research Ethics Committee in compliance with Australia’s National Statement on Ethical Conduct in Human Research (project no. 2011000603). Collectors involved in HLC took anti-malarial medication and wore long-sleeved, protective clothing.
Both adult and larval samples were collected throughout Australasia, SE Asia, Indian Ocean and Pacific Ocean islands as well as in the United States (Table 1, Fig 1 (orange dots)). Samples were stored in 70% ethanol or dried (adults) over silica beads. Samples were collected using human landing captures (HLC), human baited sweep netting, egg collections and sampling of aquatic habitats for larvae and pupae (S1 Table). For identification purposes some samples were reared to adults after field collection (S1 Table). The logistics of sourcing material across multiple international borders resulted in variability in collection methods and sample sizes (Table 1, S1 Table). Adult mosquitoes were identified morphologically [29] and for larval/pupal samples using either real-time PCR assays [30] or a PCR-restriction digest [31] to differentiate from Aedes scutellaris (Walker 1858) [32].
DNA was salt extracted [33] and diluted at 1:10 with 1X TE buffer (Tris, EDTA). Thirteen nuclear microsatellite were used in this study. These markers were previously developed [18] and include two dinucleotide and 11 trinucleotide loci (see Beebe, Ambrose [18] for loci and primers). Some samples included from the previous study were amplified using a variation of the master mix (see Beebe, Ambrose [18]), other samples were amplified in a 15.4μl reaction that consisted of 10.8μl H2O, 3μl 5X Mytaq buffer (Bioline, containing 5mM dNTPs and 15mM MgCl2), 0.1μl 10μM M13 tagged forward primer, 0.2μl 10μM reverse primer, 0.2μl M13 tagged fluorescent dye (VIC, NED, PET or FAM; S1 Table), 0.01μl (1U) MyTaq polymerase and 1μl 1:10 DNA template. PCR cycling used the same protocol as in Beebe, Ambrose [18]. Amplification was verified by running 1μl of PCR product on a 2% agarose gel stained with either GelRed (Biotium) or MidoriGreen (Bulldog Bio). Samples that amplified successfully were sent to Macrogen Inc. (Republic of Korea) for genotyping.
GeneMarker v.2.4.2 (SoftGenetics LLC [34]) was used to score alleles for each locus manually after passing the data through the standardization run wizard using the default animal fragment setting. Random selections of genotyped plates were rescored by a second person to assess consistency in scoring. In addition to the data collected in this study, we included microsatellite scores from samples in Beebe, Ambrose [18]. During the present study, it was uncovered that the Beebe, Ambrose [18] study used (in some cases) inconsistent fluorescent dyes for a given locus, which caused a dye-shift [35] resulting in inconsistently scored alleles. We regenotyped a random subset of individuals from each of the populations used in the study by Beebe, Ambrose [18] to ensure consistency with data collected from this study. The predictability of this dye-shift (based on the dyes used previously) enabled shifting of the allele scores from Beebe, Ambrose [18] for use in this study. Samples with fewer than nine scored loci of 13 total were removed before further study as we considered these poor quality samples; thus leaving 911 samples for final analyses (20% of samples were from Beebe, Ambrose [18]; Table 1). With the remaining dataset, missing values were replaced based on population allele frequencies using GenoDive v. 2.0b27 [36]–based on preliminary analyses this did not significantly alter population structure and relationships between populations. Missing values were not replaced for the STRUCTURE analysis, calculation of HWE and for checking the presence of null alleles.
Scored allele frequencies were checked for the presence of null alleles using MICRO-CHECKER [37] and for Hardy-Weinberg equilibrium using GenAlEx v.6.5 [38, 39]. Additionally, we calculated fixation index (F), allelic richness (Na), number of effective alleles (Ne) and the observed (Ho) and expected (He; unbiased estimate: uHe) values of heterozygosity using GenAlEx v.6.5. Pairwise population indices of genetic variation for Jost’s D, G”ST and FST were also calculated between populations in GenAlEx v.6.5 (S2 Table). We used 9,999 permutations and an analysis of molecular variance (AMOVA) to assess significance. A Mantel test was also performed in GenAlEx v.6.5 on geographic and genetic distance (pairwise phiPT) using 9,999 permutation [38, 39].
Population structure was investigated using the Bayesian program STRUCTURE v.2.3.4 [40] to infer the most probable number of population clusters (K). Based on our preliminary runs (S1 File, Supplementary Methods: Preliminary STRUCTURE), final analyses were run at both a lower (K = 4) and upper (K = 9) K value. For both K values we used a burn-in of 100,000 and runtime of 2,000,000 generations per iteration (20 iterations). For K = 9, cluster membership probabilities were somewhat inconsistent across runs due to multimodality; 20 iterations helped to account for this [41]. We assessed whether the burn-in period was adequate by reviewing summary statistics in STRUCTURE [41]. CLUMPP v.1.1.2 [42] was used to compile data from the 20 iterations for the independent values of K using the Greedy algorithm with 1,000 replicates. Final graphs were formed in DISTRUCT v.1.1 [43].
Discriminant analysis of principal components (DAPC) and correspondence analysis (CA) was used to further assess population structure. DAPC was implemented in R Studio v.3.2.2 (RStudio Team 2015 [44]) using the adegenet 1.4–1 package [45, 46] using the whole microsatellite dataset, where group membership was defined by the populations outlined in Table 1 (see DAPC abbreviation). These populations were more broadly defined and differed slightly from those used in STRUCTURE, to allow for easier interpretation of the results presented here. Specifically, the Torres Strait populations were split into groups based on their genetic relationship to one another and geographic/temporal information to reduce clutter in plots (Table 1; DAPC abbreviation). Only populations that were genetically similar were grouped together, which was confirmed using STRUCTURE, DAPC, CA and pairwise tests for genetic distance (FST, G”ST, Jost’s D) on the full dataset and subsets. Final DAPC analyses were performed on both a full dataset (nind = 911, npop = 23; including all populations) and a reduced dataset (nind = 458, npop = 18; excluding Jakarta, Sumba, Timor-Leste, the Torres Strait Islands and Southern Fly Region) to help discriminate genetically similar populations.
In adegenet, cross-validation was performed on each of our DAPCs independently, using a training dataset of 90% and a validation set of 10%, using 100 replicates. The number of PCs (n.pca) associated with the lowest root mean squared error (RMSE) was used as this was considered optimum [47]. Cross-validation suggested retaining 60 PCs for the full dataset and 40 PCs (n.pc) for the reduced dataset. We used five discriminant functions (n.da) for each of the analyses, but only the first three are plotted and discussed here as they explained the majority of variance (see Results). A correspondence analysis (CA) was also implemented in adegenet on the full dataset to investigate general trends and to complement DAPC, as visualization of the data is simplified in CA because within-population genetic diversity is not displayed.
The Garza-Williamson index (M-ratio) was calculated for populations using our microsatellite dataset in Arlequin v.3.5.2.2 [48]. The M-ratio was used to investigate the demographic history of populations and to test for recent bottleneck events; wherein an index statistic closer to 1 suggests the population is in a stationary state whereas very low values suggests a population has gone through a genetic bottleneck in the past (with a critical value of 0.68 indicating a bottleneck) [49, 50]. A Wilcoxon test for heterozygosity excess was also conducted on populations to detect bottlenecks using BOTTLENECK v.1.2.02 [51]. We used a two-phase model (TPM) of mutation with 10% infinite allele model and a 90% single step mutation model with 15% variance for 1000 iterations. A Wilcoxon signed rank test (two-tailed) was used to calculate significance (P < 0.05).
We tested the invasion history of Ae. albopictus in part of the study region (SE Asia/Australasia)—populations from the Indian Ocean, Fiji, Nauru and United States were not included due to insufficient sampling of these regions. Both COI sequences and the thirteen microsatellites loci were analysed together using ABC analysis in DIYABC v.2.1.0 [52]. Due to the complexity of modelling each population separately in this region, we simplified our invasion scenarios by randomly subsampling individuals from distinct geographic and genetic groups (defined using our other analyses). These representative populations included: mainland SE Asia, Indonesia, Papua, PNG, the Solomon Islands and the Torres Strait Islands/Southern Fly Region. For both the PNG and the Torres Strait/Southern Fly Region populations, we included temporal sampling in our scenarios (asterisks, Fig 2). Each of these representative populations/sampling events was made up of 30 individuals except Papua which used all 20 samples from the only sampled population, Timika. In addition, we included an unsampled ancestral population (ANC, Fig 2) in our model that split into mainland SE Asian and Indonesian populations.
Preliminary runs were carried out in accordance with Bertorelle, Benazzo [53] in order to optimise summary statistics, prior estimates and the scenarios tested. Final runs compared five invasion scenarios; the final summary statistics, prior estimates and parameter conditions used are outlined (S3 Table). Each scenario represents a plausible invasion route into the Australasian region. These were constructed using historical records regarding the timing and suspected sources of the different invasions [26, 54–58]. Priors were sampled from a wide range of distributions based on these records—less certain time priors were given a wider distribution and standard deviation whereas more likely priors were assigned narrower estimates. The upper time bound for the divergence of mainland SE Asia and Indonesia from a common ancestor was based on Porretta, Mastrantonio [59] while the lower bounds allowed for the possibility of a more recent split associated with human migration [10] (S3 Table). Because DIYABC measures time in terms of the number of generations, we assumed 10 generations per year for Ae. albopictus (which typically ranges from 5–17 generations in the tropics).
Estimates of effective population size (Ne) ranged from 10 to 1,000,000 individuals (uniform distribution) [20–22] depending on if a population was modelled as going through a change in Ne. For instance, each of the recently introduced populations was modelled to allow for a founder effect after its introduction using lower Ne ranges (Fig 2, S3 Table). We additionally allowed for a change in Ne in the Torres Strait Islands/Fly Region population due to the drastic temporal changes we observed in our other analyses—this would allow us to make a relative comparison of Ne in order to see if the population had undergone any change in Ne (indicative of bottleneck/expansion events) (Fig 2, S3 Table). For COI, we used the HKY mutation model [60] and sampled from a uniform distribution with mutation rates ranging between 7x10-10–1x10-7. For microsatellites, both di- and tri- nucleotide repeats were modelled separately due to the possibility of different repeat lengths having different mutation rates [61]. Both microsatellites used the default generalized stepwise mutation model and were assigned a loguniform distribution with mean mutation rates between 1x10-6–1x10-3 (S3 Table). All mutation rates were based on standard ranges for COI [62] and for Dipteran microsatellites [20–22, 63]. We simulated 15,000,000 datasets and each of the five scenarios was given a uniform probability. The performance of the ABC approach was assessed using multiple methods in DIYABC (S1 File, Supplementary Methods: Performance of DIYABC).
The mitochondrial protein-coding gene COI was amplified using custom designed primers (albCOIF 5’-TTTCAACAAATCATAAAGATATTGG-3’ and albCOIR 5’- TAAACTTCTGGATGACCAAAAAATCA-3’) for 259 random individuals across different populations. Each 25.3μL reaction consisted of 19μl H2O, 5μl 5X Mytaq buffer (Bioline, with pre-optimized concentrations of dNTPs and MgCl2), 0.1μl 100μM forward primer, 0.1μl 100μM reverse primer, 0.1μl MyTaq polymerase and 1μl 1:10 DNA template. PCR used an initial denature of 94°C for 3 min, 35 cycles of denaturation at 95°C for 30 sec, primer annealing at 45°C for 40 sec, and primer extension at 72°C for 30 sec. Final elongation lasted 5 min at 72°C prior to cooling to 4°C. Amplification was confirmed using gel electrophoresis (as described previously) and PCR products were purified by adding 2μl per sample of a mixture containing equal amounts of Exonuclease I and Antarctic Phosphatase (New England Biolabs, Australia) before incubation at 37°C for 20 min and denaturation at 80°C for 10 min. Samples were sequenced by Macrogen Inc. (Republic of Korea) using Sanger sequencing. Additional COI sequences of Ae. albopictus were obtained from other studies and from Genbank (1044 sequences total, 259 produced in this study, S4 Table).
Sequences were edited and aligned in Geneious v.9.0.4 (http://www.geneious.com, Kearse, Moir [64]) using the MAFFT alignment. The final alignment was trimmed to 445bp to incorporate the large number of COI sequences available from Genbank which were smaller than the ~700bp region sequenced in this study. All sequences were checked for stop codons in Geneious v.9.0.4. TCS haplotype networks [65] were constructed using 1,000 iterations in PopArt v.1.7 (http://popart.otago.ac.nz). In addition, we calculated Tajima’s D for populations with temporal data in PopArt v.1.7 to determine whether sequences were evolving randomly or non-randomly. Haplotype and nucleotide diversity was calculated using DnaSP v.5.10.1 [66].
All new COI sequences generated in this study are available on Genbank: KY907195—KY907453.
Allelic richness for microsatellites was highest in native populations of Ae. albopictus from Myanmar, Thailand, Malaysia, but also high in recently invaded areas such as some of the Torres Strait Islands and in PNG and the Solomon Islands (See Na in S5 Table). A Mantel test on the whole dataset showed a significant (P = 0.0001) positive, but weak, correlation (R2 = 0.02) between genetic (phiPT) and geographic distance (y = 0.0002x + 28.1). Pairwise estimates of FST, Jost’s D and G”ST all revealed similar results to each other and recovered mostly significant relationships between populations; here we discuss gene flow and genetic distance in regards to FST estimates but the other measures are shown in S2A–S2C Table. Lowest FST values were apparent between populations belonging to the same geographical region (for definition of regions see Table 1, Region/description), especially within mainland SE Asia (FST = 0.011–0.103) (S2A Table). However, some comparisons between regions separated by vast geographical distances also showed low FST scores, such as populations from the Solomon Islands with populations from mainland SE Asia (FST = 0.050–0.114) (S2A Table). The relationships between populations were mostly consistent with the results obtained in STRUCTURE and multivariate analyses, which are described in detail below.
Within the study region, four to nine clusters were supported by the Evanno ΔK and log likelihood methods for inferring K. While K = 4 (Fig 1) represents the simplest summary of the genetic structure of Ae. albopictus in the region, we detected substantial substructure within these four main clusters which are apparent at K = 9 (S1 Fig). We discuss the data in the context of both values of K to avoid underestimating the degree of population structure within the study region.
At K = 4, clusters mostly pertained to distinct but broad geographic boundaries although many populations and individuals show signs of admixture, despite the large geographic distances (Fig 1). The mainland SE Asian populations of Myanmar, Thailand, Malaysia and Singapore cluster with the USA (Hawaii and Atlanta; Fig 1B and 1C), La Réunion, as well as Fiji and Nauru (light purple; Fig 1). The second cluster (light green) contains populations from Indonesia (Jakarta and Sumba), Timor-Leste, the Southern Fly Region of PNG (Fig 1A) and several islands of the Torres Strait (especially collections following the first detection of Ae. albopictus in the straits in 2005 (collections between 2006–2014)) (Fig 1A). An additional cluster (purple) is prominent within the Torres Strait region (Fig 1A) and represents populations on the islands collected more recently (2013–2015), suggesting temporal shifts in population structure have occurred on some islands (see Fig 1A: Ker, War). The fourth cluster (green) is composed of historically-established PNG populations, but note that some of these populations contain admixture with the SE Asian cluster (Fig 1). The island of Daru (Fig 1A: Dar), which is less than 5km from the Southern Fly Region incursion populations (light green cluster; Fig 1A; Sig, Kul, Mbd, Kat), is distinct and clusters with the historically established PNG populations. Timika and the Solomon Islands show genetic affinity to both PNG and SE Asian clusters (Fig 1). Indian Ocean islands (Fig 1: CK and CH) appear differentiated from each other and contain a notable degree of admixture, but Christmas Is. is more similar to SE Asia whereas Cocos (Keeling) Islands appear as an admixed population made up of the PNG and Indonesian clusters. When K = 9, the same broad population patterns are observed but some populations/regions become more distinct, including the USA and Hawaii, Solomon Islands, Timika, the Cocos (Keeling) Islands and Sumba/Torres Strait Island populations (S1 Fig). Relationships for this K value are described in S1 File, Supplementary Results, STRUCTURE (K = 9).
For multivariate analyses, the DAPC on the full dataset (n.pc = 60, n.da = 5) explained 89% of variance, whereas the reduced dataset (n.pc = 40, n.da = 5) explained 79.3% of the variance in the data. Eigenvalues for these first three PCs are 206.37, 110.78 and 63.21 for the full dataset (S2A Fig) and 99.6, 69.2 and 44.99 for the reduced dataset (S3 Fig); these values correspond to the ratio of between-group over within-group variance for each discriminant function. For the correspondence analysis (CA) we plotted the first three eigenvalues (0.19, 0.15, 0.09), which indicate the proportion of variance explained by the first three PCs (S2B Fig).
DAPC and CA results of the full dataset (S2 Fig) showed similar population differentiation as observed in STRUCTURE at K = 4. Due to the large number of populations, we describe population structure based on the broad clustering—populations are color coded with geographically close populations being more similar in color. Four major clusters of populations are noticeable when the first three PCs are plotted against each other (C1-C4, S2A Fig). However, there is considerable overlap between these clusters, particularly with C4 overlapping C2 and C3, suggesting that these individuals and populations are genetically similar and show signs of admixture (S2A Fig). Cluster 1 (C1) represents recent (2012–2015) collections from Torres Strait Islands and is the most distinct from the other clusters. It is most closely related to C2, which contains earlier collections (2007–2014) from the Torres Strait Islands and populations from the Southern Fly Region, Sumba, Timor-Leste and Jakarta (S2A Fig). The relationships uncovered in the DAPC of the full dataset were recovered in the CA and are more easily visualized, where the first three principal components (PCs) of the CA are plotted in three-dimensions (S2B Fig); however, note that the large amount of within-population variation (as displayed in the DAPC plots) is not shown.
Because of the overlap of C3 and C4 clusters, we separately analysed these clusters by DAPC (referred to in the Materials & Methods as the reduced dataset) that excluded populations from the Torres Strait Islands, Jakarta, Sumba, Timor-Leste and the Southern Fly Region (S3A and S3B Fig) to explore substructure within these clusters (i.e. C3 & C4 in S2A Fig). Populations from PNG (Kiunga, Madang, Port Moresby and Daru) were similar to each other but distinct from the other populations (S3 Fig). The offshore PNG populations (Lihir Is. and Buka Is.) were somewhat differentiated from mainland PNG populations, although Buka Is. shares some overlap with both Port Moresby and Lihir Is. (S3 Fig). The Solomon Islands also appears similar to Lihir Is. and Buka Is. populations, but is somewhat distinct (S3 Fig). Mainland SE Asian populations appear genetically similar and tend to exhibit the most genetic overlap with other populations (S3 Fig). Nauru and Fiji are most similar to mainland SE Asian populations. In contrast, both USA populations (Hawaii and Atlanta) as well as La Réunion appear well differentiated from mainland SE Asian populations. Cocos (Keeling) Island and Timika share some overlap with each other, whereas Christmas Island shares overlap with both PNG and mainland SE Asia populations (S3 Fig).
The scenario with the highest posterior probability using the logistic approach was scenario 4 (P = 0.52 [95% CI: 0.44, 0.59], Fig 2). In this scenario, Ae. albopictus colonized Papua and PNG in two separate events from mainland SE Asia, established in the Solomon Islands via PNG and more recently colonized the Torres Strait Islands/Southern Fly Region via Indonesia (Fig 2). The timing of each introduction event is shown in S3 Table and corresponds with historical records for the introduction of Ae. albopictus in the tested populations, although 95% confidence intervals (CI) suggest that some introduction dates could have been earlier than first observed (see Discussion). None of the other scenarios showed overlapping 95% CI with Scenario 4 (Fig 2), however, we detected moderate levels of type I (0.45; probability that scenario 4 is rejected given that it is the ‘true’ scenario) and type II error (0.48; probability of deciding scenario 4 is the ‘true’ scenario when it is not) that suggest Scenario 1 (P = 0.34 [95% CI: 0.31, 0.37]) could provide a plausible alternative invasion scenario for our data (S3 Table). Scenario 1 is identical to scenario 4, except that the Solomon Islands is modelled as originating from mainland SE Asia, rather than from PNG (Fig 2). Consequently, we discuss the Solomon Islands invasion based on both alternative origins and posterior estimates calculated under the combined scenarios are presented (in Discussion), although show no significant difference from scenario 4 alone (S3 Table). A preliminary analysis showed low support (P = 0.001 [95% CI: 0.00, 0.12]) for a scenario where the Solomon Islands introduction was modelled as an admixture event between mainland SE Asia and PNG compared to the five scenarios compared in our final analyses (but using slightly different Ne prior ranges for all founders (10–10,000)).
Each introduced population showed no relative change in Ne due to large 95% CIs of posterior distributions, although median Ne values were smaller for founding events (S3 Table). Likewise, the duration of the modelled bottleneck had large 95% CIs, but median values generally ranged from 16–30 generations (S3 Table). Overall, the Torres Strait Islands/Southern Fly Region population showed a stable Ne since its introduction (due to overlapping 95% CIs), although there was a gradual increase in median Ne over time, potentially suggesting growth of the population as a whole. (S3 Table). Our assessment of the performance of our ABC analysis was supported as fitting our observed data well (S4 and S5 Figs).
A total of 52 COI haplotypes were identified from the 1044 individuals used for generating the TCS haplotype network, with 92% of individuals belonging to nine main haplotypes (H1-5, 11, 15, 39, 43) (Fig 3, Table 2, S6 Fig). The distribution of these haplotypes by specific population is shown in S6 Table. All new sequences generated from this study are available on Genbank (Accession no. KY907195—KY907453; see S4 Table for accession numbers of sequences from other studies). The COI haplotype network is less informative in regards to population structure than the microsatellite data, although it does highlight broader geographic relationships that are somewhat consistent with the microsatellite results. Of the nine main haplotypes, H1 has the most individuals, mostly from eastern Asia (China, Taiwan and Japan), USA (mainland USA and Hawaii), Madagascar and La Réunion (Fig 3, Table 2). Similarly, H39 is distributed in a similar temperate/subtropical region. Haplotype 3 consists primarily of individuals from the Torres Strait Islands, Fly Region, Indonesia, Timor-Leste, PNG and the Philippines. However, the majority of PNG sequences belonged to H5, which also includes individuals from the Solomon Islands, Indian Ocean islands (CK and CH), Singapore and Thailand. Another major haplotype, H4, includes the most diverse range of populations (in terms of geographic spread), although it mostly consists of mainland SE Asian populations and populations from the tropics. Of the additional COI haplotypes, many are exclusive or shared amongst close geographic regions (Fig 3, Table 2, S6 Table, S6 Fig), although others show no apparent geographic pattern.
Haplotype diversity (Hd) for the total dataset was high (Hd = 0.83) as was nucleotide diversity (π = 0.0037); however, population measures of Hd and π varied considerably (Table 3). Neutrality tests (Tajima’s D) on all populations with temporal data were not significant (S7 Table).
Past genetic bottleneck events were not consistently indicated using both Wilcoxon tests for heterozygosity excess and M-ratio, with the exception of the 2010 population from Waiben of the Torres Strait (M = 0.68, P = 0.032 (two-tailed Wilcoxon signed rank test for heterozygosity excess) (S8 Table). However, multiple populations showed signs of a bottleneck using a single method. The M-ratio indicated a bottleneck for some populations of the Torres Strait Islands (Mabuiag, Waiben, Ngurupai, Poruma, Iama), Port Moresby, Timika and Gizo, whereas the Wilcoxon test was significant for Waiben, Madang, Jakarta, Yangon and the Cocos (Keeling) Islands (S8 Table).
The population structure and genetic connectivity of Ae. albopictus within the Indo-Pacific region has been limited to a few studies that only examined regional structure or had restricted sampling within the species’ range. In addition, the genetic characteristics of some populations examined in this study have been unexplored (e.g. Solomon Islands, many Indian and Pacific Ocean Islands, Papua-Indonesia). Using multiple lines of evidence (and both microsatellite and mitochondrial markers), we show high spatial genetic structure throughout this region. We used coalescent ABC analysis to test for the first time the likely invasion route of Ae. albopictus into the Australasian region and uncovered that the species likely invaded New Guinea from mainland SE Asia and the Solomon Islands via either PNG or SE Asia. We also show the recent invasion of Ae. albopictus into northern Australia’s Torres Strait region and Southern Fly Region of PNG likely originated from Indonesia, as previously suspected [18]. Furthermore, we provide evidence of rapid temporal shifts in population structure occurring less than a decade after the Asian tiger mosquito’s introduction into Australia’s Torres Strait Islands in 2005 [26]. In contrast, historically-introduced and native populations of Ae. albopictus showed less spatial population structure at a regional level, despite large geographic distances and international boundaries between some of these populations. Importantly, this study provides a widespread sampling distribution of the species’ native range and revealed more spatial population structure than previously shown, as well as evidence for rapid temporal genetic change in newly established populations in the Torres Strait Islands.
Multiple population studies have attempted to capture the amount of genetic structure throughout the species’ native range. Allozyme studies have shown that Indonesian and Japanese populations of Ae. albopictus are likely distinct [67] and that SE Asia (Borneo, peninsula Malaysia) and southern Asian populations (India, Sri Lanka) can both be differentiated from northern Asian populations (China, Japan) [68]. While no studies have conducted a comprehensive analysis of the species’ full native range [8], the genetic differentiation of native Asian populations of Ae. albopictus may confer to both a north-south (Korea to Indonesia) and east-west (Japan to India) pattern of genetic differentiation. Our results partly support this pattern, with evidence for genetic differentiation separating northern Asia (COI data only, no microsatellite data available), SE Asia and Indonesia (both COI and microsatellite data). Within mainland SE Asia, our data revealed little to no population structure despite high genetic diversity and COI haplotype diversity, supporting the findings of other studies [59, 69, 70]. Using climatic modelling and two mitochondrial markers, Porretta, Mastrantonio [59] suggested that the low genetic structure across this mainland SE Asian region could be explained by the demographic growth between interconnected populations of Ae. albopictus preceding the last glacial maximum (LGM, occurring ~21,000 ybp), with the species’ ecological flexibility facilitating its success in the ecologically diverse Sundaland (exposed SE Asian landmass) during this period. Whilst their study lacked sampling from Indonesia, their data suggested climatically suitable habitat for Ae. albopictus existed across the southern range of Sundaland which later formed the Indonesian islands after a rise in sea levels. They hypothesized that the emergence of Sundaland during the LGM could have facilitated population connectivity across Indonesia and mainland Asia. In contrast, we found clear genetic differentiation between the Indonesian archipelago and mainland SE Asia with both our microsatellite and COI data, which could be explained by fragmentation and subsequent differentiation of Ae. albopictus populations following a rise in sea levels or potentially driven by human migration from the region [10]. Our ABC analysis supports a more historical split ~9,040 ybp (95% CI = 1,790–27,600 ybp) but further investigation with extensive native population sampling would be needed to clarify dating. The genetic homogeneity and high gene flow (FST = 0.020–0.103) in our microsatellite data across mainland SE Asia (including Malaysia, Singapore, Myanmar and Thailand) could also be explained by human-mediated gene flow associated with transportation infrastructure (land, air and sea); aircraft and road networks in particular may be major drivers in the connectivity of Ae. albopictus in this region given the inland location of many of these populations [71–73]. For our full dataset, the relationship between genetic and geographic distance was significant, however, the correlation was weak (R2 = 0.02), highlighting the potential extant that human movements have had on Ae. albopictus population structure; although it does also highlight a minor trend of isolation by distance in our study region.
A recent worldwide study that examined the mitogenome diversity of Ae. albopictus found three major haplogroups, two of which were implicated in the global spread of the species [9]. Some of these haplogroups appeared more prevalent in particular climatic and geographic regions, such as haplogroup A1a which chiefly characterized the tropics and A1a2 which is mostly distributed in temperate regions. Another haplogroup (A2) appeared important in the spread of Ae. albopictus from SE Asia toward Australasia, distinguishing many of the samples from the Philippines, PNG, Indonesia and the Torres Strait Islands [9]. A similar pattern was observed using our COI data when viewing the nine haplotypes that account for 92% of the sequences in our study. Haplotype three (H3) was found to be common in Philippines, Indonesia, the Torres Strait Islands and Southern Fly Region but also present in historically established PNG populations, Vietnam and an individual from Sepilok, Borneo (Malaysia). Consequently, the populations from Vietnam, Borneo, the Philippines and Sulawesi could represent important unsampled populations that may influence ABC results and should be considered in future studies, which could additionally explain why the posterior probability of our most likely invasion scenario was moderate (P = 0.52). We also found that H1 was more prevalent in regions that experience temperate and subtropical climates, while H4 was widespread in the tropics (Fig 3). Due to the maternal inheritance of mtDNA, these results could suggest that the movement of females has been somewhat limited to their preadaptation to certain climatic regions. For example, females originating from a temperate region may be more being likely to successfully invade other temperate regions. It is possible that these genetic patterns for COI in Ae. albopictus are associated with the photoperiod response of different populations, which could contribute to higher gene flow and invasion success between climatically similar regions [17, 74–76]. However, the photoperiodic response of Ae. albopictus in recently introduced regions within Australasia is lacking (especially for New Guinea, the Solomon Islands, Fiji and Nauru). Within the Torres Strait Islands it appears the population is of tropical, Indonesian origin and egg survival was lower in less humid conditions [77], supporting this conclusion. However, it is worth noting that there are multiple other COI haplotypes that do not conform to any obvious geographic patterns (S6 Fig), suggesting that there have been multiple introduction events into some locations from mainland Asia or potentially unsampled locations—a similar pattern which was also highlighted by other studies [9, 74]. The possibility of insertion of mtDNA in the nuclear genome [78, 79] was considered in this study and our COI sequences were assessed by examining chromatograms (no double peaks in chromatograms) and by checking for stop codons. It seems unlikely that there is nuclear insertion of mtDNA in samples from our study—but it remains a possibility and requires further research.
We included several populations from the USA and Indian Ocean to explore how these populations fit into a broader geographic analysis with our samples, which are chiefly from SE Asia and Australasia. We did not include these in our ABC analysis because of the lack of temperate Asian populations, which have been shown as the source of USA introductions and because of insufficient sample sizes in Indian Ocean populations (Cocos (Keeling) Islands, Christmas Island and La Réunion). We included them in our other analyses to assist in future studies and have discussed them in S1 File, Supplementary Discussion.
For the first time, we have used ABC analysis to compare various invasion scenarios of Ae. albopictus in the Australasian region and have additionally explored and characterised the genetic structure of a wide range of populations in the Indo-Pacific, which had not been compared under a single study. We uncovered notable temporal population structure in recently introduced Torres Strait Island populations. Importantly, this demonstrates some of the drastic changes that invading populations may undergo within a short time period (i.e. in less than a decade), which has substantial implications on the practicality and accuracy of using such genetic databases for estimating invasion sources. However, we also found that historically established populations of Ae. albopictus displayed stable population structure. Future studies that aim to address the global genetic structure of Ae. albopictus will need to consider the full native range of the species and the influence of temporal collections on population structure, especially newly established populations. Likewise, ABC analyses that account for complex scenarios (requiring thorough spatio-temporal sampling) and gene flow between populations will play a key role in better understanding the population dynamics of Ae. albopictus as well as other mosquito species that are highly associated with humans, such as Ae. aegypti. The standardisation of genotyping methods and sampling efforts will allow for more rigorous assessments of the global population structure of Ae. albopictus, given the scale of its current distribution which makes such population studies logistically challenging. This will prove essential in controlling the spread of Ae. albopictus and for assessing the health risk of different populations, given their variation in vector efficiency, physiology and behaviour [10, 86–88].
|
10.1371/journal.pgen.1003167 | Mutational Spectrum Drives the Rise of Mutator Bacteria | Understanding how mutator strains emerge in bacterial populations is relevant both to evolutionary theory and to reduce the threat they pose in clinical settings. The rise of mutator alleles is understood as a result of their hitchhiking with linked beneficial mutations, although the factors that govern this process remain unclear. A prominent but underappreciated fact is that each mutator allele increases only a specific spectrum of mutational changes. This spectrum has been speculated to alter the distribution of fitness effects of beneficial mutations, potentially affecting hitchhiking. To study this possibility, we analyzed the fitness distribution of beneficial mutations generated from different mutator and wild-type Escherichia coli strains. Using antibiotic resistance as a model system, we show that mutational spectra can alter these distributions substantially, ultimately determining the competitive ability of each strain across environments. Computer simulation showed that the effect of mutational spectrum on hitchhiking dynamics follows a non-linear function, implying that even slight spectrum-dependent fitness differences are sufficient to alter mutator success frequency by several orders of magnitude. These results indicate an unanticipated central role for the mutational spectrum in the evolution of bacterial mutation rates. At a practical level, this study indicates that knowledge of the molecular details of resistance determinants is crucial for minimizing mutator evolution during antibiotic therapy.
| Natural and laboratory populations of bacteria can readily give rise to strains with high mutation rates. The evolution of these mutator bacteria—of particular concern in clinical situations—has been understood exclusively in terms of their increased capacity to generate beneficial mutations, such as those that confer antibiotic resistance. Current models, however, have largely overlooked that each mutator allele increases only characteristic types of mutations, a prominent fact whose evolutionary consequences remain unexplored. Using laboratory Escherichia coli populations, we show that this mutational bias determines the competitiveness of different mutators across environments. Computer simulation showed that this effect can markedly influence the evolutionary fate of mutator alleles. These results indicate that this unrecognized factor can be a major determinant in the evolution of mutator bacteria and suggest future experimental approaches for improving antibiotic therapy design.
| Despite their increased load of deleterious mutations [1], [2], mutator strains of bacteria are isolated routinely in laboratory and clinical settings [3]–[8]. Theory [9]–[11] and experiments [5], [12], [13] explain these observations as a consequence of genetic hitchhiking, whereby mutator alleles reach high frequency by being co-selected with linked beneficial mutations. Mutator evolution is therefore dependent on the absence of horizontal gene transfer [14] and the availability of adaptive mutations with substantial fitness effects [11] – conditions frequently met during adaptation to a host or during antibiotic therapy, which have been invoked to explain the prevalence of mutators among pathogenic bacteria [15]–[17].
Whereas genetic hitchhiking provides a satisfactory mechanism to explain how mutator bacteria can be selected, the precise mechanistic details of this process are still a matter of research. Some authors emphasize that mutator fixation involves many consecutive hitchhiking events, prompted by frequent environmental shifts [9], [18], [19] or by the concurrence of multiple beneficial mutations [10], [11]. Other authors, in contrast, consider mutator success as primarily the result of a single step in which one hitchhiking event takes mutator frequency from rareness to fixation [20]. Despite this theoretical effort, two major observations remain to be accounted for. First, although hitchhiking probability is predicted to increase with the extent of adaptive opportunity offered by the environment (i.e., the number and effects of available beneficial mutations) [14], [20], , selection experiments with comparable adaptive opportunity report contrasting frequencies for mutator emergence [4], [12], [22], [23]. Second, mutators of different strength are predicted to emerge according to the degree of adaptive opportunity [14], [21]; nonetheless, both clinical and laboratory observations show a marked bias towards strong mutators, particularly those caused by defects in the mismatch repair system (MMR) [17], [24]. To explain these deviations, it is argued that mutator alleles of other mutator strengths might be under-represented due to putative fitness disadvantages, and that MMR mutator mutants might be over-represented because they can also be selected for their high recombination rates [24].
An additional, previously unrecognized factor is that each mutator exhibits its own mutational spectrum, that is, mutation rate increase is limited to characteristic types of mutations [25]. This bias stems from the type of mutation avoidance mechanism that is altered in each mutator genotype; depending on which mechanism is affected, its failure will lead to a preferential increase in specific types of transitions, transversions or frameshifts [25]. In line with previous suggestions [12], [26], we hypothesized that if mutators and wild-type strains have differential access to specific beneficial mutations, they might produce mutants with different fitness levels, which could influence their evolutionary dynamics. Mutator alleles that on average generate stronger beneficial mutations will have a better chance to achieve fixation; conversely, those that more often produce weaker adaptive mutations will have limited spread. The fixation probability of a mutator allele would thus depend not only on its mutation rate, but also on its mutational spectrum.
Here we test these predictions by characterizing beneficial mutations from wild-type and knockout strains of the mutT (ΔmutT) and mutY (ΔmutY) antimutator genes of Escherichia coli. mutT-defective strains show a strong mutator phenotype, leading specifically to A·T→C·G transversions [25], whereas mutY defects lead to a moderate mutator phenotype that increases G·C→T·A transversions [25]. Both mutators were reported to arise spontaneously in evolution experiments with E. coli [4], [12]. Since beneficial mutations are exceedingly rare and difficult to detect, our experimental design focused on mutations that confer antibiotic resistance, which are particularly suitable for this kind of study. We show that according to their mutational spectra, wild-type and mutator strains can generate distinct fitness distributions for antibiotic-resistant mutants. Notably, these dissimilarities can significantly alter the competitive abilities of mutators against the wild-type strain, as measured in direct competition experiments. Furthermore, using computer simulation of a simple population genetics model, we show that the hitchhiking dynamics is highly sensitive to average fitness deviations of the beneficial mutations with which mutator alleles hitchhike.
To test whether mutational spectrum differences translate into significant fitness differences, beneficial mutations were selected by plating cultures of wild-type, ΔmutT and ΔmutY strains in three antibiotics: rifampicin, streptomycin and tetracycline. Rifampicin- and streptomycin-resistant mutants showed marked colony size polymorphism (Figure 1A). Resistance to these antibiotics arises readily through alterations in their targets, the RNA polymerase [27] and the 30S ribosomal subunit [28], respectively. Since many alterations can affect the function of these essential machineries to different degrees, it is unsurprising that resistance mutants show growth differences. Consistent with the predictions of our hypothesis, wild-type and mutator strains displayed distinct colony size polymorphism in each antibiotic (Figure 1A). To test whether access to beneficial mutations with varying selection coefficients was the only factor responsible for differences among strains, we used strains with an insertion bearing a tetracycline-resistance gene and its constitutive repressor [29]. As the probability of achieving resistance to high tetracycline concentrations through point mutation in E. coli is negligible (mutant frequency <10−9), resistance arises here only through loss-of-function mutations, which relieve the repression. All of these loss-of-function mutations can be considered equivalents, and no fitness differences are thus anticipated among tetracycline-resistant mutants. Indeed, these mutants showed little growth polymorphism (Figure 1A). Small differences, apparently independent of the mutator background, probably reflect phenotypic lag or other sources of phenotypic variability.
To quantify the fitness distribution of the mutants generated by each mutator, we randomly selected 42 independent resistant colonies from each combination of genotype and antibiotic, and measured growth rate as a proxy for Darwinian fitness. It is important to remark that we are dealing with the fitness distribution after selection, not the intrinsic fitness distribution; and so it should be understood in what follows. In rifampicin, the distribution of mutants generated by the ΔmutY strain is shifted toward higher growth rates compared with both those of wild-type and ΔmutT strains (Figure 1B, centre), whereas in streptomycin the situation is the reverse (Figure 1B, right) (n = 42, P<0.0001, Kolmogorov-Smirnov's one-sided two-sample test, in all cases). These results suggested that G·C→T·A substitutions in the rpoB gene (the characteristic transversion increased in this mutator) produced alleles encoding a high-fitness rifampicin-resistant RNA polymerase; coincidentally, the same transversion generated rpsL alleles encoding a low-fitness streptomycin-resistant ribosomal protein S12. Similar reasoning could be applied to the ΔmutT strain results, which preferentially raises the A·T→C·G transversion. In the case of tetracycline, no significant differences were found between wild-type and any mutator strains (Figure 1B, left) (n = 42, P>0.18, Kolmogorov-Smirnov's two-sided two-sample test).
To confirm our interpretation, we randomly picked 10 colonies from each combination of antibiotic and strain, sequenced their rpoB and rpsL genes, and measured growth rates (Figure 2). Several G·C→T·A substitutions found in rpoB can explain the higher fitness of rifampicin-resistant ΔmutY mutants, supporting our hypothesis. The highest-fitness class (Figure 1B, centre) is likely to be composed of V146F mutants, the fastest-growing mutant detected (Figure 2A); other high-fitness mutations that resulted from this transversion were H526N and S531Y. The fitness distribution in rifampicin-resistant ΔmutT mutants can be explained, at least in part, by an A·T→C·G substitution that produces the low-fitness mutation Q513P. Among streptomycin-resistant mutants, idiosyncratic transversions in rpsL similarly help to explain the fitness differences. The low fitness of streptomycin-resistant ΔmutY mutants probably reflects predominance of the P90Q mutation (8/10 mutants tested), whereas the high fitness of ΔmutT mutants might be due to prevalence of the K42T mutation (9/10). As a brief remark, 3/10 of the streptomycin-resistant mutants in the wild-type background carry the substitution P90L. This mutation is known to prevent growth in the absence of the antibiotic [28], and offers a suggestive example of how spectra can determine the access to mutations with differences not only in fitness, but also in other related properties.
The biochemical bases of the fitness cost of both rifampicin and streptomycin resistance have been discussed elsewhere. The costs of rpoB mutations are explained by the impairment of the transcription activity of the RNA polymerase [30], [31]. Similarly, it has been long established that rpsL streptomycin resistance mutants exhibit hyperaccurate translation, resulting in a slower rate of protein synthesis and consequently, in a slower growth rate [32].
It is tempting to assume that if the fitness distribution of a given mutator is altered compared to that of wild-type, the average fitness of that mutator will change accordingly. These distributions nonetheless have distinct shapes and degrees of overlap (Figure 1B); it is thus of interest to determine the extent to which these differences translate into an overall fitness difference between each mutator and the wild-type genotypes. Direct competition experiments between mutators and wild-type strains in rifampicin and streptomycin showed significant differences in mean fitness (Figure 3) (n = 4, P<0.012, Mann-Whitney U-test, one-sided, all cases), confirming the predictions made by visual examination of Figure 1A.
Our data provide evidence that in a specific environment, distinct mutators generate their own fitness distribution among newly-arising mutants, which can influence their competitive ability. Theory predicts the fixation probability of a mutator allele to be dependent on the selection coefficient of the driver allele with which the mutator hitchhikes [11], [20]. According to our results, this coefficient can vary substantially depending on the mutational spectrum; the mutational spectrum should thus have some effect on the hitchhiking dynamics. To estimate how large this effect could be, we used simple computer simulations based on previous studies [10], [11], [20]. Briefly, we simulated the basic scenario of a non-mutator bacteria population growing in batch culture, where only one beneficial mutation is needed to achieve full adaptation. Mutators are generated at a constant rate, and produce the adaptive mutation with a 100-fold higher probability than the wild-type strain. Once the mutation is fixed in either background, the simulation ends. The mutational spectrum effect (σ) was introduced as a multiplicative factor to modify the selection coefficient (s) of the driver allele only on the mutator background (see Material and Methods).
The simulations showed that σ exerts a modest influence on the establishment of mutator genotypes bearing the adaptive mutation (i.e., on the probability that they escape random drift) (Figure 4). This effect is not surprising, as the probability of a beneficial mutation surviving drift is approximately 2s [33]. In contrast, σ had a notable effect on the fate of established genotypes en route to fixation. In the absence of mutational spectrum effects (σ = 1), a mutator genotype bearing the adaptive mutation can only succeed if it reaches fixation before any adapted wild-type bacteria escapes drift; it will otherwise always be outcompeted due to its increased deleterious mutation load (Figure 5, upper row). When the average s of the driver allele is lower on the mutator background (σ<1), there are no qualitative changes. Success is further hindered because drift is more intense and time to fixation is longer, extending the period available for the establishment and subsequent selective sweep of an adapted wild-type strain. In contrast, when σ>1 a threshold appears above which the population dynamics switches. This threshold is determined by the value of σ that offsets the increased deleterious load of mutator genotypes. Above this value, the adapted mutator is fitter than its wild-type counterpart, and therefore needs only to escape drift to reach fixation (Figure 5, lower row); as a consequence, fixation probability rises sharply (Figure 4). It is worth noting that, since the deleterious load is as small as the order of magnitude of the mutation rate [33], only a slight spectrum-dependent fitness advantage is needed to substantially increase mutator success.
Remarkably, the non-linear response of fixation probability to σ implies that previous models could have been underestimating the likelihood of mutator success by several orders of magnitude. This is clearly illustrated in Figure 4 where a change from σ = 0.56 to σ = 1.33, which is equivalent to a change in relative fitness [37] from w = 0.96 to w = 1.03, represents a ∼196-fold increase in fixation probability.
Selection of mutators has been understood exclusively in terms of the increased number of beneficial mutations they generate [5], [10]. Here we show that not only the number but also the type of these linked mutations are relevant. Our results indicate that mutator alleles can bias the average selection coefficient of the beneficial alleles with which they hitchhike. Besides, they suggest that the magnitude of this effect can easily be sufficient to drastically modify their probabilities to reach fixation. The only requirement for this bias is that the locus or loci under selection produce mutants with some variability for fitness. This is a fairly permissive condition, likely to be satisfied in several adaptive scenarios. Examples in which resistant mutants with varying degrees of fitness are commonly found include bacteriophage [34] and antibiotic resistance [35], considered major drivers of mutator evolution [8], [15]. It will be of interest to determine whether the observed relative abundance of each mutator is explained, at least in part, by the effect of its mutational spectrum on successive mutations undergone during adaptation. In the context of clinical infections, future work should address the extent to which antibiotic therapies show different propensities to select for mutator bacteria. Since mutators are recognized as a risk factor for treatment failure [15]–[17], this knowledge could help to improve the design of safer therapeutic strategies.
The fact that even slight mutational spectrum effects markedly alter hitchhiking, together with the apparent commonness of circumstances that potentially allow this to happen, lead us to conclude that the mutational spectrum is a major factor in the evolution of mutators in laboratory and clinical populations of bacteria [3]–[8], as well as in certain cancers [36].
Strains are E. coli MG1655 derivatives, obtained from Dr. I. Matic [29]. Bacteria were grown on Luria broth (LB) or LB agar plates (37°C). Antibiotics used were tetracycline (15 mg/L), rifampicin (100 mg/L) and streptomycin (100 mg/L). Incubation time was 24 h, except for streptomycin plates (42 h).
Overnight cultures of each strain were plated on each antibiotic at appropriate dilutions to ensure low colony density (∼50/plate). After incubation, independent colonies were picked at random (by proximity to an arbitrary point) and resuspended in saline solution. Population size (N) was estimated from viable counts by subsequent dilution and plating. Assuming that each colony originated from a single cell, generation number was calculated as log2N and the growth rate expressed as number of generations per hour. Note that this measurement integrates growth rate over all growth phases.
Overnight cultures of each mutator and wild-type strain were mixed at a ratio based on their mutation rates, plated on antibiotic, and allowed to compete for 24 or 42 h. Initial competitor frequency was calculated by estimating the resistant-mutant frequency of each overnight culture. To distinguish them from the wild-type bacteria, mutators carried antibiotic resistance markers [29], which entailed no significant fitness cost in these conditions (n = 4, P>0.28, Mann-Whitney U-test, two-sided, both cases). After incubation, agar plates were washed in saline solution and the final mutator∶wild-type ratio obtained by plating on LB agar and selective media. Relative fitness was estimated using a standard formula [37].
|
10.1371/journal.pmed.1002230 | Clinical applicability and cost of a 46-gene panel for genomic analysis of solid tumours: Retrospective validation and prospective audit in the UK National Health Service | Single gene tests to predict whether cancers respond to specific targeted therapies are performed increasingly often. Advances in sequencing technology, collectively referred to as next generation sequencing (NGS), mean the entire cancer genome or parts of it can now be sequenced at speed with increased depth and sensitivity. However, translation of NGS into routine cancer care has been slow. Healthcare stakeholders are unclear about the clinical utility of NGS and are concerned it could be an expensive addition to cancer diagnostics, rather than an affordable alternative to single gene testing.
We validated a 46-gene hotspot cancer panel assay allowing multiple gene testing from small diagnostic biopsies. From 1 January 2013 to 31 December 2013, solid tumour samples (including non-small-cell lung carcinoma [NSCLC], colorectal carcinoma, and melanoma) were sequenced in the context of the UK National Health Service from 351 consecutively submitted prospective cases for which treating clinicians thought the patient had potential to benefit from more extensive genetic analysis. Following histological assessment, tumour-rich regions of formalin-fixed paraffin-embedded (FFPE) sections underwent macrodissection, DNA extraction, NGS, and analysis using a pipeline centred on Torrent Suite software. With a median turnaround time of seven working days, an integrated clinical report was produced indicating the variants detected, including those with potential diagnostic, prognostic, therapeutic, or clinical trial entry implications. Accompanying phenotypic data were collected, and a detailed cost analysis of the panel compared with single gene testing was undertaken to assess affordability for routine patient care.
Panel sequencing was successful for 97% (342/351) of tumour samples in the prospective cohort and showed 100% concordance with known mutations (detected using cobas assays). At least one mutation was identified in 87% (296/342) of tumours. A locally actionable mutation (i.e., available targeted treatment or clinical trial) was identified in 122/351 patients (35%). Forty patients received targeted treatment, in 22/40 (55%) cases solely due to use of the panel. Examination of published data on the potential efficacy of targeted therapies showed theoretically actionable mutations (i.e., mutations for which targeted treatment was potentially appropriate) in 66% (71/107) and 39% (41/105) of melanoma and NSCLC patients, respectively. At a cost of £339 (US$449) per patient, the panel was less expensive locally than performing more than two or three single gene tests.
Study limitations include the use of FFPE samples, which do not always provide high-quality DNA, and the use of “real world” data: submission of cases for sequencing did not always follow clinical guidelines, meaning that when mutations were detected, patients were not always eligible for targeted treatments on clinical grounds.
This study demonstrates that more extensive tumour sequencing can identify mutations that could improve clinical decision-making in routine cancer care, potentially improving patient outcomes, at an affordable level for healthcare providers.
| Healthcare planners and oncologists require real world evidence that next generation sequencing (NGS) technologies improve gene mutation detection and enable more appropriate use of targeted drug therapies.
With a range of genomic testing options available for cancer patients, we need to know whether healthcare systems can afford to implement cancer panels in routine clinical care, even if they are effective.
This study assessed a 46-gene hotspot cancer panel assay allowing multiple gene testing of small diagnostic cancer biopsies in the context of the UK National Health Service. Tumour samples (including non-small-cell lung cancer, melanoma, and colorectal carcinoma) from 351 patients who treating clinicians thought might benefit from more extensive genetic analysis underwent NGS using the panel. A clinical report was produced with a median turnaround time of seven working days that indicated all mutations detected, including those with potential diagnostic, prognostic, therapeutic, or clinical trial entry implications.
Clinical data were collected for patients whose tumour samples underwent sequencing in order to assess changes to clinical management resulting from this test.
An accompanying detailed cost analysis was performed to determine the affordability of the panel compared to existing single gene testing options.
The panel demonstrated at least one mutation in 87% (296/342) of successfully sequenced tumours.
Forty patients in this cohort received targeted treatments on the basis of genetic data obtained using the panel. For 22 of these patients, there was no alternative genetic test available locally to produce this data.
Mutation detection with the panel costs £339 (US$449) per patient, compared with single gene testing ranging from £71 to £141 (US$94–US$187) per test, depending on the mutation type. If more than two or three genes are examined (depending on the cancer type), using the panel is less expensive than single gene testing.
The panel assay is a useful method to identify genetic mutations in tumours that can extend the range of therapeutic options available to patients.
In terms of costs and affordability, the panel may be a justifiable option if 2–3 or more genes need to be examined.
Further data need to be collected on the clinical outcomes of patients accessing drugs as a result of more extensive sequencing data outside the scope of single gene/mutation tests.
In addition to supporting routine clinical care, the panel can be used to support research studies where treatment choices are genetically determined.
| Historically, the standard approach to testing for somatic mutations in cancers has been single gene testing using methods such as Sanger sequencing. With such methods, candidate genes are examined for mutations, and, as a result, patients may become eligible to enter a clinical trial or receive targeted drug therapies [1–3]. Advances in sequencing technology, collectively referred to as next generation sequencing (NGS), mean that the entire cancer genome (whole genome sequencing [WGS]) or parts of it (via targeted panels or whole exome sequencing [WES]) can now be sequenced in hours and at great depth and increasing sensitivity. However, while NGS offers high-throughput, rapid, and accurate testing of multiple genes, it remains to be proven whether it also leads to more appropriate use of targeted drug therapies and an enhanced ability to identify patients who are more likely to benefit from treatment compared with single gene sequencing.
An increasing number of primarily privately funded laboratories are already using NGS to profile tumours for mutations in multiple cancer genes simultaneously, with designs ranging from hotspot panels to a 287-gene panel covering all exons of constituent genes and selected introns (those involved in translocations). DNA requirements vary from 10 to 750 ng, and sample types evaluated include fresh frozen tissue, formalin-fixed paraffin-embedded (FFPE) samples, and fine needle aspirate specimens [4–7]. Many NGS technologies have demonstrated good sensitivity and excellent correlation with standard genetic techniques, as well as providing potentially clinically actionable information. However, the widespread translation of NGS technologies into routine cancer diagnostics has been slow due to technical obstacles (e.g., problems with robust bioinformatics) compromising clinical-grade validation, challenges with clinical interpretation (e.g., paucity of functional data), the absence of genotype–phenotype databases for cancer [8], a lack of demonstrable clinical utility surpassing that of single gene testing, and concern over the costs of NGS to healthcare payers.
Healthcare systems are now recognising the need to understand how to efficiently use genomic technologies in the context of precision medicine and verify their safety and effectiveness with timely evidence [9]. However, there is limited empirical evidence on whether results obtained from NGS technology direct clinical management and/or improve patient outcomes and whether they represent an efficient use of healthcare resources or are just an expensive addition to cancer care.
The aim of our study was to investigate whether a targeted hotspot NGS cancer panel could be translated into routine patient care in the UK National Health Service (NHS). This study involved a clinical-grade optimisation and validation of the panel and bioinformatics pipeline for diagnostics, an assessment of the panel’s impact on clinical management, and a cost analysis of the panel compared with single gene testing.
Clinical consent was obtained for all samples prior to genetic panel testing. The validation cohort included samples from the VICTOR trial that were consented for genetic analysis (approval obtained from Oxford Research Ethics Committee B [approval number 05\Q1605\66]). For the prospective cohort analysis, we used anonymised diagnostic samples for which ethical approval for service development was not required.
The NGS technology we assessed was the Ion AmpliSeq Cancer Hotspot Panel (Thermo Fisher Scientific; 46 genes, 189 amplicons). This study was completed in two stages. Stage 1 involved technical validation of the panel using an anonymised retrospective cohort of previously genotyped tumour samples (undertaken as a service development) and comparative costings of the assay with existing technologies in use. Stage 2 was clinical implementation of the validated panel with an accompanying prospective audit of the clinical impact of this assay on treatment choice. Study design, including the genes partially covered by the panel, is presented in Fig 1.
The retrospective cohort used for technical validation (n = 108) was composed of two sequentially tested groups; cohort 1 (n = 63) and cohort 2 (n = 45). Cohort 1 included samples from colorectal carcinoma (CRC), non-small-cell lung cancer (NSCLC), melanoma, and gastrointestinal stromal tumour (GIST) patients that were tested in tandem with standard diagnostic assays (S1 Text; S1 Table). Cohort 2 included previously sequenced CRC samples from the VICTOR (Vioxx In Colorectal cancer Therapy: definition of Optimal Regime) trial [10] as well as NSCLC and GIST specimens. Cohort 1 provided a more diverse range of tumour types, whilst cohort 2 provided mutational information on a wider range of genes than was available from standard diagnostic assays.
The prospective cohort (n = 351) consisted of malignant specimens (predominantly mesenchymal tumours, melanoma, NSCLC, and CRC) consecutively submitted to the Oxford Molecular Diagnostics Centre over a 12-mo period. The decision to submit a tumour sample for testing was made at the weekly multidisciplinary team meeting, with input from the treating oncologist and reporting histopathologist. An operational policy with sample testing algorithms, designed to ensure that testing was restricted to those patients with the potential to benefit from the acquisition of more comprehensive sequencing data, was available to provide guidance on appropriate samples for testing (S1 Fig). This cohort was also used to inform the translation of the panel into routine NHS care.
Phenotypic data were obtained for patients in the prospective cohort. In addition to tumour type, number of mutations, and drugs given, socio-demographic data were collected according to the best practice guidance for clinical audit [11]. In order to evaluate turnaround times, the date of assay request was also collected.
FFPE samples from both patient cohorts were macrodissected and DNA extraction performed as described in S1 Text. All retrospective cohort samples had conventional diagnostic testing performed in tandem, e.g., Sanger sequencing, pyrosequencing, and fragment analysis depending on sample type and gene under investigation. Among the prospective cohort, 278/351 samples had tandem cobas (Roche Diagnostics) analysis performed as per DNA availability and referring clinician preference, allowing comparison of the two alternative technologies. Cohort design and testing strategy are outlined in Fig 2 and S1 Table.
The design of the panel (i.e., which genes, exons, and codons were included) was based on variants for which there were potential therapeutic, prognostic, or diagnostic implications (for full details see S2 Table). Variants with therapeutic options included both those with established treatments approved by the US Food and Drug Administration/European Medicines Agency (e.g., BRAF, KRAS, and EGFR) and those potentially targetable (e.g., IDH1). The breadth of variants covered by the assay in any gene was designed to extend that available via single gene tests (e.g., inclusion of KRAS exon 5 to cover codon [12]). The overall scope (i.e., number of targets covered by the panel) was a balance between providing maximal data of clinical utility (i.e., potentially actionable) and permitting sufficient multiplexing of samples on a single sequencing chip to render the assay affordable.
DNA from all samples in both cohorts underwent NGS using the Ion AmpliSeq Cancer Hotspot Panel (Thermo Fisher Scientific). Ten nanograms of tumour DNA was amplified using the Ion AmpliSeq Library Kit 2.0 and Ion AmpliSeq Cancer Primer Pool, designed to detect mutations in hotspots in 46 genes (Fig 1), and indexed using the Ion Xpress DNA Barcode Adaptor 1–96 Kit (all Thermo Fisher Scientific) according to the manufacturer’s instructions. Libraries were purified using the AxyPrep Mag PCR Clean-Up Kit (Axygen Biosciences) and quantified using either an Agilent 2100 Bioanalyzer with the DNA High Sensitivity Kit (both Agilent Technologies) or quantitative PCR with the Ion Library Quantitation Kit (Thermo Fisher Scientific). Individual amplified libraries were diluted to 20 pM, and four or eight libraries were multiplexed to give a final concentration of 20 pM. Template-positive Ion Sphere Particles containing clonally amplified DNA were prepared using the Ion OneTouch Template Kit v2 and enriched using the Ion OneTouch ES as per manufacturer’s instructions. Libraries were sequenced on the Ion Torrent Personal Genome Machine with four and eight barcoded samples multiplexed on 316 and 318 chips, respectively (all Thermo Fisher Scientific).
Sequencing data were analysed using Torrent Suite software, optimised in an iterative fashion (S1 Text). In a similar fashion to other investigators, we used a tier system to classify variants [13,14] (Fig 3), with all being reported to clinicians in an integrated molecular and histopathological report overseen by a senior clinical scientist and histopathologist.
The panel was integrated into diagnostic laboratory workflows to enable reporting within a clinically relevant timescale (Fig 4).
We undertook a detailed micro-costing of both the panel and the cobas system at the Oxford Molecular Diagnostics Centre to determine whether NGS was likely to be sufficiently affordable to translate into routine care. Micro-costing is a highly detailed costing approach that identifies all the underlying resources required for an intervention/activity, such as equipment, consumables, and staff time, and then calculates costs for these resources. The standard operating procedures for the alternative technologies were used to develop costing questionnaires to collect the resource use information (S2 Text). The questionnaires covered each stage in the experimental protocol from sample preparation to data interpretation and reporting. Resource information on staff time, consumables, and equipment was derived from the questionnaires. We accounted for the expected cost of errors during the testing processes. For most equipment items, the cost was spread over the item’s predicted lifetime and depreciated using equivalent annual costing with a discount rate of 3.5%. The cost of the cobas z 480 Analyzer is covered in a combined cost with the mutation kits by the machine manufacturer Roche Diagnostics. The costs of reagents were obtained from prices reported by the diagnostics laboratory and also by contacting reagent manufacturers. Commercially available, rather than any discounted prices, were used where possible. Price per sample was based on the measured yearly throughput of the sequencing platforms, which was 832 (clinical and research samples, based on sequencing 16 samples per week for 52 wk) for the Ion AmpliSeq panel and 2,340 (45 samples per week) for the cobas.
To compare the panel costs with single gene test costs, we used NSCLC, melanoma, and CRC as examples, because these cancers made up the majority of cases in our clinical study. Several alternative costing scenarios were costed, based on clinical practice and the UK National External Quality Assessment Service (UK NEQAS) 2013 guidelines for molecular pathology [15]. For example, for NSCLC we compared the following testing scenarios: (a) the Ion AmpliSeq panel, (b) cobas with single and multiple mutation kits (EGFR, BRAF, and KRAS), and (c) cobas with an EGFR mutation kit, followed by the Ion AmpliSeq panel. Scenario (c) was included to confirm with the Ion AmpliSeq panel whether a lung cancer that was EGFR negative was also KRAS and BRAF negative: if all three genes tested negative, patients went on to have ALK testing, as these mutations have been found to be mutually exclusive [16].
Detailed information concerning the technical validation of the panel is provided in S1 Text. Among the prospective cohort, comparative failure rates of the panel and cobas were examined, as were real world turnaround times. The overall panel failure rate in this cohort was 2.6% (9/351). Among those samples analysed using both the panel and cobas, the failure rates were comparable: 0.7% (2/278) and 1.1% (3/278). Fig 5 demonstrates the range of turnaround times (in working days) observed for the panel (n = 342). The median turnaround time was seven working days, with an interquartile range of 6–9 d.
Turnaround time represents the timespan from assay request in the laboratory to report generation. It should be noted that this does not include the time taken to make the original histopathological diagnosis or produce sections or punches of the tumour suitable for DNA extraction. Turnaround times of 4 d were observed when extracted DNA was already available within the laboratory (due to prior single gene tests having been previously requested). Longer turnaround times could be accounted for by occasional failed assays being repeated, the need to multiplex samples on the sequencing chip, and the fact that the assay was performed only once a week, meaning that if a sample arrived immediately after the assay was initiated, the sample had a wait of five working days before library preparation was commenced.
Although turnaround times of 2–3 working days from sample receipt in the laboratory to report generation were possible for single gene cobas tests, the laboratory practice of batching samples for this analysis and performing assays on a weekly basis meant three sequential single gene tests had a turnaround time in excess of ten working days.
The 351 prospective cohort samples sequenced had comprehensive phenotype data, a summary of which is given in Table 1; 52% were female, and the median age was 68 y (range 9–95) at the time of tumour sampling. The predominant tumour types tested were melanoma (31.1%), NSCLC (30.8%), and CRC (25.1%), all of which should have been metastatic or unresectable, correlating with the malignancies for which there are targeted therapies approved by the National Institute for Health and Care Excellence (NICE) (S7 Table) [1–3,17]. The remaining samples were mostly mesenchymal tumours, e.g., GISTs, (for which tyrosine kinase inhibitors [TKIs] may be appropriate), or were submitted to allow assessment for clinical trial entry [18].
In all, 144/351 (41.0%) patients received pharmacological therapy, either chemotherapy or targeted molecular therapy. Treatment decisions were determined at the tumour-specific multidisciplinary team meeting and followed locally endorsed guidelines in line with national and international recommendations for each tumour type. Targeted therapies were administered in accordance with NICE guidelines [1–3,17] or in the context of a clinical trial. Fig 6 demonstrates the timing of the assay in the treatment pathway of these 144 patients. All patients presented in this study had a single sample analysed once using the panel; most often, the assay was performed prior to any treatment, with only a few patients having the test done after three or more lines of therapy, usually to facilitate clinical trial entry (e.g., a trial of a PI3K inhibitor in breast cancer [19]). The proportions of these 144 pharmacologically treated patients receiving targeted and non-targeted therapies were 36.8% (53/144) and 63.2% (91/144), respectively.
Fig 7 demonstrates the number of mutations detected per sample for the various tumour types tested (Fig 7A), as well as the distribution of mutations across the genes on the panel (Fig 7B). Among the successfully sequenced samples (97.4%, i.e., 342/351), at least one mutation was detected in 86.5% (296/342) of the samples, while 48.2% (165/342) of the samples had two or more mutations identified. In keeping with published studies, there were frequent mutations of BRAF and NRAS in melanoma; BRAF, EGFR, KRAS, and STK11 in NSCLC; APC, BRAF, KRAS, and PIK3CA in CRC; KIT and PDGFR3A in GIST; and CTNNB1 in other tumours (desmoid fibromatosis). TP53 was the most frequently mutated gene across the cohort, with mutations in all tumour types.
The mutation rates detected with the panel were similar for many genes to those found in publicly available WES/WGS studies (S4 Fig) despite different portions of the genome being examined. A Fisher’s exact test was applied to these data for each tumour type on a per-gene basis, and the resulting p-values were corrected for multiplicity using a Bonferroni–Hochberg procedure [20] (S1 Data). This analysis demonstrated no statistically significant difference in the rates of mutation observed using the two different sequencing approaches for the majority of genes across all tumour types (only p-values < 0.05 were considered to be statistically significant).
Genes where a statistically significant difference in mutation rate was observed were APC (p = 9 × 10−7) in CRC, STK11 (p = 0.008) in NSCLC, and CDKN2A (p = 0.02), ERBB4 (p = 9 × 10−4), FLT3 (p = 0.02), and KDR (p = 0.01) in melanoma. This finding suggests that although the panel covers only small amounts of the genome, it is very well targeted to the most frequently mutated regions: those genes where significantly more mutations were seen using WES/WGS compared to the panel have variants distributed across the gene rather than targeted on a few codons. This means that, despite its limited scope compared to WES/WGS, the panel is likely to capture most mutations in these genes, indicating its potential utility.
Full details of the mutations detected in the prospective cohort are given in S2 Data.
Standard genetic testing of tumours is usually limited to situations where variant information will influence a specific intervention. In order to justify the cost of more extensive mutation analysis, it is necessary to demonstrate clinical utility, e.g., by analysing treatment decisions and patient outcomes based on data in electronic patient records. Given that the prescription of targeted therapies in the UK is regulated by NICE, evidence regarding clinical utility beyond NICE-approved indications is anecdotal.
Fig 7C demonstrates that the panel provided additional mutation information not detected by the cobas technology for the three “most actionable” genes: 34.2% (13/38) of BRAF mutations in melanoma, 19.0% (4/19) of EGFR mutations in NSCLC, and 11.1% (4/36) of KRAS mutations in CRC (only samples tested using both the panel and the cobas assay were included in this comparison). These mostly pertained to codons or variants outside the scope of the cobas assay, some of which are actionable. However, four additional BRAF V600E mutations in melanoma specimens were identified using the panel that should have been detected by the cobas (two failures and two false negatives). In contrast, cases where standard diagnostic assays provided additional information are entirely accounted for by panel sequencing failures.
Table 2 lists variants in the prospective patient cohort that did change, or could have changed, patient management given local availability of targeted therapeutics (assuming clinical criteria were met). In melanoma, guidance from NICE permits vemurafenib use for BRAF V600 mutated tumours, meaning one patient with a V600R mutation received the drug and others with V600G/M and V600K mutations were eligible. One patient with melanoma with a KIT V560D mutation entered a clinical trial of the TKI nilotinib, while another received the multi-targeted receptor TKI pazopanib after NRAS mutation status was determined (clinical trial entry required only knowledge of NRAS mutation status, not the detection of a mutation). NICE guidance regarding eligibility for vemurafenib in BRAF V600 mutated melanoma stipulates that patients have metastatic disease, but not all patients tested met this criterion, accounting for some of the discrepancy between actionable mutations and subsequent changes to patient management.
NICE guidance requires only the presence of activating EGFR mutations [17,21] for NSCLC patients to be eligible for EGFR inhibitors, meaning four patients with unusual EGFR mutations (M600T, S720C, V742I, and L861Q) received erlotinib as a result of the panel. Evidence for the efficacy of EGFR inhibition in these mutations is scant due to their low frequency. There is some evidence that L861Q is an activating mutation although in vitro studies suggest it is sensitive to WZ-4002 (irreversible second-generation EGFR inhibitor) rather than erlotinib [22].
During this study, eligibility for anti-EGFR monoclonal antibody therapy (cetuximab or panitumumab) in CRC evolved from requiring wild-type KRAS to requiring wild-type RAS, necessitating mutation testing of NRAS in addition to KRAS. The only method of assessing NRAS status within our laboratory was the panel, and 51/88 (58.0%) patients with CRC were found to have tumours with wild-type RAS. The audit of subsequent clinical action revealed that 5/51 (9.8%) of these patients received anti-EGFR therapy (cetuximab). Further examination revealed that many of the 46 patients who on the basis of RAS mutation status would have been eligible for the treatment did not meet the clinical criteria (i.e., did not have metastatic disease; see S7 Table). Among CRC patients with mutated RAS tumours (37/88; 42.0%), four tumours would have been classed as wild-type RAS using conventional diagnostics: two had NRAS codon 61 mutations and two had KRAS codon 146 mutations (outside the scope of the cobas assay).
The panel was also the only mutation assay available for GIST patients at the time of analysis: although use of TKIs is dictated by clinical factors (moderate/high-risk or metastatic disease), certain mutations cause tumour resistance or require a higher TKI dose [45]. Six GIST patients received standard dose imatinib due to the absence of KIT exon 9 (requires higher dose) or PDGFRA D842V mutations (confers imatinib resistance). Four breast cancer patients had their PIK3CA mutation status determined, with one receiving BLY719 (PI3K inhibitor) in a clinical trial as a result, while a patient with a wild-type RAS mucinous ovarian carcinoma received cetuximab.
In total, 122/351 prospective cohort patients (34.8%) had a mutation for which there was either a NICE-approved targeted therapy or locally available clinical trial of a targeted therapy, 40 (32.8%) of whom received a targeted therapy. Fifty-five percent (22/40) of these patients received a targeted treatment only as a result of novel information from the panel. The additional 13 patients who received targeted therapies were NSCLC patients who received EGFR inhibitors second line in the absence of a previously detected activating EGFR mutation and patients with a variety of solid tumours who received targeted inhibitors in the context of a clinical trial where demonstration of a particular mutation was unnecessary.
A concern with testing progressively larger quantities of the tumour genome is that multiple potentially actionable mutations may be detected with conflicting recommended actions. Owing to the targeted nature of this panel centred around well-characterised hotspots and the paucity of targeted agents available in the UK, variant interpretation in this study was not impacted by this possibility. Where more than one mutation was identified in an actionable gene, there was either a clear hierarchy of action—e.g., NSCLC specimen G150739T had both an exon 19 deletion and p.T790M mutation in EGFR, meaning the patient would not respond to first generation EGFR TKIs but rather may benefit from a third generation drug [46]—or the same action was appropriate for both mutations, e.g., melanoma specimen G151372L had BRAF p.V600M and p.V600G mutations, both of which are likely to benefit from a BRAF inhibitor [25,47].
In order to investigate what the potential impact on clinical management of the panel might be in the future, we analysed the mutation data from the NSCLC and melanoma samples in the prospective cohort for theoretically actionable mutations as described by Meador et al. [48] (S5 Fig). A mutation was classified as theoretically actionable if there was peer-reviewed data at any level (from in vitro cell line to phase III randomised controlled trial [RCT] data) that indicated efficacy of an available treatment (predominantly targeted inhibitors). Assuming unrestricted access to these targeted therapies, 39.0% (41/105 successfully sequenced) of NSCLC and 66.4% (71/107 successfully sequenced) of melanoma patients had a potentially actionable mutation, in contrast to 20.0% (21/105) and 32.7% (35/107) of NSCLC and melanoma patients, respectively, who had a locally actionable mutation. This suggests that, in the future, far more of the variant information generated will lead to clinical management changes.
As shown in Table 3, the total cost for testing 46 genes using the panel was £339 (US$449) per sample (patient). This is compared to the cost of mutation testing with the single gene approach (cobas): £71 (US$94) for BRAF, £104 (US$138) for EGFR, and £141(US$187) for KRAS. Table 4 shows the cost by resource category for the different tests and for the combinations of these tests for different malignancies. For all tests, most costs are attributed to consumables, followed by staff and overheads. For example, the consumable cost is £185 per sample for the panel and between £34 and £93 for the cobas, depending on the gene tested.
We have validated and clinically implemented an NGS assay that detects relevant mutations across mutational hotspots of 46 genes from minimal quantities of FFPE-derived highly fragmented tumour DNA, allowing routine testing of multiple genes from small biopsies. Its performance has been validated across a variety of tumour types for single nucleotide variants and indels and has been shown to have an enhanced sensitivity compared with conventional diagnostic techniques. Treatment data revealed that surprisingly few patients (~40%) received any pharmacological treatment, targeted or otherwise, confounding the fact that for most patients the indication for mutation analysis should be to inform whether treatment should be conventional chemotherapy or targeted agents. Examination of individual cases showed that, in contravention of operational policy, many samples were not from metastatic or unresectable malignancies. Whilst early testing in the diagnostic pathway may be desirable, many targeted therapies are indicated only for metastatic or unresectable disease, such as erlotinib in NSCLC. Additionally, differences in side effects between conventional chemotherapy and targeted agents are such that some elderly patients may be deemed suitable for the latter but not the former, meaning that if no actionable mutation is identified there is no appropriate treatment.
Over a third of the patients in our prospective cohort (122/351) had a locally actionable mutation, with 40 patients receiving a targeted therapy, of which 22 received this therapy only because of tumour testing with the panel (three patients were able to access a clinical trial, 12 patients had access to NICE-sanctioned therapies, six patients had changes to their management in line with European Societal Guidelines, and one patient was able to access a therapy via local funding arrangements). Whilst these numbers are modest, they reflect the limited availability of targeted therapies approved by NICE. For example, significant numbers of patients within the prospective cohort could potentially have benefited from targeted therapies when data from just 7/46 genes were considered (EGFR, KRAS, NRAS, BRAF, KIT, PDGFRA, PIK3CA). Although not all these patients required treatment at this stage, others were unable to access drugs due to lack of either licensing for their specific mutation and tumour type or NICE approval, reflecting a lack of prospective phase III RCT data.
The panel provides flexibility to rapidly introduce new testing as novel therapies are licensed or eligibility criteria are updated, as was demonstrated with the introduction of NRAS testing in our cohort. The panel also enables patients’ eligibility for novel agents in clinical trials to be assessed, which provides a significant contribution to available treatment options.
It is important to note that the panel testing also highlights patients who should not receive treatment due to the presence of confounding activating mutations, in particular, CRC patients with activating mutations in KRAS or NRAS, for whom treatment with anti-EGFR monoclonal antibody is not indicated, and GIST patients with activating mutations in KIT or PDGFRA, who are not eligible for imatinib. This accounted for 38/351 patients (10.8%) in the prospective cohort (37 CRC patients with activating KRAS or NRAS mutations and one GIST patient with a PDGFRA D842V mutation) and is also an important outcome of tumour testing.
The panel described here enables parallel testing of multiple genes, in contrast to conventional diagnostic techniques. However, it must be noted that the panel covers hotspot mutations only, not full gene sequencing, and in its current format does not allow testing for larger copy number variants such as HER2 gene amplification in breast cancer [49].
Furthermore, it should be recognised that although RCTs increasingly use mutation status in treatment stratification decisions, e.g., the Medical Research Council FOCUS4 trial [50], for rare mutations in frequently mutated genes, combinations of mutations, and any mutations in rarely mutated genes, it is unlikely that phase III RCTs will ever recruit sufficient patients to determine the most appropriate treatment option, particularly with ever-increasing numbers of inhibitors.
The alternative is to develop and join up worldwide repositories of genotype–phenotype data so all anecdotal experiences can be collated and statistically mined for commonality. For this approach to be informative, more molecularly directed access to novel treatment modalities of proven efficacy is required. In addition, clinical drug development has to systematically include genomic characterisation of patient samples to detect differential responses due to mutation signatures. An example of this is the National Lung Matrix trial, currently recruiting in the UK, which consists of parallel, multi-centre, single-arm phase II trials, each arm testing an experimental targeted drug in a population stratified by multiple prespecified target biomarkers employing a Bayesian adaptive design [51].
In terms of the affordability of NGS technologies for healthcare payers, Cancer Research UK has set several requirements for tumour profiling tests. One of these requirements is that the cost for such tests must be less than £300 [52]. Our cost analysis shows that the panel cost is only slightly higher than this, at £339 (US$449) per sample, and that for certain gene test combinations, such as testing for BRAF, EGFR, KRAS, and PIK3CA (Table 3), sequential single gene testing is actually more expensive than testing several genes at the same time using the panel (£27 for NSCLC and £32 for melanoma). In the context of other costing studies in this area, estimates show that there is considerable variation across studies [53]. For example, Gallego et al. reported a cost of £1,703 (US$2,700) for using a cancer panel in the diagnosis of CRC and polyposis syndromes [54], while Yorczyk et al. estimated the average cost for a single-tier hereditary 25-gene panel test using the MyRisk platform at £2,581 (US$4,099) per person [55]. Ghemlas et al. reported that the cost of NGS was £297 (US$470) per patient for genetic testing of inherited bone marrow failure [56]. There is also variation in what these studies include in the cost analysis, with some including only the costs of consumables. However, a common theme is that most of the cost estimates are substantially higher than the Cancer Research UK £300 target.
Further, our results suggest that, depending on the combination of genes tested, the panel can be less expensive than single gene testing. For example, if testing for melanoma is done using a combination of three single gene tests for BRAF, NRAS, and KIT, this costs less than the panel. However, in other contexts, if KRAS and PIK3CA or EGFR and PIK3CA were tested, then testing for only two genes would be more expensive than the cancer panel.
This study has a number of potential limitations, some of which are common to most NGS studies, and others of which are specific to our study. First, correct identification of variants requires access to high-quality tumour material; as with other studies, we used FFPE biopsies. Although sequencing was usually successful using this NGS assay, the formalin fixation process itself can cause alterations in the sequence due to deamination of cytosine to uracil [57]. Equally, detection of all the relevant variants in a tumour assumes a representative biopsy, which may not be the case given the phenomenon of clonal heterogeneity [58]. Second, as with most sequencing studies, assessment of the pathogenicity of variants whose clinical significance is not understood is challenging, and in the absence of biological data, many in silico variant effect prediction algorithms have limitations.
Specific study limitations relate to our use of real word data: submission of cases for sequencing did not always follow the suggested algorithms (see S1 Fig) and therefore, regardless of the variants detected, the patients were not eligible for targeted treatments on clinical grounds. Owing to some cases being submitted from other institutions, not all clinical information was available for all patients, reducing the overall power of the study. Finally, limited access to some targeted therapies within the UK NHS meant that even if there was a potential drug indicated for a variant in a particular tumour type, often this drug would not be funded for use. In terms of our cost analysis, due to the limited availability of single gene cost data, the comparison between the Ion AmpliSeq panel and cobas was based on only three genes, whereas the comparison should ideally be made on the cost of the nine genes that have known clinical value. At the same time, we believe that the use of real-life data in this setting was the most appropriate way to study the true clinical utility and cost-effectiveness of the panel in a given healthcare system.
In conclusion, in the context of the UK NHS, we validated and translated a cancer panel that reports clinically actionable results back to clinicians. This led to actionable mutations being identified in 122/351 (34.8%) patients in our cohort and allowed 15.3% (22/144) of those receiving pharmacological treatment to have access to targeted therapies not indicated using conventional single gene testing, thereby demonstrating the ability of the panel to impact clinical management. Our results also show that providing a mutation analysis service using a cancer panel for cancer diagnostics and treatment would provide value for money for the NHS, because if several genes are tested individually, which is often the case, then a one-stop test would require less resources (and time) and be less expensive than sequential gene testing. Using cancer panels for molecular testing will help molecular diagnostic laboratories deal with the increased demand for genetic testing and rapidly and reliably detect relevant mutations, even for a limited amount of tumour tissue, at a relatively low cost. Currently, only 9/46 genes on the panel are recognised to have clinical utility in the UK system, but as additional genes are clinically validated as targets, greater potential of NGS technologies could be realised.
|
10.1371/journal.pbio.0060033 | A Global Assessment of Salmon Aquaculture Impacts on Wild Salmonids | Since the late 1980s, wild salmon catch and abundance have declined dramatically in the North Atlantic and in much of the northeastern Pacific south of Alaska. In these areas, there has been a concomitant increase in the production of farmed salmon. Previous studies have shown negative impacts on wild salmonids, but these results have been difficult to translate into predictions of change in wild population survival and abundance. We compared marine survival of salmonids in areas with salmon farming to adjacent areas without farms in Scotland, Ireland, Atlantic Canada, and Pacific Canada to estimate changes in marine survival concurrent with the growth of salmon aquaculture. Through a meta-analysis of existing data, we show a reduction in survival or abundance of Atlantic salmon; sea trout; and pink, chum, and coho salmon in association with increased production of farmed salmon. In many cases, these reductions in survival or abundance are greater than 50%. Meta-analytic estimates of the mean effect are significant and negative, suggesting that salmon farming has reduced survival of wild salmon and trout in many populations and countries.
| The impact of salmon farming on wild salmon and trout is a hotly debated issue in all countries where salmon farms and wild salmon coexist. Studies have clearly shown that escaped farm salmon breed with wild populations to the detriment of the wild stocks, and that diseases and parasites are passed from farm to wild salmon. An understanding of the importance of these impacts at the population level, however, has been lacking. In this study, we used existing data on salmon populations to compare survival of salmon and trout that swim past salmon farms early in their life cycle with the survival of nearby populations that are not exposed to salmon farms. We have detected a significant decline in survival of populations that are exposed to salmon farms, correlated with the increase in farmed salmon production in five regions. Combining the regional estimates statistically, we find a reduction in survival or abundance of wild populations of more than 50% per generation on average, associated with salmon farming. Many of the salmon populations we investigated are at dramatically reduced abundance, and reducing threats to them is necessary for their survival. Reducing impacts of salmon farming on wild salmon should be a high priority.
| Since the late 1970s, salmon aquaculture has grown into a global industry, producing over 1 million tonnes of salmon per year [1]. The majority of this biomass is held in open net pens in coastal areas through which wild salmon migrate on their way to and from the ocean. A number of studies have predicted or evaluated the impacts of salmon farming on wild salmon through a single mechanism, in a given area. It is clear that some salmonids are infected and killed by sea lice originating from salmon farms [2–5], that other diseases have been spread to wild populations from salmonid farming activities [6,7], and there is evidence that salmon parr are at lower density in areas of Scotland where there is salmon aquaculture [8]. In addition, farmed salmon escape in all areas where salmon aquaculture is practiced, and although their breeding success may be low on average, competition for mates and hybridization with wild salmon are likely to reduce survival of wild populations [9,10].
It is well established that wild salmonids can be negatively affected by salmon farming [11], however, the importance of these interactions at the population level has rarely been determined [2]. To determine population level impacts, we examined temporal trends in the abundance and survival of wild salmonids (Figure 1 and Figure S1). Our study contrasted trends in wild populations exposed to potential aquaculture impacts with those of populations not exposed. Populations in which juvenile salmonids pass by salmon farms during their migration were considered to be exposed to impacts of salmon farming. Exposed populations were carefully paired with control populations in the same region whose migrations did not lead past farms, but which otherwise experienced similar climate and anthropogenic disturbances. Use of such paired comparisons allowed us to control for confounding factors such as climate to detect population level impacts. Using the Ricker stock recruit model [12], we performed 11 comparisons, involving many stocks from both sides of the Atlantic and from British Columbia in the Pacific (Table 1, Data section of Materials and Methods).
All estimates of the effect of aquaculture on survival or returns were negative. Both random effects estimates of the mean effect were negative and highly significant (Figure 2), indicating a very large reduction in survival and returns in populations exposed to aquaculture. Under the dynamics of Equation 1 (see Materials and Methods), percent change in survival or returns is represented by
where γ is the coefficient of aquaculture production (P) for region k. For example, the estimated change in survival per tonne of salmon farming (γk) for Bay d'Espoir in Newfoundland was estimated to be 0.026 (Figure 2). In 2003, the farmed salmon harvest from this area was 1,450 tonnes (t), so the estimated decrease in survival is
(95% CI: 44%–80%), relative to what it would be in the absence of farms. Survival and total returns of many stocks were found to be reduced by more than 50% (Figure 2), for each generation. If all exposed populations were passing by farms with a total annual harvest of 15,000 t, the mean estimated total reduction in survival would be 73% (95% CI: 29%–90%) (Figure 2). Many regions now have farmed salmon production in excess of 20,000 t/y.
Generally, Atlantic salmon populations were depressed more than Pacific salmon populations, particularly Atlantic salmon in Atlantic Canada. Irish sea trout were also estimated to have been very strongly reduced by impacts of salmon farming, whereas estimated impacts on Atlantic salmon in Scotland depended on the data used. In British Columbia (Pacific Canada), only pink salmon showed significant declines correlated with salmon aquaculture.
Results are reported for a model including autocorrelated errors and with λ set at 0.5, rather than 1 or 2, because this minimized the Akaike information criteria (AIC) for most regions [13]. The parameter λ allows for the impacts of salmon farming to change nonlinearly with the aquaculture production. A λ of 0.5 indicates that relatively small amounts of aquaculture will depress wild populations, but the effect does not increase proportionally to aquaculture production. See Tables S1 and S2 for results of alternative models.
For the New Brunswick comparison, the outer Bay of Fundy rivers are located much closer to salmon farms than the other exposed rivers. If only these outer Bay of Fundy rivers are considered exposed to salmon farming, and other Bay of Fundy rivers (inner Bay of Fundy and Saint John River) are included among the controls, the overall estimates (i.e., meta-analytic means) are still significant and negative in both versions of the analysis.
We have estimated a significant increase in mortality of wild salmonids exposed to salmon farming across many regions. However, estimates for individual regions are dependent on assumptions detailed in the Materials and Methods section, and the estimates often have large confidence intervals. Given that the data analysed are affected by considerable noise—including changes in fishing and environmental factors—the important result of this study is that we are nonetheless able to detect a large, statistically significant effect correlated with trends in farmed salmon production. The significant increase in mortality related to salmon farming that we have estimated in almost all cases is in addition to mortality that is also acting on the control populations. In most cases, control populations were also experiencing decreases in marine (and sometimes freshwater) survival, for reasons that are only partially understood. At the same time, fishing mortality has been reduced or eliminated in many areas, which may have partially masked high mortalities associated with aquaculture.
A key assumption in this study is that exposed and control areas do not differ in a systematic way across regions. We have identified three possible ways that exposed and control sites could differ systematically: first, salmon farms could be established only in areas where wild stocks have already collapsed; second, salmon farms could be established in areas where habitat is more disturbed by human activities; or, third, climate factors could differ between the exposed areas and the controls in a systematic way.
Declines in control and exposed salmonid populations preceded the growth of the salmon aquaculture industry in some regions, but inspection of the data used do not indicate that salmon populations in the majority of our regions had declined dramatically in the exposed areas only, before the start of salmon farming (averaged returns data are shown in Figure 1). In regions such as Scotland, where declines precede the start of salmon farming, the strong aquaculture effect estimated reflects a faster decline in exposed populations concurrent with the growth of salmon farming.
Areas that we consider exposed do not seem to be more developed than control areas in general. In the Atlantic, most areas have been highly altered by human activities for hundreds of years, but there is no obvious difference between the control and exposed groups in this regard. In British Columbia, all areas considered are very remote, and the main type of anthropogenic disturbance in rivers would be forestry. Comprehensive forestry records at the watershed scale are not easily available, but logging in British Columbia's Central Coast is extensive, both historically and recently [14]. It should be noted that the comparisons in British Columbia include large numbers of rivers (> 80 rivers in each case), so differences in anthropogenic effects would have to hold over many watersheds to explain the effects we estimate.
Finally, it is also very unlikely that our results are due to a climate driven trend in which more southerly populations show stronger declines than populations to the north. Although our exposed populations are to the south of control populations in three of five regions, differences in latitude are small. In New Brunswick, the control populations are to the north of the exposed populations, but by less than 200 km, and the headwaters of some of the exposed populations are adjacent to those of the controls. In Newfoundland, the difference in latitude between exposed and control populations is similarly small. In British Columbia, the control populations are also to the north, but by less than 300 km. Also, Mueter et al. [15] found that pink and coho salmon from all of the British Columbia populations we have examined respond similarly to large-scale climate trends. Thus, the pattern we found in this study does not seem attributable to a systemic difference between the control and exposed areas.
We estimated higher impacts on populations in the Atlantic than those in British Columbia, possibly because Atlantic salmon populations are conspecific with farmed salmon, and therefore susceptible to genetic effects from interbreeding with escaped farm salmon, in addition to disease or other impacts. Estimated impacts in British Columbia may also be lower because we aggregated over large numbers of populations for pink, chum, and coho salmon, because estimates of fishing mortality were only available at a very coarse scale. The individual populations may vary in their exposure to salmon farms.
The large apparent impact of Atlantic salmon farming on Irish sea trout, in contrast, can not be explained by interbreeding. In the mid-western region of Ireland (the exposed region), the total rod catch decreased from almost 19,000 sea trout in 1985 to 461 in 1990 [16]. In the few rivers where data were available, catch declines could not be explained by reduced effort [16]. Welsh sea trout catches (the controls) have remained relatively constant during the same time period, whereas fishing effort has decreased considerably [17]. Sea trout (anadromous brown trout) might be expected to experience higher mortalities, because they spend lengthy periods in coastal areas near salmon farms, relative to Atlantic salmon, thus being exposed to disease or parasites for a longer time [18].
The time period over which we are estimating impacts of aquaculture includes the establishment of the industry in each region. Improvements in management as industries mature may explain our finding that impacts of salmon farming on wild salmon do not increase linearly with the tonnage of farmed salmon. Better management should decrease the impact of salmon farming on a per tonne basis, although such improvements may not be able to keep pace with the growth of the salmon farming industry. The estimated reduction in survival of wild salmonids is large, and would be expected to increase if aquaculture production increases.
We modeled survival and, in a separate analysis, total returns to each stock, using a general linear mixed effects model for each region. To model survival, we used a Ricker model extended to include the production of farmed salmon in the area through which exposed juvenile salmon migrated, with random effects for each stock and year [19].
Let Si,y be an index of the number of fish that smolted, i.e., migrated to sea in the spring, in year y from stock i, let Ri,y be the estimated number of those fish that would subsequently return to spawn in the absence of fishing, and let Pi,y be the aquaculture production that those smolts were exposed to (in tonnes). The dynamics are assumed to be given by
where β0 is the fixed intercept for the average stock and year with no aquaculture production, ai is the random deviation of the ith stock intercept from β0, dy is the random deviation of the yth year, βi is the fixed slope of mortality (the density dependence parameter) that will vary with each stock i, and γ is the coefficient of aquaculture mortality that is assumed to scale with a possibly nonlinear function of aquaculture production, (Pi,y)λ. The random error, ei,y, is assumed to be first order autocorrelated. We assume the ai's and dy's come from normal distributions with zero mean. The autocorrelation and the random year effect are included to account for established temporal and spatial correlations (respectively) in environmental effects [20].
The effects of aquaculture are summarized by the coefficient γ for each region. The regional coefficients were combined using meta-analysis to obtain an overall estimate of the change in wild salmonid survival related to aquaculture. Because the best functional form for the aquaculture term in the model (Pi,y)λ was not known, we investigated a linear increase in impacts with aquaculture, a square relationship, and a square root relationship. We selected models by AIC, and we tested our results under alternative formulations.
To test the robustness of the conclusions, and because only returns data were available for some regions, we repeated the analysis with number of returning adults as the response variable. This analysis used Equation 1 but dropped the Si,y and βi terms. The response variables for this analysis included rod catches, rod plus marine catches, counts of salmon returning to rivers, and estimates of returns to rivers in the absence of fishing (see Data sources and treatment, below).
Outer Bay of Fundy salmon in New Brunswick, Canada, have been reduced to zero in one river and to a handful in another river. For this region only, we assumed negative binomial errors.
For the meta-analysis, we added a subscript, k, to identify each region, to γ, which summarizes the effect of aquaculture for each region. For a fixed assumption about λ, the γk's are in the same units and can be directly compared. We modeled the effects of aquaculture as a mixed effects model,
here
is the estimated value of γk, α0 is the intercept, σ2 is the among-region variance, and
is the variance of the kth estimate (which is taken from the analysis in Equation 1, and is held fixed). A fixed effects meta-analysis is obtained by constraining σ to be zero. We used maximum likelihood estimation and selected models by AIC.
For robustness, we considered five classes of models: different regions used as controls, different mixed model assumptions, different error assumptions, different functional forms for the aquaculture effect, and different autocorrelational structures, as well as performing a Bayesian meta-analysis. Overall, the results were very similar for all models. (See Tables S1 and S2 for results of alternative models and Text S1 for details of the Bayesian analysis.)
We analysed data for five species of wild salmonid in five regions: Ireland and Wales, Scotland, Newfoundland (Canada), New Brunswick (Canada), and British Columbia (Canada). There are three further regions with both wild salmonids and salmon aquaculture for which we could not carry out analyses: Norway, the west coast of Vancouver Island (Canada), and Maine (United States). We were unable to carry out analyses for Norway for three reasons. First, salmon farming in Norway is so widespread [21] that it was difficult to establish controls. Second, the adult population in many rivers has been found to contain over 50% aquaculture escapees [22], making trends in returns to rivers difficult to interpret. Third, there are confounding effects from acidification and disease [23, 24]. For the west coast of Vancouver Island, it was not possible to obtain aquaculture production data by region over time, and Maine was not included because of a lack of nearby wild populations to serve as controls.
Most populations that we considered to be exposed breed in rivers that discharge into bays or channels containing at least one salmon farm. Others breed in rivers flowing into bays without salmon farms very close to areas containing many farms. Salmon from control rivers are very unlikely to pass by salmon farms early in their life cycle, due to the direction of their migration. However, some controls may be relative, in the sense that salmon may pass by farms from a considerable distance, later during their migrations. This would tend to be conservative with respect to our study, since we would then have to detect local effects that are additional to any impacts from distant farms. Data from scientific surveys, e.g., counting fences, were used if possible; for Scottish salmon and Irish and Welsh sea trout, only catch data were available, so results are given for only the impacts on returns (not survival).
We compared rod catches of sea trout in Ireland's Western Region to rod plus in-river fixed engine catches in Wales, from 1985 to 2001 (there are no fixed engine fisheries directed at sea trout in Ireland). Salmon farming is concentrated in the Western Region (Connemara area) of Ireland, but does occur in other parts of the country [25]. Based on farm locations [25], it was estimated that all rivers considered exposed are located less than 50 km from a salmon farm, but most will enter the ocean less than 30 km from a salmon farm. There is no salmon farming in Wales. There were 16 rivers in Western Ireland considered exposed: Athry, Bhinch (Lower), Bhinch (Middle), Bhinch (Upper), Burrishoole, Costello, Crumlin, Delphi, Erriff, Gowla, Inagh, Inverbeg, Invermore, Kylemore, Newport, and Screebe [16]. The following 32 Welsh rivers served as controls: Aeron, Afan, Arto, Cleddau, Clwyd, Conwy, Dee, Dwyfawr, Dwyryd, Dyfi, Dysynni, Glaslyn, Gwendreath, Gwyrfai, Llyfni, Lougher, Mawddach, Neath, Nevern, Ogmore, Ogwen, Rheidol, Rhymney, Seiont, Taf, Taff, Tawe, Teifi, Tywi, Usk, Wye, and Ystwyth [26,27]. Trout caught and released are included in catch data from both countries. Only catch estimates were available for most of these rivers. Recruitment could not be derived, because anadromous brown trout interbreed with freshwater resident trout, about which very few data are available, so this stock was only included in the returns modeling (not survival). Farmed salmon production for all of Ireland was used in modeling [28], because the majority of farms are in the region where the exposed populations breed. This will tend to have a conservative effect, resulting in a lower estimate of the impact of aquaculture, per tonne of salmon farming.
We compared marine plus rod catches of Atlantic salmon from the east coast of Scotland to catches from the west coast of Scotland for the years 1971 to 2004. Salmon farms appear to be located in the majority of bays on the west coast of Scotland in well over 300 sites (http://www.marlab.ac.uk/Uploads/Documents/fishprodv9.pdf), so all salmon from rivers on this coast were considered exposed. There is no salmon farming on the east coast, so salmon from east coast rivers were controls. For each coast, a single time series of total catch was used in modeling. Marine catch records were from the International Council for the Exploration of the Sea (ICES) Working Group on North Atlantic Salmon [28] and rod catch records were from Fisheries Research Services of Scotland (J. MacLean, personal communication). Rod catches included salmon caught and released. These data were only used in modeling returns. Farmed salmon production for all of Scotland was used in modeling [28], because regional production data were not available.
We also used counts of Atlantic salmon of all ages returning to rivers from 1960–2001 in Scotland from Thorley et al (2005) [29]. The fish counters are maintained by Fisheries Research Services or by Scottish and Southern Energy plc. There were two exposed populations. One is from the Awe Barrage, which empties into a bay with numerous salmon farms. The other is from the Morar River, which is less than 20 km from the nearest salmon farm, in an area of the coast with many farms [8]. Salmon from the control rivers (on the east coast) do not pass by salmon farms in Scotland because of the direction of their migration routes [30], unless they approach the Norwegian coast. There were ten control populations from the following rivers: Aigas, Beanna, Torr Achilty, Dundreggan, Invergarry, Logie, Westwater, Cluni, Erich, and Pitlo. Farmed salmon production for all of Scotland was used in modeling [28] because regional production data were not available.
Estimates of marine survival to one sea winter for hatchery (and two wild) Atlantic salmon populations from Ireland and Northern Ireland (1980–2004) were collected and reported by the ICES Working Group on North Atlantic Salmon [28]. Because only survival estimates are provided, these data were only used in the survival analysis. Salmon from hatcheries on the Screebe, Burrishoole, Delphi, and Bunowen Rivers were considered exposed. Populations from hatcheries on the Shannon, Erne, Lee, Bush, and Corrib Rivers, plus wild populations from the Bush and Corrib Rivers were used as controls.
Production data were not available on a regional basis, so national values [28] were apportioned to bays into which exposed rivers empty by assuming that 30% of national production is in the Kilkieren Bay, 10% is in Clew Bay, 5% is in each of Killary Harbour and Ballinakill Bay. These proportions are based on maps of salmon farm locations from the Irish Marine Institute [25], and they approximately match stock numbers collected by the Central Fisheries Board in the years for which stock numbers are available (P. Gargan, personal communication). Years in which each bay was fallowed were obtained from the Central Fisheries Board (P. Gargan, personal communication), and in these years, the fallowed bays are assigned a production of zero. All exposed rivers empty into bays with salmon farms [25], while control rivers are at least 55 km away from the nearest farm.
Two data sets from Newfoundland were examined—marine survival estimates of wild Atlantic salmon from four rivers from 1987 to 2004 were used in the survival analysis, and grilse returns to 21 rivers from 1986 to 2004 were used in the returns modeling [31]. Salmon farming in Newfoundland is confined to Bay d'Espoir on the south coast [32] (http://www.fishaq.gov.nl.ca/aquaculture/pdf/aqua_sites.pdf). Only the Conne River (in Bay d'Espoir) was considered exposed; the Little River (also in Bay d'Espoir) was excluded because it has been regularly stocked [31]. The Exploits and Rocky Rivers were also removed from the analysis because of stocking [33]. This left three control rivers for the survival analysis: the Campbellton River, the Northeast Brook (Trepassey), and Western Arm Brook. For the returns analysis, there were 18 control rivers: Campbellton, Crabbes, Fischells, Flat Bay Brook, Highlands, Humber, Lomond, Middle Brook, Middle Barachois, Northeast Brook (Trepassey), Northeast (Placentia), Northwest, Pinchgut Brook, Robinsons, Salmon, Terra Nova (upper and lower), Torrent, and Western Arm Brook. Salmon from control rivers are very unlikely to pass salmon farms because of the direction of their migrations [34]. Farmed salmon production data are from Fisheries and Oceans Canada (DFO) Statistical Services [32].
We compared Atlantic salmon returns to six rivers in the Bay of Fundy (New Brunswick and Nova Scotia, Canada) to returns to four rivers from other areas of New Brunswick and Nova Scotia. We grouped the six exposed rivers into three groups and estimated the impact of aquaculture on each group separately, because salmon from these three groups have different degrees of exposure to salmon farming. The three groups of exposed rivers are the inner Bay of Fundy group (Stewiacke and Big Salmon Rivers), the Saint John River group (Saint John and Nashwaak Rivers), and the outer Bay of Fundy group (St. Croix and Magaguadavic Rivers). Salmon farming in New Brunswick is highly concentrated in the Quoddy region of the outer Bay of Fundy (http://www.gnb.ca/0177/10/Fundy.pdf), although some farms are also found along the Nova Scotia coast of the Bay of Fundy. Salmon from control rivers enter into the Atlantic directly (LaHave River) or into the Gulf of St. Lawrence (Restigouche River, Miramichi River, Catamaran Brook) and do not pass by farms during their migrations. The same controls are used for all comparisons in New Brunswick and Nova Scotia. The estimates of returns to the rivers are published by DFO [28,35–40]. Outer Bay of Fundy salmon must pass through an area containing many salmon farms early during their migrations [41]. Although Saint John River salmon enter the ocean in an area without salmon farms, they are known to pass through the region containing many farms early during their migrations [41]. Salmon from inner Bay of Fundy rivers are considered exposed to salmon farming despite being up to 260 km away because of historical information indicating that juvenile salmon from these populations are found during the summer and fall in the area where salmon farms are currently located [42]. However, the evidence that this region is important habitat for inner Bay of Fundy and Saint John River populations is mixed [43]. For this reason, we ran an alternative model with only outer Bay of Fundy populations considered exposed, and all other New Brunswick and Nova Scotia rivers as controls.
For all New Brunswick rivers, an estimate of egg deposition was used as an index of spawners, to account for a significant increase in the age of spawners in many rivers over the study period. The number of grilse (salmon maturing after one winter at sea) and large spawners (repeat spawners or salmon maturing after two or three winters at sea) in each year was multiplied by a river-specific estimate of fecundity for a salmon of that size. Then, the index of spawners in a given year was derived by adding up all the eggs that could produce smolts in a year y, using river-specific ages at smolting from the literature. Returning hatchery-origin spawners are also added to the “spawners” but not to “returns.” “Recruits” is the number of grilse that return to each river in year y + 1, so that
(in Equation 1) is the number of grilse returning per egg that would have smolted in year y. Estimates of returns to rivers from traps and other surveys were used in the returns analysis. No corrections were made to account for marine fisheries, but marine exploitation has been quite limited since the late 1980s, when salmon farming became a substantial industry [44]. Farmed salmon production data are from DFO Statistical Services [32].
For coho salmon in British Columbia (BC), spawner estimates are based on DFO's escapement database (NuSEDS), which includes estimates of spawning salmon of all species for hundreds of rivers and streams on the BC coast since 1950 (P. VanWill, DFO Pacific, unpublished data). We considered rivers on the east side of the Queen Charlotte and Johnstone Straits to be exposed (all rivers from Wakeman Sound to Bute Inlet, DFO Statistical Areas [SAs] 12 and 13). All rivers on the BC Central Coast from Finlayson Channel to Smith Inlet (SAs 7, 8, 9, and 10) were included as controls. In the regions considered exposed in BC, all salmon must pass by farms to get into the open ocean, although in some cases, the farms are at the end of long channels down which the salmon migrate (as far as 90 km in the most extreme case). Control populations to the north do not pass by farms, because of the direction of their migration routes [45].
Coverage in the NuSEDS database varies considerably in time and space, as does the quality of the estimates. We changed all indicators of unknown values (including “none observed” and “adults present”) to a common missing value indicator. To reduce effects of inconsistent monitoring procedures, only data since 1970 were included in the analysis. All rivers known to be regularly stocked with hatchery salmon or to contain constructed spawning channels were also removed from exposed and control areas, leaving 49 exposed and 70 control rivers. Estimates were combined for each SA, the smallest areas for which catch rates are estimated. This was done by modeling returns to each SA and year, using a generalized linear model with negative binomial errors. The predicted returns for each SA were then used as spawner estimates (Si,y in Equation 1). To derive recruitment estimates, we followed Simpson et al. (2004) [46], applying exploitation rate estimates from Toboggan Creek (J. Sawada, DFO Pacific, personal communication) to the controls, and the average of the exploitation rates for Quinsam Hatchery, Big Qualicum Hatchery, and the Black Creek wild indicator population to the exposed stocks. After 1998, only the estimates from Black Creek were used for exposed stocks. Recruitment estimates for coho were based on the assumption that coho follow a fixed 3-y life cycle.
For pink, chum, and coho salmon, aquaculture production estimates include all salmon species farmed in SAs 12 and 13 (the Queen Charlotte and Johnstone Straits) from 1990 to 2003 (H. Russell, BC Ministry of Agriculture, Food, and Fisheries, unpublished data). In years when two or fewer companies were raising salmon in either area, estimates were not available. BC salmon farm locations are made available at http://www.al.gov.bc.ca/fisheries/licences/MFF_Sites_Current.htm.
Estimates of pink salmon spawner abundance were derived in the same manner as described above for coho salmon. “Returns” are spawners plus catch for a given year, assuming a fixed two year life cycle. The same regions were considered exposed, but because enumeration varies by species, there were only 36 exposed rivers from SAs 12 and 13 (from Wakeman Sound to Bute Inlet) included. Wood et al. (1999) [47] consider the pink salmon catches in SAs 8, 9, and 10 to consist mainly of salmon returning to those areas (respectively), so catch data from DFO [48] were used in each of these SAs. Area 7 was excluded from the survival analysis because catches for SA 7 are difficult to estimate due to the adjacent regions being much larger [47], leaving 47 control rivers from Burke Channel to Smith Inlet.
For Queen Charlotte and Johnstone Straits (the exposed areas), DFO does not estimate catches at the level of individual SA. To obtain approximate returns to each exposed SA, we found the proportion of total escapement to the Straits that was in our dataset (i.e., regularly enumerated rivers on the east side of the Straits without a major hatchery or constructed spawning channel) and assumed the same proportion of the total catch would be returning to those rivers (i.e., assumed equal catchability across stocks). For odd years, we used estimates from the Pacific Salmon Commission (B. White, unpublished data) of the catch of pink salmon in Johnstone and Georgia Straits that were not returning to the Fraser River. In even years, there is no pink salmon run on the Fraser River, so total returns to the Straits could be used.
For chum salmon, we used estimates of returns (i.e., before exploitation) and spawners to large coastal areas [49]. Chum from the east side of Queen Charlotte and Johnstone Straits, from Wakeman Sound to Bute Inlet (SAs 12 and 13) were considered exposed to salmon farming, while chum from the Central Coast from Bute Channel to Seymour Inlet (SAs 8–11) were considered controls. Estimates were available as a single time series for the exposed area, and a time series for each SA for the controls. An index of recruits per spawner was generated by lining up returns with spawners according to age distributions given in Ryall et al. (1999) [50], to 1998, and then the average values from 1988–1998 for the subsequent years, to 2003.
|
10.1371/journal.pntd.0005992 | Candidate genes-based investigation of susceptibility to Human African Trypanosomiasis in Côte d’Ivoire | Human African Trypanosomiasis (HAT) or sleeping sickness is a Neglected Tropical Disease. Long regarded as an invariably fatal disease, there is increasing evidence that infection by T. b. gambiense can result in a wide range of clinical outcomes, including latent infections, which are long lasting infections with no parasites detectable by microscopy. The determinants of this clinical diversity are not well understood but could be due in part to parasite or host genetic diversity in multiple genes, or their interactions. A candidate gene association study was conducted in Côte d’Ivoire using a case-control design which included a total of 233 subjects (100 active HAT cases, 100 controls and 33 latent infections). All three possible pairwise comparisons between the three phenotypes were tested using 96 SNPs in16 candidate genes (IL1, IL4, IL4R, IL6, IL8, IL10, IL12, IL12R, TNFA, INFG, MIF, APOL1, HPR, CFH, HLA-A and HLA-G). Data from 77 SNPs passed quality control. There were suggestive associations at three loci in IL6 and TNFA in the comparison between active cases and controls, one SNP in each of APOL1, MIF and IL6 in the comparison between latent infections and active cases and seven SNP in IL4, HLA-G and TNFA between latent infections and controls. No associations remained significant after Bonferroni correction, but the Benjamini Hochberg false discovery rate test indicated that there were strong probabilities that at least some of the associations were genuine. The excess of associations with latent infections despite the small number of samples available suggests that these subjects form a distinct genetic cluster different from active HAT cases and controls, although no clustering by phenotype was observed by principle component analysis. This underlines the complexity of the interactions existing between host genetic polymorphisms and parasite diversity.
| Since it was first identified, human African trypanosomiasis (HAT) or sleeping sickness has been described as invariably fatal. Recent data however suggest that infection by T. b. gambiense can result in a wide range of clinical outcomes in its human host including long lasting infections, that can be detected by the presence of antibodies, but in which parasites cannot be seen by microscopy; these cases are known as latent infections. While the factors determining, this varied response have not been clearly characterized, the effectors of the immune responses have been partially implicated as key players. We collected samples from people with active HAT, latent infections and controls in endemic foci in the Côte d’Ivoire. We tested the role of single nucleotide polymorphisms (SNPs) in 16 genes on susceptibility/resistance to HAT by means of a candidate gene association study. There was some evidence that variants of the genes for IL4, IL6, APOL1, HLAG, MIF and TNFA modified the risk of developing HAT. These proteins regulate the inflammatory response to many infections or are directly involved in killing the parasites. In this study, the results were statistically weak and would be inconclusive on their own, however other studies have also found associations in these genes, increasing the chance that the variants that we have identified play a genuine role in the response to trypanosome infection in Côte D’Ivoire.
| Sleeping sickness, or human African Trypanosomiasis (HAT), is caused by Trypanosoma brucei gambiense and T.b. rhodesiense and is transmitted by tsetse flies (Glossina spp). T. b. gambiense is associated with a more chronic disease that can take decades to become patent, and T.b. rhodesiense causes an acute disease within months of infection. The chronic form of the disease caused by the T. b. gambiense is classically characterized by an early hemolymphatic stage (stage 1) associated with non-specific symptoms such as intermittent fevers and headaches, followed by a meningo-encephalitic stage (stage 2) in which the parasite invades the central nervous system and causes neurological disorders and death if left untreated. This chronic form is found in Western and Central Africa while the acute form caused by T. b. rhodesiense is endemic to Eastern Africa [1,2]. However, recent observations are increasingly indicating that infection by T. b. gambiense can result in a wide range of clinical outcomes in its human host [3,4]. Self-cure processes have been described and reviewed in Checchi et al. [3]. Furthermore, some authors argued that HAT is not invariably fatal [5], supporting observations made by Garcia et al. [4] who followed individuals who remained seropositive for HAT but without detectable parasites for two years. This clinical diversity is not well understood but could be due to parasite genetic diversity [6,7], human immune gene variability [8] or their interaction [9]. It has been suggested that genetic polymorphisms of the parasite could be associated with asymptomatic and very chronic infections [10]. Nevertheless, genes involved in the host immune response have been implicated in the control of infection or susceptibility to HAT [11,12] and also T. congolense infections in experimental models [13,14]. Among the genes implicated in the pathogenesis of the disease, are interleukins 1, 4, 6, 8, 10 and 12 (IL1, IL4, IL6, IL8, IL10 and IL12), tumor necrosis factor A (TNFA), interferon G (INFG), complement factor H (CFH), macrophage inhibitory factor (MIF), [8,11,15–17].
Some studies have found associations between certain polymorphisms in genes encoding cytokines. For example, IL4 plays a role in susceptibility to T. brucei infection [18], while polymorphisms in the IL6 and IL10 genes have been associated with a decreased risk of developing HAT [11,15]. On the other hand, polymorphisms in TNFA genes have been associated with an increased risk of developing the disease [11,15]. Furthermore, MacLean et al. [19] reported in their study that plasma levels of IFNG significantly decrease during the late stage of the T.b. rhodesiense disease [19]. Human blood contains trypanolytic factors 1 and 2 (TLF1 and TLF2) that are lytic to almost all African trypanosomes except T. b. rhodesiense and T. b. gambiense [20]. Human TLF1 contains two primate-specific proteins, apolipoprotein L1 (APOL1) and haptoglobin-related protein (HPR). APOL1 expression is induced by T. b. gambiense infection but expression is not associated with susceptibility to sleeping sickness [21]. MIF contributes to inflammation-associated pathology in the chronic phase of T. brucei infection [16].
Although genes directly involved in the immune response, like genes encoding cytokines, are very important candidates, genes implicated in the regulation of immunity also have a critical role. Thus, Courtin et al. [12] have shown a genetic association between human leukocyte antigen G (HLA-G) polymorphisms and susceptibility to HAT.
There is now cumulative evidence that polymorphisms in genes involved in the control of the immune response and genes implicated in the regulation of immunity could play a role in HAT infection outcome [11,12].
We report a candidate gene association study of the role of single nucleotide polymorphisms (SNP) in IL1, IL4, IL6, IL8, IL10, IL12, interleukin 4 receptor (IL4R), interleukin 12 receptor (IL12R), TNF, INFG, MIF, haptoglobin-related protein (HPR), CFH, APOL1, HLA-A and HLA-G genes on susceptibility/resistance to HAT.
A total of 100 cases, 100 controls and 33 latent infections were enrolled into the study (Table 1). The mean age (range) of the study population was 38.8 (6–84) years. The sex ratio (male: female) was 0.88 (109/124).
Three association analyses were run on the three possible pairwise combinations of the controls, latent infections and active HAT cases. After filtering out SNP loci that were not in Hardy Weinberg equilibrium (HWE) (P <0.001), or had > 10% missing genotypes or had minor allele frequencies of zero, 77 SNP loci remained in the latent infection vs active case comparison and 74 SNP loci remained in the other two comparisons. One individual was filtered out of the case sample and one out of the latent infection sample because of low genotyping rates (<90%).
FST between ethnic groups showed that allele frequency differences between the ethnic groups were small, indicating that ethnicity was unlikely to confound results (median Fst -0.00011763, Fst maximum 0.015). Scatter plots of the first two principal components between the different phenotypes (cases, latent infections and controls) and different ethnic groups show that this population is homogeneous and that the samples did not cluster by phenotype or by ethnic groups (Fig 1) and consequently the data were not stratified by ethnicity.
There were suggestive associations at three loci in IL6 and TNFA in the comparison between active cases and controls (Table 2), three loci in APOL1, MIF and IL6 in the comparison between latent infections and active cases (Table 3) and five loci in IL4 (Fig 2), and one each in HLA-G and TNFA between latent infections and controls (Table 4). After Bonferroni correction, none of these associations remained significant, however Bonferroni correction is very conservative, particularly since there was some linkage between adjacent SNP marker in some genes. The Benjamini-Hochberg false discovery rate (FDR) test, which shows the probability that an observation is a false positive, indicated that at least some of the suggestively positive samples may be true positives. Under the FDR the rs62449495 SNP in IL6 had an 80% chance of being a true positive and the TNFA-308 rs1800629 SNP had a 71% chance of being a true positive (Table 2). The strongest association was with APOL1 rs73885319, which is also known as the G1 allele, which had a 90% chance of being a true positive in the comparison between latent infections and controls (Table 3). Complete results for all tests that passed quality control are shown in supplementary data tables (S1, S2 and S3 Tables).
In this study, we investigated the role of SNPs on susceptibility/resistance to HAT. We genotyped 96 SNPs within sixteen genes that were investigated to test genetic association with HAT in Côte d’Ivoire. For consistency, all samples were commercially genotyped using two platforms: Genome Transcriptome de Bordeaux, France and LGC Genomics UK.
Our results suggest that three SNPs were associated with the development of HAT by controls, another three were associated with the development of active HAT by people with latent infections, and seven SNP were associated with the risk of developing latent infections five of them in IL4 although none of these results remained significant after Bonferroni correction. Only two of the SNP with suggestive associations were coding and missense (IL6 rs2069830 and APOL1 rs73885319) since, with some exceptions, SNP were primarily selected as tags that might be linked to functional SNP rather than for any putative function.
The most striking feature of the results is that most associations were found in the comparisons with latent infections, despite limited power, with only 33 of these samples being available compared with 100 each for the cases and controls. This illustrates the increased power to be gained from using well defined phenotypes. Given the evidence that some people in West Africa can self-cure their T. b. gambiense infections [5] it is likely that some of the control subjects are resistant to infection or have recovered from infection, whilst others are naïve and susceptible. These two groups may have very different genetic backgrounds which could confound the association studies. The data also suggest that the latent infections may represent a genetically distinctive group. Although the principle component analysis did not identify any cluster associated with latent infections (Fig 1) further studies of their genetic and immunological profiles are required to test this hypothesis.
The minor (A) allele of IL6 rs62449495 appeared to protect against progression from latent infections to active HAT (Table 3) and against the development of active HAT by controls but these associations did not remain significant after Bonferroni correction (Table 2). The major allele of rs2069830 was also protective against the development of active HAT by controls before but not after Bonferroni correction (Table 2). The rs2069830 SNP causes a proline to serine change, but this is predicted to be benign by both Sift and Polyphen according to the Ensembl Variant Effect predictor. IL6 plays a key role in the acute inflammatory response and in regulation of the production of acute phase proteins such as C-reactive protein [22]. It contributes to the inflammatory response, regulates haematopoiesis, which may contribute to the anaemia associated with HAT, and modifies the permeability of the blood brain barrier, which may contribute to the development of the stage 2 invasion of the CNS [22,23]. The results of our association study are consistent with the involvement of IL6 variants in the susceptibility to the disease as Courtin et al [15] have reported before. However, it should be noted that the polymorphisms of IL6 in our study (rs62449495 and rs2069830) are different of those found by Courtin et al. [15] with IL64339 = rs2069849. They showed that in DRC, IL64339 SNP was significantly associated with a decreased risk of developing the disease with a P-value = 0.0006 before the Bonferroni correction (0.04 after Bonferroni Correction).
Our results also indicate that subjects carrying the A allele of TNFA (rs1800629) A/G had a lower risk of developing active HAT (Table 2) or a latent infection (Table 4), suggesting the possibility of a protective effect. This SNP is also known as the TNF-308 SNP and the minor A allele is associated with higher plasma levels of TNFA [24]. In a previous study in Côte d’Ivoire, Courtin et al. [11] showed that the distribution of the TNFA-308G/A polymorphism did not differ significantly between cases and controls in the total population, but found that, under a recessive model the AA genotype was associated with risk of HAT in the 39 cases and 57 controls who had been living in the endemic area of Sinfra for less than 10 years (before Bonferroni correction). In contrast, in our study, all participants had lived in the endemic area all their lives and the association was additive rather than recessive, furthermore the MAF in Courtin’s studies was 24%, whilst in ours it was 14%, suggesting significant differences in population structure between the two studies. Varying results in TNFA associations studies are not uncommon, multiple candidate gene association studies of TNFA-308 (rs1800629) and malaria have found inconsistent results in different populations [25,26]. Another explanation could be that the TNFA-308 has no effect but is in linkage disequilibrium (LD) with another unidentified polymorphism. Varying LD across populations might lead to different findings [27]. Despite the conflicting results from association study, animal models indicate that TNFA is likely to be a key mediator in the control of T. brucei infections [28] and a direct dose dependent lytic effect of TNFA on purified T.b. gambiense parasites has been reported suggesting an involvement in parasite growth control [29,30]. Given the experimental evidence for a role for TNFA, it is possible that inconsistent results of association studies are a consequence of different TNFA alleles only making a small difference to infection outcome and that therefore larger studies would be needed to detect associations and also to heterogeneity in regulation of TNFA between populations.
Five of the seven SNP that were suggestively associated with the comparison between latent infections and controls were in IL4 (Table 4, Fig 2). Three of these SNP were adjacent to each other at the 5’ end of the gene (Fig 2), although linkage between them was modest (r2 <0.5). While the individual associations did not remain significant after Bonferroni correction the observation that five of the sixteen SNP tested in IL4 had suggestive associations increases the probability of a genuine association with IL4 and the risk of developing a latent infection. A study using three IL4 SNP in DRC did not find any associations, but that was in a comparison between cases and controls [15]. In B10.Q mice deletion in IL4 lead to increased T. brucei parasitaemia levels but longer survival time [18]. The relationship between IL4 polymorphisms and acute HAT infection requires further study.
The suggestive association with APOL1 G1 allele rs73885319 and protection against progression from latent infection to active HAT is consistent with the association found in Guinea [31,32], although in those studies this SNP was also associated with increased risk of controls developing a latent infection as well. Moreover, APOL1 expression is induced by T. b. gambiense infection but not associated with differential susceptibility to sleeping sickness [21]. APOL1 rs73885319 is also known as the G1 allele of APOL1 and is associated with kidney disease in African Americans and the relatively high frequency of this deleterious allele was assumed to be due to selection by HAT [33]. The data presented here is consistent with that hypothesis, but no support was found for the role of the G2 allele of APOL1 (rs71785313) in HAT although it is implicated in kidney disease.
Although, our data show that subjects carrying the G allele of MIF (rs36086171) G/A had a risk of developing active HAT (Table 3), we did not find a significant difference after correction (BONF = 0.52). MIF is a ubiquitously expressed protein that has proinflammatory, hormonal and enzymatic activities [34]. It is implicated in many inflammatory diseases [35]. It functions by recruiting myeloid cells to the site of inflammation [36], by inducing their differentiation towards M1 cells secreting TNF [37] and by suppressing p53-dependent apoptosis of inflammatory cells. While there are no human studies directly linking MIF to HAT, murine studies show that MIF plays a role as a mediator of the inflammation which is a key feature in trypanosomiasis-associated pathology [16,38].
HLA-G molecule plays an important role on immune response regulation and has been associated with the risk of human immunodeficiency virus (HIV) infection [39], human papilloma virus [40] and herpes simplex virus type 1 [41]. The results of our association study are consistent with the involvement of HLA-G genetic variants in the susceptibility to the disease [12]. Control subjects with rs1611139 T allele of the HLA-G gene showed a suggestive association with increased risk of leading in latent infection in case they are infected. This result could be due to differences in selection pressures possibly driven by variability in the immuno-pathology of the diseases. It is known that cytokines such as IL-10 can induce HLA-G expression by affecting mRNA transcripts and protein synthesis by human monocytes and trophoblasts, thus having a significant impact on parasitic infections [42].
The results discussed in this paper should be used with caution as no loci remained significantly associated with HAT after the Bonferroni correction for multiple testing, although some had high probabilities of being associated using a FDR test. Some of the SNP loci identified here were also significant in other studies. Multiple independent observations of marginally significant effects suggest that these effects may be genuine but that the effect size is not large enough to be detected by the numbers available to be tested.
Our data support the findings from Guinea about the role of the APOL1 G1 allele and also suggest that the polymorphisms of IL4, IL6, HLA-G and TNFA present interesting candidates for the investigation of the genetic susceptibility / resistance to HAT.
The large number of suggestive associations with latent infections despite the small number of samples indicate the people with latent infections form a genetically distinctive group that merit further investigation.
The study took place in western-central Côte d’Ivoire in the main HAT foci of Bonon, Bouafle, Zoukougbeu, Oume, and Sinfra. Samples were collected in three stages: (i) previously archived samples; existing collections of previously archived samples were centralized to Centre International de Recherche-Développement sur l’Elevage en zone Subhumide (CIRDES) in Burkina-Faso, (ii) retrospectively collected samples for which study sites were revisited for consent and resample previously diagnosed and treated patients and (iii) prospectively collected samples including new HAT cases. Three phenotypes were considered: (i) cases, defined as individuals in whom the presence of trypanosomes was confirmed by microscopy and trypanolysis test positive (TL+ve) but trypanolysis negative (TL-ve) for some cases who were retrospectively sampled for the purpose of this study and who had not been subject to the trypanolysis test (TL) when originally identified as HAT cases by microscopy; (ii) controls, defined as individuals living in the endemic area who are card agglutination test for trypanosomiasis (CATT) negative (CATT-ve) and trypanolysis test negative (TL-ve), and without evidence of previous HAT infection and (iii) latent infections, defined as subjects who were CATT positive (CATT+ve) with plasma dilution higher or equal to ¼ (CATTpl ≥1/4) but in whom parasites could not be detected by microscopy at repeated tests for at least two years.
A total of 233 individuals with 100 HAT cases, 100 controls and 33 latent infections were included in the study. Samples were frozen directly in the field at -20°C and kept at that temperature until use. For each individual, an aliquot of plasma was used to perform the immune TL test that detects Litat 1.3 and Litat 1.5 variable surface antigens specific for T.b. gambiense [43].
This study was one of five populations studies of HAT endemic areas in Cameroon, Côte d’Ivoire, Guinea, Malawi and Uganda by the TrypanoGEN consortium [44]. The studies were designed to have 80% power to detect odds ratios (OR) >2 for loci with disease allele frequencies of 0.15–0.65 and 100 cases and 100 controls with the 96 SNPs genotyped. Power calculations were undertaken using the genetics analysis package gap in R [45].
Genomic DNA was obtained from peripheral blood samples. Extraction was performed using the Qiagen DNA extraction kit (QIAamp DNA Blood Midi Kit) according to the manufacturer’s instructions and quantified by Nanodrop assay. DNA was stored at -20°C until analysis.
Ninety-six SNPs were genotyped in IL1, IL4, IL4R, IL6, IL8, IL10, IL12, IL12R, TNFA, INFG, MIF, HPR, CFH, APOL1, HLA-A, and HLA-G. The SNPs were selected by two strategies: 1) SNP in IL4, IL6, IL8, HLAG and IFNG were designed as markers for linkage disequilibrium (LD) scans of each gene [46], 2) SNPs in other genes were selected based on reports in the literature that they were associated with trypanosomiasis or other infectious diseases. The LD scans were designed using 1000 Genomes Project data [47], merged with low fold coverage (8-10x) whole genome shotgun data generated from 230 residents living in regions (DRC, Guinea Conakry, Côte d’Ivoire and Uganda) where trypanosomiasis is endemic (TrypanoGEN consortium, European Nucleotide Archive study EGAS00001002482). Loci with minor allele frequency < 5% in the reference data were excluded and an r2 of 0.5 was used to select SNP in LD.
DNA was genotyped by two commercial service providers: INRA- Site de Pierroton, Plateforme Genome Transcriptome de Bordeaux, France and ii- LGC genomics Hoddesden UK. At INRA multiplex design (two sets of 40 SNPs) was performed using Assay Design Suite v2.0 (Agena Biosciences). SNP genotyping was achieved with the iPLEX Gold genotyping kit (Agena Biosciences) for the Mass Array iPLEX genotyping assay, following the manufacturer’s instructions. Products were detected on a Mass Array mass spectrophotometer and data were acquired in real time with Mass Array RT software (Agena Biosciences). SNP clustering and validation was carried out with Typer 4.0 software (Agena Biosciences). At LGC Genomics SNP were genotyped using the PCR based KASP assay[48].
All statistical analyses were performed using the Plink 1.9 and R v 3.2.1 software. Individuals were excluded who had > 10% missing SNP data. SNP loci were excluded that > 10% missing genotypes or if the control samples were not in Hardy-Weinberg equilibrium (p<0.001). Case-control association analysis using SNP alleles/genotypes was undertaken using Fisher’s exact test. The difference between the ethnic groups was estimated using the FST, with values potentially varying from zero (no population differentiation) to one (complete differentiation) [47]. Ethnicity was not used as a risk factor for HAT because we observed no significant differences in FST between ethnic groups (FST maximum 0.015) or clustering by linguistic group by multidimensional scaling in Plink. Some studies have used ethnicity and age as risk factors for HAT but found no significant association [11,15]. 77 SNP remained in at least one comparison after filtering. P values were adjusted for multiple testing using the Bonferroni corrections and Benjamini Hocheberg false discovery rate test as implemented in Plink.
The population of the study was informed. All adult subjects provided written informed consent. For children, a parent or guardian of any child participant provided written informed consent on their behalf. The protocol of the study was approved by the traditional authorities (chief and village committee) and by National ethics committee hosted by the Public Health Ministry of Côte d’Ivoire with the number: N°38/MSLS/CNERm-dkn 5th May 2014. This study is part of the TrypanoGen project which aims to a better understanding of genetic determinism of human susceptibility to HAT and the TrypanoGen-CI samples were archived in the TrypanoGen Biobank hosted by CIRDES in Bobo-Dioulasso, Burkina Faso [44].
|
10.1371/journal.pntd.0000612 | Asymptomatic Renal Colonization of Humans in the Peruvian Amazon by Leptospira | Renal carriage and shedding of leptospires is characteristic of carrier or maintenance animal hosts. Sporadic reports indicate that after infection, humans may excrete leptospires for extended periods. We hypothesized that, like mammalian reservoir hosts, humans develop asymptomatic leptospiruria in settings of high disease transmission such as the Peruvian Amazon.
Using a cross-sectional study design, we used a combination of epidemiological data, serology and molecular detection of the leptospiral 16S rRNA gene to identify asymptomatic urinary shedders of Leptospira. Approximately one-third of the 314 asymptomatic participants had circulating anti-leptospiral antibodies. Among enrolled participants, 189/314 (59%) had evidence of recent infection (microscopic agglutination test (MAT0 ≥1∶800 or ELISA IgM-positive or both). The proportion of MAT-positive and high MAT-titer (≥1∶800) persons was higher in men than women (p = 0.006). Among these people, 13/314 (4.1%) had Leptospira DNA-positive urine samples. Of these, the 16S rRNA gene from 10 samples was able to be sequenced. The urine-derived species clustered within both pathogenic (n = 6) and intermediate clades of Leptospira (n = 4). All of the thirteen participants with leptospiral DNA in urine were women. The median age of the DNA-positive group was older compared to the negative group (p≤0.05). A group of asymptomatic participants (“long-term asymptomatic individuals,” 102/341 (32.5%) of enrolled individuals) without serological evidence of recent infection was identified; within this group, 6/102 (5.9%) excreted pathogenic and intermediate-pathogenic Leptospira (75–229 bacteria/mL of urine).
Asymptomatic renal colonization of leptospires in a region of high disease transmission is common, including among people without serological or clinical evidence of recent infection. Both pathogenic and intermediate Leptospira can persist as renal colonization in humans. The pathogenic significance of this finding remains to be explored but is of fundamental biological significance.
| Leptospirosis is a bacterial disease commonly transmitted from animals to humans. The more than 200 types of spiral-shaped bacteria (spirochetes) in the genus Leptospira are classified as pathogenic, intermediately pathogenic, or saprophytic (meaning not causing infection in any mammal) based on their ability to cause disease and on genetic information. Unique among the spirochetes that infect humans, Leptospira live both in the environment (in surface waters and moist soils), and in mammals, where they cause chronic infection by colonizing kidney tubules. Infected animals are the source of human infection, but humans have not been systematically studied as chronic Leptospira carriers. In our study, we found that more than 5% of people (in fact, only women) in a rural Amazonian village, without clinical evidence of infection by Leptospira, were chronically colonized by the bacteria. Chronic infection was not associated with a detectable immune response against the spirochete. Pathogenic and intermediately pathogenic Leptospira caused asymptomatic, chronic kidney infections. Future work is needed to determine whether such chronic infection can lead to human-to-human transmission of leptospirosis, and whether subtle measures of kidney disease are associated with asymptomatic, long-term leptospiral infection.
| Leptospirosis is a zoonotic disease caused by spirochetes of the genus Leptospira. Found worldwide, leptospirosis is more common in tropical and sub-tropical areas where environmental and socioeconomic conditions favor its transmission. It has been identified in recent years as a global public health problem because of its increased mortality and morbidity. The disease is principally transmitted to humans indirectly by contact with water or soil contaminated with the urine of domestic and wild animals with persistent renal infection by Leptospira [1]–[3].
The tropical climate of the Peruvian Amazon region of Iquitos is ideal for the maintenance and transmission of leptospirosis. In developing countries, impoverished populations typically live either in rural areas or under highly crowded conditions in urban slums. These factors increase the risk of human exposure to the urine of Leptospira-infected animals [4],[5]. In the Iquitos region, leptospirosis is common. Seropositivity as seen in cross-sectional surveys is high [6]; more than half of patients presenting to urban and rural community-based health posts with non-malarial acute febrile illness have been observed to have diagnostic levels of anti-leptospiral antibodies suggestive of acute leptospirosis [4]. The majority of patients enrolled presented with a self-resolving undifferentiated febrile illness with 70% of them having antibodies against a newly described Leptospira species, L. licerasiae [7]. These data suggest that exposure to Leptospira is common in daily life in this tropical setting [5],[8], and that, in general, Iquitos is accurately classified as hyper-endemic for leptospirosis infection.
Leptospirosis in humans is frequently misidentified because of several factors: 1) variable and nonspecific clinical presentation; 2) lack of awareness of the disease among clinicians; and 3) difficulty in access to reliable and rapid diagnostic tests. Clinical manifestations, when present, vary from a mild ‘flu-like’ febrile illness to a severe disease variably including jaundice, renal failure, pulmonary hemorrhage, refractory shock and other grave manifestations. However, many if not most people infected by Leptospira develop sub-clinical disease or have very mild symptoms, and thus do not seek medical attention [1],[2]. Asymptomatic infection, common in endemic areas, has been reported in several studies [8]–[12]. For example, in one study, 9–48% of healthy subjects were diagnosed as having asymptomatic leptospiral infection by serology (ELISA-IgM) and PCR [10]. However, in this study, the identity of the infecting strains could not be determined because of study design. We have observed in one study that patients can have asymptomatic leptospiruria for prolonged periods of time [4]. Hence an essential question about the pathogenicity of Leptospira remains: are some serovars are more likely than others to establish asymptomatic renal infection in man?
Renal colonization and persistent shedding of leptospires is characteristic of carrier or maintenance animal hosts [13]–[15]. Animals, especially rodents, are known reservoirs of pathogenic Leptospira species, but rarely develop symptoms and are not impaired by the infection of their kidneys. After infection, humans can also excrete leptospires into the urine transiently for weeks or, more rarely, months or more [1],[2],[16].
We hypothesized that like mammalian reservoir hosts, humans develop asymptomatic leptospiruria, including pathogenic Leptospira such as L. interrogans and intermediate pathogens such as the newly discovered L. licerasiae [7]. To test this hypothesis, we carried out a cross-sectional, population-based study in a rural village near the city of Iquitos to identify the presence and species of infecting Leptospira directly in the urine of healthy ambulatory people. If found, we reasoned that the high prevalence of asymptomatic urinary infection might provide fundamental insights into the nature of Leptospira-human interactions, where humans are considered to be accidental hosts. Such a finding would also provide the basis for understanding mechanisms of naturally acquired immunity in human leptospiral infection.
This study was approved by the Human Subjects Protection Program, University of California San Diego, and the Ethical Committees of Asociacion Benefica PRISMA, Lima, Peru, and Universidad Peruana Cayetano Heredia, Lima, Peru. All human subjects provided written informed consent before being enrolled in the study.
This study was carried out in the village of Padrecocha, a rural community near Iquitos, located north of the city along the Nanay River, a tributary that branches from the Amazon River 15 km downstream from Iquitos. The climate is tropical: rainfall averages 300 mm per year and temperatures range from 21.8°C to 31.6°C; the village is surrounded by a vast expanse of humid tropical rainforest. The population of this village is approximately 1,500. Most inhabitants live in brick houses, and their water supply comes from wells and local streams. These water sources harbor pathogenic and intermediate-pathogenic Leptospira [5]. Residents use water from wells or from the local streams for their daily needs (cooking, bathing and washing clothes). There is no sewage system; most households have pit latrines. Livestock (mostly chickens, pigs, and cattle) roam free through the village and its streams; the inhabitants observe rats frequently.
Using a whole-village canvassing strategy to develop a set of candidate houses from which to randomly select asymptomatic inhabitants of the rural village Padrecocha of age ≥5 years for enrollment. Subjects were excluded if they had fever within the previous 2 weeks or if they declined participation (Figure 1). All participants were clinically evaluated and subjected to an epidemiologic questionnaire. Whole blood (5 mL) and urine samples (5—50 mL) were collected from each enrollee.
Venous blood samples were drawn into tubes without anticoagulant (Becton-Dickinson, USA) and transported to the study laboratory within 4 hr at ambient temperature. Serum was separated, frozen in 1 mL aliquots at −20°C, and transported on dry ice to the National Leptospirosis Reference Laboratory at the Instituto Nacional de Salud (INS) in Lima, where the presence of anti-leptospiral antibodies was determined. An ELISA incorporating 6 pathogenic serovars (strains)–Icterohaemorrhagiae (RGA), Australis (Ballico), Bratislava (Jez Bratislava), Ballum (MUS127), Canicola (Hond Utretch IV), Cynopteri (3522 C), and Grippotyphosa (Moskva V)–was used to detect anti-leptospiral IgM antibodies. An ELISA IgM result of 11.0 IU/mL or more was considered to be positive [4],[7]. Microscopic agglutination testing (MAT) was performed using 25 leptospiral antigens, using the Centers for Disease Control and Prevention (CDC) panel [17]. MAT titers were reported as the reciprocal of the number of dilutions still agglutinating 50% of live bacterial antigen and a titer of 1∶100 or more was considered as positive.
The proportion of the seropositivity rate (at any MAT titer) and distribution of the demographic variables were compared between the subjects with and without leptospiruria using the chi-square test and Mann-Whitney U tests using Stata v8 for Windows (StataCorp, College Station, Texas) with a significance level (α) of 0.05.
In the pre-study census and sampling period, 1320 people in 225 houses were identified in the Peruvian Amazon village of Padre Cocha near Iquitos. The study enrolled 354 participants of age ≥5 years from 175 households randomly picked from a census map. Of those 354, 40 participants were excluded since 19 presented with fever within 2 weeks of enrollment and 21 did not provide urine samples (patient enrollment diagrammed in Figure 1). The study included 314 participants with a median age of 27 (range 5—64). More were female than male (212 vs. 102); 63 (20%) were children younger than 15 years old (Table 2). Men (median = 28.5 years, range (25%–75%) = (16.5 – 37) were on average slightly older than women (median = 25, range = 19–43) with borderline significance (p = 0.051).
Blood samples were available from 282 of 314 participants (89.8%). Of these, 97 were from males and 185 from females. Circulating anti-leptospiral antibodies were found, by either IgM ELISA or MAT or both, in 108 (38%) of the 281 subjects for whom serological data were available (Table 2). The most frequently observed serological reactivity (highest titer by MAT) was to serogroup Australis (34/281, 12.1%); with serogroups Djasiman (16/281, 5.7%), Icterohaemorrhagiae (13/281, 4.6%) and Cynopteri (12/281, 4.3%) also represented. Of the 108 seropositive samples, 64 had serological evidence of recent sub-clinical infection: seven had MAT titers (≥1∶800), 23 were IgM-positive but MAT-negative and 34 were IgM and MAT-positive, indicative of recent or current leptospiral infection. The proportion of MAT-positive (reflecting any previous exposure to Leptospira) and high MAT-titer (≥800, reflecting recent infection) persons was higher in men (40.6% and 9.4%, respectively) than women (24.9% (p = 0.006) and 2.2% (p = 0.006)). The difference stayed significant after adjusting for the age.
The initial qPCR screening performed on-site in Iquitos detected 63 (20%) positive samples. Further evaluation of these samples with the nested PCR assay confirmed their positivity. Among these 63 PCR-positive samples, a newly designed dot-blot assay, designed to exclude false-positive samples containing only Atopobium DNA (Figure 3), identified 13/63 (21%) pCR-positive samples as true positives. We successfully cloned and sequenced the 16S rRNA gene from 10 of these dot-blot confirmed samples (Table 1); sequence data were not obtained from 3 samples. Species assignments were made by Bayesian phylogenetic analysis of the cloned 16S rRNA gene.
Analysis of these 10 dot-blot-confirmed urine samples showed that the 16S ribosomal RNA gene sequences clustered within both the pathogenic (n = 6) and intermediate clades of Leptospira (n = 4) (Figure 4). Although asymptomatic, one inhabitant (PAD304, Table 3) had serological evidence of acute infection (IgM-positive) indicating sub-clinical infection, and consequently excreted on average one hundred-fold more Leptospira/ml compared to IgM- and MAT-negative enrollees.
Serological results were available from 281 of the 314 participants including eight of the ten with Leptospira DNA positive urine (Table 3). Ten of 13 people with DNA-positive urine had negative results in both IgM and MAT.
All thirteen participants who had leptospiral DNA in their urine were women and the proportion (100%, 13/13) of the women was significantly higher compared to that in the leptospiral-DNA-negative group (66%, 199/301, p = 0.011). The median age of the DNA positive group (43 years, range (min – max) = 9−58) was older compared to the women in the negative group (median, 24; range, 5–60; years, p = 0.005). The difference stayed significant if the men were included in the negative group (median, 27 (5–64) years, p = 0.011). Univariate analysis did not show significant association between other epidemiological factors and leptospiral DNA positivity in urine (data not shown).
Thirteen of the enrolled 314 asymptomatic inhabitants (4.1%) were confirmed to excrete Leptospira by detection of leptospiral DNA in their urine; of these, one participant may have had recent but sub-clinical leptospiral infection, based on an ELISA finding of IgM positive (Table 3). After clinical and epidemiological assessment, a group of asymptomatic participants was identified (n = 102, 32.5% of enrolled individuals) that had no evidence of recent infection (without febrile episodes in the previous year before enrollment and without anti-Leptospira IgM antibodies detected); we call them “long-term asymptomatic individuals.” Within this group, six (5.9%) excreted pathogenic and intermediate-pathogenic Leptospira (75–229 bacteria/mL of urine, Table 3).
This study has several important findings. First, asymptomatic individuals living in a region hyperendemic for leptospirosis had a high rate of seropositivity (at any level) for leptospiral infection (38% of 314 participants). Almost 60% of the seropositive individuals had evidence of recent sub-clinical infection, as indicated by MAT titer ≥1/800. Second, and of unique interest, a novel 16S rDNA hybridization assay used to screen urine samples for the presence of leptospiral DNA found that almost 5% of healthy people living in a rural Amazonian community were urinary shedders of Leptospira but did not have serological or clinical evidence of recent infection. Third, we found that both pathogenic and intermediately pathogenic Leptospira persistent infected the renal tubules of humans. Such observations have not been reported previously and are particularly notable because they demonstrate that inapparent leptospiral infection is common and frequently leads to shedding of organisms in urine. The long-term clinical significance of this finding remains to be determined.
The occurrence of leptospirosis, and indeed many infectious diseases, depends on several interacting variables. These include favorable environmental conditions, the density of local reservoir host populations, the type and frequency of exposure, exposure to infectious doses of the etiologic agent, the virulence of the infecting strain, and the lifestyle preferences and susceptibility of individuals within the exposed human population [21]. In the context of this zoonotic infection, the density of local animal reservoir populations is likely an important determinant of the extent to which the environment may become contaminated by leptospires through urine from chronically infected carriers. When environmental conditions are ideal and background contamination is prevalent, social practices that predispose to infection, and the virulence of local strains are significant factors that affect the incidence of the disease [1]. To date, there is no evidence that humans contribute to environmental contamination with Leptospira, but the data presented here do not rule out this possibility.
Exposure to Leptospira in this rural Amazonian study population was common (∼39% were serologically positive at any MAT titer) with many subjects having evidence of recent sub-clinical infection. However, the serological data presented here need to be interpreted with caution: in an endemic setting, a high individual MAT titer (≥1∶800) and/or IgM positivity are not reliable indicators of recent or current infection as antibodies may persist for prolonged periods [22]. The high background exposure rates and relative absence of severe disease in this hyper-endemic region do suggest that long-term urinary shedding may occur more frequently here than elsewhere, where natural immunity may not be as common.
It is generally accepted that humans can excrete leptospires from weeks to months after infection [23],[24]. However the data presented here indicate that humans may excrete Leptospira for periods exceeding a year; extending previous understandings of the carrier state. Ten asymptomatic individuals without clinical (no febrile episodes in more than a year) or serological evidence of recent exposure were found to be shedding either pathogenic or intermediately pathogenic Leptospira in their urine. Although these persons may have been recently sub-clinically infected and either failed to produce anti-leptospiral antibodies or all produced ‘false-negative’ serology, these explanations seem unlikely. It is more likely that they represent long-term renal asymptomatic shedders of Leptospira, regardless of whether patients were subclinically infected or had acute illness. However, a prospective study would be needed to assess this possibility. Nonetheless, prospective observational studies of such patients are required to confirm this hypothesis
Our data also suggest that women (especially mature women) are more likely to develop long-term renal carriage of Leptospira than are men; with a significant increase in incidence with age in women, possibly reflecting increased exposure with age or alternatively increased susceptibility. It is possible that the conclusion that women are more likely to be long-term asymptomatic urinary shedders than men may reflect a bias in the study, considering its relative underrepresentation of men. However, this observation may also reflect increased susceptibility of women to persistent leptospiral kidney infections; the reasons for this are unclear. However in our study population, the MAT titer was significantly lower in women than in men perhaps indicating that men are able to mount a more effective immune response than are women. Alternatively, men may have persistently higher antibody titers as a result of more frequent exposure due to work or recreational practices. Such possibilities require prospective study to address.
The long-term consequences of human renal infection by Leptospira need to be explored, in particularly the effect of persistent infection on renal function and electrolyte balance. Moreover, the nature of the infecting strains needs to be more carefully explored as some strains may be more likely than others to result in persistent renal infections in humans. Although we have identified the species of the infecting strains in the present study, other methods that are able to identify serovars, particularly isolation, will be more informative.
While humans are considered to be exclusively incidental hosts, animals can be maintenance and/or incidental hosts; maintenance hosts are defined as species in which infection is endemic, of low or no pathogenicity and (as a key factor) transmitted directly to the same species [25]–[28]. Although human-to-human transmission has been rarely documented, it is unlikely that asymptomatic infected individuals have an important role in disease maintenance and transmission [1],[11],[29],[30]. An increased risk of having leptospiral antibodies in households of leptospirosis index cases compared to controls in an epidemic setting has been shown recently [31], but this is most likely related to common environmental exposure risks or genetic susceptibilities rather than direct transmission. In light of these data, further studies should address the possibility that long-term urinary shedders may represent a source of Leptospira for their families and explore human-human transmission more carefully. Infection in carrier animals is usually acquired at an early age, and the prevalence of chronic excretion in the urine increases with age; we observed a similar trend in this population. Of note, none of the long-term urinary shedders had circulating anti-leptospiral antibodies; this is in accordance with early observations in Leptospira-carrier mammals, where chronic urinary carriage was associated with low seropositivity to urinary culture rates in asymptomatic well-established serovar-specific carriers [16]. Taken together, these observations make us speculate that in regions with high disease transmission, humans can develop some clinical and serological characteristics of asymptomatic urinary carriers, an attribute classically restricted to animals. Further longitudinal studies should address this possibility since the impact on disease transmission and in renal function of the affected individuals are unknown.
The study design had several limitations. First, we relied only on the recall of participants to define absence of fever in 1 year. Men were underrepresented; fewer men were recruited because of a lack of availability at the time of recruitment (most were away working). Thus, no leptospiruric males were detected. This observation suggests that we may have underestimated the overall number of asymptomatic shedders, as men have been typically associated with a higher risk of exposure due to work-related contact and behavioral practices. Another limitation was the initial non-efficient PCR screening strategy. The presence of other bacterial DNA hindered the identification of Leptospira-positive clones; we detected both Leptospira and Atopobium DNA in multiple samples, making the selection of colonies harboring the leptospiral 16S gene less efficient, in some instances, several hundred colonies had to be screened; we were unable to sequence the infecting strain in three enrollees due these technical limitations. Though unlikely, it is also possible that in these three instances the dot-blot gave false positive results. A third limitation of this study is that culture isolation of leptospires from urine was not attempted. Future work will be needed to further validate the molecular results presented here, and will use the PCR method to screen patients who then would have urine cultured for Leptospira. Nonetheless, the deployment of a valid molecular tool to detect leptospiruria represents a new approach to assessing chronic asymptomatic infections in humans without the need for obtaining isolates. Finally, because L. licerasiae serovar Varillal [7] had not been fully characterized nor its epidemiological implications known, this strain was not used as antigen in the MAT panel or ELISA used to study patient sera, nor are these sera available for retrospective analysis.
Few of the published Leptospira-specific PCR have been applied in clinical or field settings[32]. Furthermore, detection of bacterial DNA in urine is cumbersome because of the presence of PCR inhibitors and samples are often contaminated by multiple bacterial species whose DNA can interfere with the PCR assay [33].
Current understanding of host immune responses to Leptospira or the pathogenesis of leptospirosis remains limited. Naturally acquired immunity that protects against re-infection by Leptospira does occur and has been shown in animal models. It has been assumed that naturally acquired immunity is humorally-mediated particularly by antibodies against oligosaccharides of leptospiral LPS. Evidence also suggests that antibodies specific to Leptospira membrane-associated proteins may play a role in host defense [2],[34]. We have documented that in this hyperendemic area, in spite of the high levels of environmental exposure to Leptospira and high prevalence of seropositivity, the prevalence of severe disease is low [4],[7]. These observations suggest the possibility that protective immunity against severe disease from repeated infection may develop in areas with high leptospirosis transmission, especially if high frequency of infection leads to cross-serovar protection. Based on the finding that asymptomatic infection and urinary carriage are prevalent in this area where transmission is high and the prevalence of severe disease is low, we suggest that repeated exposure to Leptospira and asymptomatic infection could induce protective acquired immunity. Longitudinal studies are needed to test this hypothesis.
In conclusion, we have identified a long-term renal shedder group among persons asymptomatically infected with pathogenic and intermediately pathogenic Leptospira. The health implications of long-term renal colonization and whether antibiotic treatment of such patients is required remain to be determined.
|
10.1371/journal.pntd.0001473 | Successful Outcomes with Oral Fluoroquinolones Combined with Rifampicin in the Treatment of Mycobacterium ulcerans: An Observational Cohort Study | The World Health Organization currently recommends combined streptomycin and rifampicin antibiotic treatment as first-line therapy for Mycobacterium ulcerans infections. Alternatives are needed when these are not tolerated or accepted by patients, contraindicated, or neither accessible nor affordable. Despite in vitro effectiveness, clinical evidence for fluoroquinolone antibiotic use against Mycobacterium ulcerans is lacking. We describe outcomes and tolerability of fluoroquinolone-containing antibiotic regimens for Mycobacterium ulcerans in south-eastern Australia.
Analysis was performed of prospectively collected data including all primary Mycobacterium ulcerans infections treated at Barwon Health between 1998 and 2010. Medical treatment involved antibiotic use for more than 7 days; surgical treatment involved surgical excision of a lesion. Treatment success was defined as complete lesion healing without recurrence at 12 months follow-up. A complication was defined as an adverse event attributed to an antibiotic that required its cessation. A total of 133 patients with 137 lesions were studied. Median age was 62 years (range 3–94 years). 47 (34%) had surgical treatment alone, and 90 (66%) had combined surgical and medical treatment. Rifampicin and ciprofloxacin comprised 61% and rifampicin and clarithromycin 23% of first-line antibiotic regimens. 13/47 (30%) treated with surgery alone failed treatment compared to 0/90 (0%) of those treated with combination medical and surgical treatment (p<0.0001). There was no difference in treatment success rate for antibiotic combinations containing a fluoroquinolone (61/61 cases; 100%) compared with those not containing a fluoroquinolone (29/29 cases; 100%). Complication rates were similar between ciprofloxacin and rifampicin (31%) and rifampicin and clarithromycin (33%) regimens (OR 0.89, 95% CI 0.27–2.99). Paradoxical reactions during treatment were observed in 8 (9%) of antibiotic treated cases.
Antibiotics combined with surgery may significantly increase treatment success for Mycobacterium ulcerans infections, and fluoroquinolone combined with rifampicin-containing antibiotic regimens can provide an effective and safe oral treatment option.
| Buruli ulcer is a necrotizing infection of skin and subcutaneous tissue caused by Mycobacterium ulcerans and is the third most common mycobacterial disease worldwide (after tuberculosis and leprosy). In recent years its treatment has radically changed, evolving from a predominantly surgically to a predominantly medically treated disease. The World Health Organization now recommends combined streptomycin and rifampicin antibiotic treatment as first-line therapy for Mycobacterium ulcerans infections. However, alternatives are needed where recommended antibiotics are not tolerated or accepted by patients, contraindicated, or not accessible nor affordable. This study describes the use of antibiotics, including oral fluoroquinolones, in the treatment of Mycobacterium ulcerans in south-eastern Australia. It demonstrates that antibiotics combined with surgery are highly effective in the treatment of Mycobacterium ulcerans. In addition, oral fluoroquinolone-containing antibiotic combinations are shown to be as effective and well tolerated as other recommended antibiotic combinations. Fluoroquinolone antibiotics therefore offer the potential to provide an alternative oral antibiotic to be combined with rifampicin for Mycobacterium ulcerans treatment, allowing more accessible and acceptable, less toxic, and less expensive treatment regimens to be available, especially in resource-limited settings where the disease burden is greatest.
| In recent years the treatment of Mycobacterium ulcerans (M. ulcerans) has radically changed, evolving from a predominantly surgically [1], [2] to a predominantly medically treated disease [3]. This resulted from clinical experience supported by scientific data showing superior outcomes when antibiotics were used alone [4], [5], or combined with surgery [6], [7]. It is also supported by in vitro data [8]–[10] of antibiotic effectiveness against M. ulcerans. The World Health Organization now recommends combined streptomycin and rifampicin antibiotic treatment as first-line therapy for M. ulcerans infections, with surgery reserved mainly to remove necrotic tissue, cover skin defects, and correct deformities [11].
The Bellarine Peninsula in south-eastern Australia has been experiencing an epidemic of M. ulcerans since 1998. It affects local residents, but also visitors from outside the endemic region, with cases in those living locally managed at the local referral health service, Barwon Health. In our region, despite recommendations at the time against their use, adjunctive antibiotic treatment of M. ulcerans was initiated from 1998 in response to severe disease causing significant morbidity and requiring reconstructive surgery, and disease recurrences despite surgery. Fluoroquinolones (FQ) were introduced into antibiotic regimens in 2003 [6], [12] in response to perceived treatment failures and excess toxicity with other antibiotics, as well as their potential advantages in treating M. ulcerans; documented in vitro evidence of activity [9], [10], [13]–[15], good bioavailability [16], and excellent bone and tissue penetration [17]. Since their introduction they have commonly been employed in the antibiotic regimens used at Barwon Health.
FQ antibiotics offer the possibility of completely oral antibiotic regimens when combined with another active oral antibiotic, usually rifampicin. This can be especially useful where other recommended antibiotics are not tolerated, accepted, accessible, or affordable. However, although there is evidence of good activity against M. ulcerans in the laboratory [9]–[10], [13], in mouse footpad models [8], [14], [18], and small numbers of clinical cases [6], [12], [19]–[21], clinical evidence of FQ efficacy is lacking. Therefore we undertook a study during the current epidemic in the Bellarine Peninsula to describe the use of FQ antibiotics in M. ulcerans treatment and to compare their outcomes and tolerability with other antibiotics used.
Analysis was performed using prospectively collected data from an electronic database containing information on all cases of M. ulcerans infection treated at Barwon Health between 1st March 1998 and 31st May 2010. Data collected includes epidemiological details, diagnostic tests, clinical features, treatment, and outcomes. Only first lesions on initial presentation were analyzed to avoid potential confounders when analyzing recurrent cases. Data was analyzed anonymously.
A case of M. ulcerans was defined as the presence of a lesion clinically suggestive of M. ulcerans plus any of (1) a culture of M. ulcerans from the lesion, (2) a positive PCR from a swab or biopsy of the lesion [22], or (3) histopathology of an excised lesion showing a necrotic granulomatous ulcer with the presence of acid-fast bacilli consistent with acute M. ulcerans [23].
Surgical treatment was defined as the surgical excision of a lesion. Due to the paucity of cases managed without surgery, only those undergoing surgery were included. Major surgery involved the use of a split thickness skin graft or a vascularized tissue flap to close the tissue defect, whereas minor surgery involved excision plus primary closure of the defect. A positive margin was defined on histology as granulomatous inflammation or necrotic tissue extending to the margins of the excised lesion.
Medical treatment was defined as the use of antibiotics for more than 7 days, and first-line regimens were the initial antibiotic regimens used. A complication was an adverse event attributed to an antibiotic that required the cessation of that medication. Drug dosages for adults included ciprofloxacin 500 mg twice daily, moxifloxacin 400 mg daily, rifampicin 10 mg/kg/day (up to a maximum of 600 mg daily), clarithromycin 500 mg twice daily, and amikacin 15 mg/kg/day.
Treatment success was defined as complete healing of the lesion without recurrence 12 months after treatment commencement. Treatment failure was defined as those who had a recurrence with in at least 12 months of follow-up. Recurrence was defined as a new lesion appearing either in the wound, locally, or in another part of the body after the surgical excision of the initial lesion that met the case definition for M. ulcerans. Paradoxical reactions were not considered a recurrence and were defined as: initial clinical improvement followed by the clinical deterioration of the treated lesion, or the appearance of a new lesion, either locally or in a distant body site, that on histopathology showed evidence of an intense immunological reaction consistent with an immune-mediated paradoxical reaction [19].
There was no standardized treatment protocol for M. ulcerans followed in Barwon Health during the study period. The role of surgery and the use of antibiotics were determined by individual specialist practitioners involved in M. ulcerans treatment. Patients were followed up according to routine clinical practice and observed antibiotic complications recorded in clinical notes when they occurred.
This study was approved by the Barwon Health Research and Ethics Committee.
Data were collected and analysed with Epi-Info 6 (Centers for Disease Control, Atlanta). Statistical significance was determined using the 2-tailed Fisher exact test for 2×2 tables for each of the categorical values. A non-parametric cumulative failure graph using the Kaplan-Meier method and the endpoint of antibiotic cessation was plotted using the statistics package Minitab (version 15) to model the proportion of antibiotics ceased over time due to complications.
One hundred and forty seven patients with M. ulcerans were diagnosed and treated at Barwon Health over the study period 1st March 1998 to 31st May 2010. Fourteen were excluded from further analysis: 1 had no surgery, 2 died before the completion of follow-up (1 from a cerebrovascular accident 52 days and 1 of sepsis secondary to the M. ulcerans lesion 5 days post treatment commencement), 1 was lost to follow-up 85 days post treatment commencement, and for 10 treatment and follow-up was ongoing. Therefore a total of 133 patients with 137 lesions (4 patients had 2 lesions at initial presentation) were included in the analysis.
Median age of patients was 62 years (range 3–94 years); 7 (5%) were <15 years. Sixty-seven (50%) were male. Associated co-morbidities included diabetes mellitus (11), malignancy (5), connective tissue disease (4), and immunosuppressive treatment (4).
For 122 cases where the clinical type of lesion was recorded, 106 (87%) were ulcers, 9 (7%) were nodules, and 7 (6%) were oedematous lesions. Diagnosis was made by PCR in 116 (87%), positive culture in 22 (17%), and consistent histopathology in 54 (41%) cases. Eighteen of 24 (75%) PCR positive cases where culture was performed were culture positive, but no cases were culture positive and PCR negative. One case was PCR negative but positive on histopathology.
Forty-seven (34%) lesions were treated with surgical excision alone, and 90 (66%) had surgical excision and adjunctive antibiotic therapy. To close the skin defect after excision of the lesion, 65 (47%) required a split thickness skin graft and 16 (12%) required a vascularized tissue flap. The proportion of cases receiving antibiotics significantly increased pre 2005 compared to post 2005, rising from 45% to 74% [<2005 18/40, ≥2005 72/97, OR 3.52 (1.52–8.20)] (Figure 1).
The most common initial antibiotic regimens were rifampicin and ciprofloxacin (61%) and rifampicin and clarithromycin (23%). Four patients received ciprofloxacin and clarithromycin, and 2 patients received rifampicin and moxifloxacin regimens (Table 1). FQ antibiotics were used in 3 of 5 children aged <15 years who received antibiotics. Antibiotics were given for a median duration of 76 days (range 12–155 days); 21 (23%) between 12–30 days, 18 (20%) 31–60 days, 30 (33%) 61–90 days, 14 (16%) 91–120 days, and 7 (8%) 121–155 days. FQ antibiotics were not used in this study until 2004, but from then 82% of regimens were FQ containing.
Fourteen of 47 (30%) of those treated with surgery alone failed treatment compared to 0/90 (0%) of those treated with a combination of medical and surgical treatment (p<0.0001). The risk of treatment failure increased significantly with no antibiotics compared to those treated with antibiotics for major surgery (p<0.0001), minor surgery (p = 0.01), positive margins (p<0.0001), and negative margins (p = 0.05) (Table 2). If minor surgery and negative margins were present, 1/22 (5%) failed treatment. If only regimens containing an FQ (n = 64) are compared with surgery alone, the risk of treatment failure remained significantly increased when no antibiotics were used overall (p<0.0001) and for major surgery (p<0.0001) and positive margins (p<0.0001) (Table 3). There was no difference in treatment success rate for antibiotic combinations containing an FQ (61/61 cases; 100%) compared with those not containing an FQ (29/29 cases; 100%). Treatment success rates with antibiotics were also similar pre-2004 when no FQs were used (11/11; 100%) compared with post 2004 (79/79; 100%). All four cases treated with ciprofloxacin and clarithromycin combined with surgery experienced treatment success.
For those failing treatment, the recurrences occurred a median 90 days post surgery (range 14–300 days). In 9 (64%) patients this was local and in 6 (43%) patients it was distant (1 had both). Paradoxical reactions occurred in 8/90 (9%) of cases given antibiotics after a median duration of 48 days (range 14–85 days).
Fifty-eight (64%) patients had antibiotics prior to surgery for a median duration of 8 days (range 1–36 days). Of these, mycobacterial cultures were performed on 28 excised specimens. Cultures were positive for M. ulcerans in 11/20 (55%) of those who received ≤14 days of antibiotics prior to surgery, and in 1/8 (12.5%) of those who received >14 days of antibiotics prior to surgery (Table 4). All cases with positive cultures were associated with successful outcomes after a median antibiotic duration of 87 days (range 30–155 days).
Rifampicin was associated with a complication in 19/85 (22%) cases occurring at a median 27 days (range 6–94 days) and involved gastrointestinal intolerance in 15, hepatitis 4, rash 3, and hypoglycemia in 1 case. Ciprofloxacin was associated with a complication in 13/63 (21%) cases occurring at a median 24 days (range 6–90 days) and involved gastrointestinal intolerance in 10, joint or tendon effects in 3, rash in 2, and hallucinations in 1 case. Clarithromycin was associated with a complication in 10/38 (26%) cases occurring at a median 25 days (range 2–60 days) and involved gastrointestinal intolerance in 9, hepatitis in 1, and palpitations in 1 case.
By 120 days on treatment, the proportion of cases in which rifampicin [28.6% (95% CI 16.0, 41.2)], clarithromycin [29.4% (95% CI 13.8, 45.0)], and ciprofloxacin [24.9% (95% CI 12.3, 37.5)] were ceased were similar (Figure 2), and complication rates were similar between ciprofloxacin and rifampicin 17/55 (31%) and rifampicin and clarithromycin 7/21 (33%) regimens (OR 0.89, 95% CI 0.27–2.99).
Our study demonstrates that, combined with surgical excision of M. ulcerans lesions, antibiotics appear highly effective at preventing disease recurrences; we describe a reduction of recurrence rates from more than one quarter of cases with surgery alone to none if antibiotics are used. This includes 64 cases treated with FQ-containing regimens (the majority involving ciprofloxacin) which show similar efficacy to non-FQ containing regimens. Recent studies have provided strong evidence that non-FQ-containing antibiotic regimens were effective in M. ulcerans treatment; 8-week combinations of rifampicin and streptomycin cured 96% of cases in a randomized controlled trial in Africa [5], and 8 weeks of rifampicin and clarithromycin were 100% effective in an uncontrolled trial in Benin [4]. Although we have previously published small numbers of M. ulcerans cases treated with FQ-containing regimens [6], [12], [19], this is the first study large enough to provide significant evidence of the clinical effectiveness of FQ antibiotics combined mainly with rifampicin in M. ulcerans treatment.
Ciprofloxacin has been shown to have good in vitro activity against M. ulcerans with minimal inhibitory concentrations (MIC) of between 0.5 and 1 mg/l in two published studies [9], [15]; in the same studies these MICs compared favorably against the MICs of currently recommended antibiotics (rifampicin 1–2 mg/l, amikacin 1 mg/l, clarithromycin 1 mg/l). Ciprofloxacin has also been shown to have rapid bactericidal activity in humans against M. tuberculosis [24]–[26], and has been used to successfully treat other non-tuberculous mycobacterial infections including M. marinum [27], the species most closely related to M. ulcerans. Moxifloxacin similarly has good in vitro activity (MIC 0.25) [15]. In the mouse footpad model, moxifloxacin has bactericidal activity against M. ulcerans and, when combined with rifampicin, is as effective as combinations of rifampicin/streptomycin, rifampicin/amikacin and rifampicin/clarithromycin [8], [18]. Furthermore, other factors that favor the use of FQs include their high oral bioavailability (78%) [16] and excellent bone and tissue penetration [17]. Nevertheless, we would caution that fluoroquinolones should not be used as monotherapy for M. ulcerans treatment as there is the potential for the development of resistance, as has been shown to occur when FQs are used as monotherapy for M. tuberculosis [25]. Moxifloxacin may be favoured over ciprofloxacin due to slightly better published MICs against M. ulcerans, the evidence from the mouse footpad models of which there is no similar published data for ciprofloxacin, greater potential barrier to resistance if effects are similar to that in M. tuberculosis [25], and once-daily administration. A constraint at present is its significant increased cost compared to ciprofloxacin, with a cost of $671 compared with $32 Australian dollars for an 8-week treatment course at our institution.
In mouse footpad models, rifampicin has the greatest bactericidal activity against M. ulcerans [8], [18], and thus is assumed to be the most active and important antibiotic in vivo, though there are no human studies of rifampicin monotherapy to confirm this. It is possible that the second agent, including the FQs, act only as bacteriostatic agents preventing the emergence of rifampicin resistance. Therefore we recommend rifampicin as the first antibiotic choice, to be combined with another agent. Our data indicate that FQs are an appropriate combination choice, especially in cases where other antibiotics such as clarithromycin or streptomycin are not tolerated, accepted by patients, or accessible, or are contraindicated. FQs, greatest advantage may be their potential to be combined with other oral antibiotics such as rifampicin to provide completely orally administered regimens in endemic settings. This allows outpatient care to be provided close to patient homes and avoids daily intramuscular injections, potentially increasing patient willingness to present early for diagnosis and increase adherence to treatment. Furthermore, ciprofloxacin is generically produced, reducing its cost, and is readily available in many resource-limited settings.
Antibiotics were most commonly given for between 1 and 3 months (53%) in our cohort. A proportion of cases (21; 23%) had successful outcomes with less than 30 days of treatment, although 55% of cases given less than 2 weeks of antibiotics remained culture positive. One case remained positive after 36 days of antibiotics but still achieved cure, reinforcing findings from previous studies that a small proportion of cases may remain culture positive after 1–2 months of antibiotics but still achieve cure with at least 8 weeks of antibiotics [4], [5]. The proportion of patients receiving antibiotic treatment significantly increased from 2005 as it became apparent in patients treated at Barwon Health that antibiotics were associated with a reduction in disease recurrences and permitted more conservative surgery to be performed [6].
The efficacy of treatment is also determined by the tolerance of the antibiotic regimens. In our study there were no significant differences in the tolerability of the 3 main oral antibiotic choices of rifampicin, clarithromycin, and ciprofloxacin. In addition, the complications mainly involved gastrointestinal intolerance with no significant sequelae. This differs from the significant incidence of serious adverse events previously described for alternative antibiotics such as streptomycin [5], [28] and amikacin [6]. It is important to note that the complication rates were higher in our study compared to studies from African populations [4], [5], [7]. This is likely due to the older patient population in our cohort where these medications are less well tolerated, especially from a gastro-intestinal viewpoint, and there is more potential for drug interactions. Nevertheless it underlines the importance of having a number of oral antibiotic combinations available if first-line antimicrobials need to be substituted on account of intolerance.
Finally, paradoxical reactions occurred in 9% of antibiotic treated cases. To our knowledge this is the first published rate of paradoxical reactions in a cohort of patients treated for M. ulcerans. In our study the reactions occurred as early as 2 weeks and as late as 3 months after antibiotics were commenced. Recently, paradoxical reactions occurring after the cessation of antibiotics have been described [29]. It is important that paradoxical reactions are considered and recognized during treatment, and distinguished from treatment failures, to avoid unnecessary antibiotic regimens changes or further surgery [19].
There are a number of limitations to our study. First, it is an observational cohort and thus there is the potential for unknown confounders to have affected the results. Despite this possibility, the treatment outcome results are highly significant. Second, some of the infections acquired from the local endemic region occurred in visitors to the region and were not managed by our health service (Barwon Health) but in the health services where they live. Although this may have introduced a selection bias, we feel it is unlikely that this would have changed the findings of our study. Third, all cases underwent surgical excision and therefore the outcomes may not be applicable to cases where antibiotics alone are used. We advocate for further randomized studies using FQ-containing antibiotic regimens without curative surgery be performed to provide further information.
In summary, antibiotics in combination with surgery may significantly increase treatment success for M. ulcerans infections. In addition, antibiotic regimens containing oral FQs combined with rifampicin can provide an effective and safe treatment option and should be further studied in the treatment of M. ulcerans.
|
10.1371/journal.pbio.0050007 | Drug Export Pathway of Multidrug Exporter AcrB Revealed by DARPin Inhibitors | The multidrug exporter AcrB is the inner membrane component of the AcrAB-TolC drug efflux system in Escherichia coli and is responsible for the resistance of this organism to a wide range of drugs. Here we describe the crystal structure of the trimeric AcrB in complex with a designed ankyrin-repeat protein (DARPin) inhibitor at 2.5-Å resolution. The three subunits of AcrB are locked in different conformations revealing distinct channels in each subunit. There seems to be remote conformational coupling between the channel access, exit, and the putative proton-translocation site, explaining how the proton motive force is used for drug export. Thus our structure suggests a transport pathway not through the central pore but through the identified channels in the individual subunits, which greatly advances our understanding of the multidrug export mechanism.
| Bacterial resistance to antibiotics is a major challenge for the current treatment of infectious diseases. One way bacteria can escape destruction is by pumping out administered drugs through specific transporter proteins that span the cell membrane. We used designer proteins that bind to and stabilize proteins of interest in order to study the major drug efflux pump of Escherichia coli, AcrB. After selecting for designed ankyrin repeat proteins (DARPins) that inhibit this pump, we determined the crystal structure of a DARPin inhibitor in complex with AcrB. We confirmed that the AcrB is split into three subunits, each of which exhibits distinctly different conformations. Moreover, we show that each subunit has a differently shaped substrate transport channel; these variable channels provide unique snapshots of the different conformations adopted by AcrB during transport of a substrate. The structure also offers an explanation for how substrate export is structurally coupled to simultaneous proton import—thus significantly improving our understanding of the mechanism of AcrB. This is the first report of the selection and co-crystallization of a DARPin with a membrane protein, which demonstrates the potential of DARPins not only as inhibitors but also as tools for the structural investigation of integral membrane proteins.
| Drug resistance is a medical problem, ranging from cancer cells evading chemotherapy to bacteria surviving antibiotic treatment. Efflux pumps represent one class of integral membrane transport proteins in bacteria that confer antibiotic resistance [1]. These proteins actively detoxify the intracellular space by exporting drugs to the cell exterior. AcrB of Escherichia coli is such an efflux pump belonging to the subclass of resistance-nodulation-cell division transporters, which catalyze drug export driven by proton antiport [2]. AcrB associates with the outer membrane channel TolC [3] and the periplasmic protein AcrA [4] and allows direct and efficient transport of a wide range of toxic substances [5]. The structures of AcrB alone [6] and of AcrB in complex with substrates [7,8] revealed the general architecture of the transporter. However, despite all mutational and structural studies to date, the mechanism explaining how substrates are transported into the extracellular media was still unclear.
The use of antibody fragments as crystallization aids for membrane proteins has yielded a number of crystal structures [9,10]. The binding of such antibody fragments enlarges the hydrophilic extramembranal surface of integral membrane proteins, thereby providing additional surface for crystal contacts. They can also stabilize a specific conformation supporting the crystallization process. The drawback of the antibody fragment approach is that it is not always easy to get an antibody fragment that recognizes and binds to a particular conformation of a membrane protein. Further, the selected antibody fragment might be unstable or production might be difficult. To circumvent these problems, we applied an approach based on designed ankyrin-repeat proteins (DARPins) as an alternative to antibody fragments. DARPins can be selected to bind almost any given target protein with high affinity and specificity [11]. They are very stable and can be produced as soluble proteins in large amounts by bacterial expression. As DARPins interact with their target protein with an exposed interaction surface, they tend to bind to conformational epitopes rather than to peptidic ones. These characteristics make DARPins ideal tools to help the structural studies of membrane proteins.
Here we selected DARPins that not only bind to AcrB but also inhibit bacterial drug export. Crystals of a selected AcrB–DARPin complex were obtained, and the structure was determined at 2.5-Å resolution. It is the first structure of an integral membrane protein with a selected DARPin molecule binder. The structure reveals a previously unknown asymmetric conformation of the efflux pump, in which each of the three subunits has a unique well-defined conformation. The internal asymmetry of AcrB is underlined by the fact that only two of the three subunits of AcrB are recognized by DARPins in crystallo and in solution and is in contrast to the 3-fold symmetric structures reported to date [6–8]. The structural features described here together with the enhanced resolution allow us to deduce a pathway for drug export and a mechanism for the coupling of substrate export with proton import.
DARPin binders were selected in vitro by ribosome display using a designed ankyrin-repeat protein library [11,12]. The selection was performed on purified and in vitro biotinylated AcrB under native conditions. Four rounds of ribosome display were sufficient to select a pool of specific DARPins binding to AcrB. These experiments are to our knowledge the first ribosome display selections performed in the presence of detergent to select specific binders to an integral membrane protein.
The pool of specific binders was subjected to an in vivo screen, based on replica plating, to identify those AcrB-binding DARPins which lead to an inhibitory phenotype. As AcrB is crucial for the transport of certain substrates, its inhibition will consequently lead to accumulation of substrates in the bacterial cell and consequently to cell death. We chose Rhodamine 6G (R6G) as substrate for AcrB. The enriched pool after four rounds of ribosome display was cloned into an expression plasmid under the control of an isopropyl-β-d-thiogalactopyranoside (IPTG)-inducible promoter. E. coli strain XL1-blue was transformed and plated under nonselective conditions, in absence of R6G. About 1,500 colonies were subsequently replica-plated under selective conditions, in the presence of 32 μg/ml R6G. By this procedure, 18 single clones were identified which showed a rhodamine-sensitive phenotype. These selected clones were sequenced and assayed for expression. Analysis of the sequences showed that all 18 clones were different and unique.
The affinities of the AcrB inhibiting DARPins were analyzed by surface plasmon resonance using a BIAcore (http://www.biacore.com) instrument. The analysis was performed on AcrB-coated sensor surfaces with multiple concentrations of the DARPins and was evaluated with a global kinetic fit. The five inhibitors with highest affinity, having dissociation constants in the low nanomolar range (3.6 to 98.9 nM; Figure S1), were chosen for further characterization. Clone 1108_19, which was used for cocrystallization, had a dissociation constant KD of 28 nM.
We rated the inhibition efficiencies of the identified DARPins by spotting the clones onto selective plates with different R6G concentrations, ranging from 2 to 64 μg/ml (Figure 1). All inhibitors showed a significant increase of the sensitivity compared to cells expressing an unselected nonbinding DARPin, termed E3_5 [13]. This control also showed that the expression of a nonbinding DARPin per se has no influence on R6G sensitivity or bacterial growth. The acrB gene-deletion strain KAM3 [14] causes hypersensitivity to R6G and was used as a sensitivity control. In absence of R6G, DARPins conferring R6G sensitivity did not influence E. coli growth by themselves. It should be noted, however, that the affinity of the inhibitors does not always correlate with the inhibition efficacy (Figure S1). Possible factors leading to this observation include different expression levels or different binding sites of the DARPins on AcrB. Moreover, the exact inhibition mechanism is not known, even for the inhibitor 1108_19 for which the structure was solved. Different inhibition mechanisms seem plausible ranging from allosteric inhibition of the rotary export mechanism of AcrB by the DARPin to simply preventing the interaction with TolC or AcrA to form the export complex.
AcrB exists as a homotrimer, with each subunit containing 12 transmembrane (TM) helices and a large periplasmic part formed by two loops between TM helices 1 and 2 and TM helices 7 and 8. The trimer consists of three prominent domains, the TM domain, and parallel to the membrane, the adjacent pore domain and the TolC docking domain (Figure 2A). The TM domain encompasses a large central cavity, 35 Å in diameter, through the membrane which is closed toward the central funnel formed by the TolC docking domain by three helices provided by the PN1 subdomains of the pore domain (Figure 2B and 2C). The export of substrates was proposed directly through this central pore. Several structures of ligand-free AcrB and with bound substrates have been solved at moderate resolution (between 3.5 and 3.8 Å). Yu et al. [7,8] reported structures of AcrB–ligand complexes showing ligand binding to the upper wall in the central cavity and additional substrate binding in the periplasmic domain, compatible with mutational studies, which indicated that discrimination between different toxic substrates occurs in the periplasmic domain, rather than in the TM domain [15,16]. All structures determined to date imply 3-fold symmetry and thus the three AcrB subunits displayed the same conformation.
To analyze the interaction and to characterize the inhibition property of a selected DARPin, we determined the crystal structure of one inhibitor (1108_19) in complex with AcrB. The crystals diffracted to 2.5-Å resolution and belonged to space group P212121. The structure was solved by molecular replacement using the structure published by Murakami et al. [6] as a search model without using the phases of the DARPin. The results of the data collection and refinement are presented in Table 1.
The crystal structure of the AcrB–DARPin complex shows two DARPin molecules bound basically to two subunits, named A and B, in the homotrimeric transporter, while the third subunit, named C, is not bound (Figure 2A and 2B). Unlike previously determined crystal structures, there is no 3-fold symmetry of AcrB with the DARPin molecules stabilizing three distinct conformations of the subunits. The conformations of the DARPin-bound subunits A and B are similar to the known symmetric structure. A superposition on Protein Data Bank entry 1IWG [17] shows a few differences for these two subunits with root-mean-square deviations of 1.72 Å and 1.98 Å, respectively, for 1,033 Cα atoms, compared to 2.95 Å for the third subunit C. To validate in solution the 3:2 (AcrB monomer to DARPin) stoichiometry observed in the crystal structure, we performed sedimentation velocity experiments. Single molecular weight species taking the detergent micelle into consideration were observed, with molecular masses of 374 ± 1.6 kDa for AcrB alone and 408 ± 1.7 kDa for the complex (Figure S2). The mass difference indicates that AcrB in solution is bound by two DARPins, with a calculated mass of 17.8 kDa each.
As can be seen in the crystal structure (Figure 2A and 2B), the DARPins bind mainly to the β-sheet connecting the pore domain with the TolC docking domain interacting, as expected, primarily through their randomized concave surface area. The interface buries a surface of around 1,000 Å2 (see Table S1 for detailed interactions). Furthermore, each DARPin interacts with the adjacent preceding subunit of AcrB via one additional hydrophobic interaction involving Leu230 of the intersubunit connecting loop as well as two interactions involving the residues Arg263 and Lys248, again of the preceding subunit (Table S1). Subunit C is in a different orientation, and consequently no DARPin is bound. Notably, the DARPins not only stabilize the asymmetric conformation of AcrB but are also involved in direct interactions in the crystal lattice (Figure S3), resulting in a different crystal form P212121.
The three AcrB subunits are bound in three different conformations, revealing three distinct channels (Figure 3). The width of these channels is sufficient for the passage of typical AcrB substrates. In subunit A, a channel is observed, extending from the external depression through the large periplasmic domain reaching almost the central funnel at the top of the protein (Figure 4A). Here the side chains of residues Gln124, Gln125, and Tyr758 form a gate, closing the channel and therefore preventing direct access to the central funnel. The periplasmic channel entrance corresponds perfectly with the periplasmic binding site indicated in the AcrB–ligand structures of Yu et al. [8]. A similar channel, although a little wider, is present in subunit B (Figure 4B). In addition, the channel is open not only to the periplasm but also to the membrane bilayer at the periphery of the TM domain. In subunit C, the channel entrances are closed due to movements of PC2 and PN1 (Figure 4C). In contrast to subunits A and B, the gate to the central funnel at the top of the periplasmic domain formed by residues Gln124, Gln125, and Tyr758 is now open, allowing the substrate to enter the central funnel from where it most probably reaches TolC and is exported to the cell exterior.
The opening and closing of the channel entrances and the gate is coupled to structural rearrangements in the putative proton-translocation site in the TM domain (Figure 5A) involving residues in the middle of the TM helix 4 (Asp407 and Asp408) and TM helix 10 (Lys940), respectively (Figure 5B). These charged residues are localized in the center of the hydrophobic TM domain and have been shown to be essential for the proper function of AcrB, since the mutation of these residues leads to complete loss of drug resistance [18]. In subunits A and B, Lys940 forms a salt bridge to the side chains of Asp407 and Asp408, which are also involved in H-bonds with Ser481 and Thr978 presenting the same conformation described to date [6]. In subunit C, this salt bridge is not formed; instead Lys940 is tilted away, forming new polar contacts to Asn941 and Thr978.
The conformational changes described for the putative proton-translocation site are linked to the opening and closing of the channel entrances and the gate described above via rigid movements of TM helix 5 and TM helix 8 (Figure 5A). Distinct conformational differences of the polypeptide stretch that connects the N-terminal end of TM helix 8 to PC2 can be observed in each subunit (Figure 4). The functional importance of this connection is underlined by the fact that each subunit displays a different conformational state of this polypeptide stretch. In subunit A, the stretch shows a long random coil conformation, whereas in subunit B, it adopts partially a helical conformation but with still some random coil conformation. In subunit C, the polypeptide segment is in a completely α-helical conformation continuing the TM helix 8 up to the headpiece subdomain PC2.
The loss of the interaction of Asp407, Asp408 with Lys940 in the putative site for proton-translocation, the simultaneous closing of the channel entrances, and the opening of the gate are evident. This hints at a coordinated control or a coupling of drug export and proton-translocation. The coupling seems to occur via the polypeptide segment that connects TM helix 8 and PC2.
During the review process of this manuscript, the asymmetric structure of AcrB was published by two other groups [19,20]. Our experimental approach, however, differs from theirs as we apply a new technology using DARPins as crystallization aid. This lead to a new crystal form, P212121, and, finally, crystals diffracting to higher resolution. Notably, the structural comparison reveals that all three independent structure determinations of AcrB resulted all in almost identical structures. The root-mean-square deviations for 1,032 Cα atoms (Table S2 and Video S1) is around 1, which is a very good correlation at the given resolutions of the different structures (2J8S, 2.5 Å; 2DHH, 2.8 Å [19]; 2GIF, 2.9 Å [20]). Importantly, the region where the DARPins bind shows no significant differences in all the structures. This clearly demonstrates that the selected DARPins bind and stabilize an existing conformational state and do not induce a nonnative conformation upon binding. Murakami et al. [19] also crystallized the protein in the presence of two substrates, which greatly substantiates our findings. In conclusion, all three crystal structures lead to virtually identical conclusions about the drug export pathway and we all proposed similar export mechanisms.
Cocrystallization of the membrane protein AcrB and a DARPin inhibitor selected from a large combinatorial library resulted in crystals that diffract to high resolution. The significant higher resolution of our structure compared to the structures published to date allowed unambiguous modeling of the side chain conformations and the interpretation of structural rearrangements extending from the putative proton-translocation sites in the center of the membrane to the gate in the periplasmic domain. The quality of the structure is underlined by 11 resolved detergent molecules (Figure S4). Our work shows for the first time the structure of an integral membrane protein with a selected DARPin molecule, an experimental approach that has great potential for the structural biology of these exceedingly difficult proteins. DARPins may aid by stabilizing certain conformations and by extending the hydrophilic surface similar to antibody-derived protein fragments.
In contrast to the previously known R32 crystal form of AcrB with one monomer in the asymmetric unit, in our structure there are one AcrB trimer and two DARPins in the asymmetric unit, and therefore there are no constraints regarding the conformation of the individual subunits of AcrB. Indeed, the crystal structure shows different conformations for the three individual subunits and reveals channels extending either from the periplasm and the membrane bilayer to a gate near the central funnel at the top of the protein or as in one subunit extending from the gate directly into the central funnel. Moreover, we were able to confirm the 3:2 stoichiometry in solution, confirming the internal asymmetry of AcrB. We believe that the three conformations likely present intermediates in the transport cycle and suggest a rotary mechanism for drug transport where the three subunits strictly alternate their conformation in a concerted way (Figure 6). The R6G-sensitive phenotype of cells expressing the DARPin supports this hypothesis. AcrB may be locked by the DARPins and therefore the rotation of the molecule is stopped. It seems unlikely that the subunits export drugs independent of each other. If this was the case, subunit C would still be functional, and this would be in contrast to the observed inhibitory phenotype induced by the DARPin. Another possibility for the inhibitory effect of the DARPin is that either TolC or AcrA is unable to bind AcrB anymore, resulting in the loss of function of this resistance-nodulation-cell division–type transport protein.
We believe that AcrB operates through a mechanism resembling the alternating-access mechanism which is the substrate-translocation mechanism most commonly described to secondary membrane transporters [21]. We propose that substrates could enter the channel in subunit B via the transmembranal groove at the periphery of the TM domain or via the external depression in the periplasmic part of the structure close to the outer leaflet of the membrane. Substrate would then be specifically bound with the gate of the channel still closed to the central funnel at the top of the periplasmic headpiece. Upon protonation of the proton gating residues Asp407 and/or Asp408, conformational changes in the TM helix 8 take place, which in turn trigger the conformational changes in the periplasmic headpiece. The channel entrances then close through movements of subdomains PN1 and PC2 opening the gate toward the central funnel at the top of the headpiece, releasing the bound substrate into the funnel, from where it could reach TolC and finally the cell exterior.
While this model is based on the interpretation of the structural information seen in the asymmetric structure, it contains all features necessary for the export of substrates and simultaneous import of protons. It is also consistent with mutational studies available to date [8,17]. Further experiments using the crystals of the complex with drug substrates may substantiate this mechanism. In fact, similar to the experiments performed by Murakami et al. [19], we also could show binding of minocycline in the identified channel at 2.5-Å resolution (unpublished data). The structural findings most likely rule out the previously proposed export of substrates directly through the central pore of the structure.
AcrB was overexpressed in E. coli strain BL21 (DE3; Novagen, EMD Biosciences, http://www.emdbiosciences.com) overnight at 30 °C using a pET28 vector (Novagen) containing acrB with a hexahistidine-tag at the C terminus. Cells of 1-L bacterial culture were disrupted with a French pressure cell, and the membranes were collected and solubilized in buffer A (20 mM Tris-HCl [pH 7.5], 150 mM NaCl, 10 mM imidazole, 10% glycerol), containing 1% (w/v) n-dodecyl-β-d-maltoside (DDM) (Anatrace, http://www.anatrace.com). Lipids and debris were removed by ultracentrifugation at 100,000g for 1 h. The solubilized protein was purified with affinity chromatography using Ni-NTA (Qiagen, http://www.qiagen.com), equilibrated with buffer A, containing 0.03% DDM. The column was washed using this buffer and 50 mM imidazole, respectively. Purified AcrB was eluted with 200 mM imidazole added to the above buffer.
For biotinylation purposes, AcrB was eluted in 200 mM imidazole (pH 7.5) as the only buffering agent. The protein was chemically biotinylated using a 15 molar excess EZ-Link Sulfo-NHS-LC-Biotin (Pierce Biotechnology, http://www.piercenet.com) for 30 min at 4 °C. Free biotin was removed using a HiTrap Desalting column (5 ml; Amersham Biosciences, http://www.amersham.com) equilibrated in buffer B (20 mM Tris-HCl [pH 7.5], 150 mM NaCl, 0.03% [w/v] DDM).
Four rounds of ribosome-display selection were carried out using an N3C DARPin library [11,12]. In vitro biotinylated AcrB was immobilized on neutravidin-coated wells blocked with BSA. To eliminate BSA- and neutravidin-binding library members as well as further unspecific binders, two pre-panning steps were applied, where neutravidin, BSA, and biotinylated maltose binding protein (MBP) [11] were present. In the actual panning step, binders to AcrB were thus selected. The panning procedure was performed exactly as described previously [11], with the modification that all buffers contained 0.03% (w/v) DDM, to keep AcrB stable and solubilized.
The pool of specific binders was subjected to an in vivo screen, based on replica plating, to identify those AcrB-binding DARPins with an inhibitory phenotype. For this purpose, the selected pool of binders was cloned into the vector pQIA [22]. This vector carries an expression cassette for the selected DARPins under the control of an IPTG-inducible T5 promoter.
The pool of binders was transformed into E. coli XL1-blue cells (recA1 endA1 gyrA96 thi-1 hsdR17 supE44 relA1 lac [F′ proAB lacIqZΔM15 Tn10 (Tetr)]; Stratagene, http://www.stratagene.com) and plated under nonselective conditions (ampicillin, IPTG, no R6G). About 1,500 colonies were subsequently replica-plated under selective conditions (ampicillin, IPTG, R6G [32 μg/ml]). By this procedure, 18 N3C clones were identified which showed a rhodamine-sensitive phenotype. The characterization of these DARPins included sequencing, expression tests, surface plasmon resonance, MIC determination, size exclusion chromatography studies, and cocrystallization.
The susceptibility to R6G of E. coli XL1-blue cells expressing the selected DARPins were determined by sequential 2-fold dilutions with LB agar/glucose plates containing 0.1 mM IPTG. R6G was used at concentrations of 2, 4, 8, 16, 32, and 64 μg/ml. Bacterial growth was examined after 24 h at 37 °C.
Surface plasmon resonance was measured using a BIAcore 3000 instrument at 20 °C. The running buffer was buffer B. An SA chip (BIAcore) was used with 700 RU of biotinylated AcrB immobilized. The DARPin binding was measured at a flow of 50 μl/min with 5-min buffer flow, 2-min injection of AcrB-binding DARPin in varying concentrations (2.5 nM to 60 nM), and an off-rate measurement of 30 min with buffer flow. The signal of an uncoated reference cell was subtracted from the measurements. The kinetic data of the interaction were evaluated using the program BIAevaluation 3.0 (BIAcore), and global fits were used to determine KD. The possible avidity effect of the AcrB trimer was not taken into account in the fit.
Sedimentation velocity experiments were conducted using a Beckman ProteomeLab XL-I (http://www.beckmancoulter.com) with an An-50 Ti analytical rotor (Beckman). The samples were equilibrated in buffer B at a concentration of 0.42 mg/ml. All data acquired from this experiment were obtained using the UV/Vis absorbance detection system (Perkin Elmer, http://www.perkinelmer.com) on the ultracentrifuge at 280 nm and double sector 12-mm charcoal-filled Epon centerpieces. The experiment was conducted at 4 °C at a speed of 40,000 rpm. The data of 300 scans were analyzed using the LAMM equation in the program SEDFIT [23].
To obtain crystals of the AcrB–DARPin complex, AcrB and the selected AcrB-inhibiting DARPin 1108_19 were purified separately as described [12]. A small molar excess of the DARPin with regard to the AcrB monomer was used to prepare the complex, followed by purification on a Tricorn Superdex-200 (Amersham Biosciences) column equilibrated in buffer B. The peak fractions containing AcrB in complex with the DARPin were concentrated using a 30-kDa-cutoff concentrator (Amicon Ultra; Millipore, http://www.millipore.com), exchanged into 10 mM HEPES (pH 7.5), 50 mM NaCl containing 0.03% (w/v) DDM, and again concentrated to 13 to 15 mg/ml.
Initial crystallization screening was done in 96-well, sitting drop crystallization plates (Greiner Bio-One, http://www.greinerbioone.com). The initial crystals were refined using standard techniques. The crystals used for data collection appeared within 4 d in 8% PEG 4000, 50 mM ADA (pH 6.5), 200 mM (NH4)2SO4 in a hanging drop vapor diffusion experiment at 20 °C. For data collection, the crystals were soaked in six steps for about 15 s in the mother liquid containing 5% to 30% glycerol and flash frozen into liquid propane.
Data were collected at the Swiss Light Source beamline PX (http://sls.web.psi.ch/view.php/about/index.html) and processed using the program XDS [24]. The crystal belonged to space group P212121, with a Matthews coefficient VM of 3.8 Å3/Da, corresponding to an estimated water content of 67.7%.
The crystal structure was solved by molecular replacement using the program PHASER [25,26], with the structure of the AcrB monomer (Protein Data Bank code 1IWG [6]) used as search model. A PHASER search with the DARPin E3_5 (Protein Data Bank code 1MJO [13]) as search model did not yield a meaningful solution. The information obtained from the conventional PHASER protocol for the three AcrB monomers was sufficient to model one DARPin molecule into the resulting electron density with the program O [27]. The second DARPin also was positioned in O. Refinement of the structure was carried out through multiple cycles of manual rebuilding using the program Coot [28] and refinement using CNS [29] resulting in a final model with an R factor of 22.9 and an Rfree factor of 27.9. The refined structure of the AcrB–DARPin complex was validated by the program PROCHECK [30]. Three-dimensional structural figures were prepared by using PyMOL [31].
The atomic coordinates and structure factors of the described complex have been deposited in the Protein Data Bank (http://www.rcsb.org) with accession number 2J8S.
|
10.1371/journal.pgen.1003033 | Stimulation of Gross Chromosomal Rearrangements by the Human CEB1 and CEB25 Minisatellites in Saccharomyces cerevisiae Depends on G-Quadruplexes or Cdc13 | Genomes contain tandem repeats that are at risk of internal rearrangements and a threat to genome integrity. Here, we investigated the behavior of the human subtelomeric minisatellites HRAS1, CEB1, and CEB25 in Saccharomyces cerevisiae. In mitotically growing wild-type cells, these GC–rich tandem arrays stimulate the rate of gross chromosomal rearrangements (GCR) by 20, 1,620, and 276,000-fold, respectively. In the absence of the Pif1 helicase, known to inhibit GCR by telomere addition and to unwind G-quadruplexes, the GCR rate is further increased in the presence of CEB1, by 385-fold compared to the pif1Δ control strain. The behavior of CEB1 is strongly dependent on its capacity to form G-quadruplexes, since the treatment of WT cells with the Phen-DC3 G-quadruplex ligand has a 52-fold stimulating effect while the mutation of the G-quadruplex-forming motif reduced the GCR rate 30-fold in WT and 100-fold in pif1Δ cells. The GCR events are telomere additions within CEB1. Differently, the extreme stimulation of CEB25 GCR depends on its affinity for Cdc13, which binds the TG-rich ssDNA telomere overhang. This property confers a biased orientation-dependent behavior to CEB25, while CEB1 and HRAS1 increase GCR similarly in either orientation. Furthermore, we analyzed the minisatellites‚ distribution in the human genome and discuss their potential role to trigger subtelomeric rearrangements.
| All genomes contain particular DNA sequences that are prone to break and rearrange. They can be lost or rescued at the expense of sequence variations and complex rearrangements. Using a sensitive yeast model system, we examined the fragility of the HRAS1, CEB1, and CEB25 GC-rich human minisatellites (tandem repetition of motifs from 10 to 100 bp long). We observed that they all stimulate Gross Chromosomal Rearrangements but to very different extents, both in wild type and in cells deficient for the Pif1 helicase. Several intrinsic sequence features can account for these differences: the total number of repeats, the ability to form G-quadruplex secondary structures, or the ability to bind with high affinity the telomerase cofactor Cdc13. The orientation on the chromosome dictates the type of GCR (telomere addition versus other structural rearrangements) while not affecting the GCR rate in most cases. Being enriched in the subtelomeric regions of the human chromosomes, this class of GC–rich minisatellite has the potential to trigger a large variety of human genome rearrangements.
| Some chromosomal regions are more prone to rearrangement than others and thus are the source of genetic diseases and cancer. Among “at risk” sequences, tandem repeats like microsatellites and minisatellites that differ by the length of their repeat unit (1–10 nt and 10–100 nt, respectively) are prone to changes in repeat number (expansion and contraction of the array) [1]. Mechanistically, this instability can be explained by the propensity of the motifs to misalign during template-directed repair of endogenous lesions, occurring stochastically or promoted by the nucleotide sequence themselves, which, for example, can perturb replication. Consistently, their instability is exacerbated by defects of replication proteins (like Rad27 or Polδ) that ubiquitously affect genome integrity [2]–[7].
Intrinsic features of repeated sequences also play a role in the formation of rearrangements [1]. Microsatellite instability caused by hairpin formation during replication has been well documented [8] but less is known about minisatellite instability. Sequence composition and its ability to interact with endogenous factors and/or to adopt secondary structures can be invoked. Among these are G-quadruplexes. They are four-stranded structures that some G-rich nucleic acids form spontaneously in physiological salt and pH conditions in vitro [9]. A growing body of evidence implicates these structures in several biological processes, like directed genome rearrangements [10], [11], telomere capping [12], [13], and control of gene expression at the transcriptional and post-transcriptional levels [14], [15]. Recently, we showed that the GC-rich human minisatellite CEB1 forms G-quadruplexes in vitro and demonstrated that Pif1, a conserved 5′-3′ helicase, unwinds these G-quadruplexes [16]. In Saccharomyces cerevisiae, Pif1 prevents the formation of G-quadruplex-dependent CEB1 internal rearrangements during leading-strand replication and, consistently, the treatment of WT cells with the potent G-quadruplex binder Phen-DC3 mimicks the absence of Pif1 [16], [17], [18].
A different but perhaps related feature of the human GC-rich minisatellites with respect to genome stability is their clustering in the chromosomal subtelomeric regions [19], [20] that are subjected to pathological terminal truncations [21]–[24]. The genomic factors involved in the highly dynamic behavior of terminal regions being poorly identified, here we examined the fragility of the subtelomeric human minisatellites HRAS1 [25], CEB1 [26] and CEB25 [27] and the role of their specific sequence features in the induction of Gross Chromosomal Rearrangements (GCR) in S. cerevisiae. To this end, we employed the GCR assay developed by R. Kolodner and colleagues [28] that measures the rate of the yeast chromosome V terminal deletion. We showed that the three minisatellites and sequence variants stimulated the formation of GCR in WT cells to different extents depending on several factors: the number of motifs in the tandem array, the ability to form G-quadruplexes, the presence of Cdc13 binding sites, their orientation which yields different type of rearrangements, and/or the activity of Pif1 and of the homologous recombination pathway. Altogether, these results point to GC-rich minisatellites as major at-risk regions of the genome not only for changes in repeat number but also for their propensity to generate structural variants.
To study the behavior of human GC-rich minisatellites in the formation of GCR, we employed the genetic assay developed by Chen and Kolodner [28]. In this sensitive assay, the left arm of chromosome V was engineered to measure the rate of the simultaneous loss of the CAN1 and URA3 markers located in the terminal non-essential part of the chromosome V. Cells that undergo a GCR event that results in the simultaneous loss of URA3 and CAN1 are recovered on media containing canavanine and 5-fluoro-orotic acid (5-FOA). Fluctuation analysis of the number of growing colonies provide a very sensitive GCR assay (see Materials and Methods), ranging over several order of magnitude since in WT cells, the GCR rate is approximately 10−10 events per generation [28]. We inserted the minisatellites centromere-proximal to CAN1 within the non-essential NPR2 locus, together with the Hygromycin resistance gene (hphMX) (Figure 1A). Importantly, the HYGR cassette has a GC-content of 58%, does not share homology with the yeast genome, and is devoid of potential G-quadruplex-forming sequences or Cdc13 binding sites. Hereafter, to compare strains with similar size inserts, the hphMX construct constitutes our “no minisatellite” control strain.
Altogether, we examined three subtelomeric GC-rich human minisatellites: CEB1 [26], CEB25 [19], and the minisatellite located in the promoter of the HRAS1 gene [25]. They are tandem arrays with motif lengths of 39, 52, and 28 nt, respectively. The sequence of the consensus motif and additional features of these minisatellites are indicated in Table 1. Furthermore, it is known that the CEB1 and CEB25, but not the HRAS1 motifs, can form stable G-quadruplex structures in vitro [16], [27]. All three minisatellites were inserted in both chromosomal orientations at the same locus. In the orientation ‘“G”, the G-rich strands of CEB1 is on the same strand as the G-rich 3′ ssDNA overhang of the chromosome V left-arm telomere (distance is approximately 45 Kb), while in the orientation “C”, the C-rich strand is on the same strand as the G-rich 3′ overhang (Figure 1B). All the rates measured throughout this study are reported in Table S3. Hereafter, unless otherwise stated, the inserts we refer to are in the “G” orientation.
First, we examined GCR rates in the absence of minisatellite inserts. Previous studies provided an estimated rate of 3.5.10−10 events/generation for WT cells [28]. Consistently, the control strain (npr2::hphMX) exhibits the same GCR rate as the parental (NPR2+) RDKY3615 haploid strain (4.3×10−10 vs. 4.2×10−10 events/generation, respectively). Thus, adding approximately 1.8 Kb of a non-repeated GC-rich DNA to the 13 kb region permissive for rearrangements (located between CAN1 and the first centromere-proximal essential gen, PCM1) has no detectable effect. Then, we measured the consequence of the insertion of the CEB1-WT-1.7 allele containing 43 motifs [16], CEB25-WT-0.7 (13 motifs) and HRAS1-0.7 (26 motifs). Compared to the control strain (hphMX), these minisatellites strongly but differentially increased the GCR rate in WT cells: 20-fold for HRAS1 (8.48×10−9 events/generation), 1,620-fold for CEB1 (6.97×10−7 events/generation) and 276,000-fold for CEB25 (1.16×10−4 events/generation) (Figure 1C).
Pif1 is a conserved 5′-3′ helicase that suppresses GCR events by telomere healing [29], [30] through direct removal of the telomerase from DNA ends [31]. Pif1 is also involved in G-quadruplex unwinding [16]. We constructed pif1Δ cells carrying the minisatellites. Consistent with previous findings [29], [32], in the “no-insert” and in our control insert strain, the GCR rates are increased approximately 1500–2250-fold (6.63×10−7 and 1.01×10−6 events/generation, respectively) in the pif1Δ strain compared to WT. The presence of the minisatellites had various quantitative effects. Compared to the control pif1Δ strains, HRAS1, CEB1 and CEB25 stimulated the GCR rate 3.6-fold (3.68×10−6 events/generation), 385-fold (3.89×10−4 events/generation) and 120-fold (1.21×10−4 events/generation), respectively (Figure 1D). If we now compare the WT and the pif1Δ cells carrying the same minisatellite, the absence of Pif1 increases the GCR rate of HRAS1 and CEB1 approximately 500- and 558-fold, but has no effect on CEB25. This insensitivity to Pif1 reflects the already high rate of GCR induced by CEB25 in WT cells. The heterogeneous behavior of this set of minisatellites suggests that specific sequence features modulate their propensity to trigger GCR, in both WT and pif1Δ cells.
The CEB1 motif forms G-quadruplexes that are efficiently unwound by Pif1 in vitro [16], [18]. To determine the role of the G-quadruplex forming sequences of CEB1 on GCR rate, we first examined the behavior of the CEB1-Gmut-1.7 array which does not form G-quadruplex (Figure 2A) [16]. In the WT strain background, the insertion of CEB1-Gmut-1.7 yields a GCR rate of 2.06×10−8 events/generation. This is 65-fold higher than in the control strain but 30-fold lower than in the CEB1-WT-1.7 cells carrying the same number of G quadruplex forming motifs (Figure 2B). These results indicate that the effect of CEB1 on GCR rate is both G-quadruplex-independent and –dependent. Similarly, we examined the behavior of the CEB1-Gmut-1.7 allele in pif1Δ cells. The GCR rate was stimulated 6-fold (6.32×10−6 events/generation) compared to the control pif1Δ strain, but was 62-fold lower than in the CEB1-WT-1.7 cells (Figure 2C). This level is similar to the GCR rate induction observed with the HRAS1-0.7 minisatellite also devoid of G-quadruplex-forming sequence. We conclude that, in both WT and pif1Δ cells, the induction of GCR by CEB1 strongly depends on its potential to form G-quadruplexes.
To confirm the stimulating role of G-quadruplex, we compared the rate of GCR in cells treated or not with the G-quadruplex-stabilizing ligand Phen-DC3 [33]. The treatment of WT cells bearing CEB1-WT-1.7 with 10 µM Phen-DC3 yielded a GCR rate of 3.65×10−5 events/generation, 52-fold higher than in the untreated cells (Figure 2D). We verified that this induction was not due to a better growth rate of cells having performed a GCR in the presence of the ligand (Figure S1). In contrast, Phen-DC3 failed to increase the GCR rate in CEB1-Gmut-1.7 cells (3.56×10−8 events/generation) (Figure 2D). We also assayed concentration effects and treatment with Phen-DC6, a compound related to Phen-DC3 [33]. Clearly, the extent of GCR rate induction in WT cells carrying the CEB1-WT-1.7 minisatellite was stimulated by both ligands and is dependent on their concentration (Figure 2E).
Finally, since our previous studies examined G-quadruplex-dependent expansion/contraction of CEB1 in different chromosomal locations [16]–[18], we determined the frequencies of CEB1 expansion/contraction in this chromosome V location. As previously observed on Chr. III and VIII, the CEB1-WT-1.7 array was rather stable in WT cells (3/192 rearrangements) and became frequently rearranged upon treatment with Phen-DC3 (39/192, p-value vs. untreated = 8.8e−10) or PIF1 deletion (16/192, p-value vs. WT = 3.55e−3) (Table 2). This depends on the presence of the G-quadruplex-forming sequences, since the CEB1-Gmut-1.7 allele remained stable in the above conditions (Table 2).
We conclude that the impairment of the G-quadruplex unwinding capability of the cell, either by adding G-quadruplex-stabilizing ligands in WT cells or by deleting PIF1, stimulates the propensity of the G-quadruplex-prone CEB1 minisatellite to undergo a high level of expansion/contraction and to a lesser extent GCRs.
Next, we examined the relationship linking the total number of motifs in the CEB1 array and the GCR rate, both in WT and pif1Δ cells. We observed that the rate of GCR in WT cells was positively correlated to the number of repeats (p-value = 2.8×10−3, Spearman's correlation test)(Figure 2F), with rates ranging from 1.1×10−8 events/generation for the allele of 0.66 kb (17 motifs) to 1.59×10−5 events/generation (37,000-fold higher) for the longest allele of 2.7 kb (≈70 motifs). The straight slope in logarithmic scale suggests that the relationship linking the motif number and the GCR rate is roughly exponential. Similarly, CEB1-Gmut also induces the formation of GCR in a size-dependent manner (p-value = 2.8×10−3) (Figure 2F), but with a lower slope: an allele of 1.9 kb (≈49 motifs) induced a GCR rate only 4-fold higher than a 0.9 kb allele (23 motifs)(3.52×10−8 and 8.64×10−9 events/generation, respectively). In the absence of Pif1, the GCR rates also increased exponentially with the number of CEB1-WT repeats (p-value = 3.97×10−4)(Figure 2G). Hence, we conclude that the number of repetition of the minisatellite motif is an aggravating factor of the fragility of these sequences, being steeper with the G-quadruplex-forming ones.
To determine the nature of the GCR events induced by CEB1-WT-1.7, we isolated a set of Can/5FOA-resistant colonies from independent cultures to avoid sibling events and analyzed their genomic DNA by Southern blot. The DNA was digested with a restriction enzyme cutting in the centromere proximal part of CEB1 and successively visualized with a CEB1 and a telomeric probe on the same blot. In the majority of colonies isolated in the WT strain background (29/31, 94%) it revealed a smeared CEB1 hybridizing band, which co-hybridized with the telomeric probe (Figure 3A). Similar events and proportion were found for the WT strain treated with Phen-DC3 (18/18), pif1Δ cells (18/19) (Figure 4C), and WT cells carrying the CEB1-Gmut-1.7 array (8/10 events). Thus, these GCR are likely telomere addition (telomere have variable length in the cell population) associated with a variable number of residual CEB1 motifs.
Analysis of the median length of the smeared band allowed us to roughly determining the number of remaining CEB1 motifs. In untreated WT cells, the events were evenly distributed along the 43 CEB1-WT motifs with the median telomere addition at the 25th motif (Figure 3B). In contrast, upon treatment of WT cells with Phen-DC3, or deletion of PIF1, telomere addition sites shifted significantly toward small fragments, with a median of 11 (p-value = 6×10−4) and 15 (p-value = 9.1×10−3) motifs, respectively (Figure 3B). These results indicate that (i) irrespectively of the nature of the CEB1 array, most GCR events are telomere addition within CEB1, (ii) telomere addition can occur at numerous places within the CEB1 array thus leaving a variable number of CEB1 motifs, and (iii) impairing the ability of cells to unwind G-quadruplexes (Phen-DC3 and pif1Δ) is associated with an increased loss of CEB1 motifs.
To gain higher resolution mapping of the telomere healing events within the CEB1 motifs, we sequenced a set of CEB1-telomere junctions using Ion Torrent Next-Generation Sequencing technology after purification of appropriate DNA bands on agarose gel (see Materials and Methods). We identified the CEB1-Tel junctions from 15 untreated and 12 Phen-DC3-treated WT cells (Figure 3C and Figure S2, respectively).
Telomere additions occur mainly at regions of the CEB1 motif that exhibit limited homology to the yeast telomeric sequence. Precisely, 10/27 CEB1-telomere junctions lie in the longest sequence of homology between CEB1 and the telomeric sequence (GGGTGG) and 24/27 junctions have at least two nucleotides in common between CEB1 and the telomeric sequence (shown in blue in Figure 3C). This result is consistent with previous observations showing that for de novo telomere addition to occur, homology to telomeric sequence of 2-bp (TG, GG, and GT dinucleotide) is sufficient and that a longer homology facilitates telomere healing [30], [34], [35]. The fact that 62% of the telomere additions occur in, or at, the junction with the G-quadruplex-forming sequence of CEB1 (red lines Figure 3C) is consistent with the fact that 60% of the TG, GG, and GT dinucleotides overlap this sequence. The distribution of the telomere addition within the CEB1 motif is not significantly different in untreated and Phen-DC3-treated WT cells (Figure S2). Hence, although the Phen-DC3 treatment strongly increases the rate of GCR (Figure 2D) and affects the position of the telomere addition in the array (Figure 3B), the position of the CEB1-telomere junction remains unaffected and mainly lies in the G-quadruplex-forming sequence (Figure 3C and Figure S2).
Altogether, these results suggest that the G-quadruplexes present within the CEB1 array in conditions where the capacity of the cell to unwind G-quadruplexes is impaired (upon Phen-DC3 treatment or PIF1 deletion) stimulate the formation of GCR associated with a decreased number of CEB1 motifs remaining in the final repair product.
Telomere healing may occur by de novo telomere addition to a 3′ ssDNA extremity, especially in the absence Pif1 [29], [34]–[36], leaving a specific pattern of telomeric sequences [34]. However, among the 27 junctions we sequenced, we do not notice any obvious addition of a particular pattern of telomeric sequence in CEB1. On the other hand, telomere addition could occur by capture of endogenous telomeric sequences by break-induced replication (BIR) [28], [37], [38]. We examined the effect of the deletion of the RAD51 or RAD52 genes that are required for BIR [37], [38] but not for direct telomere addition by telomerase. It causes a 2-fold decrease of the GCR formation in strains bearing CEB1-WT-1.7, with rates of 2.92×10−7 and 3.54×10−7 events/generation, respectively (Figure 3D). The extent of the decrease is similar (3- to 5-fold) upon Phen-DC3 treatment, with GCR rates of 1.13×10−5 and 7.12×10−6 events/generation in the rad51Δ and rad52Δ mutants, respectively (Figure 3D). Interestingly, the molecular analyses of the nature of the events provided additional information. We found that the drop of the GCR rate in the absence of Rad52 is associated with a specific decrease of GCRs by telomere addition within CEB1 (Figure S3) while the analysis of the CEB1-telomere junction sequences recovered from untreated or Phen-DC3-treated WT cells revealed the presence of SNPs around the junction in 4/6 strains (Figure 3E). These SNPs are found either in the telomeric sequence only, or both in the CEB1 and the telomeric sequence around the junction (Figure 3E). These intriguing observations suggest that in WT cells roughly half of the telomere healing events in CEB1 occur by BIR on an ectopic telomere sharing a region of limited homology with the CEB1 motif [28]. SNPs found at the junction may result from the correction of the heteroduplex formed between CEB1 and the telomeric sequence, and/or by misincorporation of nucleotides in the early BIR steps [39].
CEB1 strands strongly differ with respect to their GC composition (GC-bias = 76.6%) and the density of TG/GG/GT dinucleotide (bias is 87%) that seeds GCR by telomere healing (Table 1). We examined the behavior of CEB1 placed in the opposite orientation (orientation C) relatively to the distal telomere (Figure 1B). Strikingly, in WT cells, the GCR rates of CEB1-WT-1.7 are similar in either orientation (6.97 and 7.47×10−7 events/generation)(Figure 4A) and alike the G-strand, the GCR rates increase according to the total size of the array (Figure S4). Similarly, although occurring at various absolute rates, there is no significant orientation-dependent difference in all the other strains and conditions that we assayed (Figure 4A and 4B, Table S3). Namely, in WT cells carrying the CEB1-WT-1.7 array treated with Phen-DC3 (3.65 and 1.66×10−5 events/generation), CEB1-WT-1.7 in pif1Δ cells (3.89 and 4.6×10−4 events/generation), CEB1-Gmut-1.7 in WT (2.77 and 2.07×10−8 events/generation) and pif1Δ cells (6.32 and 3.05×10−6 events/generation) nor in cells carrying HRAS1-0.7 in WT (8.48×10−9 and 1.1×10−8 events/generation,) and pif1Δ cells (3.68 and 3.2×10−6 events/generation) (Figure 4A and 4B, Table S3). Hence, both in the WT and pif1Δ cells, the GCR rates induced by CEB1-WT-1.7, CEB1-Gmut 1–7, and HRAS1-0.7 are not affected by the minisatellite orientation on the chromosome. However, the pattern of rearrangements in the G and C orientations is very different (Figure 3A and Figure 4C). In WT cells bearing CEB1-WT-1.7 in the orientation C, only 2/22 rearrangements are smears indicative of telomere healing. The DNA of two other colonies migrates at the size expected for an unaltered Chr. V. By PCR analysis of CAN1 and URA3, we observed that clone 12 (Figure 4C) retained both genes. Sequencing identifies a mis-sense mutation in URA3 (G411A) and a frameshift in CAN1 (del595G). It might be a rare case of two independent mutagenic events but more likely a mutagenic fill-in synthesis by BIR [39], occurring in this case on the sister chromatid to restore a full-length chromosome V. The other clone has lost CAN1 and URA3. Thus, it is a structural variant like the majority of events (19/22), which manifest themselves as discrete bands of various sizes. Among them, 15 hybridize with both the hphMX and the CEB1 probes (Figure 4C). The variable hybridization intensity of the CEB1 signal indicates that the amount of remaining CEB1 sequence in the rearranged chromosomes is different from one strain to another (for example, compare lanes 6 and 10 in Figure 4C). It is interesting to note that in some cases (4/18), two or more bands hybridizing both the CEB1 and hphMX probes are visible (clones 1–3, and 7). To gain more insights into the nature of these rearrangements, we analyzed clones 1–4 by pulse-field gel electrophoresis and Comparative Genomic Hybridization (CGH) (Figure S5). All exhibit an abnormal migration of Chr. V, while the rest of the karyotype appears normal (Figure S5A). As expected, CGH analysis revealed that the distal part of Chr. V containing URA3 and CAN1 is lost (Figure S5B). Furthermore, complex changes in copy number on other chromosomes are detected (details are reported in Figure S5). To be noticed, Ty1 elements are present in the vicinity of the breakpoints, suggesting that they are preferred sites for GCR [40]. Thus, contrary to the prominent telomere additions observed in the G orientation, GCR induced by CEB1 in the C orientation are diverse and complex, as observed among spontaneous GCR events [28], [40]. The similar rate but different product structures in the G and C orientations can be explained if they result from a similar initiating event but difference in repair; In the G orientation, BIR starting within CEB1 on a telomere substrate will process in the chromosomal distal direction and immediately heal the initiating lesion. In the C orientation, BIR on a telomere substrate will process in the proximal direction to copy the entire chromosome, thus leading to the formation of a dicentric molecule prone to secondary complex rearrangement(s) before stabilization [41].
Furthermore, to address the genetic requirements of these GCR events, we examined the role of the non-homologous end joining (NHEJ) and homologous recombination (HR) pathways. The GCR rate remains unchanged in the dnl4Δ mutant (Figure S6A) while we observed a small but significant 4-fold decrease of the GCR rate in the rad51Δ and rad52Δ mutants (Figure S6A). In the absence of Rad52, the remaining events are telomere additions (8/9 events) (Figure S6B) suggesting that the HR pathway plays a major role in the formation of the structural but not telomere addition events generated by CEB1 in the C orientation.
We next asked what could be the molecular reasons for the high GCR rate induced by CEB25 in orientation G, and the inability of Pif1 to suppress GCR induced by this construct in WT cells (Figure 1C and 1D). The GCR rate is not dependent on Rad52 (3.9×10−4 events/generation) and all events in WT cells (11/11) are telomere additions within CEB25 (Figure S7). Interestingly, we found that contrary to CEB1, the GCR rate induced by CEB25 strongly depends on its orientation: the inversion of CEB25 caused a 516-fold decrease of the GCR rate in WT cells (2.24×10−7 events/generation). In pif1Δ cells, the GCR rate of CEB25 in the orientation C was close to the “no insert” control strain (2.41×10−6 vs. 1.01×10−6 events/generation). This strong orientation-dependency prompted us to investigate the sequence composition of the CEB25 motif.
CEB25 has a GC content of 58% and exhibits an absolute GC-bias and GT/GG/TG dinucleotide bias (Table 1). Interestingly, it contains several consensus-binding sites for the 3′ telomeric overhang binding protein Cdc13 (GTGTGGGTGTG, in which the first 4 nucleotides are critical [42], underlined in Figure 5A) [43], [44]. Cdc13, together with Stn1 and Ten1, is a part of the CST complex involved in telomere capping and mutagenic DSB repair by addition of telomeric repeats at a 3′ ssDNA end [32], [45]–[47]. This unique feature, compared to CEB1 and HRAS1, led us to suspect that the recruitment of Cdc13 on CEB25 could be responsible for its GCR effect. To test this hypothesis, we conducted both in vitro and in vivo experiments. In vitro, we determined the affinity of the purified Cdc13 for the CEB25 motif upon gel shift assay (Figure 5A). Cdc13 binds with high affinity to the CEB25 motif (CEB25-WT), with a Kd = 6.4×10−11±10−11 M. Mutations of the Cdc13 binding sites present in the CEB25 motif (CEB25-Cdc13mut) resulted in a 44-fold lower affinity for Cdc13 (Kd = 2.8×10−9±3×10−10 M)(Figure 5A).
Then, to address the possibility that the high affinity of Cdc13 for CEB25 is responsible for the high GCR rate induced by this minisatellite only when the G-rich strand is in the same molecule than the distal telomere (and thus can be directly extended by telomerase), we constructed and introduced in yeast a 1.4 kb CEB25 allele mutated for its Cdc13-binding sites (CEB25-Cdc13mut-1.4, same motif as in Figure 5A) that kept the same GC content and did not change the G-triplets potentially involved in the G-quadruplex formation (see below). Remarkably, in the orientation G, this construct induced a GCR rate of 3.07×10−7 events/generation. This is 713-fold higher than in the “no minisatellite” control strain, and 377-fold lower than with CEB25-WT-0.7 in the same orientation (Figure 5B). Contrary to CEB25-WT, the GCR rate was not affected by the inversion of CEB25-Cdc13mut-1.4 (2.95×10−7 events/generation), indicating that the strong orientation dependency observed with CEB25-WT relies on the presence of the Cdc13-binding sites (Figure 5B). Additionally, in the absence of Pif1, CEB25-Cdc13mut-1.4 also shows a decreased GCR rate compared to CEB25-WT-0.7 in the orientation G (60-fold)(Figure 5C). Again, the GCR rate induced by CEB25-Cdc13mut-1.4 was similar in both the orientations G and C (3.86 and 2.89×10−6 events/generation, respectively), and close to the control pif1Δ strain (1.01×10−6 events/generation)(Figure 5C). Hence, the orientation-dependent and Pif1-independent behavior of CEB25-WT is associated with the ability of its motifs to bind the accessory telomerase subunit Cdc13 with high affinity.
CEB25 contains a consensus G-quadruplex-forming motif (Table 1) that forms a monomorphic G-quadruplex whose structure has been recently solved by NMR [27]. To investigate the potential involvement of G-quadruplexes in the fragility of CEB25, we first examined the GCR rate of CEB25-Cdc13mut-1.4 in the WT and pif1Δ strains (mutations of the Cdc13 binding sites does not change the G-triplets involved in G-quadruplex formation). We found that GCR rates were (i) similar in these strains (Figure 5C), (ii) occurred at a low level comparable to CEB1-Gmut-1.7 (Figure 4B) and HRAS1 (Figure 1D) and (iii) lower than for CEB1-WT-1.7 (Figure 4B). To investigate the potential role of the CEB25 G-quadruplex forming sequences, we synthesized a CEB25 allele mutated for both the G-tracts and the Cdc13 binding sites (CEB25-Cdc13mut-Gmut-1.4). Clearly, the Phen-DC3 treatment of WT cells bearing CEB25-Cdc13mut-1.4 and CEB25-Cdc13mut-Gmut-1.4 alleles in both orientations yielded no increase of the GCR rates (Figure 5D). This did not depend on the absence of intact Cdc13 binding sites since the CEB25-WT-0.7 allele in the orientation C also remained insensitive to Phen-DC3 (Figure 5D). Rather, the G-quadruplex-forming and the G-mutated versions of CEB25-Cdc13mut exhibited exactly the same rates of GCR in WT cells. This absence of effect of Phen-DC3 contrasts with the 22- to 52-fold inductions observed with CEB1-WT upon WT cells treatment (Figure 4A). We then combined the deletion of PIF1 to the Phen-DC3 treatment, conditions that yielded synergistic destabilization of CEB1 [18]. We observed a weak 5.5-, 2.3- and 4.6-fold induction of the GCR rates upon treatment of cells bearing CEB25-WT-0.7 in the orientation C, and CEB25-Cdc13mut-1.4 in the orientations G or C, respectively (Figure 5E). No induction was seen upon treatment of cells bearing the G-mutated version of CEB25-Cdc13mut (Figure 5E). These extreme conditions revealed a slight G-quadruplex-dependent GCR induction by CEB25.
Since the minisatellites studied here induced the formation of GCR, we wished to gain more insights into the GC-rich minisatellite representation and localization in the human genome. Using Tandem Repeat Finder [48], we determined a list of 353,460 minisatellites (Table S4). They are not evenly distributed along chromosome arms (Figure 6A) [20], being enriched in the 10 and 5% terminal arm regions (Figure 6B). Interestingly, it seems to relate to their GC-content since the 85,222 minisatellites (24%) that have a GC-content higher than 50% preferentially localize at the most terminal parts of the chromosome, whereas the other minisatellites with a lower GC-content are evenly distributed along the arms (Figure 6A and 6B). A similar bias has been previously reported for chromosome 22 [19]. Then, we examined the minisatellites having potential G-quadruplex-forming sequences. Five percent (18,906) of the minisatellites bear at least one G-quadruplex-forming sequence (see Materials and Methods), and 96% (18,191) of these G-quadruplex-forming minisatellites are GC-rich (Table S5). Among the 504 minisatellites that contain at least 30 G-quadruplex-forming sequences due to their tandem repeated structure, 60% (313/504) lie within the terminal 10% of chromosome arms, among which 80% (253/313) lie within the terminal 5%, while keeping a constant GC-content (Figure S8). Hence, GC-rich and G-quadruplex-forming minisatellites appear to preferentially cluster towards the chromosomal extremities (Figure 6C). The mutagenic behavior of HRAS1, CEB1 and CEB25 arrays in yeast described here raises the possibility that the human GC-rich minisatellites play a role in GCRs of the terminal part of human chromosomes.
In this study, we assayed the fragility of three GC-rich human minisatellites and mutant derivatives in S. cerevisiae. All these minisatellites stimulated the formation of GCRs but at rates varying by several orders of magnitude. We found that the rate depends on several intrinsic sequence features: the total number of repeats, the ability or not to form G-quadruplex secondary structures (case of CEB1) or to bind with high affinity the telomere ssDNA binding protein Cdc13 (case of CEB25). These features also explain their different levels of responsiveness to the Pif1 helicase controlling telomere elongation and G quadruplex unwinding. CEB1 and CEB25 are also differentially responsive to their orientation on the chromosome; it drastically affects the GCR rate of CEB25 but not HRAS1 or CEB1, and in all cases dictates the type of GCR (telomere addition versus other structural rearrangements). Thus, the behavior of these minisatellites is largely specific. We uncovered here their sequence features.
Spontaneous GCR in WT cells occurs at a very low rate (10−10). It yields a variety of rearrangements that delete the non-essential distal chromosomal region and rescue the chromosome by telomere addition at breaks that contain limited homology to telomere-like seed sequences as well as through more complex genome rearrangements [28]. Two factors may increase the rate of GCR: an increased number of initiating lesions or defects in the repair pathways [28], [29]. Regarding the later possibility, as previously reported, we observed that Pif1 plays an important role in suppressing the formation of GCR by telomere healing [29], [30], [32], [49]. In all but one of our minisatellite insertions, GCR rates were increased by several orders of magnitude upon PIF1 deletion. However, in sharp contrast, the extreme GCR rate stimulated by CEB25 in WT cells remained roughly the same in pif1Δ cells. This insensitivity to Pif1 depends on the orientation of CEB25 relative to the distal telomere (G-strand in the same orientation as the single-stranded telomere G-overhang is the most active) in agreement with the ability of the motif to bind the endogenous Cdc13 yeast protein with high affinity (Figure 5A and 5B). Clearly, the mutation of the three Cdc13-binding sites yields a ≈380-fold reduction in GCR, consequently abolishing the CEB25 orientation-dependent behavior. The simplest interpretation of these results is that the recruitment of Cdc13 to CEB25 is sufficient to overcome the suppressive effect exerted by Pif1 to prevent the recruitment of the telomerase [46]. This is consistent with the Pif1-independent de novo telomere addition at a long internal telomeric tract (TG)81 introduced near an unrepairable HO break [32]. In our assay, due to its motif sequence and its organization in tandem repeats, the human CEB25 minisatellite fortuitously resembles a pseudo-telomere.
On the other hand, the HRAS1, CEB1, CEB1-Gmut, CEB25-Cdc13mut and CEB25-Cdc13mut-Gmut tandem arrays devoid of Cdc13 binding sites also induce GCR but at various rates and in an orientation-independent manner. Among the parameters potentially involved in the fragility of CEB1, its ability to form G-quadruplexes appeared as an important destabilizing feature. Compared to the CEB1-Gmut-1.7 construct, the G-quadruplex-prone CEB1-WT-1.7 allele stimulates the GCR rate in WT cells 30-fold (Figure 2B) and accordingly the conditions that shift the equilibrium toward the folded G-quadruplex state increase the GCR rate : 52-fold upon treatment with the G-quadruplex stabilizing ligand Phen-DC3 and 558-fold in the absence of the G-quadruplex unwinding helicase Pif1 (Figure 2B and 2C, Figure 3D and 3E). However, it should be emphasized that a predictive G-quadruplex-dependent phenotype cannot be safely ascertained from the presence of a consensus G-quadruplex motif in a given sequence, nor its ability to form stable G-quadruplexes in vitro. Indeed, contrary to CEB1, the CEB25 array did not responded in vivo to the three conditions that affect G-quadruplex-dependent events (G quadruplex motif mutation, treatment with PhenDC3 or Pif1 deletion) except slightly, when combining the Phen-DC3 treatment to the PIF1 deletion (Figure 5E). This synergistic combination previously observed for CEB1 [18] appears as an extreme hypersensitive condition that may lead to the rare accumulation of unprocessed CEB25 G-quadruplexes. The distinct behavior of CEB1 and CEB25 may rely on different conformations of their respective G-quadruplexes affecting their folding and/or their processing in vivo.
Besides sequence affinity to Cdc13 and potential to form G-quadruplexes, a third aggravating factor stimulating the GCR rate is the total number of motifs. Thanks to the sensitivity of this GCR assay, we found that the GCR rate of CEB1 arrays increased exponentially with the number of motifs without an apparent threshold in both WT and pif1Δ cells (Figure 2F and 2G). Interestingly, a similar exponential relationship between the number of motifs and the propensity of the triplex-forming (GAA)n repeats [50] to form GCRs [51] and expansions [52] has also been reported in yeast. It suggests the intriguing possibility that the capacity of tandem arrays to form secondary structures is a relevant feature. Along this line, we know that a tandem of two and three CEB25 motifs is able to form a pearl-necklace monomorphic G-quadruplexes structure [27]. If CEB1 is also able to form a pearl-necklace G-quadruplexes structure, the size-dependent exponential increase of the GCR rate may reflect the cooperative behavior between the CEB1 motifs to fold into G-quadruplexes. Mechanistically, we recently reported that the CEB1 G-quadruplex prone array perturbs replication and lead to expansion and contraction events [17]. As we proposed, the blockage of the DNA polymerase(s) at the first G-quadruplex may be sufficient to trigger the accumulation of ssDNA between the replication forks and the polymerase and thus enhance the formation of G-quadruplexes per cell and per molecule in a manner related to the total number of repeats. This situation may be similar to the Pol2 slowdown observed at single G-quadruplex-forming motifs under treatment of Pif1-deficient cells with the replication inhibitor hydroxyurea [53].
In addition to the effect of G-quadruplexes, other non B-DNA secondary structures can be the source of sequence fragility [8]. However, we found that the HRAS1-0.7 and CEB25-Cdc13mut-Gmut-1.4 minisatellites, devoid of potential G-quadruplex or other secondary structures, also stimulated the GCR rate by 20- and 700-fold in WT cells, respectively. In addition, once the G-quadruplex-forming capacity of CEB1 was removed by site-directed mutations, we noted that the CEB1-Gmut-1.7 construct was still able to stimulate GCRs at a substantial level (≈2×10−8 events/generation), approximately 60-fold higher than in the control WT strain. Similarly, the structure-free (ATTCT)n microsatellite has been reported recently to induce chromosomal fragility in WT yeast cells, which increase with the number of motifs [54]. However, the slope of this length-dependent effect could not be derived from these experiments since only two different allele sizes have been assessed [54]. The analysis of CEB1-Gmut allele of various lengths (23–70 motifs) revealed a length-dependent fragility in WT cells in an almost linear manner (multiplying the number of motifs by two increased the GCR rate by 4), in sharp contrast with the exponential slope observed with CEB1-WT (Figure 2F). This difference suggests that the G-quadruplex-independent fraction of the CEB1 fragility does not involve a cooperative behavior between the motifs. What remaining sequence properties could account for this structure-independent fragility? The GC-richness per se can be invoked, since it has been shown to slowdown DNA polymerases in vitro [55]. In the case of our minisatellites, however, three reasons argue against its essential role. First, with similar size arrays, the GCR induction is not clearly correlated to the GC-richness: HRAS1 (GC = 67%) and CEB1-Gmut (72%) both stimulated the GCR rate ≈20-fold compared to the no insert strain, but ≈35-fold less than CEB25-Cdc13mut (GC = 56%). Second, the hphMX insert, whose size and GC content is similar to CEB25-Cdc13mut-1.4 (≈58%), did not stimulate GCR above the no-insert control strain. And third, the density of TG/GG/GT dinucleotides that can seed telomere addition is similar in the CEB25-Cdc13mut, HRAS and hphMX insertions (Table 1). These observations suggest that the GC-richness is not per se the determinant triggering GCR. The remaining shared feature of these sequences is their organization in tandem. By itself, it may perturb the normal progression of replication due to the high local concentration of homologous templates or create long range specific chromatin structures that might be processed at the expense of maintaining genome stability.
In addition to inducing truncated arrays and motifs by GCR, CEB1 also varies in size by increasing or decreasing the total number of motifs via SDSA and/or template switch without involving the flanking regions [16]–[18]. These events are extremely frequent, being detected in 8.3 and 20.3% of the cells upon deletion of PIF1 or Phen-DC3 treatment, respectively (Table 2). This is 100–1000 fold higher than the GCR rates (3×10−4 and 3.6×10−5 events/generation, respectively) of the same construct. Thus quantitatively, expansion/contraction is the major outcome of CEB1 instability with the advantage to avoid the formation of potentially detrimental structural rearrangements. This is in agreement with numerous reports that compared internal rearrangements and GCR induced by different microsatellites [2], [56]–[58]. Mechanistically, since the presence of CEB1 perturbs replication [17], GCR events might result from the rare situations in which the template directed intra-motif interactions failed, allowing break-induced replication on an ectopic telomere sequence [51] or the recruitment of the telomerase to act. Consistent with a role of the homologous recombination pathway, we observed that the deletion of the RAD51 or RAD52 genes yield a ≈4-fold decrease of the GCR rate (Figure 3D and Figure S4A). This is true in both orientations although the nature of the GCR events is different. The insufficient absolute frequency of GCR events (<10−4) prevented us to determine whether or not the variation of the GCR rates were compensated by an increase of the expansion/contraction events that can be detected by Southern blot analyses of individual or small pool of colonies.
Chromosomal rearrangements are potentially detrimental for cell functions and are the source of genetic diseases and cancer. Remarkably, subtelomeric regions are highly dynamic in primate and altered in approximately a third of the human pathologies involving chromosomal rearrangements [21], [23], [24], [59], [60]. However, the factors involved in the high propensity of these regions to break and rearrange have not been identified. The intergenic CEB1 and HRAS1, as well as the intronic CEB25 minisatellites assayed here are located 400 kb–1.4 Mb away from the telomeres (Table 1), representative of the enrichment for GC-rich and the G-quadruplex-forming minisatellites at chromosome terminal regions in the human genome (Figure 6A and 6B). In yeast, the orientation does not affect the fragility per se but the nature of the GCR. Hence, given the high number of GC-rich minisatellites clustering at chromosome ends in the human genome irrespectively of their orientation, these sequences are likely implicated in the generation of the various subtelomeric rearrangements [61], [62]. But why these harmful sequences are massively present in the human genome? And what could be the reasons of their terminal clustering? A positively selected function could be to signal defects in replicating G-quadruplex-forming sequences [17], [53]. In this regard, the arrangement in tandem of G-quadruplex-forming motifs presents at least two advantages. First, they would act as severe “tandem of problems” for replication machinery as revealed by their exponential size-dependent fragility. Hence, cells with a decreased ability to remove G-quadruplexes will experience replication difficulties preferentially at these G-quadruplex-forming minisatellites rather than at unique sequences present throughout the genome and enriched in proto-oncogenes [17], [53], [63]. Second, owing to the higher local concentration of homologous template compare to unique sequences, they will preferentially undergo internal rearrangements rather than inducing structural variations. Thus, we envision that GC-rich and G-quadruplex-forming minisatellites help signaling deficient replication machineries, and their clustering at chromosome ends and repetitive nature overall limit the potential formation of detrimental structural rearrangements.
The genotypes of the Saccharomyces cerevisiae strains (S288C background) used in this study are reported in Table S1. All strains have been derived from RDKY3615 (WT strains) [28] or RDKY4399 (pif1Δ strains) [29] by regular lithium-acetate transformation. Correct insertion of the hphMX cassette with or without minisatellite at NPR2 (position 804, BamHI site), as well as the minisatellite size, have been verified by Southern blot. The CEB1-WT-1.7 and CEB1-Gmut-1.7 minisatellites have been synthesized previously [16]. Contractions and expansions of these minisatellites have been generated during the insertion procedure at the NPR2 locus and are thus independent clones. The CEB25-WT-0.7, CEB25-Cdc13mut-1.4, and CEB25-Cdc13mut-Gmut-1.4 minisatellites have been synthesized in vitro using PCR-based method as previously described [16]. The HRAS1 minisatellite of 0.7 kb (HRAS1-0.7) has been obtained from P37Y8 (gift from D. Kirkpatrick) [64]. The motifs of the minisatellites used in this study are presented in Table S2. Deletion of RAD51, RAD52, and DNL4 has been performed by transformation of the corresponding KMX cassettes amplified from the EUROSCARF deletants collection [65]. Primer sequences are listed in Table S6.
Liquid synthetic complete (SC) and solid Yeast-Peptone-Dextrose (YPD) media have been prepared according to standard protocols [66]. Plates containing Canavanine (Sigma-Aldrich) and 5FOA (Euromedex) have been prepared according to standard protocols [67] with minor differences: because npr2Δ cells exhibit a decreased resistance to acidic pH (<4.0) [68] we adjusted the pH to 4.5–4.8 (instead of 2.8–3.0) and compensated the decreased penetration of 5FOA at this pH by using it at a slightly higher concentration (≈1.5X instead of 1X). SC liquid media containing Phen-DC3 (1, 5, or 10 µM) and Phen-DC6 (1 or 5 µM) have been prepared as previously described [18].
The GCR rate has been determined by fluctuation analysis of 5FOA and canavanine-resistant cells. A ura+ colony is used to inoculate at least 10 independent cultures at a concentration of ≈102–3 cells/mL in 2–50 mL of SC media and grown with shacking at 30°C. When they have reached saturation (2 days), cells are spread on 5FOA/canavanine-containing plates and on YPD plates. A maximum of 108 cells was spread on 85 mm plates, and 109 cells on 145 mm plates. The number of cells spread was adjusted in order not to exceed 100 colonies per plate. For G-quadruplex ligands-containing SC media, cells undergo an overnight preculture in SC prior to inoculation with the ligand, and are grown at 30°C up to saturation. For pif1Δ cells bearing CEB1-WT, which exhibit an inherently high level of CEB1 internal rearrangements, which can influence the GCR rate (strains ORT6543-1, ORT7153-9, and ORT6592-22), the size of the parental minisatellite is determined by Southern blot from individual colonies plated on YPD. The colonies bearing the parental size of CEB1-WT are directly spread on YPD and 5FOA/Can-containing plates without additional liquid culture. After 4 days at 30°C, the number of 5FOA/Can-resistant colonies (r) is counted, as well as the total number of viable cells spread (Nt) derived from the number of colonies formed on YPD. The GCR rate (M) as well as the upper and lower 95% confidence intervals (95% CI) have been calculated from r and Nt with Falcor [69] using the Lea and Coulson method of the median. For each strain and condition, 10 to 45 independent cultures have been performed, in at least two independent experiments. The rates, 95% confidence intervals, and the number of independent cultures performed are listed in the Table S3.
Colonies grown on YPD plates after the 2 days culture in SC media are inoculated in 96-well megaplaque for 24–48 hours. Pools of 4–16 colonies were made right before DNA extraction. DNA was digested with XbaI/EcoNI (leaving 414 bp of flanking sequence) and migrated O/N in a 0.8% agarose-TBE 1X gel at 50 V. Digestion products were analyzed by Southern blot using a CEB1 radiolabeled probe. Blots were scanned using a Storm Phosphorimager (Molecular Dynamics) or a Typhoon Phosphorimager (GE Healthcare), and quantified using ImageQuant 5.2 as described in [18].
In order to avoid sibling events, DNA of 5FOA/Canavanine-resistant colonies from separate cultures is extracted, digested using either SacI or XbaI, and migrated in a 0.8% agarose-TBE 1X gel overnight at 50 V. Digestion products were analyzed by Southern blot as described previously using a radio-labeled CEB1, hphMX (from pAG34), or telomeric (from pCT300) probe. The position of the telomere addition is estimated by measuring the size of the center of the smear, and subtracting both 50 bp of flanking sequence plus the mean telomere size (300 bp in WT cells and 400 bp in pif1Δ cells [30]).
DNA of colonies bearing a CEB1-telomere smear identified by Southern blot was digested by XbaI and migrated in a 0.8% agarose-TBE 1X gel overnight at 50 V. After staining of the DNA with BET, the DNA fragments containing the CEB1-telomere junction were cut and extracted from the gel using the Nucleospin Extract II (Macherey-Nagel) kit. Fragments were quantified, pooled, and precipitated. Samples were prepared for Ion Torrent Personal Genome Machine (PGM, Applied Biosystems). Sequencing has been performed according to manufacturer instructions on a 314R chip. Reads have been validated and aligned on the S288c genome (R64-1-1, 2011-02-03) and custom CEB1-telomere templates using the in-built Torrent Suite 1.5.1. Reads matching both the CEB1 and the telomeric sequences have been isolated and analyzed manually using Tablet 1.11.11.01 [70] and Microsoft Excel 2007.
The list of minisatellites (motif comprised between 10 and 100 bp) and their associated characteristics has been obtained form the Tandem Repeat Database [71] (list generated on the 2010-10-31 by the TRF algorithm [48] on the Homo Sapiens hg19 release). Overlapping duplicates of the same repeat due to uncertainties in the algorithm have been eliminated. The human minisatellites are listed in Table S4. The number of non-overlapping G-quadruplex-forming sequences per minisatellite have been determined using R software. The custom algorithm searches for 4 runs of 3 Gs in a window of 30 nt, with a minimal loop size of 1 nt, and consequently a maximal loop size of 16 nt [72]. They are listed in the Table S5.
A full length version of CDC13 WT was cloned into a pYES2 vector and expressed as a fusion with a C-terminal tag consisting of a 8 glycine linker, 5 strepII-tags (IBA, Germany) and a HAT-tag (Clontech). Cdc13 overexpression was induced in 2% galactose for 16 hours at 30°C according to the method described by P.M. Burgers [73]. Briefly, after grinding cell pellets in liquid nitrogen, the lysate was clarified from DNA by precipitation in 0.1% polyethyleneimine, and the proteins were precipitated with ammonium sulfate at 60% saturation. After resuspension in 50 mM Tris pH 8.0, 300 mM NaCl, 10% glycerol, the soluble fraction was loaded successively on a streptactin column (IBA, germany) followed by a Talon column (Clontech). Purified protein was dialysed against storage buffer 2X without glycerol, and concentrated and stored at −80°C in 1x storage buffer (25 mM tris-HCl pH 8.0, 250 mM NaCl, 0.5 mM DTT, 50% Glycerol). This procedure yielded homogeneous CDC13 estimated more than 90% pure by coomassie blue staining after protein separation by SDS-PAGE.
Gel shift was carried out by incubating 20 pM of the 52-mer CEB25 WT oligonucleotide or the 52_mer-Cdc13mut version, end-labeled at the 5′ end using γ-ATP and T4 polynucleotide kinase, with indicated amount of CDC13, in the following buffer: 5 mM Tris pH 8.0, 2.5 mM MgCl2, 0.1 mM EDTA, 2 mM DTT, 0.1 µg/µl BSA (NEB), 50 mM NaCl, 0.2 M LiCl. After incubation at room temperature for 30 minutes, binding reactions were supplemented with 3% Ficoll and run on a 6% native polyacrylamide gel (37.5∶1 acrylamide/polyacrylamide ratio), at 4°C and 8 V/cm. Gels were dried on DE81 paper and quantified using a Typhoon phosphorimager. Data were fitted to a one-site-specific binding equation (Y = Bmax*X/(Kd+X)) using Prism software (Graphpad), yielding R2 values for goodness of fit of 0.91 and 0.95 for CEB25-WT and CEB25-Cdc13mut, respectively.
Statistical tests have been performed with R software 2.13.1 [74] or Graphpad Prism 5.0b. The α-cutoff for statistical significance was set to 0.05. Rearrangement frequencies of CEB1 have been compared using a two-tailed Fisher's exact test. Correlation between the number of CEB1 motifs and the rate of GCR has been determined using the Spearman correlation test. GCR rates, as well as the distributions of the position of telomere addition in the CEB1 array have been compared using a non-parametric test (Mann-Whitney-Wilcoxon, two-tailed). A one-tailed χ2 test has been used to determined the enrichment of minisatellites in the 10 and 5 terminal percent of chromosome arms.
|
10.1371/journal.pgen.1006512 | Extensive Regulation of Diurnal Transcription and Metabolism by Glucocorticoids | Altered daily patterns of hormone action are suspected to contribute to metabolic disease. It is poorly understood how the adrenal glucocorticoid hormones contribute to the coordination of daily global patterns of transcription and metabolism. Here, we examined diurnal metabolite and transcriptome patterns in a zebrafish glucocorticoid deficiency model by RNA-Seq, NMR spectroscopy and liquid chromatography-based methods. We observed dysregulation of metabolic pathways including glutaminolysis, the citrate and urea cycles and glyoxylate detoxification. Constant, non-rhythmic glucocorticoid treatment rescued many of these changes, with some notable exceptions among the amino acid related pathways. Surprisingly, the non-rhythmic glucocorticoid treatment rescued almost half of the entire dysregulated diurnal transcriptome patterns. A combination of E-box and glucocorticoid response elements is enriched in the rescued genes. This simple enhancer element combination is sufficient to drive rhythmic circadian reporter gene expression under non-rhythmic glucocorticoid exposure, revealing a permissive function for the hormones in glucocorticoid-dependent circadian transcription. Our work highlights metabolic pathways potentially contributing to morbidity in patients with glucocorticoid deficiency, even under glucocorticoid replacement therapy. Moreover, we provide mechanistic insight into the interaction between the circadian clock and glucocorticoids in the transcriptional regulation of metabolism.
| Glucocorticoids, steroid hormones of the adrenal gland, are important regulators of metabolism and the stress response. They are also widely used as anti-inflammatory drugs. Production and release of glucocorticoids show a diurnal pattern regulated by the circadian clock. Importantly, altered daily patterns of hormone action are thought to contribute to metabolic diseases. Here, we examined diurnal patterns of gene expression and metabolism in a zebrafish model of glucocorticoid deficiency. We observed that a surprisingly large number of genes show glucocorticoid-dependent diurnal patterns of transcription. This behaviour is particularly pronounced in metabolic genes, and metabolites of various central metabolic pathways are dysregulated. Interestingly, non-rhythmic glucocorticoid treatment restored many of these metabolite changes. It also restored expression of almost half of the dysregulated genes. Using in vivo bioluminescence assays, we provide evidence that this rescue is mediated via combined glucocorticoid–circadian clock activity on a simple regulatory DNA module enriched in the genomic vicinity of these genes. Our work provides mechanistic insight into how the circadian clock and glucocorticoids cooperate in the regulation of daily patterns of gene expression and metabolism. Furthermore, it highlights metabolic pathways that may contribute to disease mechanisms in patients with a disturbed glucocorticoid hormone system.
| The circadian clock is an endogenous oscillator that regulates daily changes of behavior, physiology and metabolism [1]. The molecular basis of the circadian clock is a transcriptional-translational feedback loop, a central part of which are E-box enhancer elements [2]. To generate physiological rhythms, “peripheral” clocks in almost all tissues interact with signals produced by a “central” pacemaker, the hypothalamic suprachiasmatic nucleus. A key target of circadian clock control is metabolism, with circadian rhythms present in many metabolites and enzyme activities [3]. In addition, hormones with metabolic functions are regulated by the circadian clock. This includes glucocorticoids (GCs), steroid hormones mainly produced by the adrenal gland [4]. GC production shows higher basal levels in the morning in humans and at night in rodents. GCs were also shown to interact with clock factors in the transcriptional regulation of metabolic gene expression [5]. However, the global role of the interaction between the circadian clock and GCs in the regulation of physiology and metabolism and its underlying mechanisms are only incompletely understood.
Patients suffering from adrenal insufficiency (AI) have inadequate GC amounts either because of defects in the adrenal gland itself (primary AI) or due to deficient input from the pituitary gland (secondary AI) [6]. Patients with secondary AI have an increased risk to develop metabolic syndrome, and abnormal glucose tolerance is observed upon long-term therapy with current GC replacement regimes. This may be linked to an inadequate replication of the natural circadian GC rhythm [7]. Only limited information is available on the metabolic changes present in patients with AI [8,9], particularly with respect to their temporal dynamics. Animal models, preferentially with a diurnal lifestyle, could contribute to improve therapy by providing a mechanistic understanding of metabolic dysregulation in AI.
The zebrafish is a well-established model system for human disease including metabolic diseases [10] and has proven useful for chronobiology and endocrinology studies [11,12]. Zebrafish embryos and larvae are well-suited for in vivo bioimaging and drug screenings [13]. Similar to humans, zebrafish are diurnal and use cortisol as the main GC hormone, whereas laboratory rodents are nocturnal and use corticosterone. We previously described a mutation that leads to GC deficiency in homozygous larvae. rx3 mutants of both a weak and a strong allele show a severe eye defect [14]. The strong allele additionally presents a severe reduction of ACTH producing corticotrope pituitary cells, leading to reduced cortisol amounts which also lack a diurnal rhythm (Fig 1A)[15]. Thus, this mutant condition resembles secondary AI. Intriguingly, a clock output rhythm, the circadian fluctuation of cell proliferation, is attenuated in the mutant larvae. These rhythms can be rescued by constant treatment with the synthetic GC, dexamethasone (DEX) [15], and are thus not dependent on the diurnal glucocorticoid release pattern. It is currently not understood how a constant GC signal integrates with circadian clock function to generate such GC-dependent clock output rhythms.
Here, we examined diurnal changes in the transcriptome and metabolism of rx3 strong mutants with or without continuous DEX exposure. A surprisingly large part of diurnal gene expression is rescued by this constant DEX treatment, which also relieves specific metabolite changes in the mutants. Analysis of gene regulation revealed a combined simple enhancer element that is sufficient to mediate diurnal GC-dependent transcription. Besides providing mechanistic insight on GC-circadian clock crosstalk, our study reveals widespread changes in metabolism in an animal model of GC deficiency. These findings will help to better understand morbidities in patients with AI and identify metabolic pathways that could be used for monitoring of therapy efficiency.
To examine if and how diurnal patterns of transcription are perturbed in the GC deficiency model, we measured diurnal transcriptome changes in rx3 strong mutant zebrafish larvae and their wild-type siblings (Fig 1A) at four time points over 24 h (Fig 1B; S1 Table, for quality control and validation experiment results see S1A–S1C Fig). To control for eye absence in the strong allele, we included the equally eyeless rx3 weak mutant larvae, which have normal diurnal cortisol levels. Statistical analysis based on harmonic linear regression [16] grouped genes into models according to their rhythmic or non-rhythmic expression behaviour under the three conditions (Fig 1C–1E, S2A Fig, S1 Table; see Materials and Methods for details). Genes classified into model 1 did not exhibit rhythmicity under any condition, while genes grouped within the other 14 models showed rhythmic expression in at least one condition. Genes which did not fulfill our statistical cutoff criteria to fit within these models were named “ambiguous”. Several models were of particular interest for our aim to identify genes with GC dependent diurnal patterns of transcription. Model 11 was rhythmic in all conditions with the same rhythmic parameters (Fig 1C and 1E); we will refer to this group of genes as “unaffected”. By contrast, models 5, 6, 14 and 15 showed either a lack of rhythmic expression (model 5, 6) or a change in amplitude or phase (model 14, 15) in the strong allele (Fig 1D and 1E). For these four models, global gene expression amplitudes are not significantly different between wild-type and rx3 weak allele mutants, while they are significantly reduced (model 14 and 15) or absent (model 5 and 6) compared with the wild-type in strong allele mutants (S2D–S2G Fig). Furthermore, global comparison of phases shows that they are more perturbed when comparing wild-type and strong allele mutants than when comparing wild-type and weak allele mutants (S2H–S2K Fig). We will refer to the genes of models 5, 6, 14 and 15 as “affected”. These four models define a set of 5970 genes that are candidates for mediating GC-dependent circadian functions. Representing 43.6% of all genes showing rhythmic expression and 23.4% of all detected genes, they constitute a surprisingly large category of the diurnal transcriptome (Fig 1F). Gene Ontology (GO) analysis showed enrichment in GO terms for metabolic processes in this set (S1E Fig). Indeed, 39.9% of the temporal profiles of metabolic genes are assigned to model 5-6-14-15 genes (Fig 1F). By contrast, circadian clock and cell cycle genes are less affected in their temporal expression in rx3 strong mutants. 75% of all circadian clock genes are not altered in their rhythmicity and belong to model 11 (Fig 1F, S3A Fig). There is also enrichment for model 11 within the group of cell cycle genes (23.8%, Fig 1F, S3B Fig). Fitting this observation, model 11 genes are enriched for GO terms related to the cell cycle (S1D Fig). These findings show that many cell cycle related genes have a rhythmic expression pattern which does not change in the rx3 strong mutants. Still, 25.4% of cell cycle genes belong to model 5-6-14-15 (Fig 1F). Interestingly, there is no statistically significant alteration in oscillation amplitude between weak and strong allele mutants across both all cell cycle genes and all circadian clock genes (Fig 1G). However, the amplitudes of oscillation are significantly reduced across the metabolic genes in rx3 strong mutants (Fig 1G), further indicating a higher degree of attenuated rhythm within this group. The “affected” set of genes also encompasses a larger number of enriched metabolic pathways than the “unaffected” set (compare S1F Fig with S1G Fig), again underlining the strong effect present in this group on the diurnal expression of metabolic genes.
Next, we asked whether and to what extent affected gene expression rhythms can be restored in the mutants by constant GC treatment. To determine how the diurnal transcriptome changes under chronic DEX treatment (Fig 2A), we carried out RNA-Seq analysis of treated rx3 strong mutants and wild-type siblings. To evaluate whether the treatment leads to a rescue of rhythmic expression in the mutants, we analyzed the two treated conditions for rhythmic parameters of gene expression as done for the untreated samples. This allowed us to evaluate if differences in expression between the genotypes were abolished by the treatment, even if other general DEX effects on transcription affected both conditions similarly. Our statistical analysis yielded five models (S2B Fig), of which two (D and E) exhibit rhythmicity in both wild-types and mutants. Therefore, these models could indicate a rescue of dysregulated patterns among the “affected” gene group (Fig 2B and 2C; S4 Fig). Model D regroups those genes in which there is no difference in rhythmic expression between mutants and wild-type under DEX treatment. Model E indicates those genes in which a difference in phase or amplitude between mutant and wild-type is still present under DEX treatment. We classified as rescued all genes which were not rhythmic in the mutants before (model 5 and 6) and in which rhythms have been restored (model D or E), or genes of models 14 and 15 in which a phase or relative amplitude difference in the untreated condition has been abolished (model D) or reduced (model E). By contrast, we did not count as rescued all those genes of model E in which phase or amplitude differences were not reduced by the treatment (named model E*).
Applying these rescue criteria, 46% of the model 5-6-14-15 genes were rescued (Fig 2D, “all affected genes”). Strikingly, 58% of the metabolic genes of model 5-6-14-15 are rescued (Fig 2D). Interestingly, even though 68% of affected cell cycle genes do not show convergence of mutant and wild-type expression patterns in the DEX treated condition (Fig 2D, S3B Fig), this treatment restores cell cycle rhythms in the rx3 strong mutants [15]. In summary, chronic DEX treatment is able to restore rhythmic expression patterns matching the wild-type in nearly half of all rhythmic genes affected in the mutants and in about 60% of the metabolic genes. This is a striking finding, showing that a large proportion of affected genes do not require rhythmic GC input for the GC-dependent regulation of their rhythmic transcription.
Our RNA-Seq analysis revealed a high proportion of metabolism-related genes among the set with GC-dependent diurnal transcription. Therefore, to determine metabolite changes, we examined extracts from rx3 strong and rx3 weak mutant zebrafish larvae and their wild-type siblings at five time points over 24 h (Fig 2A) by NMR spectroscopy. We recorded 1D spectra for quantitation and additionally 2D J-resolved spectra for unambiguous identification of compounds. Principal component analysis (PCA) of the NMR spectra (Fig 3A) shows that rx3 weak and wild-type samples cluster together, illustrating that the metabolomes of rx3 weak mutants and wild-type larvae are more similar to each other than to the strong allele samples. Under DEX treatment, the wild-type and rx3 strong mutants cluster much closer together than the control samples (Fig 3B). Betaine, creatine, lactate and glutamine appear as major contributors to the main principal components (S5A and S5B Fig). Indeed, glutamine showed a strong accumulation in rx3 strong larvae at all examined time points, which was rescued by DEX treatment (S5C Fig).
Glutamine plays an important role in amino acid and central carbon metabolism, pathways which were also found to be enriched among the rescued gene set (Fig 2E). Therefore, we measured a set of amino acids and TCA cycle related metabolites by UPLC-FLR and IC-CD analysis, and evaluated rhythmicity of these data using the harmonic linear regression based model selection approach (S2C Fig, S2 Table). Rhythmicity behaviour and changes in mean levels of the metabolites indicated three groups of particular interest.
In the first group, which includes branched chain amino acids (BCAAs) and aromatic amino acids (AAAs), there is an accumulation of compounds in rx3 strong mutants which is not rescued by DEX treatment (Fig 3C, S5D–S5G Fig and S2 Table). Generally, compound levels in the wild-type show a stronger overall decrease over time than in the mutants. Also, rhythmicity behaviour in this group is not dramatically affected by GCs (models II [Tyr] and VI [Val, Leu, Ile, Phe]). These findings only partly correlate with gene expression pattern changes in the first degradation steps of the corresponding pathways. In the AAA pathway, half of the affected genes are rescued (S6A Fig), and in the BCAA pathway, all three key enzymes are rescued (S6B Fig). Here, metabolite accumulation seems to reflect other processes that are independent of GC regulation or not rescued by constant DEX application. Such processes may include posttranscriptional and -translational regulation of these enzymes, or BCAA accumulation due to increased degradation of BCAA containing proteins.
The second group of metabolites equally accumulates in the mutants, but here their levels are rescued by DEX treatment. Glyoxylate, lysine and the ornithine-urea cycle (OUC) compounds ornithine, citrulline and arginine belong to this group (Fig 3D, S5H–S5K Fig and S2 Table). Rhythmicity behavior varies in this group (models I [glyoxylate], III [Cit], VII [Arg] or ambiguous [Lys, Orn]). Patterns of gene expression in the corresponding pathways seem to be more closely correlated with metabolite changes than in the first group: glyoxylate metabolizing enzymes are downregulated, while those producing glyoxylate are upregulated in rx3 strong mutants, and changes in both pathways are rescued by DEX treatment (S6C Fig). A similar behavior is seen in the OUC pathway (S6D Fig).
The last group contains only one compound, glutamine. Here, accumulation in rx3 strong is rescued by DEX treatment, as in the second group. Additionally, glutamine shows a diurnal rhythm in the wild-type which is slightly flattened and shifted in the rx3 strong mutant (log2 amplitude 0.40 and peak at ZT22.2 compared with 0.47 and ZT17.5 in wild-type). Under DEX treatment, both mutant and wild-type exhibit strong rhythmic glutamine concentrations and the phase difference is reduced (model IX, Fig 3E, S5C Fig, S2 Table). Remarkably, glutamine is the only compound of the set which shows such a rescue of both overall levels and circadian rhythmicity by DEX treatment.
Glutamine forms part of several pathways enriched in the rescued gene set. For example, it is a required source of nitrogen for purine and pyrimidine synthesis. Interestingly, an entire chain of enzymes downstream of glutamine entry into the purine biosynthesis pathway shows dysregulation in rx3 strong mutants, which is rescued by DEX (S6E Fig). These enzymes act upstream of IMP (inosine 5'-monophosphate) dehydrogenase 2 (impdh2), which has recently been suggested to be involved in the regulation of circadian rhythms of cell proliferation [18]. impdh2 expression is also dysregulated in the mutants and rescued by DEX. As genes in many other branches of the purine synthesis pathway are equally rescued (S1 and S3 Tables), GCs seem to regulate a large part of the diurnal transcription within this pathway.
Glutamine is also important for refilling (anaplerosis) of the TCA cycle with α-ketoglutarate when it is deprived of intermediates (Fig 4A). We chose this glutamine-TCA cycle connection for a proof-of-principle analysis. Examination of the cumulated levels of six TCA intermediates shows that citrate levels are higher in the mutants, while succinate levels are reduced (Fig 4B). This finding is consistent with reduced anaplerosis at the level of α-ketoglutarate, leading to an upstream block and downstream depletion of cycle intermediates. In DEX treated conditions the levels are normalized, indicating restored flow. The TCA cycle connects glutaminolysis, glycolysis and gluconeogenesis, and indeed, many TCA cycle and glycolysis enzymes as well as key enzymes of gluconeogenesis show dysregulated expression rescued by DEX treatment (Fig 4A, 4C and 4D). Among them, phosphoenolpyruvate carboxykinase 1 (soluble, pck1) removes OAA from the TCA cycle (cataplerosis) and channels it into gluconeogenesis. It has been suggested that this cataplerotic PCK1 function balances anaplerotic refilling of the TCA cycle by glutamine metabolism [19]. Expression of dysregulated genes in this anaplerotic pathway, such as glutaminase 2 (gls2), is also restored by chronic DEX treatment (Fig 4A, 4C and 4E). Therefore, DEX treatment likely restores the cataplerosis/anaplerosis balance of the TCA cycle, thereby normalizing TCA compound and glutamine levels.
PCK1 has long been described as a GC target gene and was reported to show circadian expression in mammals [5]. Interestingly, there are two Glucocorticoid Response Elements (GREs) and an E-box in the zebrafish pck1 promoter region, which are conserved across evolution (Fig 5A). Our RNA-seq study identified gls2 as another GC inducible circadian gene, and it equally contains both GREs and E-boxes. To test whether this is a typical feature of GC regulated circadian genes, the putative promoter sequences (-1000 +500bp) of the model 5-6-14-15 genes were examined for concomitant presence of both E-box and GRE elements. We observed an enriched co-occurrence of E-boxes and GREs in the rescued genes (hypergeometric test, p = 0.04), while no significant enrichment was seen in the non-rescued ones (p = 0.97; Fig 5B). Importantly, this enrichment is also observed in the mouse orthologues of the zebrafish genes, indicating evolutionary conservation of this regulatory module. These findings indicate that E-box/GRE modules, which allow for direct transcriptional regulation by both the circadian clock and GCs, are a characteristic feature of GC-dependent diurnal genes.
Can a simple combination of GREs and E-box enhancer elements drive GC dependent circadian patterns of gene expression? Exploiting the direct light sensitivity of zebrafish cells [20], we transfected cells with luciferase reporter constructs driven by different combinations of E-boxes and GREs and exposed them to light-dark cycles at different concentrations of DEX in otherwise GC-depleted culture medium (Fig 5C–5H). Of all combinations, only the E-box/GRE combination (Fig 5G) drove GC dependent circadian transcription, while the other constructs were not sensitive to DEX (1x E-box, 4x E-box) or did not show rhythmic expression (1x E-box, 1x GRE, 4x GRE). Thus, a synergistic interaction between a single E-box and a GRE is sufficient to drive GC-dependent circadian gene expression.
To begin to explore mechanistic aspects of this interaction between E-Boxes and GREs, we examined the bioluminescence patterns driven by the combined E-box/GRE element upon entrainment by different light intensities of the light part of the LD cycle and upon changing distances between the two elements. Interestingly, at the highest light intensity examined, oscillation behaviour was less pronounced (S7A and S7B Fig). Two-way ANOVA analysis indicated that there was no interaction between GC dose (10–60 nM) and light intensity effects (150–1500 lux, S7B Fig). Examination of constructs with different E-box/GRE spacings revealed a slightly less robust oscillation behaviour when the two elements were separated by 10 bps, as in the original construct, than at the other distances (5, 15 and 20 bp; S7C and S7D Fig). These results indicate that the E-box/GRE module is sensitive to light conditions and that interaction between the two elements might be hampered when they are placed relatively nearby on the same side of the DNA helix.
To examine whether the interaction between E-boxes and GREs also occurs in vivo, we generated larvae carrying stable genomic insertions of the 1xE-box/GRE reporter. Both in wild-type siblings and in rx3 strong mutants, the construct drove rhythmic expression (Fig 5I), indicating that the low levels of cortisol present in the mutants [15] are still sufficient for activation of the construct. Importantly however, the amplitude of expression was clearly reduced in the mutants (Fig 5J). Strikingly, the amplitude difference between mutants and wild-type siblings disappeared when the larvae were treated with DEX during the experiment (Fig 5I and 5J). Thus, E-box/GRE driven bioluminescence rhythms mimic the rhythmicity behavior of many metabolic genes from the RNA-Seq analysis (Fig 1F), suggesting that the E-box-GRE combinations enriched in the GC-dependent diurnal gene set mediate highly efficient regulatory inputs for this expression pattern.
Our data reveal a strong impact of GCs on diurnal gene transcription and key metabolite levels. About 44% of all genes with diurnal expression patterns are dysregulated in GC deficient larvae, and almost half of these can be rescued by constant chronic DEX treatment. This result shows that a surprisingly large part of the GC-dependent diurnal transcriptome does not depend on the diurnal pattern of GC levels themselves. Thus, GCs may have a permissive rather than instructive role in the diurnal expression of these genes. This mechanism of regulation may prevent these genes from being inappropriately regulated by irregular GC rises, such as those during acute stress responses. Recent studies have described direct interactions of circadian clock factors with GRs, leading to circadian modulation of GC signaling [21,22,23]. For example, circadian Pck1 expression was attributed to direct inhibitory regulation of the GR by CRY binding [22]. Interestingly, the Pck1 promoter also contains an E-box element which is conserved between fish and mammals. Our in vivo reporter gene analysis shows that a simple E-box-GRE combination is sufficient to mediate GC dependent circadian luciferase expression. However, we did not observe circadian regulation of reporter expression driven by four concatemerized GREs. Endogenous Cry levels in zebrafish fibroblasts may not be high enough to mediate transcriptional repression of the 4xGRE reporter in vivo, in contrast to the elevated CRY levels upon overexpression used in mouse fibroblasts studies reported previously [22]. Remarkably, GR binding to the GRE in the mouse Pck1 gene is highest at the time when CRYs reach peak levels and reduced in CRY double knock-out mice [22]. Stable CRY-GR interactions may occur at the promoter itself, where the Clock-Bmal1 bound to the E-box may increase the local concentration of CRY proteins and thereby facilitate their interactions with GRs binding a neighbouring GRE. The increased co-occurrence of E-boxes and GREs in the genes rescued by constant DEX treatment argues in favor of such a mechanism. Of note, zebrafish Cry1a has been implicated in mediating a negative regulatory influence of light on the circadian clock by its interaction with the Clock-Bmal dimer [24], a function that may underlie our observation of reduced oscillations upon exposure to light-dark cycles with higher light intensity. Furthermore, our data indicate that interactions between the GR and the clock machinery may be hampered when the two elements are located nearby on the same side of the DNA helix, as elements separated by 10 bp show slightly less robust oscillations than those separated by 5 bp or 15 bp. Further studies targeted at identifying the full set of factors involved in regulation of transcription by the variants of the element will likely reveal more mechanistic details of the interactions between GCs and the clock in transcriptional regulation.
While the constant DEX treatment clearly influenced gene expression and metabolite patterns, it did not lead to any visible perturbation of the larval phenotype, nor did it affect circadian rhythms of S-phase in the wild-type [15]. Effects of prolonged DEX treatment may appear only at later developmental stages or in adulthood, and only be noticeable upon challenges to the organism. This is seen in humans or other mammals exposed to high DEX levels during development, which later in life show increased disease susceptibility and altered responses to stress [25,26]. The DEX-treatment related changes to metabolites or gene expression patterns revealed within our data set may provide interesting starting points for understanding the mechanistic principles of such long-term effects.
Constant DEX treatment rescues circadian rhythms of S-phase in rx3 strong mutants [15]. Our transcriptome analysis reveals that a number of cell cycle genes, related to all phases of the cell cycle, show a similar pattern, as do numerous genes acting in metabolic pathways implicated in the regulation of cell proliferation. Surprisingly, fewer cell cycle genes than metabolic genes were affected by the loss of GCs. Also, more of the affected metabolic genes were rescued by constant DEX treatment. This indicates that GC dependent metabolic control may play a more important role in circadian cell cycle rhythms than GC regulation of cell cycle genes. Strikingly, of the metabolites examined, only glutamine showed a restoration of both overall levels and rhythmicity by constant DEX treatment. Glutamine plays a major role in cell proliferation related pathways such as purine synthesis or anaplerosis of the TCA cycle. Thus, it emerges as a potential key metabolite in the circadian orchestration of cell proliferation. The presence of a circadian rhythm of glutamine levels in human blood [27] and in rat liver [28] suggests that such functions are conserved across evolution.
Intriguingly, we observed a strong accumulation of glyoxylate in rx3 strong mutants, which was prevented by DEX treatment. This glyoxylate accumulation may reflect the bypassing of part of the disturbed TCA cycle by the so-called glyoxylate cycle. However, apart from C. elegans, the presence of this cycle in metazoans is controversial, and orthologues of the glyoxylate cycle enzymes isocitrate dehydrogenase and malate synthase are reported to be absent or pseudogenes in vertebrates [29]. Interestingly, the zebrafish genome contains a potentially functional malate synthase-like sequence (ENSDARG00000074684), which showed DEX rescue of a dysregulated expression pattern in rx3 strong mutants (S1 Table). Also peroxisomal and mitochondrial glyoxylate detoxification is dysregulated in the mutants, involving decreases in expression of glyoxylate metabolizing enzymes and increases in glyoxylate producing enzymes. Importantly, three of the dysregulated genes encode enzymes linked with human genetic disorders of glyoxylate detoxification (alanin-glyoxylate amino transferase [agxtb], glyoxylate reductase/hydroxypyruvate reductase [grhprb] and 4-hydroxy-2-oxoglutarate aldolase 1 [hoga1], which are involved in human primary hyperoxalurias type I-III, respectively [17]). With one exception (grhprb), constant DEX rescues expression patterns of all these genes. Thus, GCs may be a promising candidate for treatment of patients in which insufficient amounts of functional protein are produced [17,30].
Another intriguing finding is our observation that the aberrant accumulation of both BCAA and aromatic AA levels in the mutants could not be restored by DEX treatment. Dysregulation of these metabolites may reflect a perturbation of other regulatory inputs in addition to GCs in the secondary AI model. Alternatively, our temporally constant GC replacement by DEX may not be sufficient for proper functional restoration of all pathways influencing BCAA and AAA levels. It will be interesting to explore if other treatment schemes are more efficient in restoring BCAA and AAA levels, and whether these amino acids also show disturbed regulation in human patients.
In summary, our study establishes zebrafish larvae as an easily accessible model for studies targeting metabolic aspects of GC related disease and GC therapy. In addition, our work reveals a massive impact of GCs on the diurnal patterns of gene transcription. Surprisingly, a large part of this diurnal regulation does not require changing levels of GCs themselves, and we provide a model based on simple enhancer element interactions that can explain this behavior.
Animal experiments were conducted in accordance with German animal protection regulations and approved by local regulatory authorities (Regierungspräsidium Karlsruhe, approval number Aktenzeichen 35–9185.81/G-83/14 and 35–9185.81/G-242/15).
Fish (AB wild-type and the mutant lines rx3t25327/t25327 [rx3 strong] and rx3t25181/t25181 [rx3 weak] [14]) were raised and bred as described [31].
Total RNA extraction, cDNA synthesis and qPCR were carried out as described [32]. Primer sequences used were: β-actin: fw: 5’-gcctgacggacaggtcat-3, rv:’ 5’-accgcaagattccataccc-3’; apoa1: fw: 5’-cttgacaacctggacggaac-3, rv: 5’-gcatattcctggagcttggt-3; arntl1a: fw: 5’-tagagcgctgtttgctgatg-3’, rv:5’-gacccgtggacttcagtgac-3; cyp3a: fw: 5’-ccaaagacaacacgaagcag-3’, rv: 5’-acaagatctcgtggtcactcag-3’; rbp4: fw: 5’-ccgaagatccagctaagttca-3, rv:’ 5’-caatgatccagtggtcgtca-3’; si:dkey-18a10.3: fw: 5’-ctttgtgcgccaactcaac-3’, rv: 5’-tttaggcaagccggagtcta-3’; pck1: fw: 5’-tgacgtcctggaagaacca-3’, rv: 5’-gcgtacagaagcgggagtt-3’; gls2a: fw:5’-gacatgacagcagctcttgact-3’, rv: 5’-tgcctgactcacatgtcacc-3’.
Synthesis of probes against lhx4, apoa1, and zgc:158293 and whole-mount in situ hybridization was carried out following Armant et al. [33].
1 dpf rx3 mutant and wildtype sibling embryos from pooled clutches were separated and transferred into cell culture flasks containing E3 medium (5 mM NaCl, 0.17 mM KCl, 0.33 mM CaCl2, 0.33 mM MgSO4, 0.1% methylene blue). For GC treatment, the medium was supplemented with either 25 μM Dexamethasone (DEX) solved in 0.1% DMSO or 0.1% DMSO alone as a control. Embryos/larvae were kept at 28°C in an incubator under a 12 h light:12 h dark (LD) exposure. Larvae were sampled in liquid nitrogen at the indicated Zeitgeber times (ZT, ZT3 = 3 hours after lights on) starting at 5 dpf and stored at -80°C until further processing.
To ensure efficient and reproducible sample preparation for NMR metabolomics of zebrafish larvae, we have established an extraction protocol that generates highly reproducible data. 50 larvae per sample were collected in homogenization tubes (PeqLab, #91-PCS-CK14), snap-frozen in liquid nitrogen and lyophilized overnight to avoid recovery of enzymatic activity. Lyophilized samples were stored at -80°C for a maximum of 2 days. For extraction, 1 ml of acetonitrile/water (1:1) and ceramic beads (#91-PCS-CK14, PeqLab) were added to the larvae and extracted with a liquid nitrogen cooled cell shaker (Precellys 24, PeqLab) according to the manufacturer’s instructions, with the following settings: 6,000 rpm, 4x20 s, 120 s. Homogenates were vortexed for 1 min and incubated on ice for 10 min. Next, the samples were transferred into fresh vials, and were briefly centrifuged at 4°C to remove debris. 750 μl of supernatant were transferred into fresh vials, and 620 μl of ultrapure water (HPLC grade) were added. Samples were vortexed before lyophilization overnight. For measurements, 650 μL of D2O/buffer (1.5 M KH2PO4, 2 mM NaN3, 0.1% (v/v) TSP (= 3,3-(trimethylsilyl)-2,2',3,3'-tetradeuteropropionic acid) in D2O) (9:1) were added to the extracts, and 600 μl of this mixture were transferred into a 5 mm standard NMR tube. Spectra were recorded on a Bruker Avance III 600 spectrometer equipped with a 1H,13C,15N-TCI triple resonance cryoprobe. 1D spectra were recorded with 64k data points, 90.5 receiver gain and 32 scans for comparing the phenotypes and 64 scans for comparing the DEX treatment at 300 K using a 1D NOESY experiment with presaturation for water suppression. A mixing time of 10 ms and a prescan delay of 10 s were used. Pulse length was determined automatically by the Bruker AU program pulsecal and presaturation was set corresponding to a 25 Hz pulse. Irradiation frequency for water suppression was optimized prior to acquisition. Spectra were processed identically with an exponential apodization function with line broadening of 0.3 Hz. Automatic phasing, baseline correction and referencing were done by the Bruker AU program apk0.noe. Additionally, J-resolved spectra were recorded with identical pulse lengths, presaturation, irradiation frequency and receiver gain as the corresponding 1D NOESY. Spectra were recorded with 8192 data points in the direct dimension and 40 increments in the indirect dimension. One scan was acquired per increment. Spectra were processed with 16k × 256 data points and a sine window function in both dimensions.
We employed ultra-performance liquid chromatography with fluorescence detection (UPLC-FLR) for targeted quantification of amino acids and ketoacids (S8 Fig) and ion chromatography with conductivity detection (IC-CD) for quantification of other organic acids. 30 larvae per sample (in triplicates) were collected for absolute quantification of amino acids and α-ketoacids (glyoxylate) and for organic acid content each. For extraction of free amino acids and α-ketoacids 300μl 0.1M HCl was used. Derivatization and separation of amino acids was performed as described by Yang et al. [34]. For derivatization of α-ketoacids 150 μl of the acidic extract was mixed with an equal volume of 25 mM OPD (o-phenylendiamine) solution and incubated at 50°C for 30 min. The derivatized α-ketoacids were separated using an Acquity HSS T3 column (100 mm x 2.1 mm, 1.7 μm, Waters) connected to an Acquity H-class UPLC system. Prior separation, the column was heated to 40°C and equilibrated with solvent A (0.1% formic acid in 10% acetonitrile) at a flow rate of 0.55 ml/min. Separation of α-ketoacid derivates was achieved by increasing the concentration of solvent B (acetonitrile) in solvent A as follows: 2 min 2% B, 5 min 18% B, 5.2 min 22% B, 9 min 40% B, 9.1min 80% B and hold for 2min, and return to 2% B in 2 min. The separated derivates were detected by fluorescence (Acquity FLR detector, Waters, excitation: 350 nm, emission: 410 nm) and quantified using ultrapure standards (Sigma). For quantification, a linear seven-point calibration curve ranging from 0.3–15 pmol on column was used (R2 > 0.99). Data acquisition and processing was performed with the Empower3 software suite (Waters). Organic acids were extracted with 700 μl ultra-pure water for 20 min at 95°C. These compounds were separated using an IonPac AS11-HC (2mm, ThermoScientific) column connected to an ICS-3000 system (Dionex) and quantified by conductivity detection after cation suppression (ASRS-300 2mm, suppressor current 95–120 mA). Prior separation, the column was heated to 30°C and equilibrated with 5 column volumes of solvent A (ultra-pure water) at a flow rate of 0.38 ml/min. Separation of anions and organic acids was achieved by increasing the concentration of solvent B (methanol) and solvent C (100mM NaOH) in buffer A as follows: 8 min 4% C, 11 min 10% C, 18.2 min 20% B / 18.1% C, 27.5 min 20% B / 21% C, 32 min 24% C, 43 min 30% C, 47 min 40% C, 48 min 90% C for 8 min, and return to 4% C in 9 min. A linear three-point calibration curve was used for quantification of organic acids (0.5–5nmol on column; R2 > 0.99) Data acquisition and processing was performed with the Chromeleon 6.7 software (Dionex). Rhythmic properties of metabolite levels were assessed with the model selection based method described in Atger et al. [16] (see also below, Rhythmicity assessment in different genotypic backgrounds).
Library preparation and sequencing were performed by the IGBMC Microarray and Sequencing Platform and by the Next Generation Sequencing and Genomics facility of the BioInterfaces research programme at KIT. RNA integrity numbers measured with a 2100 Bioanalyzer (Agilent) were 10 for all samples. cDNA libraries were generated using the directional mRNA-seq sample preparation kit (#15018460, Rev.A, October 2010, Illumina). Single-end 54 nt reads were obtained with a Genome Analyzer IIx.
Zebrafish cells were maintained as described [44] in Leibowitz’s (L-15) medium (Life Technologies, #11415–049) supplemented with antibiotics and 15% (PAC2 cells) or 17% (AB.9 GRE:Luc cells) Fetal bovine serum (%[v/v], FBS, Biochrom AG, #S0115). Bioluminescence studies for cells and larvae were carried out as reported previously [32,44,45].
In case of multiple comparisons, p-values were adjusted using the Benjamini-Hochberg method [40].
|
10.1371/journal.pcbi.0030156 | Mechanisms of Firing Patterns in Fast-Spiking Cortical Interneurons | Cortical fast-spiking (FS) interneurons display highly variable electrophysiological properties. Their spike responses to step currents occur almost immediately following the step onset or after a substantial delay, during which subthreshold oscillations are frequently observed. Their firing patterns include high-frequency tonic firing and rhythmic or irregular bursting (stuttering). What is the origin of this variability? In the present paper, we hypothesize that it emerges naturally if one assumes a continuous distribution of properties in a small set of active channels. To test this hypothesis, we construct a minimal, single-compartment conductance-based model of FS cells that includes transient Na+, delayed-rectifier K+, and slowly inactivating d-type K+ conductances. The model is analyzed using nonlinear dynamical system theory. For small Na+ window current, the neuron exhibits high-frequency tonic firing. At current threshold, the spike response is almost instantaneous for small d-current conductance, gd, and it is delayed for larger gd. As gd further increases, the neuron stutters. Noise substantially reduces the delay duration and induces subthreshold oscillations. In contrast, when the Na+ window current is large, the neuron always fires tonically. Near threshold, the firing rates are low, and the delay to firing is only weakly sensitive to noise; subthreshold oscillations are not observed. We propose that the variability in the response of cortical FS neurons is a consequence of heterogeneities in their gd and in the strength of their Na+ window current. We predict the existence of two types of firing patterns in FS neurons, differing in the sensitivity of the delay duration to noise, in the minimal firing rate of the tonic discharge, and in the existence of subthreshold oscillations. We report experimental results from intracellular recordings supporting this prediction.
| About 25% of the neurons in the mammalian neocortex are inhibitory, namely reduce the activity of neurons they contact. These inhibitory neurons exhibit diversity of morphological, chemical, and biophysical properties, and their classification has recently been the focus of much debate. Even neurons belonging to a single class of “fast-spiking” (FS) display a large variety of firing patterns in response to standard square current pulses. Previous works proposed that this class is in fact a discrete set of neuronal subtypes with biophysical properties differing in a discontinuous way. In this work, we propose an alternative theory, according to which the biophysical properties of FS neurons are continuously distributed, but distinct firing patterns emerge due to highly nonlinear dynamics of these neurons. We ascertain this theory by exploring with mathematical techniques a biophysically based model of FS neurons. We demonstrate that variable firing responses of cortical FS neurons can be accounted for if one assumes heterogeneity in the strength of some of the ionic conductances underlying neuronal activity. Our theory predicts the existence of two main firing patterns of FS neurons. This prediction is verified by direct recordings in cortical slices.
| Among inhibitory neurons in the neocortex, the “fast-spiking” (FS) compose the most prominent type. These neurons are characterized by brief action potentials with a width smaller than 0.5 ms followed by a deep monophasic afterhyperpolarization (AHP) [1,2]. Delayed rectifier currents of the types Kv3.1–Kv3.2 are responsible for these characteristics of FS action potentials [3,4]. The firing patterns of FS cells in response to a step of injected current are highly variable. Depending on the neuron and on the amplitude of the current pulse, FS cells fire action potentials immediately after the onset of the current step or after a prolonged delay, which can be on the order of several hundreds of milliseconds [2]. Interestingly, voltage-dependent subthreshold oscillations in the gamma range have been reported during the delay period. They typically occur in a narrow voltage range just negative to threshold [4,5]. The steady-state firing pattern is reached after an adapting [6], non-adapting, or accelerating [2] transient. This steady state can be tonic or bursting. In the latter case, the neuron fires rhythmic irregular bursts of action potentials; this activity pattern is named “stuttering” [7]. The goal of the present modeling study is to address the origin of this variability.
Bifurcation theory, which classifies how the behavior of dynamical systems changes as their parameters vary, reveals that in strongly nonlinear systems, qualitatively different dynamical regimes can emerge as a result of a continuous variation of some parameters [8]. Hence, heterogeneities in biophysical parameters of neurons can induce distinct classes of firing patterns even if their distributions are smooth. In the present paper, we propose that a great deal of FS electrophysiological variability is a consequence of heterogeneities in the maximal conductance of a slowly-inactivating d-type K+ current [9,10], known to be present in FS cells [11,12], and in the strength of the Na+ window current. This window current, governed by the overlap between the activation and inactivation curves of the Na+ current, affects the ability of the neuron to fire at low rates [13]. To assess this proposal, we consider a minimal, conductance-based neuronal model that incorporates these two ionic currents and a fast delayed rectifier K+ current. We investigate this model using techniques from nonlinear dynamical system theory. We find that as the conductance of the d-current and the overlap of the activation and the inactivation curves of the Na+ current are varied within a range compatible with experimental data [14], the model neuron displays a “”variability of firing patterns similar to those observed experimentally in FS cells.
Our study leads to several predictions that can be tested experimentally. In particular, we predict that FS cells that fire tonically at low rates do not exhibit subthreshold oscillations during the delay period. We present experimental results consistent with this prediction.
We study a minimal, single-compartment conductance-based model that incorporates a Na+ current, INa, a fast delayed rectifier K+ current of the Kv3.1–Kv3.2 type, IKdr, and a d-type K+ current, Id. The details of the model are given in the Model section of Materials and Methods.
In all of this section, we assume that the system is noiseless. We will change the Na+ window current by varying the half-maximum potential, θm, of the activation curve of INa. As shown in Figure 1, the amplitude of the window current decreases as θm is more depolarized. The strength of the d-current is controlled by its maximal conductance gd. The amplitude of the current step injected into the neuron is denoted by Iapp.
For very large window current, i.e., small θm (θm < −31.4 mV for gd = 0 and θm < −32.9 mV for gd = 2 mS/cm2), the neuron is spontaneously active. In contrast, for very small window current, i.e., sufficiently large θm (θm > −15.2 mV for gd = 0 and θm > −16.4 mV for gd = 2 mS/cm2), the neuron remains quiescent for all amplitudes of the step current. In the intermediate range of θm, the neuron is quiescent for Iapp smaller than a threshold Ith, while it fires spikes for Iapp > Ith. Then, depending on θm, two qualitatively different behaviors of the neuron occur.
In this section, we clarify the mechanisms underlying the different firing patterns described above and the dependency of the phase diagram on the Na+ window current. The time constant of the inactivation variable b of the d-current is τb = 150 ms (see the Model section in Materials and Methods). Hence, b varies much more slowly than all the other dynamical variables of the model. The full dynamical system describing the neurons can subsequently be separated into fast (variables V, h, n, and a) and slow (variable b) subsystems. This allows us to dissect the dynamics of our model using the “fast–slow method” [13,16–18]. The first step in this method is to study how the attractors of the dynamics of the fast subsystem depend on the value of b, taken as a time-independent parameter. In a second step, one derives the dynamics of the full system taking into account the slow variations of b.
The minimal firing rates in the tonic discharge following the delay period is broadly distributed among FS neurons [22–24]. The theoretical results described above established correlations between the minimal firing rate of a neuron and the existence of subthreshold oscillations during the delay period. Neurons that fire at high rates display subthreshold oscillations during that period [22,23]. In contrast, neurons that can fire at low rates lack those oscillations. We report here experimental results supporting this prediction.
The responses to steps of depolarizing current pulses were recorded intracellularly in FS neurons (n = 20, average ratio of the spike's rising phase dV/dt to falling phase dV/dt = 1.17 ± 0.29), as described in the section “Whole cell recordings and analysis” in Materials and Methods. Eight neurons responded to the depolarization onset almost instantaneously with a tonic non-adapting spike train. The other 60% of the neurons (n = 12) displayed a prolonged delay period, which was preceded by a transient firing of 1–3 spikes in some cases (Figure 11A and B), similar to the simulations of our model for low Na+ window current (Figure 7C). One cell from this group of 12 cells fired irregularly (Figure 11C). Eight out of those 12 cells had properties as reported in previous experimental studies of FS neurons (see also [22,25]). These eight cells exhibited a delay period followed by a tonic, high-frequency regular discharge (Figure 11A) or stuttering discharge with large, instantaneous intra-burst firing rate (Fig 11B). The minimal average firing frequencies in this group of neurons ranged from 25 to 220 Hz (average 81 ± 59 Hz) and the voltage threshold to action potential was −31.6 ± 5.0 mV. All the computed trial-average power spectra of the membrane fluctuations during the delay period (n = 4) displayed a peak within the gamma range (20–100 Hz). Two examples are shown in Figure 12A and 12B. These features are similar to those exhibited by our model neuron for small window current (Figure 9A and 9B).
Different properties were found in the remaining three neurons (out of these 12 cells) that displayed delay to firing. First, their minimal firing frequencies were under 10 Hz (average 6.8 ± 2.6 Hz) (Figure 11D), substantially lower than in the other group of eight neurons. Their spike threshold was −39.6 ± 6.4 mV, a significantly more hyperpolarized value than in the other group (p = 0.01, student t-test). These neurons did not exhibit stuttering behavior. Finally, spectral analysis of the fluctuations during the delay period failed to reveal subthreshold oscillations (Figure 12C). These properties are as predicted in our model for large Na+ window current (Figure 9C).
The minimal model of FS neurons studied in our work displays four types of behavior in response to a current step, depending on the Na+ window current and on the strength of the conductance of the K+ current Id.
1. When the Na+ window current and the conductance gd are small, the neuron exhibits tonic, high-frequency firing that follows the current step onset almost immediately, even if the step amplitude is just above firing threshold.
2. When the Na+ window current is small and the conductance of the d-current is of intermediate strength, delayed high-frequency tonic firing occurs for just suprathreshold step amplitudes. The delay duration decreases as the step amplitude increases and abruptly jumps to zero at some critical value. Noise dramatically reduces this duration. Noise also induces subthreshold oscillations during the delay and can also induce stuttering.
3. For small Na+ window current but large values of the d-current conductance, the response to just suprathreshold input is delayed stuttering with high-frequency firing within the bursts. As the current step amplitude increases, the response becomes tonic firing, first delayed and subsequently and abruptly non-delayed. Other properties are as in 2.
4. For large value of the Na+ window current, the neuron responds with delayed tonic firing for small amplitude of the current step and non-delayed tonic firing when the amplitude of the current step is large. In contrast to what happens in 2, the average firing rate is low near firing threshold; noise very weakly affects the delay duration and does not induce subthreshold oscillations during the delay period.
A large spectrum of K+ currents with different activation and inactivation properties and kinetics has been reported in FS neurons [7,11,26]. Delayed rectifier K+ channels from the Kv3.1–Kv3.2 types are responsible for the fast spike repolarization and strong AHP of these neurons [3,4,27]. Slowly inactivating K+ channels from the Kv1.1, Kv1.2, and Kv1.6 types have also been found in FS cells [12]; blockade of these currents with DTx-I eliminates delays to firing present in the control situation. The d-current incorporated in our model can be thought of as representing these slow channels.
We are aware of a single experimental study where the activation and inactivation properties of the Na+ channels in FS neurons were measured [14]. One conclusion of that study is that the overlap between the activation and inactivation curves of this current is small. Clearly this does not mean that the Na+ window current has no effect, since this will depend on the maximum conductance of this channel, a value which is not known. As a matter of fact, in our model, the inactivation mid-point potential, and the gain of the activation and inactivation functions are in accordance with the data provided by Martina and Jonas [14]. We take the same value of the Na+ conductance gNa as in [3,28]. With this value, we find variability of firing patterns while varying the activation mid-point potential in a range compatible with the data provided by Martina and Jonas [14].
Throughout this article, we use the half-activation curve of INa, θm, to quantify the strength of the Na+ window current. Effects of modifying the window current by depolarizing or hyperpolarizing θh are similar to the effects of hyperpolarizing or depolarizing θm, respectively (Figure S2). We did not include a persistent Na+ current in our minimal model because this current was not found in FS neurons [14]. However, if added to our model, this current would have an effect similar to increasing the Na+ window current (unpublished data).
Synchronization properties of neuronal networks are tightly related to single neuron properties [29–33]. For instance, increasing the strength of the Na+ window current transforms the bifurcation of the rest state from a Hopf type to an SN type. This change may switch an inhibitory-coupled network of FS neurons from an asynchronized state to a synchronized state [34]. Furthermore, Skinner et al. showed that networks of FS cells that possess both sufficiently strong Na+ window current (or persistent Na+ current) and Id, and that are coupled by both inhibitory and electrical coupling [35], may exhibit collective bursting oscillatory behavior [36]. Hence, the variability in single-cell properties, presented in this article, is very relevant to the network's behavior.
The firing patterns exhibited by our model include “classical” non-delayed tonic firing, delayed tonic firing, and delayed stuttering. These three patterns of firing are consistent with those described in recent experimental studies of FS cells (e.g., [7,22]) as well as in the experimental results reported in the present study. Increasing Iapp in the model eventually causes the disappearance of the delay. This is also consistent with experimental observations [2].
A large variability is observed between FS neurons in their minimal firing rates in response to steady current. Whereas many FS cells have high steady-state minimal firing frequencies on the order of tens of Hz or more [2,23], the minimum firing rate of other FS neurons can be as low as 20 Hz [22] or even less than 10 Hz [37]. Especially, FS neurons with neurogliaform morphology can fire at low rates [24,38]. Although in our experimental data we classified neurons as FS based on their spike width and repolarization rate [1], and did not examine their morphology, our experimental results are consistent with such variability in the minimal firing rate. Further experimental work is needed to verify whether only FS neurons with neurogliaform morphology can fire at low rates. Relying on our modeling study, we propose that heterogeneities in the Na+ window current contribute strongly to variability in the minimal firing rate.
In our model, a delay in action potential firing is induced by a slow crossing of a bifurcation driven by the slowly inactivating d-current. Depending on the window current of the Na+ current, this bifurcation can be of Hopf or SN types. As a consequence, the properties of the neuron during the delay period depend also on the Na+ window current. In particular, subthreshold oscillations are found during that period (and also during the quiescent periods in stuttering patterns). This is consistent with the observation of subthreshold oscillations at frequencies in the gamma-range (20–100 Hz) in FS neurons in cortex [22] during delay or interburst periods. Moreover, in our model, these oscillations exist only when the INa window current is small (Figure 9A and 9B). Subsequently, we predict that subthreshold oscillations in an FS neuron are more likely to be observed in neurons with a large minimal firing frequency. Our experimental results (Figures 11 and 12) are consistent with this prediction.
Finally, noise can induce irregular stuttering in our FS model (Figure 10). Similar patterns were found in previous experiments (Figure 1C in [23]), as well as in the experiments reported here (Figure 11C).
There are several physiological observations that the model does not replicate. The AHP and the spike amplitude are larger in the model (Figures 2 and 7) than in FS neurons recorded in slice experiments (Figure 11). Changing the reversal potential of the K+ current reduces the AHP of the model neuron to some extent. It may also happen that in FS cells, the spike-generating area is distant from the soma, and therefore action potentials recorded in the soma are filtered by cable properties. This effect, which cannot be included in our single compartment model, may contribute to the reduction of the AHP.
Another limitation of our model is that it can account neither for the substantial accommodation observed in some FS cells (AC cells in Figure 5 in [7]) nor for the burst of action potentials that precedes tonic firing sometimes observed in FS neurons (“b” cells in Figure 5 in [7]). Although our model may exhibit some adaptation, it is a result of the strong AHP, which causes Id to inactivate during firing, but this inactivation is only weak. However, accommodation and initial bursting can probably be accounted for if one incorporates additional slowly activating K+ currents into the model.
Finally, stuttering FS cells that do not exhibit an initial delay in response to the injection of a step current are observed experimentally [6]. Such a behavior is not present in the phase diagram of Figure 2A. However, it can occur in the framework of our model if the reversal potential of the leak current is taken to be more depolarized (e.g., VL = −60 mV in Figure S1B) than the reference parameter set VL = −70 mV (Figure 1).
To our knowledge, our work is the first to propose a minimal conductance–based model incorporating ionic channels known to exist in FS cortical interneurons and which accounts in a comprehensive way for the variability of firing patterns these cells display. The analysis we have made of this model builds on previous theoretical works. The role of the window INa in achieving low firing rates was considered in [13,18]. The fact that slowly inactivated K+ currents can induce delay to action potential firing and bursting was also described in [13,39]. The stuttering pattern displayed by our model is an example of “elliptic bursting” [18,20,40] (also named “SubHopf/Fold cycle” [19]). In addition, the present paper relates the appearance of subthreshold oscillations and the dependence of the delay duration on the levels of noise and applied current with the bifurcation structure of the fast subsystem.
Marder and colleagues proposed that for a specific pattern of activity, one can find parameter subspaces within which the model displays qualitatively, and even quantitatively, similar behavior [41,42]. They also proposed that neuronal function can be stabilized by homeostatic mechanisms ensuring that the neuron always remains in those subspaces [43]. Clearly, the bifurcation point cannot exist in such subspaces. There are, however, directions in parameter space along which the qualitative behavior of the neuron varies via bifurcations of the dynamics, as shown in our paper. These bifurcations can underlie the variability observed in the electrophysiological properties of FS cells.
We predict that for FS cells that exhibit delay before firing, the delay duration, tdelay, decreases with the amplitude of the current step, Iapp, and disappears at a non-zero value as Iapp is elevated (Figure 4A). When the neuron displays stuttering in response to just suprathreshold oscillations, we predict that elevating Iapp will first increase the average number of spikes during the stuttering state (Figure 2D), and then will transform the cell into a tonic firing cell (Figure 2A). A depolarizing pre-pulse shortens tdelay and even eliminates it if the pre-pulse is large enough, but does not affect the stuttering behavior (Figure S1B). Similarly, a hyperpolarizing pre-pulse increases tdelay. These predictions can be tested by current-clamp experiments.
Our theoretical work and the experimental results presented here suggest the existence in FS cells of two types of responses to step current pulses. They differ in the minimal firing frequencies, the properties of the membrane potential fluctuations during the delay period, and the sensitivity of the delay duration to noise. More specifically, we predict that the minimal firing rate, the sensitivity of tdelay to noise, and the presence of subthreshold oscillations of the membrane potential during the delay period are negatively correlated to the strength of the Na+ window current. Furthermore, we predict that FS neurons that can fire at low firing rates cannot stutter, and that increasing gd artificially via dynamic clamp may convert a tonic-delay response into stuttering. These predictions can be tested in a detailed population study of electrophysiological properties of FS neurons.
The modeling results presented in this paper can be applied to understand the effect of some neuromodulators on the firing patterns of FS cells. Dopamine attenuates the d-type K+ current Id in a subgroup of FS neurons [44]. Consistent with the results of our modeling study, dopamine also transforms the firing pattern of an FS cell from a tonic-delay type to the tonic-no delay type [44] (see also [45]). Na+ currents are affected by metabotropic glutamate receptor subtype 1 (mGluR1). As shown by [46], it shifts θh to more hyperpolarized potentials in pyramidal neurons, decreasing their Na+ window current. It also facilitates the persistent Na+ current INaP by shifting its activation curve leftward [46]. Similarly, serotonin makes θh more negative in pyramidal cells, and also reduces the maximal conductance of INaP [47]. If one assumes that these modulators have similar effects on Na+ currents in FS cells, our results suggest that they may modify qualitatively the firing patterns of these neurons.
Our model of FS cells is based on that of [3,28], with several modifications based on voltage clamp data. The current balance equation is
where V is the membrane potential of the neuron, C = 1μF/cm2 is the membrane capacitance, and the parameters of the leak current are gL = 0.25 mS/cm2 and VL = −70 mV. The external current injected into the neuron is denoted by Iapp.
The Na+ current INa is given by:
where the gating variables, h and m, follow:
The parameters are: gNa = 112.5 mS/cm2, VNa = 50 mV, σm = 11.5 mV, θh = −58.3 mV, σh = −6.7 mV, θth = −60 mV, σth = −12 mV [14]. In this work, we study the effect of the strength of the Na+ window current, controlled by the parameter θm, on the dynamics of the neuron.
The delayed rectifier K+ current IKdr is of the Kv3.1–Kv3.2 type. It is responsible for the brief duration of the spike, about 0.5 ms [2,48], and for the high firing frequency [3,49]. It is given by:
with:
All the parameters of the delayed rectifier current are fixed: gKdr = 225 mS/cm2, VK = −90 mV, θn = −12.4 mV, σn = 6.8 mV, θtn = −27 mV, σth = −15 mV [50].
The K+ current Id incorporated in the model [10,11] has fast activation and slow inactivation. It is defined by:
Throughout the paper, all the parameters of the d-current but gd are fixed: θa = −50 mV, σa = 20 mV, τa = 2 ms, θb = −70 mV, σb = −6 mV, τb = 150 ms [51,52]. The parameter gd is varied to study the effect of the strength of this current.
Finally, to study the effect of noise in the external input on the firing pattern of the neuron, we add an additional external input, Inoise, of the form:
where ξ(t) is a Gaussian white noise with an average 0 and a unit variance, and D has the units of μA2 × ms/cm4 .
Numerical methods. Simulations were performed using the fourth-order Runge-Kutta method with a time step of 0.01 ms implemented as a C program or within the software package XPPAUT [53], which was used also for computing bifurcation diagrams.
Delay. The delay duration tdelay is defined to be the time from the onset of current injection, or, if the neuron fires transient 1–3 spikes, from the last transient spike to the first spike of the sustained firing. We define that the neuron shows a delay if tdelay is at least twice as large as the inter-spike interval during steady-state spiking tISI, or if it is larger than both 100 ms and 1.2 tISI.
Fourier spectrum. Discrete Fourier transforms of subthreshold oscillations were calculated numerically over a time window of TFT ending TBS = 5 ms before the first spike of the steady-state firing. The absolute values of the Fourier components were averaged over nR repetitions of the same stimulus. Parameters for Figure 9 are: TFT = 120 ms, nR = 20. Parameters for Figure 12 are: TFT = 90 ms, nR = 5 (A), TFT = 120 ms, nR = 13 (B), TFT = 120 ms, nR = 11 (C).
The bifurcations of the fast subsystem when b varies depend on the shape of the function VFP(b), where VFP is the value of the membrane potential of the neuron at the fixed point of the dynamics for fixed b. Equivalently, one can relate the bifurcations to the shape of the curve b = b(VFP), in the b-VFP plane (the “b-VFP curve”), which is defined by (see Equations 5, 6, 11, and 15):
where
and
The denominator in Equation 21 is positive and increases with VFP. The numerator therefore is positive in the relevant range of VFP for which b > 0. The functions IL(VFP) and IKdr(VFP) are increasing with VFP. Only the function INa(VFP) may decrease with VFP. Therefore, for a small Na+ window current, b decreases monotonously with VFP. This happens for instance for θm = −24 mV (Figure 5A). In contrast, if the overlap between the activation and the inactivation curves of the Na+ current is sufficiently large [13], the term −INa(VFP) in Equation 21 can contribute substantially to make the function b(VFP) be non-monotonous. This happens for instance for θm = −28 mV (Figure 5D).
We estimate tdelay, the duration of the delay to firing of action potentials, using the “fast–slow method,” and derive the dependence of tdelay on Iapp near the current threshold Ith in the noiseless case. During the delay period, the fast subsystem is at its fixed point, and b decreases slowly. We use this fact to compute the scaling of the divergence of tdelay with Iapp−Ith for Iapp ≳ Ith for the two bifurcation scenarios.
Hopf bifurcation of the fast subsystem. The evolution of b is given by Equation 17, and it becomes very slow when b approaches b∞(V). For large τb, V follows the curve VFP(b) during the delay period (Figure 5A). We denote by bx̃ the solution of the equation
, and define Vx̃ ≡VFP( bx̃) Note that the fixed point of the fast subsystem for b = bx̃ is unstable if the neuron fires following the delay. When Iapp is near Ith, VFP(b) is approximated by Vx̃ [13], and therefore
Equation 17 for the evolution of b is approximated by
Using Equation 22 for computing tdelay is justified because tdelay is determined mainly by the slow dynamics of b when it is near bx̃. Before the current step is applied at time t = 0, the system is at rest with b = brest. The subsystem converges immediately to its fixed point on the slow time scale, and the solution to Equation 22 is
According to Equation 3.13a from [20], tdelay is determined by the equation
Near a Hopf bifurcation,
where bHopf is the value of b at the Hopf bifurcation and α is a constant. Substituting Equation 23 in Equation 24 and using Equation 25, we obtain
Near the threshold current Ith, tdelay >> τb, and Equation 26 becomes
Generically, when Iapp−Ith is small, bx̃ depends only weakly on Iapp and bHopf − bx̃ depends linearly on Iapp−Ith. Therefore, tdelay scales as (Iapp−Ith) −1.
Saddle-node bifurcation of the fast subsystem. The dynamics are very slow near the SN bifurcation, occurring at (bSN, VSN). Neglecting the changes in VFP during the evolution, Equation 23 becomes
where bISN = b∞(VSN). Namely,
Generically, when Iapp−Ith is small, bSN−bISN depends linearly on Iapp−Ith. Therefore, tdelay scales as −log(Iapp−Ith).
We calculate the dependence of tdelay on the noise variance, D, for weak noise and weak window INa. In this case, the delay ends because the fast subsystem is destabilized via a Hopf bifurcation. According to Theorem 4.1 in [40], assuming that the variance of the noise, D, is neither too large nor too small, tdelay is determined by the equation
where Ax̃ and Bx̃ are constants. As above, tdelay is determined mainly by the slow evolution near bx̃, and therefore one can use the approximation VFP(b) ≈ Vx̃. Substituting Equations 23 and 25 in Equation 30, we obtain
Solving this equation for tdelay, in the limit tdelay >> τb, we obtain
where
and
.
We consider the case that, for a certain value of gd, gd = gd1, and a step current with an amplitude Iapp, there is a solution to the equation F(b) = b. We prove here that for a large enough gd, a solution to this equation does not exist for this value of Iapp. The function F(b) is defined only when the limit cycle exists, namely only for b ≤ bSNP. Since the current Id depends on gd and b only through the product gdb (Equation 15), bSNP for any other value of gd is
This means that bSNP(gd) is very small for large gd.
We continue by noticing that: (1) F(b) is the time-average of the function b∞(V(t)) over LC(b) (Equation 2); (2) b∞(V) is a positive, decreasing function of V (Equation 19); and (3) V ≤ VNa. Therefore, F(b) ≥ F(b∞(VNa)). From the fact that bSNP(gd) decreases with gd (Equation 33), one finds that, for large enough gd, F(b∞(VNa)) > bSNP(gd). Since F(b) is defined only for b ≤ bSNP(gd), we obtain that F(b) > b, and there is no solution to the equation F(b) = b if gd is large enough. Therefore, the full system cannot exhibit a tonic firing state. If the rest state of the neuron is unstable, the neuron stutters. In practice, F(b) is much larger than F(b∞(VNa)) because the membrane potential spends a large fraction of its period in subthreshold values, and therefore gd should not be extremely large to prevent a solution of the equation F(b) = b.
Mice (CD1, 21–28 d old) were deeply anaesthetized with pentobarbital, decapitated, and their brains quickly removed into cold (5 °C) physiological solution. Coronal cortical slices (400 μm thick) were cut with a vibratome (Campden Instruments, http://www.campdeninstruments.com) and then transferred to a holding chamber where they were kept at room temperature for at least 1 h before recording, continuously bubbled with 95% O2, 5% CO2. Recording was done in a chamber mounted on an upright microscope equipped with IR/DIC optics (Nikon physiostation EC-600), where they were held at 32–34 °C and constantly perfused. The normal bathing solution contained (in mM): 124 NaCl, 3.5 KCl, 2 MgSO4, 1.25 NaHPO4, 2 CaCl2, 26 NaHCO3 and 10 dextrose, and was saturated with 95% O2, 5% CO2 (pH 7.4). Whole-cell recordings were made from neurons in the barrel field. Patch recording micropipettes (4–6 MΩ) were filled with a solution containing (in mM) 125 K gluconate, 5 NaCl, 2 MgCl2, 10 EGTA, 10 HEPES, and 2 Na2-ATP, pH 7.2, 280 mOsm). Voltages were recorded with a patch clamp amplifier (AxoPatch 2B, Axon Instruments, http://www.axon.com), and digitally sampled at 10 kHz. Data acquisition and analysis were performed with Labview (National Instruments, http://www.ni.com). Series resistance was typically <15 MΩ. During all recordings, 50 μM DL-2-amino-5-phosphopentanoic acid (AP5, Sigma, http://wwwsigmaaldrich.com) and 6,7-dinitroquinoxaline-2,3-dione (DNQX; 20 μM, Sigma) were present in the bath to block excitatory transmission.
Identification of FS neurons. Non-pyramidal neurons were targeted by their soma and proximal dendrites image under the IR/DIC microscope. Among those, FS neurons were identified according to their electrophysiological properties. A neuron was classified as an FS neuron if: 1) it fired brief spikes with fast, deep, monophasic AHPs [1]; and 2) the ratio of the spike's rising phase dV/dt to falling phase dV/dt was smaller than 2. Previous studies revealed that the morphological correlate of FS neurons can be either “basket” [7] or “neurogliaform” [24,54,55]. Most FS neurons express parvalbumin [56], but others express somatostatin [57]. Thus, this type of interneuron may be heterogeneous in terms of its morphology or chemical content, but we did not use these criteria for our classification. |
10.1371/journal.ppat.1005697 | The WOPR Protein Ros1 Is a Master Regulator of Sporogenesis and Late Effector Gene Expression in the Maize Pathogen Ustilago maydis | The biotrophic basidiomycete fungus Ustilago maydis causes smut disease in maize. Hallmarks of the disease are large tumors that develop on all aerial parts of the host in which dark pigmented teliospores are formed. We have identified a member of the WOPR family of transcription factors, Ros1, as major regulator of spore formation in U. maydis. ros1 expression is induced only late during infection and hence Ros1 is neither involved in plant colonization of dikaryotic fungal hyphae nor in plant tumor formation. However, during late stages of infection Ros1 is essential for fungal karyogamy, massive proliferation of diploid fungal cells and spore formation. Premature expression of ros1 revealed that Ros1 counteracts the b-dependent filamentation program and induces morphological alterations resembling the early steps of sporogenesis. Transcriptional profiling and ChIP-seq analyses uncovered that Ros1 remodels expression of about 30% of all U. maydis genes with 40% of these being direct targets. In total the expression of 80 transcription factor genes is controlled by Ros1. Four of the upregulated transcription factor genes were deleted and two of the mutants were affected in spore development. A large number of b-dependent genes were differentially regulated by Ros1, suggesting substantial changes in this regulatory cascade that controls filamentation and pathogenic development. Interestingly, 128 genes encoding secreted effectors involved in the establishment of biotrophic development were downregulated by Ros1 while a set of 70 “late effectors” was upregulated. These results indicate that Ros1 is a master regulator of late development in U. maydis and show that the biotrophic interaction during sporogenesis involves a drastic shift in expression of the fungal effectome including the downregulation of effectors that are essential during early stages of infection.
| The fungus Ustilago maydis is a pathogen of maize which induces tumor formation in the infected tissue. In these tumors huge amounts of fungal spores develop. As a biotrophic pathogen, U. maydis establishes itself in the plant with the help of a large number of secreted effector proteins. Many effector proteins are important for virulence because they counteract plant defense reactions. In this manuscript we have identified and characterized Ros1, a master regulator for the late stages of U. maydis development. This transcription factor is expressed late during infection and controls nuclear fusion, hyphal aggregation and late proliferation. ros1 mutants are still able to induce tumor formation but these are a dead end because they do not contain any spores. We show that Ros1 interferes with the early regulatory cascade controlled by a complex of two homeodomain proteins. In addition, Ros1 triggers a major switch in the effector repertoire, suggesting that different sets of effectors are needed for different stages of fungal development inside the plant.
| The basidiomycete Ustilago maydis is a biotrophic pathogen colonizing maize. The resulting disease, the so-called smut disease, is characterized by the formation of large tumors on all aerial parts of the plant. In these tumors fungal hyphae proliferate profusely and eventually produce massive amounts of dark pigmented, diploid teliospores.
U. maydis is a dimorphic fungus which can grow by yeast-like budding in the absence of a host. On the leaf surface compatible haploid yeast-like cells mate and generate a dikaryon. The dikaryon switches to a filamentous form which is cell cycle arrested [1–3]. Upon perception of surface cues [4] the dikaryon differentiates infection structures and penetrates the maize epidermis. Following penetration, the cell cycle arrest is released [2], the dikaryon invades the plant tissue, proliferates with the help of clamp formation and triggers the development of large tumors [2, 5]. In the late stages of infection, after tumors are formed, the sporogenesis program is initiated. Although the chronology of events leading to teliospore formation is not yet fully understood, the first step is likely to be the fusion of the two haploid nuclei, followed by extensive mitotic divisions of the diploid hyphal cells leading to the formation of large hyphal aggregates [5, 6]. Concomitantly, a mucilaginous matrix of undefined composition and origin is formed embedding the fungal cells during the subsequent hyphal fragmentation and maturation stages. Eventually, the tumors rupture and release the diploid spores in the environment. The cycle is completed when the teliospores germinate and give rise to haploid progeny after meiosis [6]. To overcome PAMP-triggered plant defense responses, and to establish a biotrophic interaction, U. maydis secretes a large panel of effector proteins which may function in the apoplast (e.g. Pep1) or be translocated to the host cells (e.g. Cmu1, Tin2) [7–9]. Expression of the vast majority of the about 300 effectors lacking known protein domains is tied to the biotrophic stage [10]. About 25% of all effectors are arranged in gene clusters and many of these affect virulence either generally or in an organ-specific manner [10–13]. So far, the molecular basis for the virulence function of only a few U. maydis effectors (Pep1, Pit2, Cmu1, Tin2, See1) has been elucidated [8, 9, 14–16].
U. maydis pathogenic development requires fusion of haploid cells and is initiated by the a and b mating type genes. Their expression is induced by the pheromone response factor Prf1 in response to pheromone and host derived signals transmitted via a cAMP-dependent and a MAPK pathway. The a locus encodes a pheromone/receptor system mediating cell-cell recognition and fusion. The b locus encodes two homeodomain proteins which, when derived from different alleles, form the bE/bW heterodimer which acts as master regulator for the switch to filamentous growth, host tissue colonization and tumor induction [10, 17]. bE/bW induces a regulatory cascade which influences the expression of over three hundreds genes [3, 17]. The majority of these genes are regulated via the zinc-finger protein Rbf1, a direct target of bE/bW which acts as a central node in the b-regulatory cascade [3]. Because effector gene expression coincides with pathogenic development, the induction of most effector genes was initially considered to depend on the b cascade [10]. However only few effector genes are subject to direct regulation via components of this cascade. Rbf1 induction in axenic culture leads to activation of only a small subset of effector genes suggesting that plant signals and additional regulators are required for their expression [3, 17]. The membrane proteins Sho1 and Msb2 induce effector gene expression in response to surface cues prior to penetration. They act via the transcription factors Biz1 and Hdp2, two Rbf1 targets, with a specific function in appressorium development [1, 18]. Biz1 might also regulate effector gene expression in conjunction with Mzr1 in the later stages of colonization [19].
Proteins of the WOPR family constitute a novel class of fungal-specific transcriptional regulators that bind DNA via their N-terminal WOPR box. The WOPR box consists of two highly conserved domains, WOPRa and WOPRb, predicted to adopt a globular structure and which are separated by a linker region of variable length and sequence. Both WOPR domains are required for DNA binding activity [20]. Most fungal genomes contain two paralogous WOPR genes that phylogenetically fall into two distinct clades [21, 22]. To date, WOPR proteins have been studied exclusively in ascomycetes where they fulfill a conserved function in the control of developmental processes. Mit1p in Saccharomyces cerevisiae is a core regulator of invasive growth in haploid cells as well as pseudohyphal growth in diploid cells [23, 24]. Wor1, the best characterized member of the WOPR family, is the master regulator of the white-opaque phenotypic switching allowing Candida albicans to adapt to niches in the human host [20]. Similarly, the Ryp1 protein in Histoplasma capsulatum is a key regulator of the temperature-dependent mycelia-to-yeast transition critical for virulence [25].
Sge1 was the first WOPR protein studied in a plant pathogenic fungus. Sge1 supports parasitic growth of the tomato wilt pathogen Fusarium oxysporum f. sp. lycopersici by inducing the expression of at least four of the SIX effector genes [26]. WOPR proteins have later been linked to virulence in other ascomycete plant pathogens [27–31]. Several members of the WOPR family in plant pathogens also positively regulate the production of secondary metabolites potentially involved in pathogenicity [27–29]. In addition to their role in plant colonization, most WOPR proteins regulate sexual/asexual reproduction in phytopathogenic fungi [26–28, 30–32].
Here we investigate the function of a WOPR regulator in the pathogenic development of the basidiomycete pathogen U. maydis. We show that Ros1 (Regulator of sporogenesis 1) is not required for plant colonization but is essential for teliospore production occurring late during the biotrophic life cycle. ros1 deletion strains are locked in the dikaryotic, filamentous stage of infection. They fail to undergo karyogamy, subsequent mitotic cell divisions and are unable to form the mucilaginous matrix in which teliospores are embedded. We show that Ros1 affects the expression of many genes via direct interaction with their promoter regions. Remarkably, Ros1 triggers a dramatic switch in gene expression of the vast majority of effector genes.
The genome of U. maydis is predicted to encode two members of the WOPR family, pac2 (UMAG_15096) and ros1 (UMAG_05853). As the deletion of pac2 (UMAG_15096) was previously shown to not have any effect on U. maydis virulence or reproduction [33], only ros1 was investigated here.
A BLAST search revealed that Ros1 is conserved in other smut species belonging to the four genera of the class Ustilaginomycetes (Ustilago, Sporisorium, Pseudozyma, and Melanopsichium). Outside of the Ustilaginomycetes, conservation is restricted to the N-terminal part comprising the WOPR box (S1 Fig). In Ros1 the N-terminal WOPR box (amino acids 8 to 305) contains both WOPRa (amino acids 8 to 90) and WOPRb (amino acids 240 to 305) domains separated by a rather long linker region of 156 amino acids (S1A Fig). With the exception of Thr210, all residues critical for DNA binding in Wor1 of C. albicans are conserved in Ros1 (S1B Fig) [34]. With a length of 879 amino acids Ros1 is the longest WOPR protein described so far (S1A Fig).
Several members of the WOPR family share a conserved nuclear localization signal (NLS) (PGEKKRA) (S1 Fig). This motif is absent in Ros1 and we could not identify any other canonical NLS. However, bioinformatic tools predict Ros1 to localize in the nucleus. In addition Ros1 displays an unusually long polyglutamine stretch (46 residues, amino acids 707–752) in its C-terminal domain. Several other members of the WOPR family possess one or more glutamine-rich regions (S1A Fig) which might mediate the interaction with components of the transcriptional machinery [35].
To study the localization of Ros1, a haploid FB1 strain was generated which constitutively expresses a C-terminal mCherry fusion of Ros1 together with the nuclear envelope marker Nup107eGFP. The red signal from Ros1mCherry was detected in the nucleus, surrounded by the green fluorescence of the nuclear envelope, demonstrating that Ros1 is targeted to the nucleus (Fig 1A). The nuclear localization and the presence of the long polyQ stretch are consistent with a potential function of Ros1 as transcriptional regulator.
To study its function we deleted ros1 in the two compatible haploid strains FB1 and FB2. FB1Δros1 and FB2Δros1 could successfully mate and produce dikaryotic filaments on charcoal containing medium (S2 Fig). When injected into maize seedlings, nearly all plants infected with the ros1 deletion strains developed tumors. However, compared to the FB1 x FB2 mixture, ros1 mutants exhibited reduced virulence (Fig 1B). In particular, only 4.8% of the plants infected with the ros1 mutant mixture were dead after 12 days compared to 53.7% for the FB1 x FB2 infection. Remarkably, dark pigmented teliospores were absent in the tumors induced by the ros1 deletion strains (Fig 1C) and softening of the tumor tissue which usually becomes evident when teliospores accumulate in wild type infections did not occur. Spore formation could be restored by reintroducing ros1 in single copy in the ip locus of FB1Δros1 and FB2Δros1 strains (Fig 1B). Because constructs containing a promoter region of 1 kb did not complement, the entire 7.6 kb region separating ros1 from the upstream gene UMAG_05850 was included in the complementation construct. While the complementation strains had regained the ability to produce spores (Fig 1C), virulence was only partially complemented (Fig 1B). We speculate that this reflects a position effect resulting from integrating the construct into the ip locus. Alternatively, the expression of ros1 might partially depend on distal regulatory elements that are missing in the complementation construct.
To determine at which stage of development ros1 deletion strains are affected, we stained fungal hyphae with wheat germ agglutinin-Alexa Fluor 488 and followed the sequence of events leading to the formation of mature teliospores in wild-type infections by confocal microscopy (Fig 2, left panel). Until 4 days after infection, growth of the ros1 deletion strains was indistinguishable from the growth of wild type strains. In both cases uniform spreading of fungal hyphae within the leaf was observed (Fig 2). After 4 days wild type hyphae started to form aggregates which became visible at 6 dpi, indicating that the mucilaginous matrix in which the cells are embedded during spore formation [37] was produced (Fig 2). Between 6 and 8 dpi the hyphal aggregates continued to expand reaching diameters of up to 250 μm at 8 dpi. Around 10 dpi, hyphae in these aggregates underwent fragmentation and individual cells entered the spore maturation process. Finally, at 12 dpi groups of mature teliospores with their characteristic ornamentation became clearly visible. This sequence of events parallels what has been described [6]. By contrast, in plants infected by ros1 deletion strains neither hyphal aggregates nor fragmented hyphae could ever be observed (Fig 2, right panel). Thus, in the absence of ros1, U. maydis development was locked in the filamentous stage.
To see how this failure to aggregate was linked to expression of the ros1 gene, the expression pattern of ros1 in haploid strains in axenic culture and of FB1 x FB2 mixtures during plant infection was analyzed using quantitative RT-PCR (Fig 3). ros1 was expressed at a very low basal level in axenic culture and during the early steps of maize infection. Expression was then upregulated 35-fold at 6 dpi when sporogenesis was initiated and reached a maximum 70-fold induction at 8 dpi as hyphal aggregates expanded. Between 8 and 12 dpi, ros1 transcript abundance slowly decreased, concomitantly with the accumulation of mature teliospores (Fig 3).
To evaluate in more detail the contribution of Ros1 to the regulation of spore development, we analyzed karyogamy and matrix formation in the ros1 mutant. To be able to visualize nuclei, the nuclear envelope marker Nup107eGFP as well as the plasma membrane marker Sso1mCherry were introduced into FB1, FB2 and the corresponding ros1 deletion strains. Infected plant tissues were then analyzed by confocal microscopy between 6 and 8 dpi. At this stage wild type filaments contained only one nucleus while pairs of nuclei were visible in filaments of the ros1 deletion strains (Fig 4A), indicating that karyogamy had not occurred (even at later time points). In addition, the mucilaginous matrix observed in tumors induced by wild type strains was absent in tumors induced by the ros1 deletion strains (Fig 4B).
Teliospore differentiation takes place in large hyphal aggregates [5, 38]. As such aggregates were absent in tissue infected by ros1 mutant strains, we also studied the accumulation of fungal biomass in plants infected with wild type and ros1 deletion strains by quantitative RT-PCR (Fig 4C). In line with the microscopic data (Fig 2), until 4 dpi wild type strains as well as ros1 mutants showed a comparable small increase in fungal biomass (Fig 4C), likely reflecting coordinated mitotic divisions of the dikaryon with the help of clamp connections [2, 3]. However, while in wild type infected tissue the ratio U. maydis/plant biomass increased dramatically (7.5 to 42.3) between 6 dpi and 12 dpi, the ratio remained at the 4 dpi level in plants infected with the ros1 deletion strains (Fig 4C). These data show that Ros1 affects karyogamy as well as matrix formation and is needed for massive late proliferation in the infected tissue.
To further characterize the function of Ros1, we studied the effect of expressing ros1 prematurely. To mimic early pathogenic development of U. maydis, we used strain AB33 in which filamentous growth and cell cycle arrest can be triggered in liquid culture via nitrate-inducible expression of bE1 and bW2 homeodomain genes [39, 40]. Ros1 was placed under the control of the arabinose-inducible crg1 promoter [41]. In nitrate minimal medium supplemented with glucose (NM + Glucose), the b genes were expressed and cells switched from budding to filamentous growth (Fig 5A). However, when ros1 was induced simultaneously (NM + arabinose), cells failed to filament, increased their diameter and formed septa (Fig 5A). DAPI staining revealed that each section contained one nucleus, indicating that mitotic divisions had resumed. When filamentation was induced prior to ros1 expression, the resulting filaments stopped elongating and became septated with one nucleus per segment (Fig 5B). This shows that Ros1 counteracts the activity of the bE/bW heterodimer, inhibits filamentation and triggers mitotic divisions.
To study the effect of premature ros1 expression during plant colonization, we generated compatible ros1 deletion strains expressing ros1 under the control of the mig2-6 promoter. mig2-6 is an effector gene whose expression is strongly upregulated shortly after penetration [42]. Compared to wild type infections and infections with ros1 deletion strains (Fig 6A) the strains expressing ros1 prematurely caused severely attenuated disease symptoms ranging from chlorosis to very small tumors (Fig 6C). These strains penetrated the plant surface but invasion of the plant tissue rapidly stopped and hyphae with an abnormally high number of septa developed (compare Fig 6B and 6D). These results show that the timing of ros1 expression is critical for biotrophic development and tumor induction of U. maydis.
To identify genes regulated by Ros1, we combined two approaches: RNA sequencing to evaluate the global effect of Ros1 on gene expression and ChIP sequencing to identify which of the differentially regulated genes are direct targets. Both experiments were carried out on samples collected at 8 dpi when the ros1 expression level is maximal.
For the transcriptomic analysis, we compared maize tumor tissue infected by strains FB1 x FB2 and the corresponding ros1 deletion strains. RNA-seq data showed that 2005 genes were differentially regulated (fold change (FC) ≥ 1,5; p-value < 0.01). Of these genes, 1091 were expressed at lower levels in the wild type strains compared to ros1 mutants and 914 were higher expressed in wild type strains compared to ros1 mutant strains (S1 Table). This shows that about 30% of the 6766 protein-encoding U. maydis genes are differentially regulated by Ros1. RNAseq results were confirmed by qRT-PCR for 15 genes encoding two glycoside hydrolases (UMAG_05550, UMAG_04503), a trehalase (UMAG_02212), a cyclopropane fatty acid synthase (UMAG_01070), a polyketide synthase (pks1 [43]), transcription factors (UMAG_04101, biz1 [1], rbf1 [3], fox1 [44], UMAG_02775) and secreted effectors (mig2-3 [45], UMAG_04096, dik1 [46], UMAG_02473 [10], UMAG_03046) (S3 Fig).
The deletion of ros1 mostly affects metabolic processes and cellular transport (Fig 7). Functional categories “C compound and carbohydrate metabolism”, “lipid fatty acid and isoprenoid metabolism” as well as “secondary metabolism” were predominantly enriched in both upregulated and downregulated gene sets (Fig 7). In contrast, categories related to mitochondrial function (respiration, electron transport, mitochondrion biogenesis and mitochondrial inner membrane) as well as protein synthesis (translation, ribosome biogenesis) were enriched only among Ros1-upregulated genes. About 170 genes belonging to “cell cycle and DNA processing”and 55 genes belonging to “cell growth and morphogenesis” were differentially regulated. Although they did not show significant enrichment, both categories were highlyrepresented in the upregulated gene set. Unexpectedly, Ros1 was also shown to cause a massive shift in secreted effector gene expression. 128 effectors genes were downregulated including 126 effector genes without functional domains, cmu1 [8] and UMAG_01130 [13], two effectors containing a functional domain. In addition, 70 effector genes were upregulated by Ros1: 68 without functional domains and two with functional domains, UMAG_03615 [10] and UMAG_11763 [13]) (Fig 8, S1 Table).
To identify the genes directly targeted by Ros1, we carried out a ChIP-seq analysis. Maize seedlings were infected with a compatible pair of complemented ros1 deletion strains expressing an HA-tagged version of Ros1. A pair of compatible strains complemented with the native version of Ros1 without an HA tag was used as negative control. After sequencing the output DNA from three biological replicates, 1907 peaks showed high reproducibility, a significant peak shape score (> 20) and a low p-value (p < 0.01) (S2 Table). 1441 distinct intergenic regions including 620 intergenic regions for divergently transcribed genes were found to be targeted by Ros1. Only considering promoter regions, this brings the number of genes potentially targeted by Ros1 to at least 1913 (S3 Table).
Of the 2006 genes which were differentially expressed in the RNA-seq data, only 790 (40%) displayed at least one ChIP peak in their upstream region (S1 Table) suggesting that a significant part of Ros1 regulation depends on intermediate regulators. In total 80 transcription factor-encoding genes were differentially expressed in the RNA-seq dataset and of these 42 were downregulated and 38 were upregulated by Ros1. 25 upregulated transcription factor genes were predicted to be direct targets from the ChIP-seq analysis (S1 Table) and six of these including ros1 displayed multiple Ros1-ChIP peaks in their promoters (S2 Table). For example, in the long intergenic region between ros1 and UMAG_05850 (Fig 9A) we detected six regions bound by Ros1, suggesting a rather complex regulation including autoregulation. The ChIP analysis also revealed that Ros1 binds the promoters of genes previously identified as regulators of sporogenesis rum1, hgl1, tup1 and ust1, suggesting that they could be direct targets. These genes did not show differential regulation by Ros1 in the RNA-seq dataset generated at 8 dpi. However, a time-resolved analysis of their expression pattern showed that rum1, hgl1 and ust1 are slightly but significantly induced by Ros1 at 10 and 12 dpi (S4 Fig). In general, Ros1 binding is detected more frequently upstream of upregulated genes than upstream of downregulated genes (50% of the upregulated genes against 30% of the downregulated genes) (S1 Table). This suggests that Ros1 acts primarily as a transcriptional activator but can also function as a repressor. A more detailed discussion of the RNA-seq and ChIP-seq analysis is found in the discussion to avoid redundancy.
The WOPR regulators Wor1, Ryp1 and Mit1 all recognize a similar DNA binding motif [20, 24]. A search for the corresponding 14 bp consensus sequence in Ros1 ChIP-seq data (FIMO online tool, http://www.meme-suite.org), identified 975 motifs (p-value < 0,001) corresponding to 763 ChIP peaks (S4 Table). To test whether one of these regions is directly bound by Ros1, we carried out electrophoretic mobility shift assays (EMSA) using a recombinant His-tagged version of Ros1 containing the WOPR domain only (Ros1WOPR-His). As target we used a 237 bp biotinylated probe corresponding to the ros1 promoter region between 2854 and 3090 bp upstream of the ros1 gene (Fig 9A and 9B) containing three putative binding motifs (WT-probe). In presence of Ros1WOPR-His in a molar ratio of 2700:1 of protein to DNA the probe was completely shifted. This shift could be abolished by addition of a 500-fold excess of unlabeled WT-probe competitor, indicating that Ros1 interacts specifically with the WT-probe (Fig 9C). In comparison, incubation of Ros1WORP-His with a fragment of the same length from the ros1 open reading frame (ORF-probe) did not lead to a mobility shift (Fig 9C). To narrow down the region bound by Ros1, mutations were introduced in the probe sequence. Two of the predicted Wor1-like binding motifs (m1 and m2) located at the center of the peak (Fig 9A), were mutated in probes mut-m1 and mut-m2 (Fig 9B) and tested for binding by Ros1WOPR-His. For both probes we observed discrete, significantly smaller shifts than for the WT-probe and the interactions could again be competed by an excess of unlabeled WT-probe (Fig 9C). When a fragment containing both mut-m1 and mut-m2 (mut-m1+2) mutations was used, an even less shifted complex was observed which could be competed (Fig 9C). This strongly indicates that the predicted m1 and m2 sites are indeed bound by Ros1 and suggests furthermore, that the WT-probe fragment contains an additional binding site, most likely the m3 site (Fig 9B). Using similar conditions, we also tested the binding of Ros1WOPR-His to other promoter regions identified by ChIP which contain at least one predicted binding site. Probes were designed for promoters of a gene encoding a transcription factor (UMAG_02775) upregulated by Ros1, four effector genes downregulated by Ros1 (UMAG_02854, UMAG_04040, UMAG_02538 [10], cmu1 [8]) and three effector genes upregulated by Ros1 (UMAG_03138, UMAG_12258, UMAG_03046). All probes were specifically bound by Ros1WOPR-His (S5 Fig), confirming that these genes are direct targets of Ros1.
RNA-seq and Chip-seq analysis had shown that 47 transcription factor genes may represent direct targets of Ros1 while 33 may represent indirect targets. The expression pattern of two of these, UMAG_02775 (presumed to be directly regulated by Ros1) and UMAG_01390 (presumed to be indirectly regulated by Ros1) was additionally determined in maize plants infected with FB1 x FB2 or the corresponding ros1 deletion strains in a time course experiment (S6 Fig). Results show that Ros1 is responsible for the late upregulation (between 8 dpi to 12 dpi) of these two genes.To follow up on these two transcription factors, the corresponding genes were deleted in FB1 and FB2 and mutant strains were then tested for virulence in maize seedlings and for their ability to produce teliospores (Figs 10 and S7). Tumor induction was not affected by the deletion of UMAG_02775 and the mutant hyphae produced aggregates (Fig 10). However, in these aggregates only few spores developed and these were misshaped and were missing the ornamentation characteristic of mature spores (Fig 10).The deletion of UMAG_01390 attenuated virulence to a comparable extent to what had been observed in the ros1 mutant strains (Fig 10). Contrary to the ros1 mutant, UMAG_01390 deletion strains showed hyphal aggregation and reached the fragmentation stage of spore development (Fig 10). However, fragmented hyphal cells failed to enter the maturation process and did not give rise to ornamented teliospores (Fig 10). Taken together these results illustrate that UMAG_02775 and UMAG_01390 genes both affect discrete steps in spore development downstream of Ros1.
In this study we demonstrate that late biotrophic development in U. maydis is coordinated by Ros1, a member of the WOPR family of fungal regulators. Ros1 does not influence the ability of U. maydis to induce tumor formation, but is the key regulator for switching from b-dependent filamentation to hyphal aggregation and spore formation. This development is accompanied by a dramatic shift in the expression of 60% of the putative effector genes without functional domain [47, 48]. The processes regulated by Ros1 are depicted schematically in Fig 11.
Previous light microscopy studies had seen paired nuclei commonly in hyphae outside the aggregates but not in the aggregates of sporogenous hyphae [5]. This is consistent with our analysis using fluorescent nuclear markers which shows that hyphae in aggregates are monokaryotic. In contrast to wild type strains, ros1 deletion strains are unable to form such aggregates, fail to accumulate matrix material and remain dikaryotic. This could suggest that karyogamy precedes the formation of hyphal aggregates and may be required to initiate the synthesis of the mucilaginous matrix (Fig 11). In wild type strains hyphal aggregates develop within the plant intercellular space and considerably expand over time due to a dramatic increase of fungal biomass which is not observed in plants infected with ros1 deletion strains. When expressed ectopically in axenic culture, Ros1 triggered mitotic divisions suggesting that aggregate expansion during colonization is due to multiple rounds of mitotic divisions of the diploid cells as was earlier hypothesized [5]. Moreover, the phenotype of cells expressing ros1 prematurely in hyphae indicates that the resulting cell divisions do not involve clamps. We speculate that without clamps diploid cells proliferate faster than the dikaryon which could explain the rapid and massive increase of fungal biomass late in infection.
Starting at six dpi U. maydis cells begin to aggregate and form large, ball-like structures. What glues cells together is presently unknown, but based on the finding that the matrix can be stained with basic fuchsin [49], it is likely that the matrix contains polysaccharides. In addition, it was reported that hyphae containing diploid nuclei are partially refractory to chemical fixation and this was attributed to lysis of the cell wall and its conversion to a gelatinous material [5]. Incidentally we noticed that during sporogenesis staining of the cell wall, but not of the septa, with wheat germ agglutinin-Alexa Fluor 488 becomes fainter with time (Fig 2, 8 dpi time point) which could reflect chitin degradation or modification of the cell wall. Cell wall loosening might facilitate the changes in cell morphology which are associated with hyphal fragmentation and spore maturation. The matrix could connect the cells and shield them against biotic and abiotic stresses. Among the 255 Ros1-regulated genes belonging to the functional category “C compound and carbohydrate metabolism”, 55 are predicted to be involved in polysaccharide metabolism. Most of them encode glycoside hydrolases. They are enriched in both up and downregulated gene sets. The two most upregulated genes encode enzymes targeting the fungal cell wall, an EXG1 beta-glucanase (FC = 831) and a chitinase A (FC = 441) (S1 Table), and both could conceivably be involved in the gelatinization process. The strong Ros1-dependent upregulation of an UDP glucose dehydrogenase (UMAG_00118) suggests an increased glucuronic acid production at the onset of sporogenesis. Bacterial extracellular matrices as well as the capsule of Cryptococcus neoformans contain highly polar glycosaminoglycans which are rich in glucuronic acid [50]. Among the genes upregulated by Ros1, we found a gene related to C. neoformans CAP59. This C. neoformans gene is involved in capsule synthesis by supporting polysaccharide export [51]. It is conceivable that the related gene in U. maydis (UMAG_11017) could fulfill a similar function in the export of matrix material. We also observed that the repellent gene rep1, which is already upregulated in hyphae [52] is further induced by Ros1 during sporogenesis. Repellents are structural proteins forming amyloid fibrils at the cell surface which mediate hyphal adhesion to hydrophobic surfaces [53]. These amyloid fibrils could be of importance during sporogenesis, could aid in connecting hyphae with each other and assume a structural function in the formation of the matrix. Consistent with this hypothesis hydrophobins which are functionally similar to repellents [54] play a structural role in the formation of fruiting bodies [53]. Since rep1 deletion mutants still produce viable teliospores [52] Rep1 is unlikely to be essential for teliospore formation in U. maydis. However, it cannot be excluded that the efficiency of spore formation is affected in rep1 mutants.
In addition to genes involved in cell wall modification we observed the Ros1-dependent upregulation of several genes involved in the synthesis / modification of membrane lipids (Fig 10). Among them were genes encoding ergosterol biosynthetic enzymes, sphingolipid biosynthetic enzymes and fatty acid synthesizing / modifying enzymes like cyclopropane fatty acid synthase (S1 Table). The likely ensuing alterations of plasma membrane composition might reflect that the plasma membrane in spores has a different composition from that of vegetative cells. In Schizosaccharomyces pombe and S. cerevisiae, sporulation involves de novo synthesis of the forespore membrane within the cytoplasm of mother cells, which subsequently becomes the plasma membrane of the developing ascospores [55]. Cyclopropane fatty acid synthesis was also reported to be essential for fruiting body development in the basidiomycete Coprinus cinerea [56]. Sphingolipids have important roles in membrane and lipoprotein structure and in cell regulation as signaling molecules for growth and differentiation. They have been shown to be required for proper cell growth and morphology in U. maydis [57] and the upregulation of sphingolipid synthesis by Ros1 might be prerequisite for teliospore differentiation. Many genes involved in fatty acid beta oxidation were downregulated by Ros1. This may reflect that the predominant spore storage fatty acids of U. maydis are linoleic and palmitic acid [58]. In line with this, a gene encoding a caleosin-like protein (UMAG_02753) is strongly upregulated by Ros1 (FC = 220). Caleosins are involved in the structural maintenance and turnover of lipid storage organelles, so-called lipid droplets [59]. The strong upregulation of this gene at 8 dpi might thus indicate lipid storage in spores.
Most WOPR regulators characterized so far in plant pathogenic ascomycetes regulate both plant invasion and sexual / asexual spore production [26–28, 30–32]. Ros1 is the first WOPR protein characterized in a basidiomycete. In comparison to Wor1 from C. albicans and Ryp1 from Histoplasma capsulatum for which ChIP-chip identified only about 200 and 700 targets [60, 61], Ros1 might directly regulate a much larger set of genes (1900 identified by ChIPseq). Binding sites for all WOPR proteins characterized to date are conserved [20, 24], and we have shown here that Ros1 can also bind the 14 bp consensus sequence identified for Wor1 [20]. However, not all Ros1-bound regions identified by ChIP-seq (765 out of 1913) harbor this motif. This could indicate that Ros1 can recognize additional sequences more distantly related to the Wor1 binding site or that it can bind additional sequences via interaction with other transcription factors. Recent studies in C. albicans and H. capsulatum have provided evidence that WOPR proteins can bind promoters in complex with other core regulators and have suggested that the formation of these complexes might be mediated by glutamine-rich regions [61, 62] which serve as protein interaction domains also in other proteins [63]. The presence of an exceptionally long poly-glutamine tract in the C-terminal domain of Ros1 might reflect such a role. Similarly to Wor1 [60] and based on finding that direct Ros1 targets include both up and downregulated genes Ros1 may function both as an activator and a repressor.
In C. albicans, S. cerevisiae and H. capsulatum, Wor1, Mit1 and Ryp1 are part of core regulatory networks in which each transcription factor regulates and is regulated by the others [24, 61, 62]. In these networks, WOPR regulators fulfill a critical function because they bind most of the target promoters [23, 24, 61, 64]. Similar to the genes encoding core regulators in these species, ros1 exhibits an unusually long promoter and binds to its own promoter, most likely reflecting autoregulation. Moreover, Ros1 directly regulates many transcriptional regulators which remain to be investigated and which could potentially be core regulators.
WOPR regulated genes have functionally diverged considerably during evolution and show very poor overlap even in closely related species [24]. However, several classes of genes / processes regulated by WOPR proteins in plant pathogenic fungi appear conserved: these include spore formation as well as secondary metabolism and effector gene expression (all discussed below).
Premature expression experiments in hyphae showed that Ros1 is antagonistic to bE/bW and inhibits b-dependent filamentation (Fig 11). Analysis of Ros1 dependent gene expression revealed that the negative effect of Ros1 on filamentous growth could originate from an alteration of the bE/bW regulatory cascade. About 50% of the 345 genes which are regulated by overexpressing bE1/bW2 in axenic culture [3] were found to be affected by Ros1. 75 genes upregulated by bE/bW are repressed by Ros1 and this includes several regulators of the b cascade: Rbf1, the central regulator of pathogenic development responsible for inducing the majority of the 345 b-regulated genes, and two of its downstream targets, Biz1 and Hdp1, which modulate the cell cycle and regulate the growth of filaments [3, 65]. Conversely, 40 genes downregulated by bE/bW are induced by Ros1 and 36 genes upregulated by bE/bW were also induced by Ros1. Among the 170 b-dependent genes differentially expressed, Ros1 is likely to directly regulate 71. Interestingly, bE, bW and prf1, the main regulators of the b cascade, are only slightly repressed by Ros1, suggesting that the Ros1 induced inhibition of filamentation does not re-establish the budding program This is also apparent when b-expressing cells are microscopically observed after premature expression of Ros1. In these cases we observe the formation of septated, compartmentalized cells each containing a single nucleus. Such structures, where cell segments containing a single nucleus become deliminated by thick septa, are reminiscent to structures in sporogenous hyphae [5]. This suggests that Ros1 targets the b cascade at certain nodes without downregulating the entire cascade and this may then be prerequisite for entering this septation program. During infection, inhibition of parts of the b cascade would occur at a specific stage of biotrophic development, when the cell cycle has been released and the U. maydis dikaryon is proliferating in the infected tissue by clamp formation. This developmental stage is different from the filamentous stage achieved by overexpressing bE1/bW2 in axenic culture. Therefore, data sets generated by Heimel et al. (2010) [3] and the 8 dpi time point studied here after infection with the dikaryon are not fully comparable.
Our studies have illustrated that ros1 mutants do not initiate nuclear fusion and fail to trigger massive proliferation. This shows for the first time that the strong increase in fungal biomass late in infection may require karyogamy to be completed. An inspection of the RNA-seq data revealed no differential regulation by Ros1 of the four U. maydis genes related to genes KAR7, KAR2, KAR3 and KAR4 implicated in karyogamy in S. cerevisiae. While 168 genes involved in DNA processing and cell cycle (Fig 11) were differentially regulated by Ros1, we did not observe any significant enrichment for these categories in the ros1 upregulated gene set. One likely explanation is that at the 8 dpi timepoint chosen for the RNA-seq analysis genes involved in karyogamy are already shut off. Among the 89 Ros1 upregulated genes in the category “DNA processing and cell cycle” we detect clear indicators for proliferation like DNA helicases, DNA primase, PCNA and several DNA mismatch repair proteins (S1 Table). However, the ros1 deletion strain is also able to replicate its DNA in the dikaryotic phase, and this could explain why no significant enrichment is observed for the category “cell cycle and processing” (Fig 7). In line with this interpretation, protein synthesis and mitochondrial function (respiratory chain) were both overrepresented functional categories among the genes upregulated in sporogenous hyphae compared to the dikaryon of the ros1 mutant. Late proliferation of diploid hyphae might rely primarily on plant sugars as suggested by the strong upregulation of several glycoside hydrolases located at the surface of the cells, e.g. a secreted trehalase (UMAG_02212), a membrane located glucoamylase (UMAG_04064), the secreted invertase SUC2 and several other non characterized secreted glycoside hydrolases (e.g. UMAG_00102, UMAG_06434).
Infection experiments in maize seedlings showed that ros1 mutants cause significantly fewer dead plants compared to the wild type. We speculate that the massive late proliferation of the wild type in hyphal aggregates might negatively affect plant fitness and account for the high percentage of plants not surviving under our glasshouse conditions.
We have shown that Ros1 controls the early events of spore development by inhibiting b-dependent filamentation and inducing karyogamy and hyphal aggregate formation as well as the initiation of spore formation. Among the upregulated genes in wild type infections are 38 putative transcription factors which are Ros1-regulated. Mutants in two of these transcription factors genes (UMAG_02775 and UMAG_01390) revealed that both mutant strains were able to reach the gelatinization stage and to trigger hyphal fragmentation but failed at discrete subsequent steps of spore maturation. This suggests that these genes are regulators of the spore maturation process downstream of Ros1. Other regulatory proteins previously shown to interfere with teliospore formation are Hgl1, the histone deacetylase Hda1, and the transcription factors Rum1, Ust1, and Tup1 [33, 43, 66–68]. Hgl1, Rum1 and Tup1 are all required for proper spore development after the fragmentation stage [33]. Ust1 (UMAG_15042) is a transcriptional repressor in haploid cells which represses filamentation as well as formation of spore-like structures. [68, 43]. The ChIP analysis revealed that Ros1 binds the promoters of rum1, hgl1, tup1 and ust1, suggesting that they could be direct targets. rum1, hgl1 and ust1 were slightly but significantly induced by Ros1 at 10 and 12 dpi (S4 Fig). The small effect of Ros1 could indicate that these genes are regulated by a combination with other TFs. For rum1 and hgl1 this is in line with their requirement for spore formation. To explain the upregulation of the negative regulator ust1 by Ros1, we consider its role in controlling the budding program [68] may be required during sporogenesis when we observe its upregulation. This would imply that the observed formation of spore-like structures in haploid cells of the ust1 mutant [68], could be a default pathway and not reflect what is happening during the sporulation program after infection. Consistently, we do not observe expression levels of ust1 during the U. maydis life cycle which are below the levels in axenic culture.
tup1 expression levels were not influenced by Ros1. As we did not find evidence that Ros1 differentially regulates the divergently transcribed UMAG_10827 gene (S1 Table), we speculate that Ros1 binding to the promoter is not essential for the activity of this promoter under the tested conditions. Having shown that Ros1 participates in regulating hgl1 and rum1 expression and having shown that Ros1 is required at an earlier stage that precedes the formation of the sporogenous hyphal aggregates prior to karyogamy places Ros1 upstream of these genes and processes and highlights the importance of Ros1 for the regulation of sporogenesis. The functional analysis of the other Ros1-regulated transcription factors identified here is a promising avenue to elucidate the entire regulatory network controlled by Ros1 during sporogenesis.
Secondary metabolism is commonly associated with sporulation processes in microorganisms, including fungi [69]. Among the genes upregulated by Ros1, we found pks1 and laccase I responsible for the synthesis of the melanin pigment in teliospores [43]. In addition, the most upregulated gene involved in secondary metabolism is related to versicolorin B synthase, an enzyme involved in the synthesis of aflatoxine in Aspergillus sp. [70]. In U. maydis this gene belongs to a newly identified gene cluster (E. Reyes-Fernàndez and M. Bölker, personal communication) and we speculate that it is responsible for an intermediate step in the biosynthesis of an antimicrobial compound and / or a pigment.
A gene cluster involved in the synthesis of itaconic acid [71] is also upregulated by Ros1 and in this case it appears to be the regulatory gene ria1 which is directly targeted while the promoters of the other genes in the cluster are not bound by Ros1. Itaconic acid inhibits the bacterial glyoxylate shunt essential for many bacteria to survive during infection of mammalian hosts. The gene cluster for the production of the mannosylerythritol lipids which have biosurfactant and antimicrobial activity [72] is also upregulated by Ros1. The antimicrobial properties of itaconate and mannosylerythrol lipids might serve to combat against competing microbes during late stages of U. maydis development.
The modulation of effector gene expression seems to be a common trend shared by many WOPR proteins from various plant pathogenic fungi [26, 29–32]. However, in C. fulvum and V. dahliae the deletion of the respective WOPR protein is associated already with growth defects in axenic culture [31, 32], and in Zymoseptoria tritici abnormally swollen cell structures are observed during axenic growth in the wor1 mutant [30]. This suggests that in these cases the respective WOPR protein might be more involved in developmental processes than in specific regulation of effector genes. Furthermore, in cases where effector gene expression has been analyzed, effectors are usually upregulated by the respective WOPR protein and the upregulation concerns a relatively small number of effectors (six in F. oxysporum, 14 in F. verticilloides, six in V. dalhiae) [26, 29, 32]. One of the main findings emerging from our transcriptional analysis is that Ros1 in U. maydis induces a massive switch in the effector repertoire affecting about 60% (194 genes) of the predicted set of effector genes without functional domains [47, 48] as well as 4 genes encoding effectors with functional domain. 82 of the 198 differentially regulated effector genes are likely to be directly targeted by Ros1. These include cmu1 [8], pit2 [15, 73], mig1 [74], 23 genes residing in effector clusters (2A, 2B, 5A, 6A, 10A, 19A, 22A,) [10] as well as 45 of the late-induced effectors. The differential regulation of the genes which are not direct targets likely results from the alteration of the bE/bW cascade by Ros1. Contrary to Sge1 which induces effector genes required for virulence in F. oxysporum [26], Ros1 mostly downregulates effector genes. 26 of the 128 downregulated effector genes (Fig 7) reside in clusters where the cluster deletion causes a virulence defect (5A, 5B, 6A, 10A, 19A) [10] and 10 are members of the eff1 effector family that also contributes to virulence [75]. Unexpectedly, among the effectors downregulated by Ros1 are also critical effectors involved in the inhibition of plant defense responses like Stp1, Cmu1 and Pit2 [8, 15, 73, 76]. In addition, of the 14 leaf-specific effector genes without functional domain [13] (of which many contribute to virulence) all are downregulated by Ros1. While it is easy to conceive that see1, an effector involved in cell expansion associated with tumor formation [16] is downregulated because tumor formation has happened already at the stage when ros1 is induced, the downregulation of effectors with critical functions in plant defense suppression is more difficult to explain. Either plant defenses at this late stage of U. maydis development could be distinct from the early stages of infection and require a different effector set for suppression. Or alternatively the massive production of matrix material which is associated with the formation of the fungal aggregates, could shield the aggregated hyphae from plant detection or from the action of antimicrobial compounds associated with plant defenses. Reduced effector expression could then be sufficient to maintain the inhibition of plant defense at the periphery of the aggregates.
As a third possibility it could be the combined action of late induced effectors plus matrix that is effective. In total, there are 70 effector genes that are induced by Ros1 (68 without functional domains). Four of them (UMAG_03138, UMAG_05926, UMAG_03046, UMAG_12258) exhibited a 50-fold higher expression in infections with wild type compared to the ros1 mutant strains. 13 of these late-induced effector genes reside in effector clusters (19A, 10A, 2B, 5A, 6A, 9A) including clusters involved in virulence (5A, 6A, 10A, 19A) [10]. Since the cluster deletions were all generated in a solopathogenic haploid strain [10] in which spore formation does not follow the same sequence of events as in the dikaryon, we cannot presently assess whether these late effectors affect processes facilitating U. maydis sporulation, inhibit late plant defense responses and / or are participating to the formation of the matrix. Alternatively, these late effectors might, together with secondary metabolites induced at that stage, be used as a cocktail to defend the spores against other microbes which could colonize tumor tissue when it dries up and ruptures, releasing the spores. One could also consider that late effectors could fulfill a signaling function inside the aggregates to control the spore maturation process.
In conclusion, Ros1 emerges as the central regulator of a major developmental reprogramming leading to teliospore production and completion of the life cycle in U. maydis (Fig 10). Of particular interest is how U. maydis can survive in the hostile plant environment with reduced expression of a large set of effector proteins of which many have a critical virulence function early during colonization. In addition, elucidating the role of the “late effectors” which are specifically induced by Ros1 promises to provide new insights into how this facultative biotrophic fungus has established itself in its natural environment. Moreover, we are confident that deciphering the structure and dynamics of the regulatory network in which Ros1 functions will provide an understanding how this master regulator achieves such a broad control over gene expression. Another yet unresolved task is the identification of the upstream signals triggering ros1 induction during biotrophic development and thereby inducing late development including spore formation.
The Escherichia coli strain Top10 (Life technologies) and BL21(DE3)pLysS (Promega) were used for cloning purposes and for expression of recombinant Ros1 protein respectively. U. maydis strains used in this study are listed in S5 Table, they are derivates of haploid strains FB1 and FB2 [77] or AB33 [39]. Cells were grown in liquid YEPSL (0.4% yeast extract, 0.4% peptone, 2% sucrose) at 28°C on a rotary shaker at 220 rpm. For virulence assays, compatible haploid strains were grown separately in YEPSL to an OD600 of 1.0, transferred to the same volume of sterile water and mixed in equal amounts prior to injection into maize seedlings.
For premature ros1 expression studies, AB33 and strains derived from AB33 were grown in complete medium (CM) [78] supplemented with glucose (2%) to an OD600 of 0.5. Cells were collected by centrifugation, washed with H2O and resuspended in nitrate minimal medium (NM) [78] containing arabinose (2%) as sole carbon source. Cells were subsequently grown for 12h for microscopic observation. To induce ros1 after the switch to filamentous growth, AB33 or AB33 derived strains were incubated for 6 h in NM + glucose and then shifted to NM + arabinose for 6 h. All chemicals used for media preparation were of analytical grade and were obtained from Sigma-Aldrich.
PCR reactions were performed using the Phusion High-Fidelity DNA Polymerase (New England Biolabs). Templates were either FB1 genomic DNA or indicated plasmid DNAs. Point mutations were generated using the Quick change lightning kit (Agilent Technologies). Restriction enzymes were all supplied by New England Biolabs. U. maydis was transformed by protoplast-mediated transformation [79] Gene replacements and integrations into the ip locus [80] were verified by Southern blot analysis.
All primer sequences used to generate plasmids are listed in S6 Table. To generate the Ros1mCherry fusion construct, plasmid p123 [81] conferring resistance to carboxin and allowing integration into the U. maydis ip locus was used. mCherry-HA was amplified from plasmid p1139 (kindly provided by A. Djamei) using primers mCherry_EcoF / mCherry_NotR and cloned in place of gfp between EcoRI and NotI sites of p123 to generate pPotef-mCherry-HA. The ros1 open reading frame was amplified with primers ros1_XmaF / ros1_EcoR and cloned between sites XmaI and EcoRI of pPotef-mCherry-HA to generate pPotef-ros1-mCherry-HA. pPotef-ros1-mCherry-HA was linearized by SspI prior to transformation of U. maydis. For the deletion of ros1, a PCR-based strategy [82] and the SfiI insertion cassette system [79] were used. 1 kb long left border and right border fragments adjacent to ros1 were PCR-amplified using primer pairs Dros1LB_F/R and Dros1RB_F/R and FB1 genomic DNA as template. The resulting fragments were ligated to the hygromycin resistance cassette of pBS-hhn [82] via SfiI restriction sites and cloned into pCRII-TOPO (Life technologies) to generate pDros1. The deletion construct was PCR amplified from plasmid pDros1 using primers Dros1_F/R and transformed into U. maydis strains FB1 and FB2 to generate FB1Δros1 and FB2Δros1. The drag and drop cloning method [83] was used to generate plasmids pDUMAG_02775 and pDUMAG_01390 for the deletion of UMAG_02775 and UMAG_01390, respectively. Primer pairs D02775LB_F/R and D02775RB_F/R and primer pairs D01390LB_F/R and D01390RB_F/R were used to amplify left and right border fragments from UMAG_02775 and UMAG_01390, respectively. Left and right border fragments and the hygromycin resistance cassette were integrated in plasmid pSR426 (kindly provided by S. Reissmann) by homologous recombination in S. cerevisiae. The deletion constructs were excised from plasmids pDUMAG_02775 and pDUMAG_01390 after cleavage by Bsu36I and used for transformation of FB1 and FB2 to generate strains FB1ΔUMAG_02775, FB2ΔUMAG_02775, FB1ΔUMAG_01390 and FB2ΔUMAG_01390.
For complementation of ros1 mutant strains, plasmid p123-Bsu was generated from plasmid p123 [81] by introducing a silent point mutation in the cbx gene to create a Bsu36I site using primers p123Bsu_F/R. The 7.5 kb long genomic region separating ros1 from the upstream gene UMAG_05850 was used as promoter in complementation constructs. This region might contain an additional gene (UMAG_05852). To make sure that complementation constructs would not have two copies of this gene, the putative start codon of UMAG_05852 was deleted. To this end, the 6 kb intergenic region between UMAG_05850 and UMAG_05852 was amplified from genomic FB1 DNA with primers Cros1_KpnF / Cros1_SbfR and a second fragment containing UMAG_05852 and ros1 open reading frame was amplified using primers Cros1_SbfF / Cros1_NotR. Both fragments were ligated and cloned into pCRII-TOPO (Life technologies). In the resulting plasmid pCRII-TOPOros1 the sequence GCTGACGCATG containing the start codon of UMAG_05852 is changed to TAGCATAG. Using primers TOPOros1_mF / TOPOros1_mR, a silent point mutation was introduced in pCRII-TOPOros1 to remove a KpnI site located in UMAG_05852. The complementation construct was then excised from pCRII-TOPOros1 and cloned into p123-Bsu between KpnI and NotI sites. The resulting plasmid pCros1 was linearized with Bsu36I and transformed into ros1 deletion strains to generate strains FB1Δros1-ros1 and FB2Δros1-ros1.
For complementation of UMAG_02775 mutant strains, fragments corresponding to the UMAG_02775 promoter and open reading frame were PCR amplified from FB1 genomic DNA using primer pairs P02775_F/R and orf02775_F/R respectively and cloned in p123 between KpnI and NotI Sites. The resulting plasmid pCUMAG_02775 was linearized by SspI and transformed into FB1ΔUMAG_02775 and FB2ΔUMAG_02775 to generate strains FB1ΔUMAG_02775-UMAG_02775 and FB1ΔUMAG_02775-UMAG_02775.
For complementation of UMAG_01390 mutant strains, a fragment corresponding to UMAG_01390 promoter followed by the open reading frame was PCR amplified from FB1 genomic DNA using primer pairs C01390_F/R and cloned in p123 between KpnI and NotI sites. The resulting plasmid pCUMAG_01390 was linearized by SspI and transformed into FB1ΔUMAG_01390 and FB2ΔUMAG_01390 strains to generate strains FB1ΔUMAG_01390-UMAG_01390 and FB1ΔUMAG_01390-UMAG_01390.
For the generation of strains expressing a C-terminal Ros1HA fusion for the ChIP analysis, the ros1 open reading frame without the stop codon was amplified from FB1 genomic DNA using primers Ros1_XmaF / Ros1_NotR and subcloned in vector p1306 (Kindly provided by A. Djamei) between XmaI and NotI sites upstream of a triple HA sequence to generate pRos1-3HA. A fragment containing the last 1 kb of ros1 orf and the triple HA tag sequence was then amplified from pRos1-3HA using primers Endros1_NcoF / Endros1_PspR and cloned in pCros1 between NcoI and NotI sites. The resulting vector pCros1-3HA was linearized by Bsu36I and transformed in ros1 deletion strains to generate FB1Δros1-Ros1HA and FB2Δros1-Ros1HA.
To allow expression of ros1 from the crg1 promoter, Pcrg1 was integrated upstream of ros1 in the native locus of AB33 using pRU11 [39]. Right border (corresponding to the first 1000 bp of ros1) and left border fragments (corresponding to the 1000 bp upstream of the ros1 start codon) were amplified from FB1 genomic DNA using primer pairs crg-ros1LB_F/R and crg-ros1RB_F/R, respectively. The generated fragments were cloned between NdeI and EcoRI sites of plasmid pRU11. The resulting plasmid pPcrg1-ros1 was linearized with NcoI and transformed into AB33 to generate strain AB33Pcrg1Ros1.
To express ros1 from the mig2-6 promoter, Pmig2-6 [42] was amplified from FB1 genomic DNA using primers Pmig2-6_NdeF / Pmig2-6_XmaR and cloned into NdeI and XmaI sites of p123 to generate pPmig2-6. The ros1 gene was amplified with primers Ros1_XmaF / Ros1_NotR and cloned via XmaI and NotI sites downstream of Pmig2-6 in pPmig2-6. The resulting plasmid pPmig2-6-ros1 was linearized by SspI and transformed into ros1 deletion strains to generate strains FB1Δros1Pmig2-6Ros1 and FB2Δros1Pmig2-6Ros1.
To allow expression of the Sso1 nuclear membrane marker fused to mCherry, Psso1 was amplified from FB1 DNA with primers Psso1_KpnF and Psso1_NcoR. Potef was replaced by Psso1 in p123 using KpnI and NcoI sites to generate pPsso1. A fragment containing the mCherrySso1 fusion gene followed by the terminator of sso1 (Tsso1) was amplified from plasmid pBS-otef-mCherry-sso1-hyg [84] using primers McSso1_NcoF and McSso1_HpaR. The fragment containing mCherrySso1-Tsso1 was then integrated between NcoI and HpaI sites in place of GFP-Tnos of pPsso1 to generate pPsso1-mcherry-sso1. This vector was linearized by SspI and transformed into FB1, FB2, FB1Δros1 and FB2Δros1. To introduce the nuclear marker nucleoporin Nup107 fused to GFP in the resulting strains, pNup107GFP-ble [85] was used. The insert from this plasmid encoding the Nup107GFP including the regulatory sequences was amplified with primers Pnup_F/R and integrated in the native nup107 locus to generate strains FB1Pnup107Nup107eGFP-Psso1Sso1mCherry, FB2Pnup107Nup107eGFP-Psso1Sso1mCherry and the corresponding ros1 deletion strains. The sequence of all PCR-amplified regions was verified.
Haploid strains were grown in YEPSL medium to an OD600 of 1.0, washed and resuspended in sterile H2O. Compatible strains were mixed in a 1:1 ratio and syringe-inoculated into seven-day-old maize seedlings of the variety Gaspé Flint (originally provided by B. Burr, Brookhaven National Laboratories). Three independent infections were carried out for each strain and disease symptoms were evaluated after 12 days according to established disease rating criteria [10]http://journals.plos.org/plospathogens/article?id=10.1371/journal.ppat.1004272—ppat.1004272-Kamper1. Significant virulence differences between strains were assessed by one-way ANOVA applying the Tukey-Kramer test [36].
Total RNA was extracted from cells grown in axenic culture or from infected plant samples. Leaf material was ground in liquid nitrogen to a fine powder while cells in culture were pelleted by centrifugation and frozen in liquid nitrogen. Samples were resuspended in TRIzol reagent (Life technologies) and homogenized in a FastPrep-24 (MP Biomedicals). Total RNA was isolated according to the manufacturer's protocol. Genomic DNA contaminants were eliminated using the Ambion Turbo DNA free Kit (Life Technologies). For RNA sequencing RNA samples were further purified using the RNeasy Mini Kit (Qiagen) and the RNA quality was controlled using an Agilent 2100 Bioanalyzer.
Total RNA samples from plants infected in three biological replicates with FB1 X FB2 strains or the corresponding ros1 deletion strains were used to prepare sequencing libraries with the Illumina TruSeq RNA sample preparation Kit. Library preparation started with 2 μg total RNA. After poly-A selection (using poly-T oligo-attached magnetic beads), mRNA was purified and fragmented using divalent cations under elevated temperature. The RNA fragments were reverse transcribed using random primers. A second strand cDNA synthesis was carried out with DNA Polymerase I and RNase H. After end repair and A-tailing, indexing adapters were ligated to the cDNA. The products were then purified and amplified (14 PCR cycles) to create the final cDNA libraries. After library validation and quantification (Agilent 2100 Bioanalyzer), equimolar amounts of library were pooled. The pool was quantified using the Peqlab KAPA Library Quantification Kit and the Applied Biosystems 7900HT Sequence Detection System. The pool was sequenced using an Illumina TruSeq PE Cluster Kit v3 and an Illumina TruSeq SBS Kit v3-HS on an Illumina HiSeq 2000 sequencer with a paired-end (101 x 7 x 101 cycles) protocol.
Sequence reads were mapped to U. maydis protein encoding genes (ftp://ftpmips.gsf.de/fungi/Ustilaginaceae/Ustilago_maydis_521/) using CLC Genomics Workbench 7.5 (CLC bio). The unique gene reads for all of the 6970 annotated U. maydis genes from the 6 libraries were combined and analyzed in R using the Differentially Expressed Genes (DEG) algorithm edgeR [86]. Differentially expressed genes between FB1 x FB2 and FB1Δros1 x FB2Δros1 were selected on the basis of their fold change (FC ≥ 1.5) and p-value (< 0.01).
Expression data were submitted to GeneExpressionOmnibus (http://www.ncbi.nlm.nih.gov/geo/) under the accession number GSE76231.
7-day-old maize seedlings of the variety Gaspé Flint (originally provided by B. Burr, Brookhaven National Laboratories) were infected with mixtures of U. maydis strains FB1Δros1-Ros1HA x FB2Δros1-Ros1HA expressing an HA tagged Ros1 protein or FB1Δros1-Ros1 x FB2Δros1-Ros1 expressing a non-tagged Ros1 protein as negative control. Leaf samples (from 5 different plants) were collected at 8 dpi and incubated in fixation buffer (50 mM HEPES pH 7.5, 1% formaldehyde) for 10 min under vacuum. Excess formaldehyde was quenched by addition of 2 M glycine. Samples were then ground to a fine powder in liquid nitrogen. For chromatin preparation, powder was resuspended in lysis buffer (50 mM HEPES pH 7.5, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 5 mM benzamidine, 2 mM PMSF, 1 X Roche complete EDTA free protease inhibitors) and further treated by sonication to lyse the remaining cells using a microtip sonifier (Branson). Chromatin was then sheared in a bioruptor sonication bath (Diagenode) for 10 cycles (30 s on / 30 s off) at high power setting. An aliquot was saved to serve as input control and the rest of the chromatin solution was incubated with ChIP grade Protein A/G magnetic beads (Life technologies) coupled to a monoclonal anti-HA antibody (Sigma) for 10 h at 4°C. Beads were washed three times in lysis buffer, two times in high salt buffer (50 mM HEPES pH 7.5, 300 mM NaCl, 1 mM EDTA, 1% Triton X-100, 1 X Roche complete EDTA free protease inhibitors) and once in TE buffer (10 mM Tris-HCl, 1 mM EDTA, pH 7.5). DNA / Ros1HA complexes were eluted from the beads in TE SDS buffer (10 mM Tris-HCl, 1 mM EDTA, 1% SDS, pH 7.5) by incubation for 15 min at 65°C. Samples and input controls were de-crosslinked for 8–10 h at 65°C in TE SDS buffer containing 200 mM NaCl and 0.65 μg/μL proteinase K. DNA was purified using the ChIP DNA clean and concentrator kit (Zymo research). The experiment was done in three biological replicates which were sequenced separately.
DNA libraries were prepared using the Illumina TruSeq RNA sample preparation Kit v2 starting from the end repair step of the protocol. Up to 100 ng ChIP DNA was used as starting material. After end repair and A-tailing, indexing adapters were ligated. The products were then purified and amplified for 18 PCR cycles to create the final libraries. After validation (Agilent 2200 TapeStation) and quantification using the KAPA Library Quantification Kit (VWR) and the Applied Biosystems 7900HT Sequence Detection System, equimolar amounts of library were pooled. The pool was sequenced using the Illumina TruSeq PE Cluster Kit v3 and the Illumina TruSeq SBS Kit v3-HS (101 x 7 x 101 Cycles) on an Illumina HiSeq 2000 sequencer with a paired-end (101 x 7 x 101 cycles) protocol.
Sequencing data were mapped to U. maydis genome (ftp://ftpmips.gsf.de/fungi/Ustilaginaceae/Ustilago_maydis_521/) and analyzed using the ChIP-seq tool of the CLC genomics workbench 7.5 software (CLC bio). ChIP-peak discovery was based on read coverage enrichment (p-value) and shape of read distribution (peak shape score). ChIP-seq data were submitted to GeneExpressionOmnibus (http://www.ncbi.nlm.nih.gov/geo/) under the accession number GSE76231.
Gene expression analysis from infected plant material was performed as described in Brefort et al. (2014) [87], with some modifications. Briefly, samples of U. maydis cells grown in YEPSL and of infected tissue were ground to powder on liquid nitrogen and RNA was extracted with TRIzol (Life technologies). After extraction, the first-strand cDNA synthesis kit (Life technologies) was used to reverse transcribe 1–2 μg of total RNA with oligo(dT) Primers. The qPCR analysis was performed using the SYBR Green Supermix (Life technologies) and an iCycler (Bio-Rad). Cycling conditions were 2 min 95°C, followed by 45 cycles of 30 s 95°C/30 s 62°C/30 s 72°C. The experiment was done in three biological and three technical replicates and gene expression levels were calculated relative to the expression levels of the constitutively expressed fungal gene encoding peptidyl prolyl isomerase (ppi). Primers used were ppi_qF / R for the reference gene ppi and primer pairs ros1_qF / R, rum1_qF / qR, ust1_qF / qR, hgl1_qF / qR, tup1_qF / qR, 05550_qF / qR, 04503_qF / qR, 02212_qF / qR, 01070_qf/qR, pks1_qF / qR, 04101_qF/qR, biz1_qF / qR, rbf1_qF / qR, fox1_qF / qR, mig2-3_qF / qR, 04096_qF / qR, dik1_qF / qR, 02473_qF / qR, and 03046_qF / qR for ros1, rum1, ust1, hgl1, tup1, UMAG_05550, UMAG_04503, UMAG_02212, UMAG_01070, pks1, UMAG_04101, biz1, rbf1, fox1, mig2-3, UMAG_04096, dik1, UMAG_02473, UMAG_03046 respectively. All primer sequences are listed in S6 Table. Relative expression was determined using the ΔΔCt method [88]. t-tests were used to assess statistically relevant differences between expression levels at different time points (p ≤ 0.05).
Quantification of relative fungal biomass in infected maize leaves was performed as described previously [87], with some modifications. 2 cm long sections from 10 leaves with the most prominent symptoms were harvested from 10 different plants at the indicated time points. For genomic DNA extraction leaf material was frozen in liquid nitrogen, ground to a fine powder, and extracted using a phenol-based protocol modified from Hoffman and Winston (1987) [89]. The qPCR analysis was performed using the Platinum SYBR Green Supermix (Life technologies) in an iCycler (Bio-Rad). Cycling conditions were 2 min 95°C, followed by 45 cycles of 30 s 95°C / 30 s 62°C / 30 s 72°C. U. maydis biomass was quantified with primers ppi_qF/R amplifying the fungal ppi gene. Maize glyceraldehyde 3-phosphate dehydrogenase was amplified with primers Gapdh_qF/R and served as reference gene for normalization. The experiment was done in three biological and three technical replicates. t-tests were used to assess statistically relevant differences among strains (p ≤ 0.05).
Wheat germ agglutinin-Alexa Fluor 488 / propidium iodide staining of infected leaf material was performed as described previously [38]. For staining of the mucilaginous matrix, leaf tumors from maize seedlings infected with strains FB1 x FB2 or FB1∆ros1 x FB2∆ros1 were collected at 10 dpi. Samples were fixed in 4% glutaraldehyde and embedded in Epoxy resin. 1–2 μm thick sections were generated with a microtome and stained with methylene blue-azure II-basic fuchsin following the protocol described by Humphrey and Pittman (1974) [49].
To examine fungal colonization of leaf tissue, samples from infected plants were fixed in ethanol, transferred to 10% KOH, incubated at 85°C for 4 hours, washed twice with PBS buffer (140 mM NaCl, 16 mM Na2HPO4, 2 mM KH2PO4, 3.5 mM KCl, and 1 mM Na2-EDTA, pH 7.4), and incubated under vacuum in staining solution (10 μg/mL propidium iodide and 10 μg/mL WGA Alexa Fluor 488 in PBS, pH 7.4) according to Doehlemann et al. (2008) [38]. WGA Alexa Fluor 488 was purchased from Life technologies. To visualize the septa in strains expressing ros1 in axenic culture, cell walls were stained with calcofluor (Fluorescent brightener 28). Nuclei were stained with 4',6-diamidino-2-phenylindole (DAPI).
For microscopy, an Axioplan II microscope (Zeiss) with differential interference contrast optics was used. Fluorescence of GFP, mCherry, calcofluor and DAPI was observed using GFP (ET470/40BP, ET495LP, and ET525/50BP), rhodamine (HC562/40BP, HC593LP, and HC624/40BP), and DAPI (HC375/11BP, HC409BS, and HC447/60BP) filter sets (Semrock). Pictures were taken with a CoolSNAP-HQ charge-coupled device camera (Photometrics). Image processing was done with MetaMorph software (Universal Imaging).
Confocal microscopy was performed using a TCS-SP5 confocal microscope (Leica Microsystems). GFP and wheat germ agglutinin-Alexa Fluor 488 were excited at 488 nm and emitted fluorescence was detected in the 495–530 nm range. Propidium iodide and mCherry were excited at 561 nm and emission was detected in the 580–630 nm range. Images were processed using LAS-AF software (Leica Microsystems).
To allow the recombinant expression of the Ros1WOPR domain comprising amino acids 1–321 of Ros1, this domain was C-terminally fused to a His-tag (Ros1WOPR-His). To this end ros1 was amplified with primers WOPR_NdeF / WOPR_XhoR, cloned in pET28 (Novagen) to generate pRos1WORP-His and transformed into E. coli strain BL21(DE3)pLysS (Promega). Expression of Ros1WOPR-His was induced in exponentially growing cell cultures for 4h at 28°C in dYT medium supplemented with 0.15% glucose, 1 mM MgSO4 and 0.5 mM IPTG. To achieve cell lysis cell pellets were resuspended in BugBuster Master Mix reagent (Merck Millipore) supplemented with protease inhibitors (complete EDTA-free tablet, Roche) and incubated for 20 min at room temperature. The crude cell extract was then centrifuged (30000 x g; 30 min) and the supernatant was loaded on a Ni-NTA column (HisTrap FF Crude, GE Healthcare) equilibrated in wash buffer (50 mM Na2HPO4, 300 mM NaCl, 20 mM imidazole, pH 8.0) using an Äkta FPLC system (GE Healthcare). After unbound protein was washed off the Ros1WOPR–His protein was eluted with elution buffer (50 mM Na2HPO4, 300 mM NaCl, pH 8.0, 135 mM imidazole). The Ros1WOPR-His containing fractions were pooled and loaded on a Superdex75 gel filtration column (SuperdexTM 75 10/300GL, GE Healthcare) equilibrated in gel filtration buffer (20 mM HEPES, 20 mM NaCl, 0.1 μM PMSF and 1 mM DTT). The Ros1WOPR-His containing fractions were pooled and incubated in batch with source 15Q anion exchange beads (GE Healthcare) equilibrated in gel filtration buffer. Ros1WOPR-His remained in the supernatant, which was concentrated via Amicon Ultra-4 centrifugation units with an Ultracel-3 membrane (Merck Millipore). The protein was stored at 4°C for up to five days.
For ros1 WT-probe and ORF-probes, pCros1 was used as template. For the mutated versions of the ros1 WT-probe, corresponding fragments were generated by annealing several overlapping oligonucleotides carrying the desired mutations and cloning them into p123 [81] between restriction NdeI and BamHI sites The inserts in the resulting plasmids pm1, pm2 and pm1+2 were sequenced and plasmids were subsequently used as template for the respective PCR reactions using 5’ biotin labeled primers WT-probe_F/R for the WT-probe and the mutated versions m1, m2 and m1+2 and primers ORF-probe_F/R for the ORF_probe. Wild type competitor DNA corresponding to the same sequence as WT-probe was PCR amplified using unlabeled WT-probe_F/R primers. Probes corresponding to promoter regions upstream of UMAG_02854, UMAG_04040, UMAG_02538, cmu1, UMAG_03046, UMAG_03138, UMAG_12258, and UMAG_02775 were generated by PCR using 5’ biotin labeled primer pairs Probe-02854_F / R, Probe-04040_F / R, Probe-02538_F / R, Probe-cmu1_F / R, Probe-03046_F / R, Probe-03138_F / R, Probe-12258_F / R, Probe-02775_F1 / R1 and Probe-02775_F2 / R2 respectively. Competitors were generated by PCR using corresponding non labeled primers. All primer sequences are listed in S6 Table. EMSAs were performed using the Lightshift Chemiluminescent EMSA Kit (Thermofischer). Binding reactions were carried out in 20 mM HEPES buffer, pH 8.0 supplemented with 1 mM DTT, 50 mM NaCl, 25 ng/μL poly dI/dC, 2.5% glycerol 0.05% NP40, 5 mM MgCl2 and 1 μg/μL BSA. 10 fmol of biotin labeled dsDNA probe and 3 μg purified Ros1WOPR-His protein were used per binding reaction. For competition reactions, competitor fragment was added in a 500-fold molar excess. Binding reactions were incubated for 30 min at room temperature. In competition experiments Ros1WOPR was incubated for 20 min with the non-labeled competitor prior to addition of the probe. Binding reactions were separated on native 4% polyacrylamide 0.5 X TBE gels. Gels were transferred to a nylon membrane and biotin-labeled DNA fragments were detected using a streptavidin horseradish peroxidase conjugate and a highly sensitive chemiluminescent substrate as recommended by the manufacturer (Thermofischer).
ros1 (UMAG_05853): XM_011393913; UMAG_02775: XM_011390829; UMAG_01390: XM_011388973
|
10.1371/journal.ppat.1002958 | Molecular Basis for Genetic Resistance of Anopheles gambiae to Plasmodium: Structural Analysis of TEP1 Susceptible and Resistant Alleles | Thioester-containing protein 1 (TEP1) is a central component in the innate immune response of Anopheles gambiae to Plasmodium infection. Two classes of TEP1 alleles, TEP1*S and TEP1*R, are found in both laboratory strains and wild isolates, related by a greater or lesser susceptibility, respectively to both P. berghei and P. falciparum infection. We report the crystal structure of the full-length TEP1*S1 allele which, while similar to the previously determined structure of full-length TEP1*R1, displays flexibility in the N-terminal fragment comprising domains MG1-MG6. Amino acid differences between TEP1*R1 and TEP1*S1 are localized to the TED-MG8 domain interface that protects the thioester bond from hydrolysis and structural changes are apparent at this interface. As a consequence cleaved TEP1*S1 (TEP1*S1cut) is significantly more susceptible to hydrolysis of its intramolecular thioester bond than TEP1*R1cut. TEP1*S1cut is stabilized in solution by the heterodimeric LRIM1/APL1C complex, which preserves the thioester bond within TEP1*S1cut. These results suggest a mechanism by which selective pressure on the TEP1 gene results in functional variation that may influence the vector competence of A. gambiae towards Plasmodium infection.
| Anopheles mosquitoes transmit malaria, the world's most devastating parasitic disease, of which Anopheles gambiae is the principal vector for malaria in Sub-Saharan Africa. Different populations of mosquitoes vary widely in how readily they become infected with malaria parasites, while some strains do not transmit malaria at all. The mosquitoes' innate immune system is a significant factor that may influence the level of malaria infection; in particular the thioester-containing protein 1 (TEP1) targets malaria parasites for destruction during their initial invasion of the body cavity. The TEP1 gene varies significantly across mosquito populations with two major classes of alleles, TEP1*S and TEP1*R. We report the three-dimensional molecular structure of the TEP1*S1 protein and compare it to the previously determined TEP1*R1 structure. Differences between the structures are localized around the active site and thioester bond, and correlate with a difference in stability of this bond within the two proteins and their interaction with a heterodimer of two other immune genes, LRIM1 and APL1C. These results shed light on the mechanism of mosquitoes' natural immunity to malaria infection.
| Thioester-containing proteins (TEPs) are a major component of the innate immune response of insects to invasion by bacteria and protozoa [1], [2]. Anopheles gambiae thioester-containing protein 1 (TEP1) is a complement-like protein that plays a central role in the opsonization of gram-negative bacteria in the hemolymph [3]. TEP1 also binds to the surface of Plasmodium ookinetes that traverse the midgut epithelium following ingestion of an infectious blood meal, targeting those ookinetes for lysis and, in certain mosquito strains, melanization [4]. TEP1 activity has been demonstrated against both P. berghei [4] and P. falciparum [5], [6]. Deciphering the molecular basis of the TEP1-mediated immune response is relevant to understanding the determinants of vector competence and a potential source of novel vector-based malaria control strategies.
The crystal structure of full-length TEP1 revealed significant structural homology to complement factor C3 [7]. TEP1 is composed of a series of eight macroglobulin (MG) domains, the β-sheet CUB domain and α-helical thioester domain (TED). The TED contains an intramolecular β-cysteinyl-γ-glutamyl thioester bond that is protected from inadvertent hydrolysis by sequestration within a protein-protein interface formed by the TED and MG8 domains. Based upon the known mechanism of complement factors [8], activation of the thioester in TEP1 is presumed to involve a large conformational change causing dissociation of the TED-MG8 interface in the direct proximity of a pathogen, whereupon the thioester may react with nucleophilic groups on, and covalently attach TEP1 to, the surface of the pathogen.
TEP1 lacks two additional domains that are present in complement factors, the anaphylatoxin (ANA) and C345C domains. The ANA domain in particular plays a key role in the activation of complement factor C3, which is cleaved intracellularly in a protease-sensitive region between the MG6 and ANA domains prior to secretion. The ANA domain contacts both the MG3 and MG8 domains in the structure of mature, circulating C3 [9]. Activation of C3 occurs after regulated proteolysis immediately following the ANA domain whereby the anaphylatoxin C3a is released. Dissociation of C3a destabilizes the remaining C3b fragment and leads to a large-scale conformational change and rapid activation of the thioester bond [10], [11].
In contrast, TEP1 is secreted as a full-length protein into the mosquito hemolymph where it is cleaved by as yet unknown protease(s). Cleavage of TEP1, producing TEP1cut, does not instantly lead to activation of the thioester [12], suggesting that full-length TEP1 is a pro-form [13] that must undergo conversion to an active species following cleavage within the protease-sensitive region. TEP1cut is meta-stable in solution and precipitates over time. This precipitation is concomitant with hydrolysis of the thioester bond and is prevented in vivo by formation of a ternary complex between TEP1cut and a heterodimer of two leucine-rich repeat proteins, LRIM1 and APL1C [12], [14]–[16]. The ternary complex TEP1cut/LRIM1/APL1C was formed in vitro only after chemical inactivation of the thioester bond of TEP1cut by treatment with methylamine (MeNH2) [15]. This raised the question as to whether LRIM1/APL1C stabilizes a conformation of TEP1cut that contains an active thioester, or a distinct conformation in which the thioester has either reacted with substrate or been hydrolyzed by water.
The TEP1 gene is highly polymorphic, with distinct alleles conferring variable levels of protection from pathogens. Two alleles were originally identified in laboratory mosquito strains (indicated in brackets) as being susceptible (G3) and refractory (L3–5) to infection with P. berghei [4]. Recently, additional alleles were identified from laboratory strains conforming to two major classes S and R: TEP1*S1 (PEST), TEP1*S2 (4Arr), TEP1*S3 (G3), TEP1*R1 (L3–5) and TEP1*R2 (4Arr) with TEP1*S2 and TEP1*R2 alleles displaying intermediate phenotypes with respect to P. berghei infection [17]. The refractory allele, TEP1*R1, has been expressed in vitro and utilized in structural and functional studies [7], [12], [15]. The TEP1*S1 and TEP1*R1 proteins share 93% sequence identity with the majority of amino acid differences being confined to three hypervariable loops within the TED domain [3], [4]. Two of these loops, the pre-α4 loop and the catalytic loop, are situated in close proximity to the thioester itself at the TED-MG8 domain interface [7] and are complemented by amino acid differences within the MG8 domain that interact with the pre-α4 and catalytic loops and also conform to the TEP1*S/R division of alleles.
A recent study of wild mosquito populations from five locations in West, Central, and East Africa detected three similar sets of TEP1*S/R alleles as observed in the laboratory strains; s (TEP1*S), rA (TEP1*R2) and rB (TEP1*R1) [6]. Furthermore, specific geographical variation in allelic frequencies and a statistically significant decrease in P. falciparum oocysts within s/rB heterozygous vs. s/s homozygous mosquitoes were observed. The concordance of laboratory and field studies prompted us to further investigate the structure and properties of TEP1*S1 in comparison to TEP1*R1. Here we report the crystal structure of full-length TEP1*S1. We also report the relative reactivity of the thioester bond in TEP1*S1 and TEP1*R1 to hydrolysis and the association of TEP1*S1cut with LRIM1/APL1C. These results suggest a potential mechanism by which allelic variation in TEP1, particularly in the pre-α4 and catalytic loops, may translate to functional variation towards distinct pathogens.
Full-length TEP1*S1 crystallized in space group P43 and the structure was determined at 3.7 Å resolution (PDB 4D93), (see Materials and Methods, Table S1 and Figure S1). To facilitate comparison with recent studies of TEP1 alleles the structure of TEP1*S1 is numbered according to the complete protein sequence. The previously determined structure of TEP1*R1 [7] has been re-refined (PDB 4D94) to correct some errors in the original model and was used as a reference structure for refinement of TEP1*S1. Residues in the new TEP1*R1 and TEP1*S1 models are numbered according to the complete peptide chain (including signal peptide) for comparison with other reports.
The refined model of TEP1*S1 has three molecules, two of which (chains A, C) comprise residues 22–1338 with the exception of four gaps; residues 561–562, 575–582 in the linker (LNK) domain, 606–628 within the protease-sensitive region, and 822–829 in the CUB domain. A third molecule (chain B) with higher B-factors has poor or absent density for much of domains MG1, MG2, MG4, MG5 and MG6, but is otherwise complete for residues 629–1338 with the exception of 822–829 in the CUB domain. No significant difference in conformation is apparent between the three molecules. Further description of the structure is based upon molecule A.
The overall structure of TEP1*S1 is very similar to that of TEP1*R1 (Figure 1A). The first six MG domains form a super-helical quaternary structure, with MG6 split by the insertion of the linker and protease-sensitive regions (585–607). Thus the TEP1cut N-terminal fragment (β chain) and C-terminal fragment (α chain) are interleaved within the MG6 domain. Following the MG6 domain two additional domains, MG7 and MG8, are divided by the nested insertions of the CUB and TED domains. SDS-PAGE of redissolved crystals confirms TEP1*S1 within the crystal to be full-length protein (Figure 1B). Some protein domain motion is evident between TEP1*S1 and TEP1*R1 (Figure 1C), confirmed by analysis using the program DYNDOM [18] (Table S2). The TEP1*S1 MG3, MG7, CUB, TED, and MG8 domains are superimposable as a rigid body upon TEP1*R1. The MG1, MG2, MG5 and MG6 domains also form a rigid body but are rotated 11° relative to TEP1*R1. One hinge for this movement is the MG2–3 linker (217–222) a short sequence identically conserved with human complement factor C3. The second hinge is the MG4 domain itself which is rotated 26° relative to the two other rigid domains. As there is no sequence variation between TEP1*S1 and TEP1*R1 at the interface of the MG2, MG6 and TED domains or the MG3–MG4 domain interface, these rearrangements likely reflect inherent flexibility and different packing constraints within the TEP1*S1 crystal vs. TEP1*R1.
Amino acid variation between TEP1*R1 and TEP1*S1 was previously noted to be largely confined to domains surrounding the TED [4], [7]. Analysis of alleles TEP1*R1–2 and TEP1*S1–3 [17] confirms and extends this observation. No amino acid substitutions that separate TEP1*R and TEP1*S alleles occur within domains MG1–MG6, and except for five substitutions to similar residues, all variation between TEP1*S and TEP1*R alleles are confined to the TED, CUB and MG8 domains. These polymorphisms are hereafter described as mutations to the TEP1*R1 allele, i.e. R{res#}S.
We focused on amino acid differences between TEP1*S1 and TEP1*R1 that are preserved in all S and R laboratory alleles [17] and wild mosquito populations [6]. Of 42 such polymorphisms within the TED (Table S3), 18 occur within three previously identified hypervariable loops termed the pre-α4 loop (914–920), the catalytic loop (966–974) and the β-hairpin (1054–1069). An additional polymorphism S1108R was noted from the crystal structure of TEP1*R1 as potentially significant [7]. The remaining 23 polymorphisms are generally localized in short loops between the TED α-helices and introduce no significant alteration to the structure (Table S3), with the possible exception of four (F960S, E1005V, K1009V, T1012N) on the face of helix α7 and adjacent to the post-α5 turn that form a crystal contact within the TEP1*R1 structure.
The pre-α4 and catalytic loops form part of the TED-MG8 domain interface (Figure 2A) that protects the thioester from premature activation or hydrolysis. Both loops are ordered in the TEP1*S1 structure. The TEP1*R1 catalytic loop contains a sequence of five residues including Lys 966 and Glu 970 (966KAGAE970). These charged residues also occur in the TEP1*S1 catalytic loop but their positions are switched (966ETGKV970). TEP1*S1 Glu 966 adopts a different conformation than Lys 966 in TEP1*R1 (Figure 2B), directed into a pocket occupied by a Cl− ion in TEP1*R1, within hydrogen bonding distance of Ser 921 Oγ and Phe 923 N (the same conformation observed for Glu 1098 in complement factor C3). In contrast Ser 921 is within hydrogen bonding distance of TEP1*R1 Tyr 971 but not TEP1*S1 Trp 971. These differences impart a ∼1.8 Å displacement of residues 966–968 in TEP1*S1 relative to TEP1*R1.
In the pre-α4 loop the substitution N919G permits a different backbone conformation for TEP1*S1 with Gly 919 O within hydrogen bonding distance of Val 914 N (Figure 2C). The L914V and S1108R substitutions were previously noted as potentially affecting the environment of the thioester [7]. The displacement of the catalytic loop in TEP1*S1 leads L914V to introduce a small cavity within the TED-MG8 interface between Trp 971 and the thioester bond. The S1108R substitution does not cause any perturbation in the interface however, the Arg side chain adopts a conformation within hydrogen bonding distance of the carbonyl oxygen of Tyr 1307 instead of a water molecule as seen in TEP1*R1.
Substitutions in the TED domain at the TED-MG8 interface are complemented by substitutions within the MG8 domain (Table S4). Three pairs of substitutions noted in the TEP1*R1 structure [7] are preserved between the TEP1*S and TEP1*R alleles, two of which produce significant differences in the TEP1*S1 structure (Figure 2D). The K1260N substitution preserves the hydrogen bonding distance to Gly 858 N in the thioester motif but not to Tyr 884. The N1275Y substitution is no longer compatible with hydrogen bonding to Trp 915 in the pre-α4 loop, and the conformation adopted by TEP1*S1 Tyr 1275 forms is neither favorable for alternative hydrogen bond formation nor π-stacking interactions with nearby aromatic residues. Though the substitution N1276K appears to introduce a repulsive electrostatic interaction with Lys 970 in the TEP1*S1 catalytic loop we note that (i) the density for this side chain is poor, (ii) the nearby substitutions R1227S/R1228Q compensate for the introduction of this charge and (iii) Asn is conserved at this position in TEP1*S2–S3 [17] (the corresponding region was not sequenced for wild alleles reported by White et al. [6]).
An additional 11 polymorphisms conserved between S and R alleles occur in the MG8 domain but introduce no discernible alterations (Table S4). Amino acid variation within the CUB domain is localized to peripheral residues, none of the S/R-conserved polymorphisms are observed in the central β-strands β5, β6, β7 or β10 (Table S5). Two pairs of substitutions, V797A/R800K and V1183I/N1187D, are located on adjacent strands linking the CUB domain to the MG7 and MG8 domains, respectively, with the substitution T831K adjacent at the end of the β4–β5 turn. This is a site of large structural changes in the conversion of complement factor C3 to C3b [10], [11], and the site of C3b cleavage by factor I [19].
To assess the functional role of TEP1 polymorphisms in vitro, we sought to determine the relative stability of the TEP1*S1 compared to TEP1*R1. In addition to wild-type alleles we generated the following TEP1 variants: (i) TEP1*R1 with thioester cysteine mutation C859A, (ii) TEP1*R1-sTED2, in which residues 878–1108 in the TED and 1227–28, 1260–61 and 1275–76 in MG8 were replaced with TEP1*S1, and (iii) TEP1*R1 with MG3 glycosylation mutant N312D (Figure 3A). We previously observed that limited proteolysis of TEP1 in the protease-sensitive region leads to slow hydrolysis of the thioester bond and, in the absence of the LRIM1/APL1C complex, hydrolysis of the thioester leads to precipitation [15]. We purified TEP1*R1-C859A and observed that, while it was a stable full-length protein, the protein precipitated rapidly following proteolysis (Figure S2), suggesting that thioester hydrolysis is the rate-limiting step in the precipitation of TEP1cut.
We therefore measured the rate of precipitation of TEP1*R1cut and TEP1*S1cut to determine the rate of thioester hydrolysis in TEP1cut at 20°C, the same temperature used for in vivo studies of P. berghei infection. The half-life of TEP1*S1cut is 8.5 h (Figure 3B), significantly shorter than the half-life of TEP1*R1cut (6.5 days), suggesting that TEP1*S1cut is more susceptible to hydrolysis of the thioester bond than TEP1*R1cut (Figure 3C). The soluble fractions of the TEP1cut proteins analyzed with silver-stained SDS-PAGE also reflects the shorter half-life of TEP1*S1cut (Figure S3). The half-life of TEP1*R1-sTED2cut is 12 h (Figure 3B), confirming that the increased reactivity of TEP1*S1 towards hydrolysis is largely due to variation within the TED domain and the TED-MG8 interface in particular. The glycosylation site Asn 312 was previously noted to form a significant fraction of the interface between the MG3 and MG8 domains [7]. The half-life of the glycosylation mutant TEP1*R1cut-N312D is 8 days however (Figure 3C), indicating that removal of this glycosylation site does not affect the stability of the thioester in TEP1*R1cut.
Precipitation of TEP1cut is an indirect effect of thioester hydrolysis and may not correlate quantitatively with the rate of reaction of the thioester. We therefore measured the fraction of TEP1*S1cut containing an intact thioester by treatment of samples with MeNH2 as a function of time. Methylated and hydrolyzed TEP1cut were simultaneously quantified by monitoring the modification of Gln 862 with quantitative mass spectrometry (see Materials and Methods). The fraction of methylated TEP1*S1 decreased with time with an estimated half-life of 9 h (Figure 3D), in close agreement with the rate of precipitation of the protein. This supports the conclusion that hydrolysis of the thioester bond is the rate-limiting step in precipitation of TEP1cut.
We previously observed the ternary complex TEP1cut/LRIM1/APL1C was formed only after chemical inactivation of the thioester bond of TEP1*R1cut by MeNH2 [15], demonstrating that LRIM1/APL1C interacted with a reacted form of TEP1*R1cut without an intact thioester. To test whether this was also the case for TEP1*S1, we prepared TEP1*S1cut and incubated for 36 h at 20°C in the absence or presence of LRIM1/APL1C (Figure 4A). TEP1*S1cut incubated without LRIM1/APL1C precipitated (Figure 4A, lanes 1–2), while TEP1*S1cut mixed with LRIM1/APL1C remains soluble (Figure 4A, lanes 5–6).
The presence of an intact thioester bond in thioester-containing proteins can be determined by heating under denaturing conditions in the absence of reducing agent, promoting autolytic cleavage of the peptide chain at the site of the thioester bond [20]. The thioester bond is hydrolyzed in precipitated TEP1*S1cut (Figure 4B, lane 1). In contrast, soluble TEP1*S1cut in complex with LRIM1/APL1C possesses an intact thioester, as shown by heat-induced fragmentation of the C-terminal fragment (Figure 4B, lane 6). We conclude that the conformational changes in TEP1 required for the binding of LRIM1/APL1C is distinct from that involving reaction of the thioester. Hence the complex between LRIM1/APL1C and TEP1*S1cut is a distinct species from the complex of LRIM1/APL1C and TEP1*R1cut(MeNH2) [15].
The preceding experiments suggest that formation of the ternary complex between TEP1*S1cut and LRIM1/APL1C is due to a conformational change with a similar half-life as the measured rate of thioester hydrolysis. Thus previous attempts to produce a ternary complex between TEP1*R1cut and LRIM1/APL1C were unsuccessful simply because the period of incubation was too short. Accordingly FLAG immunoprecipitation assays were performed with 6×His-tagged TEP1cut proteins and FLAG-tagged LRIM1/APL1C. TEP1*S1cut co-immunoprecipitated with LRIM1/APL1C within 24 h (∼3× t1/2 for thioester hydrolysis) whereas TEP1*R1cut remained in the supernatant after 48 h (Figure 4C). However, after incubation at 20°C for 24 days (∼4× t1/2 for thioester hydrolysis) TEP1*R1cut remained soluble and was co-immunoprecipitated with LRIM1/APL1C (Figure 4D). Thus the conformational change following limited proteolysis in vitro that allows TEP1*S1cut and TEP1*R1cut to bind LRIM1/APL1C is comparable to their respective rates of thioester hydrolysis and precipitation in the absence of LRIM1/APL1C.
As a central component of humoral immunity in A. gambiae, the TEP1 gene is under selective pressure. Significant variation within two major allelic forms, TEP1*S and TEP1*R, are found in both laboratory and wild mosquito populations. Comparison of the structures of TEP1*S1 and TEP1*R1 reveals the consequences of this variation on the pro-form of TEP1 and stabilization of the intramolecular thioester bond. We observe distinct side chain and backbone conformations of two hypervariable loops within the thioester domain and two complementary substitutions within the MG8 domain that directly influence the TED-MG8 interface and the surrounding environment of the thioester bond.
An important caveat in analysis of the present structures is that the role of specific polymorphisms may be relevant to another conformation of TEP1 than is represented in the full-length protein. At present three soluble forms of TEP1cut have been identified in vitro ([12], [15] and this study). The first form contains an intact thioester and does not bind LRIM1/APL1C, (e.g. TEP1*R1cut 0–48 h post-cleavage). The second form contains a thioester but requires LRIM1/APL1C for stability in solution (e.g. TEP1*S1cut 24–36 h post-cleavage). The third form does not contain a thioester and also requires binding of LRIM1/APL1C for stability in solution (e.g. TEP1*R1cut(MeNH2) 12 h post-cleavage).
Distinct phenotypes for TEP1*S and TEP1*R alleles are observed for the response to both P. berghei [4] and to P. falciparum [6]. Our results provide the first evidence for a distinct chemical property of TEP1*S and TEP1*R proteins; the rate of thioester hydrolysis and precipitation in the absence of LRIM1/APL1C. Furthermore this difference affects the relative amount of the three in vitro soluble TEP1cut forms arising from cleavage in the protease-sensitive region. Within 24–36 h post-cleavage at 20°C the major soluble form of TEP1*S1cut has an intact thioester and binds LRIM1/APL1C, whereas TEP1*R1cut has an intact thioester but does not bind LRIM1/APL1C. In the absence of LRIM1/APL1C ∼90% of TEP1*S1cut has undergone hydrolysis of the thioester and precipitated from solution within 24 h at 20°C, whereas ∼90% TEP1*R1cut has an intact thioester bond and is soluble [12]. Hence, our results suggest that phenotypic variation in TEP1 alleles can result not only by activity in a single pathway but by distinct mechanisms arising from different forms present in the hemolymph.
Our in vitro studies may directly pertain to in vivo studies of P. berghei infection that are also conducted at ∼20°C [21] with microscopic analysis of TEP1 binding at 24–48 hours post-infection [4], [12], [22]. Our results are consistent with a model for activation of TEP1*S as proposed by Fraiture et al. (2009) [12] (Figure 5). Full-length TEP1*S represents a pro-form. Cleavage within the protease-sensitive region produces a meta-stable species similar to the pro-form that does not interact with LRIM1/APL1C. A slow (8 h) spontaneous conformational change generates a mature form of TEP1*S and exposes a cryptic binding site for LRIM1/APL1C. In the absence of LRIM1/APL1C however, the thioester bond in the mature form is susceptible to hydrolysis, presumably coupled to a large conformational change, producing a reacted form that rapidly aggregates and precipitates from solution. This model is consistent with the roles of TEP1*S3, LRIM1 and APL1C in the immune response of A. gambiae G3 to P. berghei ookinetes [4], [12], including the concept of basal immunity [22], as spontaneous formation of the active immune complex TEP1*S3cut/LRIM1/APL1C at 20°C is slow relative to the residence of ookinetes beneath the basal lamina.
TEP1*R1cut also forms a complex with LRIM1/APL1C that presumably contains an active thioester by a spontaneous conformational change with a half-life of 6.5 days at 20°C (Figure 5). We previously observed that TEP1*R1cut retained an active thioester after 60 h incubation with LRIM1/APL1C and a small component of a high molecular weight complex with LRIM1/APL1C [15]. Originally interpreted as hydrolysis of TEP1*R1cut, this may now be considered to be slow formation of the same complex formed by TEP1*S1cut. Such a slow rate of complex formation however, cannot account for the LRIM1/APL1C-dependent activity of TEP*R1 against P. berghei ookinetes [14] that traverse the midgut epithelium 18–24 h post blood-meal. Hence additional factor(s) must exist that accelerate the conformational change of TEP1*R1cut in vivo.
In the context of the present model, such factor(s) could act in three ways. First, protease(s) that cleave TEP1 may comprise or recruit chaperone(s) that accelerate the maturation of TEP1, revealing the binding site for LRIM1/APL1C; TEP1 activation would remain LRIM1/APL1C-dependent. Second, factors could accelerate both maturation and activation of TEP1 directly in an LRIM1/APL1C-independent manner. Third, factors may interact with a complex of reacted TEP1cut and LRIM1/APL1C to activate other TEP1 molecules, i.e. a “TEP1 convertase” as proposed previously to explain the interaction of TEP1*R1cut(MeNH2) with LRIM1/APL1C [15] (Figure 5).
Distinct TEP1*S/R phenotypes are observed for both the LRIM1/APL1C-dependent response to P. berghei [12], [14] and the response to P. falciparum that is LRIM1/APL1C-independent in Yaoundé and Ngousso strains [23], [24] that carry TEP1*S alleles [6]. This suggests a functional role of TEP1*S/R polymorphisms in the active form of TEP1, i.e. direct interaction of the pre-α4 and catalytic loops with the thioester and pathogen surfaces at the point of covalent attachment. The selective pressure that has given rise to these polymorphisms is not only (even unlikely) Plasmodium, but environmental pathogens such as bacteria encountered in both adult and pre-adult stages. Our results suggest a possible trade-off between selection for reactivity of the thioester upon activation and steady-state stability of the thioester in circulating TEP1cut. This may be relevant to immune responses based upon basal immunity, as is indicated in the case of Plasmodium [22], compared to responses based upon infection-induced upregulation of TEP1 expression.
Many outstanding questions remain regarding the mechanism of TEP1-mediated immune responses. The structure of the TEP1cut/LRIM1/APL1C ternary complex and the interaction of LRIM1/APL1C with reacted TEP1*R1cut [15], the source of phenotypic differences between different TEP1*S and TEP1*R alleles, and the role of polymorphism in the TED β-hairpin remains unknown. The interaction of LRIM1/APL1C with MeNH2-treated TEP1*R1cut [15] and with distinct TEP proteins TEP3 and TEP4 [16] suggest additional roles for LRIM1/APL1C in TEP1-mediated immunity besides stabilization of a re-circulating active immune complex. Further structural and functional studies of TEP1, LRIM1/APL1C and the identification of additional factors are required to address these questions.
TEP1*S1 was generated by total gene synthesis (Genscript) and subcloned into pFastbac1 with a C-terminal 6×His tag. TEP1*R1-sTED2 was constructed as follows using QuickChange site-directed mutagenesis (Stratagene). An SphI site was inserted into TEP1*R1-pFastbac1 corresponding to TEP1*S1 H878Y and removed from the pFastbac1 MCS. Digestion of both vectors with SphI/AfeI allowed replacement of TEP1*R1 residues 878–1108 with the corresponding sequence from TEP1*S1. Finally (i) 1227–8, 1260–1 and 1275–6 in the MG8 domain (TED-MG8 interface) were mutated to the corresponding residues in TEP1*S1, and (ii) residues 960, 1005, 1009, 1012 in the TED were mutated back to the corresponding residues of TEP1*R1. All TEP1 and LRIM1/APL1C constructs were expressed using the baculovirus expression system. Purification, limited proteolysis, thioester autolytic cleavage assay and immunoprecipitation experiments were performed as previously described [7], [12], [15].
Following limited proteolysis and re-purification TEP1 samples were concentrated to an OD280 of 0.5–1.0 and stored at 20°C. To measure rate of precipitation as a result of thioester hydrolysis upon proteolysis, samples and matching blank (filtrate from concentration) were centrifuged at 17,000×g, 20°C for 10 min and A280-A330 recorded in a standard UV spectrophotometer (Shimadzu UV1800). Separate time points are all derived from the same protein batch and purification and qualitatively similar results derived from 2–3 independent biological replicates. Half-lives were calculated from samples with a decay to <25% initial value and fit to log-linear plot with R2>0.99 (except TEP1*R1-N312D, final value 30% initial, R2 = 0.95).
To determine the rate of thioester hydrolysis by quantitative mass spectrometry, TEP1*S1 was cleaved as before and MeNH2 was added at specific time points to react with intact thioester bonds, methylating Gln 862. Samples were TCA precipitated and redissolved in 0.4 M NH4HCO3 containing 8 M urea followed by reduction and alkylation with DTT and iodoacetamide, respectively. Trypsin digestion was performed for 16 h at 37°C at a 10-fold molar excess of protein to trypsin. TFA and acetonitrile was added to final concentrations of 0.5% and 5%, respectively, followed by purification with C18 spin columns (Pierce) and elution in 80% acetonitrile. Tryptic peptides corresponding to hydrolysis (deaminated, [Dea] = +1 m/z) or methylation ([Me] = +14 m/z) of TEP1*S1 Gln 862 were characterized by time-of-flight LC-MS. Three specific fragments selected for quantitative analysis on an AB SCIEX 5500 Q-TRAP instrument coupled to an online Waters nanoACQUITY Ultra High Pressure Liquid Chromatography system and analysis with Multiquant 2.0 software. Assuming the deaminated and methylated peptides have similar ionization efficiency, the fraction of intact thioester is equal to IMe/(IMe+IDea). Reported data is the average of three fragments, two instrument replicates.
TEP1*S1 (PDB 4D93), TEP1*R1 (PDB 4D94).
|
10.1371/journal.ppat.1006168 | Identification of several high-risk HPV inhibitors and drug targets with a novel high-throughput screening assay | Human papillomaviruses (HPVs) are oncogenic viruses that cause numerous different cancers as well as benign lesions in the epithelia. To date, there is no effective cure for an ongoing HPV infection. Here, we describe the generation process of a platform for the development of anti-HPV drugs. This system consists of engineered full-length HPV genomes that express reporter genes for evaluation of the viral copy number in all three HPV replication stages. We demonstrate the usefulness of this system by conducting high-throughput screens to identify novel high-risk HPV-specific inhibitors. At least five of the inhibitors block the function of Tdp1 and PARP1, which have been identified as essential cellular proteins for HPV replication and promising candidates for the development of antivirals against HPV and possibly against HPV-related cancers.
| Human papillomaviruses are causative agents of many different cancers; they are most commonly associated with cervical cancer which leads to about quarter of a million deaths each year. Regardless of extensive studies for decades there is no specific cure against HPV infection. During this research, we have engineered modified HPV marker genomes that express Renilla luciferase reporter gene which expression level correlates directly with viral genome copy number. We have used such modified HPV genome in high-throughput screening of NCI Diversity Set IV chemical library and have identified a number of novel high-risk HPV-specific chemical compounds and drug targets. Such Renilla-expressing marker genomes could be used in various cell systems suitable for HPV replication studies to conduct high-throughput screens and quantify viral genome copy number quickly and effectively.
| Human papillomaviruses (HPVs) are small, double-stranded DNA viruses that infect the epithelium of the skin and mucosa. To date, at least 202 HPVs have been characterized, but studies suggest that the true number is considerably higher[1,2]. HPVs induce benign lesions in the mucosal and cutaneous epithelia, and most of the infections are cleared by the immune system within a year after infection. However, a small fraction of infections become persistent and may lead to the transformation of cells and the development of invasive cancers. The vast majority of HPV-associated cancer cases are related to oncogenic mucosal high-risk HPVs from genus alpha (types 16, 18, 31, 33, 35, 39, 45, 51, 52, 56, 58, 59, and 68); types 16, 18, 31, 33 and 45 are estimated to cause more than 99% of cervical cancers [3–5]. Cervical cancer was the seventh most common cancer and the fourth most common among women in the year 2012. There are an estimated 528,000 new cases per year, with 80% in developing countries, leading to 266,000 deaths[6]. In the United States alone, 6 million new HPV cases are diagnosed every year[7]. In addition to alpha PVs, infection with cutaneous beta PVs is also prevalent in the population. These viruses have not been as thoroughly studied as alpha PVs, but an increasing number of studies suggest their association with skin cancer[8,9]. In addition to serious health problems, HPV-related infections and cancers are a serious economic burden: in the United States, a total of 3.4 billion is spent annually on the diagnosis and treatment of HPV-related cancers, which does not even account for the cost for treating various warts and other benign papillomas[10]. These numbers suggest that there is a clear need for better prevention and treatment solutions regarding HPV-related diseases.
Regardless of being studied for decades, there is still no effective cure for an ongoing HPV infection. There are approved invasive treatments, such as cryotherapy, larger excision procedures, laser therapy and electrosurgery, which do not eliminate HPV DNA completely, leading to a 40% chance of recurrence of infection[11,12]. Immune system stimulants (imiquimod for example) as well as trichloroacetic acid and podophyllotoxin have 50% efficiency and a relatively high recurrence rate[13,14]. In addition to therapy, three vaccines against HPV are available: Gardasil (against subtypes 6, 11, 16 and 18), Gardasil 9 (against subtypes 16, 18, 31, 33, 45, 52, 58, 6 and 11) and Cervarix (against subtypes 16 and 18). These vaccines have proven to be very useful tools in the prevention of HPV infections[15,16], but they are prophylactic, and their availability is limited, especially in developing regions, which have the highest cervical cancer prevalence[17].
Although no effective HPV inhibitors have been developed, several compounds and targets have been analyzed. E1 and E2 are the only two viral proteins necessary for HPV genome replication. The first attempts in the development of HPV inhibitors were focused on E1, specifically targeting its ATPase and DNA helicase activities[18,19]. These inhibitors were never approved, presumably due to a lack of specificity for E1. For efficient replication, E1 interacts with E2, which directs it to the origin of replication. As the crystal structure of the E1-E2 complex has been described[20], several compounds—the first HPV-specific compounds, inhibiting complex formation, have been developed[21,22]. While these inhibitors effectively reduced HPV replication, they were only effective against low risk HPV types 6b and 11. It is known that both E1 and E2 interact with numerous cellular (replication) proteins, and these interactions are crucial for successful HPV genome replication, gene expression regulation and maintenance. Therefore, designing molecules specifically targeting these interactions would perhaps lead to the desired goal (reviewed extensively in[23]). Or perhaps HPV even uses pathways based solely on cellular proteins to facilitate its infection, which are essential for the virus but dispensable for the cells (see one example in the Results section below).
In addition to targeting HPV proteins, the viral genome itself could be the target as well. Several sequence-specific DNA binding compounds belonging to the pyrrole-imidazole polyamide class have been developed[24,25]. This type of compounds bind to AT-rich regions near the E1 and E2 binding sites in the HPV replication origin and effectively reduce the stably maintained episomal viral genomes, presumably by affecting the binding of the E1-E2 complex to the origin.
The current situation in HPV therapy strongly suggests that there is a need for specific drugs targeting HPV infection, preferably without restriction to specific subtypes.
HPV infects epithelial cells in the cutaneous and mucosal epithelia, and its life cycle is dependent on keratinocyte differentiation. Thus, most of the work regarding the HPV life cycle has been conducted in human primary epithelial keratinocytes; using these cells is relatively time-consuming and expensive[26–31]. This is particularly true for drug development, since currently, high-throughput screening (HTS) of available chemical libraries is a widely used technique to identify new inhibitors of various diseases[32]. There are number of systems for conducting HTS to target the HPV life cycle. Fradet-Turcotte et al. have developed a model system where they use heterologous E1 and E2 expression vectors and monitor the replication of an HPV origin-containing plasmid using luciferase as a reporter[33]. This model system is definitely useful in the identification of novel drugs. However, it only allows researchers to study the E1- and E2-dependent initiation of replication, and since it does not use the full HPV genome as model, it does not take into account the functions of other HPV proteins as well as the regulation of viral gene expression, maintenance, vegetative amplification and segregation. Moreover, as E1 and E2 are expressed from heterologous vectors, both protein and replication levels may significantly differ from HPV genome replication. Two other model systems based on HaCat cells have also been developed. One of them allows researchers to monitor the stable maintenance of the wt HPV11 genome, but it can only be used for a limited number of passages[34]. Another HaCat cell-line system uses HPV VLPs that express different reporter genes. The cells are infected with these VLPs, and reporter gene expression can be monitored. This system is extremely powerful when screening compounds inhibiting the very early stages of HPV infection[35].
U2OS cells, which are derived from moderately differentiated osteosarcoma, have an adherent epithelial morphology and carry wild-type p53 and pRB genes, which have proven to be useful for studying various aspects of the HPV life cycle[36–38]. It has been shown that the replication mechanism of HPVs in U2OS cells is identical to other cell lines, such as HaCat cells[39,40]. In addition to describing replication, transcriptome analyses of mucosal HPVs 11 and 18 and cutaneous HPV 5 have been published[41–43]. These studies indicate that HPV gene expression in the U2OS cell line is very similar to that in keratinocytes[44,45]. Replication and transcriptome studies suggest that U2OS cells provide an adequate cellular environment to study HPV, and since they can be rapidly and cost-efficiently grown, they would be useful in conducting high-throughput screens to identify novel HPV inhibitors.
In this study, we report the generation of various HPV 18, 16 and 5 marker genomes containing reporter genes for rapid and easy quantification of viral copy number. First- and second-generation marker genomes contained reporter genes in various configurations in the late region of the HPV genome. All of these versions had reduced replication ability. Based on the transcription map of HPV18[41], we next turned to the early region of the HPV genome and added the Renilla luciferase gene into the ORF of E2 immediately after the overlapping region of E1 and E2, which added 20 (in the case of HPV5 and 16) or 22 amino acids (in the case of HPV18) at the N-terminus of the luciferase protein. In the C-terminus, luciferase carried the 2A peptide sequence of FMDV and was followed by the complete sequence of the full-length E2 protein starting from the methionine. This configuration resulted in the production of two functional proteins during the translation of the fusion mRNA: full-length E2 and functional Renilla luciferase with the short E2 protein 20- or 22-amino-acid tag at the N-terminus. Since the expression levels of Renilla luciferase are controlled by viral transcription, its expression level correlates with the HPV copy number. In addition to validating this novel model system, we used it for HTS of the NCI Diversity Set IV public chemical library and a customized library and identified several novel high-risk HPV-specific inhibitors with IC50 values ranging from 2.5–60 μM. These inhibitors effectively blocked the replication of HPV18, HPV16, HPV31, HPV33 and HPV45 but not HPV11 or HPV5.
U2OS cells, which were obtained from the American Type Culture Collection (ATCC no: HTB-96), the modified cell lines U2OS GFP2-Fluc #10.15, U2OS-EBNA1 (Icosagen Cell Factory Ltd) and the HPV18 Rluc-E2-positive U2OS #10.15 subclones #2G10 and #2B3, were grown in Iscove’s modified Dulbecco’s medium (IMDM) supplemented with 10% fetal calf serum (FCS). The U2OS cells were transfected with the indicated amounts of different HPV minicircles or 500 pmol of HPV18-specific siRNAs by electroporation (220 V, 975 μF) with a Bio-Rad Gene Pulser II that was supplied with a capacitance extender (Bio-Rad Laboratories). The U2OS GFP2-Fluc #10.15 cell line was generated by transfecting U2OS cells with a linearized expression vector containing both the GFP2 and Firefly luciferase (Fluc) ORFs as well as a puromycin resistance gene. Individual puromycin-resistant clones were picked and analyzed by GFP and Fluc expression. HPV18 Rluc-E2 positive subclones were generated similarly to those described in[36]. The CIN 612E cells (kind gift from Dr. Frank Stubenrauch) were grown in Keratinocyte-SFM Medium Kit (Gibco, cat# 17005075).
The parental plasmid pMC-HPV18 was constructed for the production of HPV18 genome miniplasmids. A recognition site for the BglII restriction endonuclease was introduced into the HPV18 genome between nt 7473 and nt 7474 (herein, the numbering of the HPV18 genome is according to the NCBI Reference Sequence NC_001357.1). These sites were used previously without observing changes in gene expression compared to unaltered HPV18[44]. The modified HPV18 genome was cloned into the minicircle production plasmid pMC.BESBX. The pMC backbone, derived from pMC.BESBX, permits the production of the HPV18 genome from pMC-HPV18 as a supercoiled minicircle that contains a 92-bp-long non-HPV sequence. pMC-HPV5 was generated by linearizing the HPV5 genome with XmaJI (site in L2) and cloning the DNA into the XbaI site in the pMC.BESBX vector. pMC-HPV16 was made by inserting the BamHI-linearized HPV16 genome (site in L1) into the BglII site in the pMC.BESBX vector. pMC-HPV33 was cloned into the pMC.BESBX vector SalI sites introduced into the HPV33 genome after nt. 6915. pMC-HPV45 was cloned into the pMC.BESBX vector using the BglI sites introduced into the HPV45 genome after nt. 7569. Miniplasmid production was performed in E. coli strain ZYCY10P3S2T according to a previously published protocol[46]. Finally, the HPV genomes were purified from E. coli as supercoiled minicircles using the QIAfilter Plasmid kit (Qiagen). A vector containing the HPV31 genome sequence was obtained from the International Human Papillomavirus Reference Center, digested with EcoRI and religated before the transfection as described in[47]. The first-generation HPV18 marker genomes were constructed by substituting the late region of the genome different expression marker genes (Renilla luciferase, Rluc; Gaussia luciferase, Gluc; red fluorescent protein, RFP; and destabilized RFP, TurboRFP-dest1) driven by the Rous Sarcoma Virus (RSV) LTR promoter. The second-generation HPV18 marker genomes were constructed by introducing the marker gene Renilla luciferase into the L2 ORF. A full-length VCIP IRES element that has been demonstrated to be active in U2OS cells[48] was added in front of the Rluc marker gene to promote its translation, as described in[49]. Two variants of the second-generation marker genome were constructed that differed in the absence or presence of the heterologous polyadenylation region after the Rluc coding sequence (HPV18L2-Rluc, HPV18L2-Rluc-pA). The HPV18 Rluc-E2 marker genome was constructed by adding the Renilla luciferase-encoding cDNA (from pRL-TK, Promega) with the C-terminal 2A sequence of FMDV between the E1 and E2 ORFs. The length of the overlapping region between the 3’ end of the E1 ORF and the 5’ end of the E2 ORFs in HPV18 is 71 nt (nt 2818–2887). This resulted in a novel ORF of a single polypeptide that starts from the native start codon of E2. The HPV5 marker genome HPV5-RlucE2 and the HPV16 marker genome HPV16-RlucE2 were generated similarly to HPV18-RlucE2. The HPV18-RlucE2 K490A marker genome contains a mutation (amino acid lysine 490 is mutated to alanine in the E1 protein) that abrogates the E1 protein ability to hydrolyze ATP. This results in replication-deficient viral genome [50]. This mutant genome was generated by PCR mutagenesis using primer CTTAGTATCTGTTAACGGTTCCAACCAAAAATGACTAGTGGAATTCACAAATGATATTACTGCTCCTTGTATAAAGTGTATAAAACTCATTCCAAAATATGAcgcTCCTGTATTTGC. The expression vectors for HPV18 E1 and E2 and the origin-containing HPV18-URR minicircle plasmid are described in[39]. The Epstein Barr Virus oriP plasmid p994 was a kind gift from B. Sugden, described in[51]. HPV18 E5-, E6-, E7- or E6-E7- mutant genomes and their replication properties were described in [39]. ShRNA expression was under the control of RNA polymerase III promoter U6.
Tdp1 shRNA sequence: GCACGATCTCTCTGAAACAAACTCGAGTTTGTTTCAGAGAGATCGT
PARP1 shRNA sequence: GGACTCGCTCCGGATGGCCTTCAAGAGAGGCCATCCGGAGCGAGTCC
Tdp2-1 shRNA sequence: GTACAGCCCAGATGTGATACGAATATCACATCTGGGCTGTAC
Tdp2-2 shRNA sequence: GAAGGATATTTCACAGCTACGAATAGCTGTGAAATATCCTTC
The U2OS-GFP-Fluc #10.15 cells were transfected with 2 μg of the HPV18-Rluc-E2 minicircle, and the cells were seeded onto 100 mm plates. On the next day, the cells were detached and seeded onto 96-well plates (5000 cells per well). Forty-eight hours after the transfection, the screened compounds were added to the media in 5 μM and 1 μM concentrations. The cells were grown for three days, and both Firefly luciferase (shows cellular viability) and Renilla luciferase (shows HPV copy number) were measured using the Dual-Glo luciferase assay system (Promega) according to manufacturer’s protocol with the GloMAX-96 luminometer (Promega). The results were blotted on a XY-scatter diagram, and HPV-specific hits were chosen.
The Diversity set IV and the additional compounds NSC9782, NSC 88915, NSC 82269, NSC 109128 and NSC 305831 were obtained from the Drug Synthesis and Chemistry Branch, Developmental Therapeutics Program, Division of Cancer Treatment and Diagnosis, National Cancer Institute. Camptothecin (CPT) (sc-200871) and ABT-888 (sc-202901) were purchased from Santa Cruz Biotechnology.
The viral genome copy number in U2OS cells during replication was analyzed by quantitative real-time PCR (qPCR). Genomic DNA was extracted from U2OS cells, and the samples were linearized by digestion with the appropriate enzyme (see above), and the bacterially produced input DNA was fragmented by digestion with DpnI. For each qPCR reaction, 3 ng of DNA was used, and the reactions were performed with EvaGreen qPCR Mix Rox (Solis BioDyne) according to the manufacturer’s protocol on a 7900 HT Fast Real-Time PCR System (Applied Biosystems). The HPV18 replication signal was amplified with the following oligonucleotides (300 nM of each per reaction): 5’-GCGCTTTGAGGATCCAAC-3’ and 5’-GTTCCGTGCACAGATCAG-3’. For HPV5, the following oligonucleotides were used: 5’-GGTTGCAGGAACTGTGAGGT-3’ and 5’-TCCGCGACAGTCGGGGCACAGG-3’. For HPV16, the oligonucleotides were 5’- CCCACAGCTACAGATACAC-3’ and 5’- GCAGGTGTGGTATCAGTTG-3’. The analysis was performed according to the comparative threshold cycle (ΔCt) method. The results were calculated from the PCR cycle number in which the HPV signal exceeded the threshold value (CtHPV). The CtrDNA was detected as a normalization standard from the ribosomal DNA gene sequence in the U2OS genome with the following oligonucleotides (300 nM of each): 5’- GCGGCGTTATTCCCATGACC-3’ and 5’- GGTGCCCTTCCGTCAATTCC-3’. The relative value CN, which reflects the average viral genome copy number per cell, was calculated from the data with the formulas ΔCt = CtHPV − CtrDNA and CN = 2-ΔCt.
For first- and second-generation marker genome analyses, episomal DNA was extracted as described in[52]. For Southern blot analyses, U2OS, U2OS#10.15 or U2OS-EBNA1 cells were transfected with 2–5 μg of the indicated HPV genome or 1 μg of oriP plasmid (only in the case of U2OS-EBNA1 cells). HPV DNA was linearized and digested with DpnI to cleave non-replicated, bacterially produced DNA. The samples were resolved in an agarose gel, blotted, and hybridized with an HPV genome or oriP sequence-specific probe labeled with [α-32P]dCTP using random priming (DecaLabel kit; Thermo Scientific). Specific HPV replication signals were detected by Typhoon TRIO (GE Healthcare).
The cells were lysed with passive lysis buffer by freezing and thawing. Both Renilla and Firefly luciferase expression was measured with a Glomax 20/20 luminometer using the Dual-Luciferase reporter assay system following the manufacturer’s protocol (all equipment and components from Promega Corporation).
The cells were lysed with Laemmli buffer (4% SDS, 20% glycerol, 120 mM Tris-Cl (pH 6.8), and 200 mM DTT) and boiled for 10 minutes at 100°C. The samples were resolved in an SDS-PAGE gel and transferred to Immobilon-P PVDF membrane (Millipore). Tdp1 and PARP1 were detected with their specific antibodies from Santa Cruz Biotechnology: sc-365674 and sc-56197, respectively. Tubulin (used as a loading control) was detected using Sigma Aldrich antibody T9026. Anti-mouse peroxidase-conjugated secondary antibody (LabAS) and Amersham ECL Western blotting Detection Kit (GE Healthcare) were used for visualization. Signals were exposed on an X-ray film (AGFA).
The cells were detached and collected by centrifugation 300g 5 minutes. The pellet was resuspended 100μl 1X PBS and 900μl ice-cold 80% ethanol was added drop-wise. The cells were next incubated on ice for at least 30 minutes, collected by centrifugation, resuspended in 1X PBS containing 50μg/ml propidium iodide and 200μg/ml RNase A and incubated at 37°C for 45 minutes. Cell cycle profiles were analyzed by flow cytometry (LSR II from Becton Dickinson) and by FlowJO 10 software using Watson (pragmatic) model.
To evaluate cellular viability and the potential toxic effects of the compounds during the HT screen, we generated a monoclonal U2OS cell line stably expressing both the GFP and Firefly luciferase (Fluc) reporter genes, called U2OS-GFP2-Fluc #10.15. Insertion of these reporter genes allows to measure cell growth and viability as well as HPV genome replication (S1 Fig). To measure HPV genome replication rapidly and quantifiably, we added reporter genes (like luciferase or fluorescent protein) to the viral genome so that their expression would be controlled by viral gene expression. First, we generated first- and second-generation marker genomes (schematic maps in S2 Fig), where modifications were made in the late region of HPV. However, all modified genomes had a reduced replication capacity (S2C Fig). We next turned to the early region of HPV genome and tried to add the Renilla luciferase coding sequence into the ORF of E2. The portion of the ORF of E2 protein encoding the first 22 amino acids of E2 protein overlaps with the ORF of E1 protein in case of HPV18. Thus, the Renilla luciferase reporter gene was added in frame after 22 N-terminal amino acids of the HPV18 E2 ORF, immediately after the overlapping region with E1 ORF. The Renilla luciferase reporter coding sequence was followed in frame by the FMDV 2A peptide sequence, which induces single peptide bond pausing during the translation of the mRNA. The 2A peptide sequence was followed by the full-length E2 protein sequence starting from the E2 methionine. This configuration results in two proteins: the Renilla luciferase containing first 22 amino acids of HPV E2 protein in its N-terminus and the full-length HPV E2 protein. The first 22 amino acids of E2 protein are thus synthesized twice. In the case of HPV5 and HPV16, similar additions were made in frame after 20 N-terminal amino acids of the respective HPV genomes. Schematic representations of the Rluc-E2 marker genomes and their working principle are shown in Fig 1. Detailed descriptions of the marker genomes are shown in the materials and methods section. To characterize the replication properties of the marker genomes, U2OS #10.15 cells were firstly transfected with HPV18 wt, HPV18-RlucE2 and HPV18-RlucE2-K490A genomes. Genomic DNA was extracted 3, 5 and 7 days after the transfection, transient replication was analyzed by Southern blot (Fig 2A) and the HPV copy number was quantitated (Fig 2B). It is clear that addition of the Renilla luciferase coding sequence into the ORF of E2 protein did not alter the replication properties of HPV18 significantly compared to the wt genome. In the qPCR analyses, ~150bp fragment of HPV genome that contains only one DpnI restriction site, is amplified. The relatively small signal obtained for HPV18-RlucE2-K490A in the qPCR analyses comes likely from the fact that DpnI restriction is not 100%. Both Firefly (from U2OS genome) and Renilla (from HPV marker genomes) luciferase expression were also measured in the same experiment and are presented as the ratio of Rluc/Fluc, as measured by the dual-luciferase assay described in the materials and methods section. The results in Fig 2B show that for the HPV18, Renilla luciferase expression could be used to describe changes in the viral copy number during replication. The wt and K490A genomes served as negative controls for Renilla luciferase expression. The replication-deficient K490A marker genome shows relatively high Rluc/Fluc ratio the in 3-day timepoint likely because there is still large portion of transfected, transcriptionally active DNA present in the cells.
The marker genomes of HPV16 and HPV5 were analyzed exactly as HPV18 (Fig 2). In case of HPV16, the marker genome replicated similarly to the wt genome, for HPV5, the marker genome replication is 2–3 times more effective than the wt genome. Renilla luciferase expression was reproducibly detectable, the net luminescence from 1 second integration time ranged from 10 to 25 million from approximately 400 000 cells. Using HPV18-specific siRNAs, we showed that the expression level of Renilla luciferase correlated with the HPV18 copy number during initial amplification (S3 Fig). We have characterized the transcription map for HPV18 in U2OS cells[41], and the results shown in S4 indicate very similar RACE PCR patterns for wt HPV18 and HPV18-RlucE2. This confirms that the promoter activity and splicing in the viral early region of the marker genome are similar to wt HPV18.
After initial amplification, the stable maintenance phase of HPV replication is turned on. To study this phase in U2OS cells, stable cell lines carrying episomal HPV genomes must be generated. We generated two monoclonal cell lines (#2B3 and 2G10) based on U2OS-GFP2-Fluc #10.15 that contain an episomal HPV18-RlucE2 genome, as described in[36]. First, both Renilla and Firefly luciferase readouts were measured from these cell lines on two different days. The results in Fig 3A show that both #2G10 and #2B3 indeed express Renilla luciferase, indicating that these cell lines contain the HPV18-Rluc-E2 marker genome and express Firefly luciferase, which could be used to evaluate cell growth and viability. Next, genomic DNA was extracted from these cell lines, and the viral copy number was measured by qPCR analyses. Line #2G10 had approximately 200 copies of HPV18-Rluc-E2 per cell, whereas #2B3 had 70. To further characterize the cell lines, we evaluated their stability to maintain the HPV18-RlucE2 genome. The cells were thawed from the generated cell banks and grown in subconfluent conditions for 30 days; Rluc/Fluc ratios and HPV copy number was measured in every 5 days. As shown in Fig 3B, the #2B3 is stable for at least 30 days while HPV18-RlucE2 genome copy number rapidly decreased from the #2G10 after 15 days of cultivation. It is possible that the high copy-number (200 copies) per cell is the reason of instability. During extensive passaging of the cells, the cells divide relatively quickly. In some HPV-positive U2OS cell clones, the segregation of the viral genome could be less effective resulting in decrease of the HPV copy number during extensive passaging of the cells. Still, both cell lines are suitable for identifying the compounds capable of inhibiting the stable maintenance phase of HPV18.
The third replication stage of HPVs is late amplification, which is triggered by the initiation of keratinocyte differentiation. This stage could be studied using organotypic raft cultures or by inducing keratinocyte differentiation with high concentrations of calcium[53,54]. It has been shown that when U2OS cell lines containing episomal HPV18 are grown in confluent culture for at least three days, late amplification is turned on, and the viral copy number can increase by 10-fold. Although the U2OS cell line is a immortalized cell line, the cultivation in dense cultures likely induces a differentiation-like state as early differentiation marker K10 expression is elevated [36]. The cell lines #2B3 and #2G10 were seeded in 6-well plates (150 000 cells per well) and grown for 4, 8 and 12 days with regular feeding but without passaging of the cells. Both genomic DNA and luciferase samples were collected at each timepoint and used to analyze the viral copy number by qPCR and the expression of the Renilla and Firefly luciferase reporter genes. As seen in Fig 3C, the viral copy number increased by 5-fold, reminiscent of late amplification. Fig 3D shows that the increase in the Renilla expression is very similar to the viral copy number change. In case of late amplification studies, the cells are not passaged and thus the viral genome is not lost from the more unstable #2G10.
Collectively, the data shows that cell lines #2B3 and #2G10 can be used in high-throughput screens to identify compounds capable of inhibiting the stable maintenance and/or late amplification of HPVs.
We screened (as described in the Materials and Methods section) NCI Diversity set IV using the model system described in the previous sections for compounds inhibiting the initial amplification of the HPV18 genome in U2OS cells. HT screening was performed with all the compounds in this library in 5 μM and 1 μM concentrations, and 80 compounds were selected for further analyses (approximately 5% of the analyzed compounds). After validation of the hits on the HPV18 wt genome replication, five most effective compounds were chosen for detailed analyses (schematic structure and expected activities of the compounds are shown in S7 Fig). To further characterize the compounds, we performed a replication assay, as described in the Materials and Methods section. The inhibition relative to the vehicle control, DMSO, and the logarithmic inhibition curves are shown in Fig 4. All five compounds, NSC 9782, NSC82269, NSC 88915, NSC 109128 and NSC 305831, exhibited concentration-dependent inhibition of HPV18 initial amplification, with IC50 values ranging from 2.5 to 60 μM. To ensure that the inhibition of HPV18 replication is specific, the effects of the compounds on cell cycle progression was evaluated by the exact assay set up described in Fig 4. Both non-transfected and HPV18-transfected U2OS cells were used, the results in S1 Table and S8 Fig clearly demonstrate that the compounds identified here do not alter the cell cycle progression significantly. We additionally analyzed the effect of the compounds on HPV18 ori—URR replication supported by the HPV18 E1 and E2 proteins expressed from heterologous expression vectors. None of the compounds inhibited the replication of URR plasmid facilitated by E1 and E2 expressed from heterologous expression vectors (S5 Fig). These compounds were also analyzed for their ability to inhibit the stable maintenance and late amplification of HPV18 in the U2OS cell line #1.13, which contains episomal viral genomes, as described in[36,41]. Compounds 109128, 305831 and 82269 clearly inhibited both the stable maintenance and vegetative amplification of HPV18, while 88915 only inhibited the stable maintenance phase, and 9782 did not inhibit any of the later replication stages (S6 Fig).
Two of the identified compounds, 88915 and 305831, are known inhibitors of Tdp1[55,56]. To see if Tdp1 is actually necessary for HPV18 initial amplification, we transfected U2OS cells with an HPV18 minicircle genome together with various concentrations of the shRNA_Tdp1 plasmid. Empty shRNA plasmid was used as a mock control. Genomic DNA was extracted 3 and 4 days after the transfection a replication assay was performed and the results quantified. Results in Fig 5A (compare lanes 1 and 2 with 3–8) show at least 50% decrease in HPV18 replication when Tdp1 expression was downregulated.
It was recently shown that the PARP1 protein activates Tdp1 through PARylation and recruits it to sites of DNA damage[57]. PARP1 involvement in initial HPV18 amplification was assessed similarly to Tdp1 using shRNAs or the PARP1-specific inhibitor ABT-888 (Veliparib)[58]. The results in Fig 5 show that similarly to Tdp1, PARP1 is essential for HPV18 replication.
Next, the involvement of Tdp1 and PARP1 in the initial amplification of other HPV types was evaluated by co-transfection of shRNA-s with HPV 31, 33, 11 and 5 minicircle genomes. Genomic DNA was extracted three days post-transfection, analyzed by Southern blot and the signals were quantified (Fig 5D). Unexpectedly, Tdp1 and PARP1 seem to be necessary for high-risk HPV replication only as the downregulation of these proteins decreases the replication of HPV31 and 33 but not 11 and 5.
Effects of Tdp1 and PARP1 downregulation on cell cycle progression were also evaluated in a similar assay setup. Results in S2 Table demonstrate that loss of Tdp1 or PARP1 expression has no significant effect on cell cycle progression of U2OS cells.
It has been previously shown that instead of Tdp1, Tdp2 is actually necessary for HPV replication [25]. In this study, downregulation of Tdp1 actually resulted in increase of the replication. We designed two shRNA-s specific to Tdp2 and firstly analyzed their effect on cell cycle progression of U2OS similarly to Tdp1 and PARP1. The downregulation of Tdp2 significantly altered the cell cycle progression, one shRNA caused the block in G1 and the other one in G2/M phase of the cell cycle (S2 Table). Thus, it was not possible to estimate the potential involvement of Tdp2 in HPV replication in U2OS cells specifically.
CPT is a Topoisomerase I inhibitor that stabilizes entrapped Top1cc complexes on DNA[59,60]. Since Tdp1 is responsible for cleaving entrapped Top1cc complexes from DNA, a synergistic effect occurs between CPT and Tdp1 inhibitors. This type of synergistic treatment between Topoisomerase I inhibitors and NSC 305831 was shown to be effective against Murine Lupus Nephritis significantly improving survival of the infected mice[61]. Topoisomerase I is also shown to interact with HPV E1 and E2 proteins which stimulate its activity to ensure effective replication [62,63] In the experiments shown in Fig 6, we used U2OS cells constitutively expressing the Epstein-Barr virus (EBV) EBNA1 protein—U2OS-EBNA1 cells. The expression of the EBNA1 protein allowed us to monitor the replication of the plasmids containing EBV latent origin—oriP, reminiscent of genomic DNA replication[51,64], [65]. Since HPV amplificational replication is initiated multiple times per cell cycle, it should differ significantly from oriP replication, making the latter suitable as an internal control to determine the specificity of the inhibitors tested. U2OS-EBNA1 cells were co-transfected with an HPV18 minicircle genome, and the oriP plasmid and the cells were grown in the presence of various concentrations of our identified HPV inhibitors alone or together with 2 nM CPT for 5 days, with DMSO used as a vehicle control. A replication assay was performed, and the signals were quantified with a phosphoimager. The results shown in Fig 6 show that all of the HPV inhibitors are specific, as none of them inhibited EBNA1-dependent replication of the oriP plasmid significantly. Treatment with 2 nM CPT alone did not inhibit HPV18 initial amplification. Fig 6 shows that in the case of compounds 9782, 88915, 109128 and 305831, modest concentration-dependent synergistic effects with CPT occurred, as HPV replication was slightly more efficiently inhibited (compare lanes 1–7 with 8–12 in all experiments). Compound 82269 (Fig 6B, compare lanes 1–7 with 8–12), however, showed no synergistic effect together with CPT, suggesting that it inhibits HPV18 replication through some other currently unknown pathway. The compound 9782 acts like a Tdp1 inhibitor but it does not inhibit the stable maintenance and late amplification of HPV18 (S6 Fig) like other compounds do. Perhaps this observation that the compound 9782 is Tdp1 inhibitor is not true, and it inhibits some other protein in this pathway or causes synergistic effects with CPT through some other (unknown) pathway or protein (Tdp2). Tdp2 may be a likely candidate as it has been shown to be capable of carrying out the same functions as Tdp1 [66] The compound 88915 also does not inhibit the late amplification of HPV18 (S6 Fig) although it is a Tdp1 inhibitor. But this could be explained by the low potency of this compound, perhaps very effective inhibition of Tdp1 is necessary during late amplification where HPV genome replication rate is very high.
Thus far, the experiments for describing the inhibitors were conducted using the high-risk HPV18 genome. Since Tdp1 and PARP1 seem to be necessary for the replication high-risk HPVs but not low-risk or beta HPVs (Fig 5D), we expected the compounds to behave the same. U2OS cells were transfected with HPV types 5, 11, 16, 31, 33 and 45, and the effect of the compounds on these HPV types was evaluated similarly to HPV18. The results in Fig 6 show that indeed the compounds do not inhibit HPV5 or HPV 11 initial amplification. As expected, they are however effective against HR-HPV types 16, 31, 33 and 45.
When analyzing chemical compounds, there is always a possibility that the effects may be cell-line specific. In order to ensure that the compounds identified here are true HPV inhibitors, we tested them on HPV31-positive CIN 612E cells (kind gift from Dr. Frank Stubenrauch). The CIN 612E cells were seeded onto 6-well plates and grown in subconfluent conditions in the presence of compounds for 6 days. Genomic DNA was extracted, analyzed by Southern blot and quantified by Phosphoimager. Results in Fig 7G demonstrate that the compounds NSC 88915, NSC 109128 and NSC 305831 decrease the HPV31 copy-number in a concentration-dependent manner in CIN 612E cells. The compound NSC 82269 did not have an effect on HPV31 replication in these cells and the compound NSC 9782 was extremely toxic.
Taken together, these data suggest that the inhibitors identified in this study are specific HR-HPV inhibitors, blocking the replication of HPVs that cause more than 99% of all cervical cancer cases worldwide.
U2OS cells, an osteosarcoma cell line, are suitable for studying the replication properties of various types of HPVs from different phylogenetic genera[36,37]. Studies performed in our lab show that the transcriptomes of both alpha and beta papillomaviruses are very similar to that described in keratinocytes. Therefore, U2OS cells seem to be an adequate model for studying replication, gene expression and other aspects of the HPV life cycle[67,68].
There are numerous well-characterized chemical compound libraries available, meant for the identification of novel drug candidates via high-throughput screening using appropriate model systems. To date, there was no such system for analyzing the complete HPV genome replication cycle. Thus, the first major goal for this work was to construct a model system for HPV research. It has been shown that HPV genomes that lack the late region replicate similarly to wt genomes[36]. Therefore, our initial attempts at constructing modified HPV genomes containing easily measurable reporter genes were focused on substituting the late genes of the genome with genes encoding either fluorescent proteins or various luciferases (first- and second-generation marker genomes described in S2A and S2B Fig). The results demonstrated that insertion of the reporter gene cassettes into the late region of the HPV18 genome greatly interferes with gene expression and/or replication properties, as the replication of these marker genomes was almost undetectable (S2C Fig).
Based on the transcriptome analyses of HPV18 in U2OS cells[41], we next tried to insert the Renilla luciferase coding sequence into the beginning of the ORF of the E2 protein and generated marker genomes for HPV18, HPV16 and HPV5 (Fig 1). The coding sequence of Renilla luciferase is inserted into the ORF of the E2 protein directly after the overlapping region of the E1 ORF so that the expression of Renilla luciferase is regulated by the modulation of viral promoters driving the transcription of E2 protein mRNAs. The transcription of Renilla luciferase begins with the native E2 start codon and proceeds for 22 amino acids for HPV18 or for 20 amino acids for HPV16 and 5 before coding the Renilla luciferase. The mRNA produced contains the beginning of E2, the Renilla luciferase and FMDV 2A followed by the full-length E2 (the principle is described in Fig 1). This configuration results in the translation of two functional proteins due to the FMVD 2A protein-induced ribosome skipping: Renilla luciferase that also contains 20–22 N-terminal amino acids of the E2 protein and full-length E2 protein with an additional amino acid proline in its N-terminus. The HPV18 and HPV16 marker genomes replicate slightly more efficiently than the respective wt genomes (Fig 2), perhaps due to the more stabilized E2 protein (due to the additional proline in the N-terminus). The HPV5 marker genome replicated at a considerably higher level that the wt genome (Fig 2). Since gene expression is different between alpha and beta papillomaviruses, addition of the Renilla coding sequence into the HPV5 E2 ORF may enhance splicing or change the ratios of different viral transcripts, leading to higher expression levels of the E1 and E2 proteins. Nonetheless, to our knowledge, this system is the first suitable system for HTS for beta papillomavirus inhibitors. In addition to analyses of initial amplification, U2OS cells, more specifically cell lines harboring episomal HPV genomes, also allow to study the stable maintenance and late amplification of HPV. For that, stable cell lines containing episomal HPV18 marker genomes were generated (Fig 3). Analyses of these cell lines showed that regardless of the insertion of the Renilla luciferase coding sequence, the HPV18 marker genome displays similar replication characteristics as the wt genome during the stable and late amplification phases, at least in U2OS cells. The previously described systems for HPV drug development usually model very specific steps during viral infection (E1-E2 interaction, initiation of viral replication), which significantly reduces the probability of identifying potential inhibitors. As the model system described here allows to study all three HPV genome replication stages and harbors the expression of all viral proteins, more suitable drug targets and inhibitors could be identified, and new mechanistic insights into HPV pathogenesis could be revealed.
Although there are currently three vaccines against various types of HPVs that effectively prevent infection with various HPV types, no approved effective cure for an ongoing infection is available. Mostly identification of novel drugs against various targets begins with high-throughput screening of chemical libraries containing thousands of compounds[69]. Thus, using the model system described here, we conducted a HTS of the NCI Diversity Set IV public chemical library, which consists of different classes of compounds that have shown some type of biological activity. The first round of screening gave ~80 positive hits, out of which 5 compounds (the structures of the compounds is shown in S7 Fig) inhibited HPV18 initial amplification in the low-micromolar range (Fig 4). Many studies regarding HPV replication have been carried out by measuring the E1- and E2-dependent replication of an HPV URR plasmid, which contains the origin of replication. Even an HTS model system for measuring URR replication has been developed[33]. This type of model uses heterologous expression vectors for E1 and E2 expression and could be used for studying the E1/E2-dependent initiation of HPV replication. However, the compounds identified in this study did not inhibit the replication of a URR-containing plasmid (S5 Fig), and compound 88915 even slightly activated the URR plasmid replication. It is possible that this compound stimulates cell growth, prolongs the S-phase or stabilizes replication proteins. Therefore, to identify HPV replication inhibitors, it is important to analyze full-length HPV genome replication because it includes the expression of other viral proteins besides E1 and E2, appears to include different cellular proteins and/or uses different mechanisms for replication.
In almost all HPV-related cancers, the viral genome is integrated into the host genome. Integration of the viral genome usually occurs during persistent infection, and thus to prevent integration and tumor progression, already-established infections must be targeted. Therefore, we next tested these compounds on the stable maintenance and late amplification of HPV18. Four out of five compounds successfully inhibited the stable maintenance phase of viral replication (S6A Fig), and three compounds inhibited late amplification as well (S6B Fig).
It has become clear in recent years that HPVs activate the DNA damage response network during their replication to “invite” cellular replication and DNA repair proteins to their replication foci[50]. Moreover, it has been shown that HPV uses homologous recombination to efficiently replicate its genome[39,40]. The exact mechanism and all the necessary cellular partners are not yet known. Inhibitors targeting cellular DDR proteins are valid candidates for cancer therapy; for example, the PARP1 inhibitor ABT-888 is in clinical trials[70]. Two of the compounds (NSC 88915 and NSC 305831, also known as Furamidine) identified in this study are known inhibitors of the DDR network protein Tdp1[55,56]. Tdp1 is not an absolutely essential protein for normal cellular replication; however, it is needed for repairing certain types of DNA damage. Due to specific DNA damage or replication fork collapse, Topoisomerase 1 and 2 cleaving complexes (Top1/2cc) become entrapped on the DNA and will interfere with normal replication/transcription fork progression. Tdp1 releases these entrapped Top1cc and Top2cc proteins from the DNA[71,72]. Inhibitors of Tdp1, together with its activator, PARP1, and topoisomerase I inhibitors are considered viable drugs in cancer therapy[73]. Here, we have shown that Tdp1, together with its regulator/activator protein PARP1, are essential cellular proteins in HPV18 replication, since downregulation of these proteins decreases the HPV genome copy number (Fig 5), thus making those proteins good targets for developing HPV inhibitors. Moreover, as Tdp1 and PARP1 are necessary for cancer cell survival, their inhibitors may also be effective against HPV-induced cancer. Thus, the same compounds could be used to cure all HPV-related cancers. Camptothecin (CPT) is a Topoisomerase I inhibitor that stabilizes the interaction between Topoisomerase and DNA[59,60]. Thus, by simultaneously inhibiting Tdp1/PARP1 and using CPT to further stabilize the entrapped Top1cc, more effective inhibition should be achieved[74]. We show here that four out of five compounds indeed have modest synergistic effects; up to five times more efficient inhibition of HPV18 initial amplification was observed when CPT was supplemented, suggesting that these compounds may inhibit Tdp1 or have some other target related to releasing Top1cc complexes from DNA. The inhibition seems to be relatively specific, as the compounds did not decrease the levels of the EBV EBNA1-dependent replication of the oriP plasmid significantly, which, in contrast to HPV amplification, takes place only once per cell cycle, reminiscent of cellular DNA replication (Fig 6). These results suggest that at some point during HPV replication, entrapment of Top1cc occurs on the viral genome. Tdp1 seems to then be activated (probably by PARP1[57]) and recruited to HPV DNA, where it releases those complexes, allowing replication/transcription to continue. When Tdp1 is inhibited, abnormal replication intermediates could emerge, and HPV replication cannot be completed (a proposed model of the pathway is in Fig 8). A study by Edwards et al. has shown (through siRNA library screening) that instead of Tdp1, Tdp2 might be necessary for HPV replication [25]. We analyzed the effects of Tdp2 downregulation on cell cycle progression of U2OS cells by two different shRNA-s and saw significant enrichment of cells in G1 or G2/M phase of the cell cycle (S8 Fig and S2 Table). Thus, it is difficult to assess the role of Tdp2 in HPV replication specifically. The reduction in HPV copy number by downregulation of Tdp2 shown by Edwards et al. may very well be due to the cell cycle block. Additionally, it is known that both Tdp1 and Tdp2 can release the Top1cc complexes from DNA [66]. It is also possible that in different experiment setups or in different cellular systems, either Tdp1 or Tdp2 are used during HPV replication. Regardless, it is still fascinating how HPV replication machinery uses cellular pathways to facilitate its genome replication.
It is unclear how Tdp1 is recruited for the HPV replication process and if it is necessary for all HPV types. More than 99% of all cervical cancer cases are caused by five HR-HPV types: 18, 16, 31, 33 and 45. We show here that all five inhibitors successfully block the initial amplification of all of those HR-HPV types (Figs 4 and 7). Unexpectedly, however, none of the compounds were able to block the replication of low-risk HPV type 11 or cutaneous beta HPV type 5 (Fig 7A and 7B). It is possible that there is a (minor) difference between the replication mechanism of these viruses. Thus far, the two known major differences between HR and LR HPVs are the sequence of the genome and oncogenic potential of the E6 and E7 proteins, which interact with various DNA repair proteins[75,76]. The latter seems not to be the reason for this difference, as the compound NSC305831 still inhibited HPV18 replication when the oncoproteins E5, E6 and E7 alone or in combination were mutated (S9 Fig). It was also clear that the DDR was activated during the replication of all HPVs tested. Since Top1/2cc-s only become trapped on DNA due to specific DNA damage, the nature of DDR activation could differ between high-risk and other HPV types. In addition to the genome sequence and oncoproteins, the viral replication proteins E1 and E2 may differ between HPV types. For example, similar mutations in the C-terminus of the E1 protein completely abolished BPV1 replication but not HPV11 replication[77]. As none of the inhibitors blocked the E1/E2-dependent replication of the 1-kbp-long HPV origin-containing plasmid URR (S5 Fig), it seems that Tdp1 is not crucial for the initiation of HPV replication. In this “URR-assay”, the levels of E1 and E2 protein are significantly higher compared to the HPV viral genome replication and thus the rate of replication initiation is also higher. This may mean that even if the Top1cc is entrapped on URR plasmid, the high rate of replication may not affect the URR plasmid copy number. However, this does not rule out the possibility that E1 and/or E2 interaction with key proteins in the pathway responsible for the release of Top1cc is critical in the case of 8-kbp HPV genome replication.
In addition to the involvement in replication, the inhibitors described in this study may (also) affect HPV gene expression. It is known that Top1 is necessary to relax tensions in DNA during both replication and transcription. Entrapment of Top1cc affects transcription in several ways, perhaps most important of which in the context of HPV life cycle could be the negative impact on splicing[78–81]. Since there are significant differences in mRNA splicing between LR and HR HPVs (reviewed extensively in [82]), for example LR-HPVs do not express E6*. Thus, entrapment of Top1cc may be more deleterious to oncogenic HPV E1/E2 gene expression. The potential effects on HPV gene expression may be another explanation, why these compounds did not work in the “URR assay” as in this case E1 and E2 proteins are expressed from heterologous expression vectors containing CMV promoter.
Regardless, further analyses should be performed to describe the involvement of Tdp1 and PARP1 in HR-HPV replication or gene expression. It would be interesting to identify the specific location(s) in the HPV genome where Top1cc entrapment occurs. Studying the differences in the involvement of Tdp1 between high-risk, low-risk and cutaneous HPV replication would perhaps describe some yet unknown differences between the replication mechanisms of these viruses. It is possible that aside from E6 and E7, differences in replication could be the reason why HR-HPVs, but not LR-HPVs, integrate more readily into the host genome and are therefore causative agents of various cancers.
Collectively, we have engineered modified HPV genomes that express Renilla luciferase as a marker that could be used to monitor viral replication in various assays. The gene expression and replication properties of these marker genomes are almost identical to wt genomes and such genomes could be thus used in primary keratinocytes or other suitable cell lines for rapid HPV genome copy number quantification. We used this system in high-throughput screening and have identified several novel HR-HPV-specific inhibitors. Importantly, we have demonstrated that three compounds (88915, 109128 and 305831) inhibit the HPV replication in the cells derived from human cervical intraepithelial neoplasia (CIN612E; Fig 7). This suggests that the target(s) of these compounds are also active during HPV replication in vivo. Analyses of the inhibitory properties of these compounds led to the discovery of Tdp1 and PARP1 as promising targets for the development of new anti-HPV drugs.
When Tdp1 and/or PARP1 are inhibited and Top1cc is stabilized by Camptothecin (CPT), replication forks will collide, HPV genome replication is blocked, and eventually, aberrant DNA replication intermediates form.
|
10.1371/journal.ppat.1002063 | Suboptimal Activation of Antigen-Specific CD4+
Effector Cells Enables Persistence of M. tuberculosis In
Vivo | Adaptive immunity to Mycobacterium tuberculosis controls
progressive bacterial growth and disease but does not eradicate infection. Among
CD4+ T cells in the lungs of M.
tuberculosis-infected mice, we observed that few produced IFN-γ
without ex vivo restimulation. Therefore, we hypothesized that one mechanism
whereby M. tuberculosis avoids elimination is by limiting
activation of CD4+ effector T cells at the site of infection in
the lungs. To test this hypothesis, we adoptively transferred Th1-polarized
CD4+ effector T cells specific for M.
tuberculosis Ag85B peptide 25 (P25TCRTh1 cells), which trafficked
to the lungs of infected mice and exhibited antigen-dependent IFN-γ
production. During the early phase of infection, ∼10% of P25TCRTh1
cells produced IFN-γ in vivo; this declined to <1% as infection
progressed to chronic phase. Bacterial downregulation of fbpB
(encoding Ag85B) contributed to the decrease in effector T cell activation in
the lungs, as a strain of M. tuberculosis engineered to express
fbpB in the chronic phase stimulated P25TCRTh1 effector
cells at higher frequencies in vivo, and this resulted in CD4+ T
cell-dependent reduction of lung bacterial burdens and prolonged survival of
mice. Administration of synthetic peptide 25 alone also increased activation of
endogenous antigen-specific effector cells and reduced the bacterial burden in
the lungs without apparent host toxicity. These results indicate that
CD4+ effector T cells are activated at suboptimal
frequencies in tuberculosis, and that increasing effector T cell activation in
the lungs by providing one or more epitope peptides may be a successful strategy
for TB therapy.
| Mycobacterium tuberculosis causes persistent infection even in
human or animal hosts that develop antigen-specific CD4+ and
CD8+ T cell responses. To understand this phenomenon, we
tested the hypothesis that the CD4+ effector T cells that are
generated in response to M. tuberculosis infection fail to
encounter their antigens at the site of infection in the lungs. Using mice
infected with M. tuberculosis, and an assay of in vivo
antigen-dependent activation of CD4+ T cells, we found that both
polyclonal CD4+ and T cell receptor transgenic
CD4+ T cells specific for antigen 85B peptide 25 are
activated at low frequencies in the lungs. We found that this is due in part to
downregulation of antigen gene expression by M. tuberculosis,
as forced expression of the antigen gene resulted in higher frequency activation
of CD4+ T cells, as well as CD4+ T
cell-dependent reduction in bacterial burdens and prolonged survival of infected
mice. We also found that administration of antigen 85B peptide 25, which is
recognized by a high proportion of M. tuberculosis-specific
CD4+ T cells, reduced the bacterial burden in the lungs,
indicating that stimulation of existing antigen-specific CD4+ T
cells may be a promising approach to therapy of TB.
| Even though its etiologic agent was discovered over 125 years ago, tuberculosis
remains a global scourge, killing 1.7 million people in 2009, at least ¾ of
whom were immunocompetent [1]. Long-term persistence of Mycobacterium
tuberculosis, which resides principally in phagocytic cells within the
lungs, results in a chronic infection despite the presence of an apparently
appropriate adaptive immune response. In mice infected with virulent M.
tuberculosis, the early phase of infection proceeds with unchecked
bacterial growth until day 17–21 post-infection, when adaptive immunity
finally exerts control of bacterial growth in the lungs. Control of infection in
both humans and mice critically depends on M. tuberculosis-specific
CD4+ Th1 cell responses, which include production of IFN-γ
[2], [3]; however
adaptive immune responses do not eradicate the infection.
Several potential mechanisms may account for the failure of adaptive immune responses
to eradicate the bacteria in tuberculosis. Generation of M.
tuberculosis-specific CD4+ effector T cells is delayed
compared with responses to other pathogens [2], [4]. In addition, certain
individuals, or strains of mice, may develop inappropriate (e.g.,
Th2) [5], [6] or imbalanced
effector phenotypes such as Th1/Th17 [7] in response to infection. However, even in humans or mice
that develop Th1 responses, a failure of CD4+ effector T cells to
recognize infected cells may preclude their optimal activation and limit induction
of effector functions in the lungs. Prevention of effector T cell activation could
result from impaired antigen presentation by lung APCs containing M.
tuberculosis
[8], [9], [10] or because the
antigens that effector T cells recognize are not expressed or otherwise available in
the lungs. Furthermore, host regulatory mechanisms that limit immune pathology, such
as T regulatory cells [11], production of inhibitory cytokines [12], and, possibly, onset of T cell
exhaustion [13], [14] may inhibit the
activity of effector T cells at the site of infection. Finally, even when
CD4+ effector T cells are activated, the efficacy of these
responses may be limited by the impaired ability of infected cells to respond to
IFN-γ [15],
[16], [17], induce phagosome
maturation [18],
[19], or
undergo apoptosis [9], [20], [21], [22]. Understanding the contribution of each of these
potential mechanisms limiting adaptive immunity to M. tuberculosis
is an essential prerequisite for vaccine design and other immunologic approaches to
tuberculosis prevention and therapy.
Here, we report that CD4+ effector T cells are activated at
submaximal and suboptimal frequencies in the lungs during M.
tuberculosis infection, that this is due in part to bacterial
modulation of antigen expression, and that increasing the availability of a single
antigen results in improved immune control of M. tuberculosis.
We hypothesized that M. tuberculosis evades adaptive immunity by
modulating the activation of CD4+ effector T cells at the site
of infection in the lungs. Since in vitro studies have revealed evidence that
M. tuberculosis modulates MHC class II antigen presentation
[10],
[23], [24], [25], [26], we
focused on in vivo activation of CD4+ T cells in the lungs. We
reasoned that, if M. tuberculosis-infected cells do not present
antigens efficiently to effector T cells in the lungs, then the frequency of
activation of effector functions of CD4+ cells would also be low
at the site of infection. To test this, we used direct intracellular cytokine
staining of lung cells from infected mice for IFN-γ, without ex vivo
restimulation. We found that that the frequency of IFN-γ expression by
CD4+ T cells in the lungs varied with the time of infection
(Figure 1B).
IFN-γ+ CD4+ cells were undetectable in
the lungs at day 14, increased in frequency beginning by day 21 to a peak at day
35 post-infection, and then markedly decreased afterward; no more than 7%
of the bulk population of CD4+ T cells expressed IFN-γ at
any time point after infection, and fewer than 1% expressed IFN-γ
during the chronic phase. Other studies investigating IFN-γ production by
CD8+ T cells in vivo have used treatment of mice with
brefeldin A or inclusion of brefeldin A during cell isolation and staining [27], [28]. However, we
determined that these methods did not improve detection of intracellular
IFN-γ by CD4+ T cells during M.
tuberculosis infection (Figure S1). These data indicate that a small
minority of polyclonal CD4+ T cells recruited to the lungs of
M. tuberculosis-infected mice are activated to produce
IFN-γ at a given time, and are consistent with defective antigen
presentation, costimulation, and/or inhibition of effector T cell activation at
the site of infection.
Since the low frequency of CD4+ T cell expression of IFN-γ in
the lungs of M. tuberculosis-infected mice could be due to the
presence of effector cells that traffic to the lungs but are not specific for
M. tuberculosis antigens, we performed the remainder of our
studies using CD4+ TCR transgenic T cells that specifically
recognize a well-characterized immunodominant M. tuberculosis
antigen. To quantitate the frequency of activation of M.
tuberculosis antigen-specific effector cells in the lungs, we
prepared CD4+ Th1 effector cells (P25TCRTh1 cells) from
transgenic mice with a TCR specific for peptide 25 (amino acids 240–254)
of Ag85B. When P25TCRTh1 cells were incubated with irradiated splenocytes in the
absence of peptide 25, <1.0% of the cells expressed IFN-γ as
detected by intracellular staining and flow cytometry, whereas addition of
peptide 25 in vitro induced IFN-γ expression in ∼90% of cells
(Figure 2A). This result
demonstrated that the frequency of IFN-γ staining in P25TCRTh1 cells can
specifically assay antigen dependent stimulation of P25TCRTh1 cells.
Since Day 21 post-infection corresponds to an acute stage of infection when
adaptive immune effector mechanisms have been initiated and reduce the rate of
bacterial population growth in the lungs, and since it resembles the stage of
LCMV infection in which a high frequency of antigen-specific
CD8+ T cell responses are observed [28], we chose this time point for
initial characterization of P25TCRTh1 cell responses in vivo. We verified that
adoptively transferred P25TCRTh1 cells traffic to the site of infection by
examining sections of lungs from infected mice that had received
CFP+ P25TCRTh1 cells. CFP+ cells were
abundant in the lung parenchyma, and were concentrated in granulomas (Figure 2B). Furthermore, we
determined that >85% of the transferred cells were protected from
labelling by an i.v. injection of PerCP-labeled anti-CD4 antibody, indicating
that adoptively transferred P25TCRTh1 cells efficiently migrate out of the lung
vasculature into the parenchyma of infected lungs (Figure
S2A).
To determine the frequency of activation of antigen-specific CD4+
effector T cells in the lungs early in infection, we adoptively transferred
P25TCRTh1 cells on day 18 and harvested them on day 21 after infection of
wild-type mice with wild-type M. tuberculosis H37Rv. The
frequency of IFN-γ+ P25TCRTh1 cells isolated from the lungs
was unexpectedly low at Day 21 post-infection (Figure 2C and 2D). Approximately
1–2% of the transferred P25TCRTh1 cells were stimulated to produce
IFN-γ in vivo at that time point (Figure 2C), and this percentage was similar
to the frequency of total endogenous lung CD4+ T cells
expressing IFN-γ on day 21 post-infection (Figure 1B, 2D). Moreover, after intravenous injection of
PerCP-labeled anti-CD4 antibody, the only IFN-γ+ P25TCRTh1
cells identified were PerCP negative (Figure S2B), indicating that the responding
cells were those that had migrated out of the vasculature into the lung
parenchyma and were protected from staining by the in vivo injection of
antibody.
We verified that stimulation of P25TCRTh1 cells to express intracellular
IFN-γ is due to recognition of Ag85B peptide 25 by transferring P25TCRTh1
cells into mice infected with an Ag85B-null strain of M.
tuberculosis (ΔAg85B), which is equivalent to wild-type H37Rv
in virulence [2]. A lower mean percentage (0.74%) of P25TCRTh1 cells
isolated from ΔAg85B-infected mice expressed IFN-γ than those from
H37Rv-infected mice (Figure 2C and
2D). This indicates that in vivo IFN-γ production by P25TCRTh1
cells is antigen-dependent and not the consequence of inflammatory cytokines
present at the site of infection. We also evaluated several alternative
approaches to detecting effector T cell activation in the lungs. P25TCRTh1 cells
expressed both CD25 and CD44 prior to adoptive transfer, which excluded their
use in evaluating effector cell activation in vivo. Surface expression of CD69
was induced after adoptive transfer of P25TCRTh1 effector cells into
H37Rv-infected mice; however, we found similar induction of CD69 in mice
infected with ΔAg85B, indicating that it did not specifically reflect
antigen-dependent effector cell activation. This result, together with evidence
that CD69 can be induced by costimulation and by certain cytokines present at
the site of M. tuberculosis infection [29], [30], [31], indicates that expression of
intracellular IFN-γ is the most accurate reporter of antigen specific Th1
effector cell activation in the lungs. Together, these results indicate that
even though they traffic efficiently to the site of infection, Ag85B peptide
25-specific CD4+ effector cells are activated to execute their
Th1 effector function at low frequency in the lungs of M.
tuberculosis-infected mice.
Although IFN-γ production by P25TCRTh1 cells at day 21 was dependent on
Ag85B, the frequency of IFN-γ+ cells was surprisingly low in
H37Rv infected mice. One possible explanation for the low frequency of
activation of effector cells is that their cognate antigen is not available for
recognition at the site of infection. To test this hypothesis, we provided
antigen in vivo by injecting peptide 25 intravenously into mice that had been
infected 21 days earlier. When P25TCRTh1 recipient, H37Rv-infected mice received
peptide 25 six hours prior to lung cell harvest, the frequency of
IFN-γ+ P25TCRTh1 cells increased to 20–50%
(Figure 2C and 2D).
Similarly, peptide 25 injection stimulated a higher frequency of IFN-γ
expression by endogenous CD4+ T cells from mice infected with
H37Rv (Figure 2C and 2D),
consistent with prior evidence that peptide 25 of Ag85B is a dominant antigen in
C57BL/6 mice infected with M. tuberculosis
[32],
[33].
P25TCRTh1 cells transferred into ΔAg85B-infected recipients were also
stimulated at a higher frequency after intravenous peptide 25 treatment, while
endogenous CD4+ T cells from ΔAg85B-infected mice did not
respond to peptide 25 with increased IFN-γ expression (Figure 2D). The failure of endogenous
CD4+ T cells from ΔAg85B-infected mice to respond to
peptide 25 injection reflects the absence of Ag85B peptide 25-specific effector
T cells generated in response to this infection. These results indicate that the
frequency of IFN-γ+ P25TCRTh1 cells is an accurate and
specific measure of CD4+ effector T cell stimulation in response
to presentation of Ag85B peptide 25 in vivo. The observation that in vivo
IFN-γ responses to peptide 25 injection depend on the presence of
previously-generated (endogenous or transferred) peptide 25-specific effector T
cells indicates that the responses are not due to a nonspecific effect of the
epitope peptide on costimulation or responses of CD4+ T cells
with specificity for other antigens. In addition, they demonstrate that if
antigen is made available to them, adoptively transferred P25TCRTh1 cells can
respond to antigen in the infected lungs, and they provide evidence against an
exclusive role for T regulatory cells and/or suppressive cytokines in limiting
the activation of CD4+ effector cells at the site of M.
tuberculosis infection in the lungs.
To further characterize the in vivo assay system, and to evaluate the possibility
that low frequencies of P25TCRTh1 responses are attributable to either
competition for antigen by endogenous CD4+ T cells and/or a
dominant effect of T regulatory cells, we specifically ablated endogenous T
cells from M. tuberculosis-infected CD4-DTR mice [34] prior to
assaying P25TCRTh1 responses in vivo. Compared to untreated mice, DT treatment
reduced the fraction of endogenous CD4+ T cells in the lung by
an average of 48.9%, p = 0.0053 (Figure
S3A). However, this had no effect on the percentage of P25TCRTh1 cells
activated to produce IFN-γ (Figure S3B). These results strongly suggest
that the low frequency of activation of P25TCRTh1 cells is caused neither by
competition for peptide 25:MHC II complexes by endogenous CD4+ T
cells, nor by the influence of T regulatory cells in the lungs. We therefore
conclude that the response of adoptively transferred P25TCRTh1 cells is an
accurate reflection of MHC II presentation of Ag85B peptide 25 by lung APCs
during infection.
Adaptive immunity restricts progressive growth of M.
tuberculosis, but it does not eliminate the bacteria from the
lungs, which results in chronic infection in mice and latent infection in
humans. To determine whether suboptimal activation of M.
tuberculosis-specific T cells contributes to the ability of the
bacteria to persist, we first asked whether activation of P25TCRTh1 cells in the
lungs changes as infection progresses to a chronic phase. To compare the
frequency of effector T cell stimulation at various stages of infection, we
transferred P25TCRTh1 cells into H37Rv-infected mice on day 11, 18, 25, 32, or
39 post-infection. Lung cells were harvested 72 hours after transfer (day 14,
21, 28, 35, or 42 post-infection) and analyzed by flow cytometry for
intracellular IFN-γ without ex vivo restimulation. The proportion of
P25TCRTh1 cells producing IFN-γ was highest (∼10%) on day 14
(Figure 3A and 3C).
These results indicate that during the acute stage of infection, adoptively
transferred P25TCRTh1 cells are stimulated in the lungs at a frequency
comparable to that of TCR transgenic CD4+ effector cells at the
site of injection of a protein antigen and adjuvant [35]. In contrast, expression
of IFN-γ by endogenous (CD45.2−) CD4+ cells
was rare (<0.1%) at that time point (Figure 1B and 1C). The difference between
transferred and endogenous cell responses on day 14 is consistent with our
previous observation that initiation of adaptive immunity to M.
tuberculosis is delayed until day 11–14 post-infection, and
consequently, endogenous CD4+ effector T cells specific for
M. tuberculosis antigens are first detected in the lungs on
day 17 post-infection. [2].
The frequency of IFN-γ production by P25TCRTh1 cells progressively decreased
from day 14 to day 42 post-infection, indicating a decrease in the efficiency of
peptide 25-specific T cell stimulation as infection enters its chronic phase
(Figure 3A and 3C).
These results with TCR transgenic CD4+ effector cells closely
mimic the results observed with endogenous polyclonal CD4+ T
cells after day 14 post-infection (Figure 1B and 1C). Although Ag85B peptide 25-specific responses
reached an earlier peak and decreased earlier than did those of endogenous
polyclonal CD4+ T cell responses, the results with the two cell
populations were similar, with endogenous CD4+ effector T cell
responses also diminishing by day 42 post-infection.
To determine whether activation of naïve Ag85B peptide 25-specific
CD4+ T cells is also diminished in the later stages of
M. tuberculosis infection, we assayed the response of
adoptively transferred naïve P25 TCR-Tg T cells in the lung-draining
mediastinal lymph nodes of H37Rv-infected mice at various time points
post-infection. 7 days after transfer, we harvested lymph node cells and
measured in vivo T cell proliferation by flow cytometry using a CFSE dilution
assay. The rate of naïve P25TCR-tg T cells was highest upon transfer into
mice on day 18 post-infection, while fewer cells exhibited CFSE dilution at days
24 and 48 post-infection (Figure
3B). These results indicate that decreased stimulation of P25TCRTh1
effector cells is also accompanied by decreased generation of peptide 25
specific effector T cells from naive cells at later stages of infection.
Since treatment of infected mice with exogenous peptide 25 enhanced T cell
responses, indicating that adoptively-transferred P25TCRTh1 cells are capable of
responding to antigen stimulation in the lungs, we hypothesized that
availability and/or presentation of antigen is a limiting factor in the
activation of CD4+ effector T cells at the site of M.
tuberculosis infection. To test this hypothesis, we first
investigated whether changes in the expression of the M.
tuberculosis gene that encodes Ag85B influence the frequency of
activation of P25TCRTh1 effector cells. We found that the frequency of in vivo
activation of P25TCRTh1 cells mimicked the temporal pattern of expression of
fbpB (which encodes Ag85B) by M.
tuberculosis in vivo (Figure 3C). This suggests that reduced expression of Ag85B
contributes to the low frequency of activation of Ag85B-specific
CD4+ effector cells in the lungs, thus resembling
previously-reported observations with Salmonella FliC expression and
FliC-specific CD4+ T cell responses [36].
To test the hypothesis that fbpB down-regulation contributes to
the submaximal frequency of CD4+ effector cell activation and
the limited efficacy of the Th1 response in vivo, we constructed a recombinant
strain of M. tuberculosis to express fbpB at
high levels during chronic infection. Using the ΔAg85B strain as a
background, we introduced a wild-type fbpB allele under control
of the hspX/acr/Rv2031c promoter to the M.
tuberculosis chromosome via the pMV306 integrating vector.
hspX is expressed at high levels during chronic phase
infection in an expression pattern inverse to fbpB
[37], [38]. This strain
(hspXp:fbpB, termed “CPE85B” for
chronic phase
expressed Ag85B) exhibited
higher fbpB expression compared to H37Rv in the lungs of mice
after aerosol infection (Figure
4A). The expression of fbpB measured by RT-qPCR was
approximately 10-fold higher (normalized for the abundance of 16S rRNA) at day
21 post-infection for CPE85B than for H37Rv. As the infection progressed to
chronic phase (day 28–42 post-infection), fbpB expression
from the native promoter declined by approximately 100-fold while
fbpB expression driven by the hpsX
promoter remained at nearly constant, higher levels (Figure 4A). Increased fbpB
gene expression in the CPE85B strain was accompanied by markedly enhanced
expression and secretion of Ag85B protein when the hspX
promoter was induced in stationary liquid culture (Figure 4B).
To determine whether forced expression of fbpB in M.
tuberculosis results in increased presentation of Ag85B peptide 25
to CD4+ T cells, we infected bone marrow-derived dendritic cells
(BMDC) with either H37Rv or CPE85B and compared their ability to activate
P25TCRTh1 cells in culture. At all APC∶T cell ratios examined, DCs
infected with CPE85B induced significantly greater amounts of IFN-γ
secretion from P25TCRTh1 cells than did DCs infected with H37Rv (Figure 4C). To determine
whether forced expression of fbpB can increase the frequency of
P25TCRTh1 stimulation during H37Rv infection in vivo, we compared the frequency
of P25TCRTh1 cell activation in the lungs of mice infected with either H37Rv or
CPE85B. Compared to cells from H37Rv-infected recipients, P25TCRTh1 cells from
CPE85B-infected mice produced IFN-γ with a 2-fold (day 21) to 5-fold (day
42) higher frequency (Figure 4D and
E). These findings indicate that forced expression of
fbpB by M. tuberculosis increases the
proportion of P25TCRTh1 cells that are activated to produce IFN-γ in the
lungs. By suppressing fbpB expression after the initial stages
of infection, wild-type M. tuberculosis can reduce the
frequency of activation of Ag85B-specific effector T cells. Although expression
of fbpB was maintained at high levels from day 14 to day 42
post-infection, P25TCRTh1 cell stimulation in CPE85B-infected mice was only two-
to five-fold higher than in mice infected with H37Rv, and decreased as infection
progressed to chronic stage, indicating that other mechanisms, such as
inhibition of antigen presentation and/or induction of regulatory T cells, exist
to limit the activation of CD4+ effector T cells in the
lung.
We reasoned that, if diminishing fbpB expression during chronic
infection limits effector T cell activation and thereby enables M.
tuberculosis to evade adaptive immunity, then constitutive
expression of fbpB throughout infection should improve immune
control of infection. To test this hypothesis, we infected mice with either
H37Rv or CPE85B and quantitated M. tuberculosis CFUs in the
lungs throughout the course of infection. The rates of bacterial growth for the
two strains were indistinguishable prior to day 14 post infection (Figure 4F), indicating that
expression of fbpB by the hspX promoter does
not attenuate M. tuberculosis in vivo during the innate immune
stage of infection, prior to recruitment of CD4+ effector T
cells to the lungs. Indeed, the in vivo generation time of the CPE85B strain
(23.0 h) was slightly shorter than that of H37Rv (26.4 h) during days 1–14
of infection (these are not significantly different by nonlinear curve fit and
F test). However, at times corresponding to the adaptive
immune phase of infection, the bacterial burden of the CPE85B strain in the
lungs was approximately 10-fold lower than that of H37Rv (Figure 4F). These results suggest that forced
expression of fbpB partially overcomes the antigen deficit that
limits the activation of CD4+ T cells in the lung during chronic
infection and allows greater antimycobacterial efficacy of the adaptive immune
response.
The observation that CPE85B demonstrates a growth pattern indistinguishable from
H37Rv during the first two to three weeks of infection, prior to onset of
adaptive immunity, suggested that CPE85B was not inherently attenuated for
growth in vivo. However, we considered the possibility that over-expression of
fbpB could cause attenuation of M.
tuberculosis as a result of gene dysregulation or toxicity of an
overabundant Ag85B protein. Notably CPE85B demonstrated a similar growth pattern
to H37Rv during in vitro shaking culture. Furthermore, under conditions of
hypoxic stationary culture, when Ag85B protein is strongly expressed by CPE85B
compared to H37Rv, the survival of the CPE85B strain is not impaired compared
with that of wild-type bacteria (Figure 4B). Taken together, these findings imply that impaired
persistence of M. tuberculosis CPE85B in vivo is the
consequence of increased antigen presentation and activation of
CD4+ T cells, and not due to intrinsic attenuation of the
CPE85B strain in vitro or in vivo. We reasoned that if the decreased lung
bacterial burden of CPE85B compared with that of H37Rv is attributable to
increased antigen presentation and recognition by CD4+ T cells,
then the attenuated phenotype of CPE85B should be abrogated in mice lacking
CD4+ T cells. Indeed, whereas wild type C57BL/6 mice
infected with CPE85B survived significantly longer than those infected with
H37Rv (median survival >300 and 239 days, respectively;
p = 0.0062), MHCIIKO mice, which lack CD4+
T cells, exhibited indistinguishable susceptibility to infection with the CPE85B
and H37Rv strains (median survival 79 and 81 days, respectively;
p = 0.425), (Figure 5A), clearly establishing that in vivo attenuation of the
CPE85B strain depends on MHC II antigen presentation and CD4+ T
cell responses. These results also indicate that increased antigen expression,
accompanied by increased antigen-specific T cell activation, can enhance control
of M. tuberculosis without detectable detrimental effects,
since wild-type mice infected with the CPE85B strain survived longer than mice
infected with H37Rv.
Since MHC II-deficient mice are highly susceptible to M.
tuberculosis infection, this could potentially mask any
hypothetical CD4+ T cell-independent mechanisms of attenuation
of the CPE85B strain. We reasoned that, if mechanisms other than increased
CD4+ T cell recognition contribute to the lower burdens of
CPE85B, then this strain would not recover and grow normally in the lungs when
CD4+ T cells are depleted during the chronic phase of
infection. We infected mice with H37Rv or CPE85B and allowed the infection to
proceed for 28 days, when initial lung CFUs were measured for each group. As
expected, bacterial CFUs for CPE85B were ∼3 fold lower than H37Rv at this
time point (Figure 5B). The
remaining mice in each infection group were then treated with monoclonal
antibody GK1.5 every 6 days until day 50 post-infection to deplete
CD4+ T cells. After an initial lag, in which neither
bacterial strain expanded, both CPE85B and H37Rv resumed growth in the lungs at
indistinguishable rates (Figure
5B). Taken together, these data provide strong evidence that improved
control of the CPE85B strain is attributable to increased activation of
Ag85B-specific CD4+ T cells, although we cannot exclude the
possibility that other factors contribute to the lower lung burdens of CPE85B
that appear after the development of adaptive immunity.
Our observation that forced expression of fbpB increased the
frequency of Ag85B peptide 25-specific CD4+ T cells and reduced
the bacterial burden in the lungs (Figure 4D, 4E, and 4F), together with our observation that injection
of peptide 25 also increased activation of CD4+ effector T cells
at the site of infection (Figure 2B
and 2C) suggested that providing antigen by injection of peptide 25
might also result in improved immune control of infection. We first determined
the duration of increased IFN-γ production by adoptively transferred
P25TCRTh1 cells or endogenous CD4+ T cells after peptide 25
injection. The frequency of IFN-γ cells was highest in both 6 hours after
treatment, and decreased to approximately 20% of maximal levels by 24
hours after peptide injection for both endogenous CD4+ and
P25TCRTh1 cells (Figure 6A).
By 72 hours post-treatment, the frequency of IFN-γ+ cells
returned to levels observed in the absence of peptide 25 injection, indicating
that the activating effect of peptide 25 treatment is remarkably transient,
entirely dissipating within 3 days of the treatment. Despite the transient
nature of this effect, we found that treatment of M.
tuberculosis H37Rv-infected mice with peptide 25 (in the absence of
adoptively transferred P25TCRTh1 cells) every 2–3 days from day 28 to day
45 post-infection reduced lung bacterial burdens by
1.05±0.40×106 bacteria
(p = 0.018) compared with that in mice treated with OVA
peptide, an unrelated MHC II epitope (Figure 6B). Neither group of mice displayed
any signs of toxicity, even after repeated peptide injections. These results
indicate that during M. tuberculosis infection,
CD4+ effector T cells are not stimulated at their maximum
potential frequency at the site of infection in the lungs. Because effector T
cell responses progressively decrease during chronic infection, and enhancing T
cell responses with exogenous peptide antigen improves immune clearance of
M. tuberculosis, we conclude that failure to optimally
activate effector T cells at the site of infection is an important determinant
of the limited efficacy of adaptive immunity in tuberculosis.
M. tuberculosis evades adaptive immunity to persist in the lungs,
often for the lifetime of the host. Here, we have characterized one mechanism by
which this impressive feat of immune evasion is accomplished in vivo. We found that,
of the large number of CD4+ effector T cells recruited to the lungs
of infected mice, few are stimulated to produce IFN-γ (Figure 7A). While there are few precedents
available for comparison, our findings are in stark contrast to those found in
C57BL/6 mice infected with the Armstrong strain of LCMV [28]. In that context, which results in
CD8+ T cell-dependent resolution of infection, >20% of
virus-specific CD8+ T cells are activated to produce IFN-γ
during the acute stage of infection when viral burdens and antigen availability are
highest, and the frequency of in vivo-activated virus-specific CD8+
T cells does not decrease until the viral burden is reduced. We found that the
initially low proportion of CD4+ T cells producing IFN-γ in the
lungs of M. tuberculosis-infected mice diminishes further as
infection progresses to chronic phase, even though the bacterial burden in the lungs
remains high. Our studies using adoptively transferred Ag85B-specific P25TCRTh1
cells revealed that the decreasing responses of CD4+ effector cells
are caused in part by decreasing expression of fbpB by M.
tuberculosis. By reducing fbpB expression during
chronic infection, M. tuberculosis restricts the availability of
Ag85B, an immunodominant antigen, and thereby prevents infected APCs from optimally
activating CD4+ effector T cells. Consistent with this model, we
found that a recombinant strain of M. tuberculosis engineered to
maintain the expression of fbpB at high levels during chronic
infection (CPE85B) was attenuated during the chronic phase of infection in a
strictly CD4+ T cell dependent manner, indicating that
down-regulation of fbpB and limitation of antigen availability is
important for evasion of adaptive immunity by M. tuberculosis.
Treatment of infected mice with synthetic Ag85B peptide 25 also increased
CD4+ effector T cell IFN-γ responses and significantly
reduced the bacterial burden in the lungs. We conclude that suboptimal effector T
cell activation enables M. tuberculosis to evade elimination by
adaptive immunity during the chronic stage of infection, and that some of this
suboptimal effector T cell activation is attributable to restricted antigen
expression by the bacteria. In addition, other mechanisms that limit effector T cell
activation, such as interference with the MHC class II antigen processing and
presentation pathway and/or the action of regulatory T cells, likely contribute to
the remarkable survival of M. tuberculosis in vivo.
Infection with M. tuberculosis induces a robust T cell response
involving CD4+ and CD8+ T cells and the effector
cytokines IFN-γ and TNF [3], which are all essential for control of infection [5], [39], yet adaptive
immunity fails to eradicate M. tuberculosis. Mechanisms for the
limited efficacy of the adaptive immune response in tuberculosis fall into two
general (not mutually exclusive) categories: either the effector functions that T
cells perform (e.g. IFN-γ production) are not effective because
of failed responses by the infected cells targeted by effector T cells; or the T
cells recruited to the site of infection do not optimally perform the effector
functions required for immune clearance. Regarding the former, the ability of
M. tuberculosis to resist and inhibit the TNF- and
IFN-γ-induced microbicidal responses of the phagocytic cells it infects is one
documented component of its immune evasion strategy in vivo [40]. However, our observation that
only a small fraction of the CD4+ effector T cells in the lungs is
activated to synthesize IFN-γ provides new support for the latter explanation.
The potential causes of this mechanism include bacterial factors and host regulatory
mechanisms that directly impair effector T cell function. As an example of a direct
bacterial effect, mycobacterial cell wall glycolipids have been found to impair
CD4+ T cell responses in vitro [41]. With regard to host regulatory
mechanisms, during mouse infection, T regulatory cells limit the ability of adaptive
immunity to restrict the bacterial population size in the lungs [11], [42].
Interleukin-10 (IL-10), whether expressed by myeloid cells or T cells, provides an
additional host regulatory mechanism that inhibits T cell effector functions in
tuberculosis, as transgenic over-expression of IL-10 in infected mice impaired T
cell responses and caused an increase in bacterial CFUs [12], while deletion of IL-10 causes
enhanced control of infection [43], indicating that T cell-directed suppressive factors can
limit the success of the adaptive immune response to M.
tuberculosis. On the other hand, CD4+ effector T cells
at the site of infection may not recognize or become activated optimally by APCs
bearing M. tuberculosis-derived peptide:MHC II complexes, a process
that is required for IFN-γ production in peripheral tissues [35]. Recent
observations using live imaging revealed that a small fraction of Leishmania
major-infected macrophages interact with
Leishmania-specific CD4+ T cells in vivo [44]
indicating that in certain infections, effector T cells may not recognize infected
cells efficiently, and this may contribute to slow clearance or persistence of
infection. Suboptimal stimulation of CD4+ T cells could occur via
direct targeting and inhibition of MHC II antigen presentation pathways in infected
APCs, or as a result of the limited availability of peptide T cell epitopes, a
consequence of bacterial suppression of antigen encoding genes, or a combination of
these mechanisms.
In this study, we first determined that the frequency of endogenous polyclonal
CD4+ T cells producing IFN-γ in the lungs was surprisingly
low, and varied during the course of infection, with the highest responses during
the acute stage and the lowest responses observed as infection reached the chronic
stage. These reduced responses occur despite the presence of similar numbers of
bacteria in the lungs during these stages of infection. To further understand the
underlying mechanisms of the low frequency of effector T cell activation in the
lungs, we quantitated CD4+ effector T cell responses to the peptide
25 epitope of M. tuberculosis Ag85B, a secreted protein targeted by
a large number of M. tuberculosis-specific CD4+ T
cells [45].
Ag85B is targeted by 5 of the 9 novel tuberculosis vaccine candidates currently in
clinical trials [46], thus understanding its behavior and responses to it in
vivo has considerable importance for TB vaccine development. The reduced expression
of fbpB we observed is consistent with regulation by the state of
bacterial growth, though it may be indirectly triggered by the onset of Th1
immunity, since expression of fbpB is maintained in mice lacking
IFN-γ [38].
Because Ag85B is a cell wall biosynthesis enzyme, down-regulation of
fbpB has been interpreted as a consequence of transition by
M. tuberculosis into a relatively stationary state.
Alternatively, fbpB suppression during chronic infection may also
be an evolved bacterial immune evasion mechanism that enables long-term persistence
of M. tuberculosis by limiting T cell activation. In support of
this, we found that forced expression of fbpB by the CPE85B strain
during chronic infection resulted in a higher proportion of P25TCRTh1 cells
producing IFN-γ than in H37Rv-infected mice. Other studies have suggested but
not directly examined the possibility that over-expression of certain M.
tuberculosis proteins (including Hsp70 and ESAT-6) may cause
attenuation of bacterial persistence by increased immune recognition [47], [48]. Our finding that
polyclonal CD4+ effector T cell responses diminish in chronic
infection suggests that this may be a general phenomenon in tuberculosis.
Importantly though, the higher frequency of P25TCRTh1 cell activation observed in
CPE85B-infected mice diminished at a later time point as it did in H37Rv infection,
implying that other mechanisms, especially impairment of MHC II antigen presentation
by M. tuberculosis, exist to limit effector T cell activation
during chronic infection in vivo.
Several in vitro studies have found that M. tuberculosis subverts or
impairs antigen presentation by the cells it infects, limiting the capability of
infected APCs to activate antigen specific T cells [8], [10]. Initial observations include
the finding that M. bovis BCG survives in primary human macrophages
that CD4+ T cells fail to recognize [26] and that M.
tuberculosis-infected THP-1 cells express low amounts of surface MHC II
[25]. Several
mechanisms for inhibition of MHC II antigen presentation have been characterized
using a spectrum of mycobacterial strains and cell components. Among these, impaired
phagosome maturation, a well-characterized component of the ability of M.
tuberculosis to survive in phagocytic cells [19], has been found to limit
activation of cathepsin D for efficient processing of mycobacterial antigens [24], while inducing
autophagy with rapamycin was recently found to improve the efficacy of BCG and other
live mycobacterial vaccines, by enhancing presentation of mycobacterial antigens
[49].
Impaired expression of MHC II by macrophages after IFN-γ treatment was also
observed after in vitro infection or treatment of macrophages with certain
mycobacterial cell components [23], [50], [51], [52], [53]. This effect may involve prolonged signals received
through bacterial pattern recognition receptors (PRRs) including TLR2, although we
recently reported a TLR2 independent mechanism for impaired MHC II expression in
response to IFN-γ [53], [54].
These and other in vitro studies are consistent with our present results and lend
support for the hypothesis that APCs do not efficiently stimulate
CD4+ effector T cells in the lungs during M.
tuberculosis infection in vivo. Attempts to verify and explore the
significance of these in vitro findings with in vivo infection models have been
limited thus far, until the present paper. One study of mouse infection with
GFP-expressing M. bovis BCG found a modest decrease in surface
expression of MHC II on some populations of lung APC that harbored intracellular
bacteria when compared to those that did not contain bacteria [55]. In contrast, in a low dose
aerosol infection of mice with GFP-expressing H37Rv, we did not detect a difference
in surface MHC II expression between infected and non-infected APCs at various time
points post-infection; we also found that M. tuberculosis-infected
APCs isolated from the lungs expressed high levels of the costimulatory molecules
CD80 and CD86 [54]. Nonetheless, there is evidence that the activation of
M. tuberculosis-specific T cell responses is impaired during in
vivo infection, indicating that M. tuberculosis may specifically
impair presentation of its antigens without decreasing overall surface expression of
MHC II. One recent study found that mice provided with CD4+
TCR-transgenic effector T cells specific for the M. tuberculosis
antigen ESAT-6 prior to infection can restrict bacterial population size to a lower
level but cannot prevent establishment of infection [56]. Despite the presence of this
effector T cell population in the lungs from the onset of infection, control of
bacterial growth was delayed until 7 days post-infection. Likewise, despite mounting
apparently normal anti-M. tuberculosis CD4+ T cell
responses, infected mice and humans treated with anti-mycobacterial drugs to
eliminate primary infection remain susceptible to reinfection [33], [57]. These studies indicate that
susceptibility to persistent tuberculosis is more likely due to failure to activate
antigen-specific effector T cells, rather than to insufficient development of
antigen specific T cells in response to infection.
We observed increased survival of wild type, but not CD4+ T
cell-deficient mice infected with the CPE85B strain when compared to those infected
with H37Rv, highlighting the importance of enhanced T cell stimulation to the
long-term outcome of infection, and indicating that enhanced effector T cell
activation, through increased antigen availability, can be accomplished without
detrimental effects. Moreover, our finding that sustained expression of Ag85B during
the adaptive immune phase of infection was associated with a 2- to 5-fold increase
in antigen-specific CD4+ T cell activation, yet reduced the
bacterial burdens approximately 10-fold implies that a massive increase in effector
T cell activation is not necessary to significantly improve immune control of
tuberculosis. Future efforts to develop tuberculosis therapies should therefore aim
to bypass or overcome factors that limit effector T cell activation including direct
T cell suppression, impaired antigen presentation, and bacterial gene regulatory
mechanisms. For example, we found that the chronic phase antigen deficit resulting
from bacterial suppression of fbpB could be overcome by systemic
treatment of infected mice with synthetic peptide 25, which strongly but transiently
enhanced CD4+ T cell responses specific for this epitope and reduced
the bacterial burden. This result implies that the endogenous CD4+ T
cells generated in response to infection with M. tuberculosis and
recruited to the infected lungs can be stimulated to perform their effector
functions if they are provided antigen, resulting in improved bacterial clearance
(Figure 7B). The potential
for anti-tuberculosis therapies that aim to enhance existing T effector cell
responses in infected individuals with synthetically produced peptides encoding
known T cell epitopes remains unexplored; however, given the steadily increasing
prevalence of drug resistant M. tuberculosis, such
immunotherapeutic approaches to tuberculosis are an attractive option. Although the
consequences of increasing the activation of existing T cell responses have not been
widely tested, in the context of certain highly monoclonal T cell responses,
administration of epitope peptides has caused rapid mortality of infected or
previously immunized mice [58], [59]. However, despite these findings and concerns about
possible immunopathology induced by hyperactivation of effector T cells in
tuberculosis [60],
we observed no morbidity or mortality in infected mice repeatedly treated with
peptide 25, a result that encourages the continued exploration of this therapeutic
strategy. Future studies should also aim to determine the host and bacterial
regulatory mechanisms that account for chronic phase suppression of
fbpB and whether genes encoding other immunodominant M.
tuberculosis antigens behave similarly. Identification of the elements
of this host-pathogen interaction may lead to the development of therapies that
target antigen gene suppression and inhibition of antigen presentation and provide a
novel strategy for overcoming bacterial persistence in vivo, leading to better
outcomes in M. tuberculosis-infected individuals.
C57BL/6, B6.SJL-Ptprca Pepcb/BoyJ
(CD45.1+), and MHCII KO mice for aerosol M.
tuberculosis infection experiments were either bred in the New York
University School of Medicine Skirball animal facility or purchased from Taconic
Farms, Inc. P25TCR-Tg mice, whose CD4+ T cells express a
transgenic T-cell antigen receptor that recognizes the complex of peptide 25 (aa
240–254) of M. tuberculosis Ag85B and the mouse MHC II
allele I-Ab were prepared on a C57BL/6 background, as previously
described [2],
[61]. All
animal experiments were done in accordance with procedures approved by the NYU
School of Medicine Institutional Animal Care and Use Committee and in strict
accordance with the recommendations in the Guide for the Care and Use of
Laboratory Animals of the National Institutes of Health under the Assurance of
Compliance Number A3435-01.
Wild type M. tuberculosis H37Rv was originally obtained from
ATCC. Frozen stocks for aerosol infection and in vitro use were prepared and
stored at −80°C. GFP-expressing H37Rv and Ag85B null (ΔAg85B)
strains of M. tuberculosis were generated as previously
described [2],
[62].
M. tuberculosis cultures were grown in 10 mL Middlebrook
7H9 liquid medium supplemented with 10% v/v albumin dextrose catalase
enrichment and incubated under shaking conditions at 37°C. Mice at
8–12 weeks of age were infected with ∼100 CFU of M.
tuberculosis via the aerosol route using an Inhalation Exposure
Unit (Glas-Col) as previously described [62]. To verify inoculum size,
3–5 infected mice were euthanized 24 hours after infection and lungs were
homogenized and plated on Middlebrook 7H11 medium supplemented with 10%
v/v albumin dextrose catalase enrichment. To determine bacterial population size
at time points post-infection, lungs were homogenized, diluted in
PBS+Tween-80 (0.5%), and added to 7H11 plates. Plates were incubated
at 37°C for 3 weeks and single colonies were counted. To determine
M. tuberculosis survival in stationary culture, 7H9 medium
was inoculated with H37Rv or CPE85B, grown in shaking conditions to saturation
(O.D.600>1.0), and initial CFUs were measured. Cultures were
then placed in stationary incubator at 37°C for 17 days, and final CFUs were
measured.
C57BL/6 mice were infected with M. tuberculosis H37Rv and on day
25 post-infection received 1×106 CFP+ P25TCRTh1
cells via adoptive transfer. On 28 post-infection, lungs were perfused and
frozen in OCT before 5 µm sectioning and fixation in cold acetone.
Sections were stained with DAPI to label nuclei and analyzed on a Leica DMRB
fluorescent microscope (objective: Leica PL Fluotar 20×/0.50) equipped
with a Spot RT digital camera. Separate images for DAPI and CFP fluorescence
were acquired and merged using Spot software.
P25 TCR-Tg CD4+ Th1 effector cells were generated in vitro as
follows: naïve CD4+ T cells were magnetically isolated from
lymph node cell suspensions of P25 TCR-Tg mice (or for fluorescent microscopy, a
P25TCR-Tg mouse expressing CFP under control of the ubiquitin promoter) using
CD4 (L3T4) microbeads and an AutoMACS (Miltenyi Biotech). P25TCR-Tg
CD4+ T cells were co-cultured with irradiated C57BL/6
splenocytes in the presence of mouse IL-12p70 (10 ng/ml), mouse IL-2 (5 ng/ml),
anti-IL-4 neutralizing antibody (50 ng/ml), and synthetic peptide 25 (0.5
µM). Cells were cultured at 37°C with 5% CO2. On
days 3 and 5 of culture, cells were split 1∶3 with fresh media containing
IL-12p70, IL-2, and anti-IL-4, but no peptide 25. Cells were washed with PBS and
counted on day 7 of culture before use for in vitro or in vivo assays. For in
vitro restimulation, P25TCRTh1 cells were co-cultured with irradiated C57BL/6
splenocytes for 24 hours in RPMI-10 in the presence or absence of peptide 25
(0.5 µM) or bone marrow derived dendritic cells infected with M.
tuberculosis (MOI: 0.1). Cells were collected and analyzed by flow
cytometry for intracellular IFN-γ, or culture supernatants were analyzed for
IFN-γ by ELISA. For in vivo experiments, 1×106 P25TCRTh1
cells were injected via tail vein or retro-orbital sinus into recipient mice at
various time points post-infection. Cells were routinely isolated from lungs of
recipient mice 72 hours after adoptive transfer and analyzed by flow
cytometry.
3×106 CFSE-labeled CD4+ T cells, harvested from
the lymph nodes of P25TCR-Tg mice were adoptively transferred into infected
recipients at various time points post-infection. 7 days after adoptive
transfer, mediastinal lymph nodes were harvested from recipient mice and cells
were analyzed for CFSE dilution by flow cytometry.
The Ag85B null strain of M. tuberculosis (ΔAg85B),
previously created by our lab from wild-type H37Rv [2], was used as a background
strain for generating CPE85B. Both the hspX promoter sequence,
consisting of 254 bp directly 5′ of the hspX start codon,
as well as the fbpB open reading frame were amplified by PCR
from H37Rv genomic DNA. Each of these fragments was ligated into the pMV306
integrating vector to create a recombinant construct, whose sequence was
verified by Sanger sequencing performed by the NYU DNA sequencing facility.
ΔAg85B was grown in 7H9 liquid media and transformed with this construct via
electroporation. The reaction was plated on 7H11 plates containing 25
µg/ml kanamycin to select for bacteria incorporating the construct into
the M. tuberculosis chromosome. Presence of the construct in
kanamycin resistant colonies was verified by PCR. Expression and secretion of
Ag85B by CPE85B was confirmed by SDS-PAGE and anti-Ag85B western blot of
supernatants from 7H9 liquid medium after stationary culture. For stationary
culture-induced expression of Ag85B by the CPE85B strain, 10 mL cultures were
grown to late phase (OD600∼1.0) in normal shaking conditions,
then flasks were sealed and transferred to a stationary incubator for >1 week
before supernatants were collected.
To quantitate expression of M. tuberculosis genes during mouse
infection, lungs of infected mice were rapidly placed into a solution of
RNAlater (Ambion) and stored overnight at room temperature in accordance with
manufacturer recommendations to allow permeation of the tissue. Thereafter,
samples for RNA isolation were stored at −80°C. When comparing
expression of genes at various time points, tissues were transferred to TRIzol
(Invitrogen) and quickly homogenized using a Tissue Tearor homogenizer to
disrupt mouse cells. Lung homogenates were centrifuged to pellet intact
bacterial cells, and supernatants discarded. M. tuberculosis
pellets were disrupted with zirconia/silica beads, RNA was extracted, and
RT-qPCR was carried out as previously described [37] with fbpB
copy number normalized to the constitutively expressed 16S rRNA and multiplied
by a factor of 105. The following RT-qPCR primers were used in this
study. 16S rRNA: RT 5-ATTACGTGCTGGCAACATGA-3, qPCR For 5-GCCGTAAACGGTGGGTACTA-3, qPCR Rev
5-TGCATGTCAAACCCAGGTAA-3;
hspx/acr/Rv2031c: RT 5-GAATGCCCTTGTCGTAGGTG-3, qPCR For 5-AGATGAAAGAGGGGCGCTAC3,
qPCR Rev 5-TAATGTCGTCCTCGTCAGCA3; fbpB/Rv1886c: RT
5-TCCTGGAACTTCAGGTTGCT-3,
qPCR For 5-ACCCCCAGCAGTTCATCTAC-3, qPCR Rev 5-TTCCCGCAATAAACCCATAG-3.
To isolate cells from infected tissues for flow cytometry, mice were euthanized
with CO2 followed by cervical dislocation. Tissues were removed and
mechanically disrupted by mincing in RPMI as previously described [62] or using a
gentleMACS dissociator (Miltenyi Biotec) in the manufacturer-recommended HEPES
buffer. Lung suspensions were incubated in Collagenase D and DNase at 37°C
with 5% CO2 for 30 minutes and cells were isolated by forcing
suspensions through a 70 µM cell strainer. RBCs were removed by ACK lysis
and live cells counted by trypan blue exclusion. Cell suspensions were stained
using the following fluorescently-labeled antibodies (Biolegend, BD Pharmingen,
or eBioscience): anti-CD3 PE, anti-CD4 (L3T4) FITC, anti-CD45.2 PerCP,
anti-CD45.1 Pacific Blue, anti-IFN-γ (XMG1.2) APC, and rat IgG1 APC isotype
control. Flow cytometry was performed using a FACSCalibur or LSR II (BD
Biosciences) at the NYU Cancer Institute Flow Cytometry and Cell Sorting
facility. Analysis of flow cytometry data was performed using FlowJo
software.
To detect intracellular IFN-γ produced by cells in vivo, a protocol was
developed based on a previous study [28]. In contrast to this study,
however, optimal detection of IFN-γ producing cells from the lungs of mice
infected with M. tuberculosis did not require treatment of mice
with i.v. brefeldin A or inclusion of brefeldin A in tissue processing buffers.
Instead, after euthanasia, tissues were rapidly placed on ice and all cell
isolation steps except collagenase/DNase digestion (37°C for 30 minutes) and
ACK lysis (room temperature for 5 minutes) were carried out quickly and on ice.
Cells were stained for surface markers at 4°C for 30 minutes followed by
permeabilization and fixation with Cytofix/Cytoperm (BD Biosciences) at 4°C
for 20 minutes. Finally, fixed cells were stained with anti-IFN-γ or a rat
IgG1 isotype control at 4°C for 30 minutes. Flow cytometry dot plot gates
for IFN-γ+ cells were set based on comparison with isotype
control and unpermeabilized cells stained for IFN-γ.
Mice were treated with an intra-peritoneal dose of 500 µg of either
monoclonal antibody GK1.5, which depletes CD4+ T cells, or a rat
IgG2b isotype control (LTF-2) every 6 days from day 28 to Day 50 post-infection.
Efficiency of CD4+ T cell depletion 6 days after GK1.5 treatment
was determined to be >95% by flow cytometry of cell suspensions from
lungs, spleen and blood. In mice treated with LTF-2 isotype control, no
differences were observed in CD4+ T cell number or bacterial
burden when compared to untreated mice.
To determine the influence of endogenous CD4+ T cells on the
response of adoptively transferred P25TCRTh1 cells in vivo, a system was
developed to deplete endogenous CD4+ T cells selectively from
infected mice. Mice expressing Cre recombinase under control of the CD4 promoter
were crossed with those carrying an inducible Diphtheria Toxin Receptor (iDTR)
allele, whose baseline expression is prevented by a stop codon flanked by
loxp sites [34]. Progeny of this cross (CD4-DTR) carry
CD4+ T cells that are sensitive to Diphtheria Toxin mediated
ablation. CD4-DTR mice were infected with H37Rv and received daily
intraperitoneal doses of DT (100 ng) to ablate endogenous CD4+ T
cells from day 21 to day 28 post-infection. The efficiency of
CD4+ T cell ablation in the lungs was determined by flow
cytometry to be 48.9%. P25TCRTh1 cells were adoptively transferred on day
25 post-infection and the frequency of IFN-γ production was assessed on day
28 post-infection.
On day 25 post-infection, wild-type mice infected with M.
tuberculosis H37Rv received P25TCRTh1 cells via adoptive transfer.
On day 28 post-infection, mice were treated intravenously with 800 ng (at 4.0
ng/µL) PerCP-labeled anti-CD4 (RM4-5). Fifteen minutes later, mice were
euthanized and total lung cells were stained with FITC-labeled anti-CD4 (GK1.5).
Lung cells stained by anti-CD4-PerCP were considered to be CD4+
T cells residing in the intravascular compartment at the time of antibody
injection. Cells staining positive for anti-CD4-FITC and negative for PerCP were
considered to be CD4+ T cells residing in an extravascular or
parenchymal lung compartment protected from labeling with intravenous antibody.
IFN-γ production in vivo was assessed by intracellular staining of all cells
with APC-labeled anti-IFN-γ as previously described.
Mice were intravenously treated with 100 µg of Ag85B peptide 25
(FQDAYNAAGGHNAVF) or OVA peptide control (ISQAVHAAHAEINEAGR) in 100 µl
sterile PBS via tail vein or retro-orbital sinus. Peptides were synthesized by
EZBiolab or Peptides International to a purity of >95%.
Data shown are representative of 2 or more experimental replicates. In all
figures, error bars indicate mean ± SEM. To determine statistical
significance when comparing experimental values from two groups of mice, one- or
two-tailed student's t-tests were routinely used, each
where appropriate. To compare the growth rate of H37Rv and CPE85B in vivo, a
non-linear regression analysis (curve fit) with F-test was used to determine
whether a single curve could account for both data sets. In mouse survival
experiments, Logrank test was used to evaluate statistical significance when
comparing survival of one mouse strain after infection with either of the two
bacterial strains. * = p<0.05;
** = p<0.005; n.s = not
significant.
|
10.1371/journal.pntd.0001974 | Leishmania donovani Develops Resistance to Drug Combinations | Drug combinations for the treatment of leishmaniasis represent a promising and challenging chemotherapeutic strategy that has recently been implemented in different endemic areas. However, the vast majority of studies undertaken to date have ignored the potential risk that Leishmania parasites could develop resistance to the different drugs used in such combinations. As a result, this study was designed to elucidate the ability of Leishmania donovani to develop experimental resistance to anti-leishmanial drug combinations. The induction of resistance to amphotericin B/miltefosine, amphotericin B/paromomycin, amphotericin B/SbIII, miltefosine/paromomycin, and SbIII/paromomycin was determined using a step-wise adaptation process to increasing drug concentrations. Intracellular amastigotes resistant to these drug combinations were obtained from resistant L. donovani promastigote forms, and the thiol and ATP levels and the mitochondrial membrane potential of the resistant lines were analysed. Resistance to drug combinations was obtained after 10 weeks and remained in the intracellular amastigotes. Additionally, this resistance proved to be unstable. More importantly, we observed that promastigotes/amastigotes resistant to one drug combination showed a marked cross-resistant profile to other anti-leishmanial drugs. Additionally, the thiol levels increased in resistant lines that remained protected against the drug-induced loss of ATP and mitochondrial membrane potential. We have therefore demonstrated that different resistance patterns can be obtained in L. donovani depending upon the drug combinations used. Resistance to the combinations miltefosine/paromomycin and SbIII/paromomycin is easily obtained experimentally. These results have been validated in intracellular amastigotes, and have important relevance for ensuring the long-term efficacy of drug combinations.
| Leishmania is a protozoan parasite that infects human macrophages to produce the neglected tropical disease known as leishmaniasis. Chemotherapy is currently the only treatment option for leishmaniasis. First-line therapies include pentavalent antimonials, except in some regions in the Indian subcontinent, the liposomal formulation of amphotericin B, miltefosine and paromomycin. The WHO has recently recommended a combined therapy in order to extend the life expectancy of these compounds. However, resistance could be induced in Leishmania if this approach is not applied in a controlled and regulated way, thus resulting in a rapid loss of efficacy of not one but two therapeutic options. In light of this, we have designed relevant experimental studies in order to determine whether Leishmania parasites are able to develop resistance to the different potential anti-leishmanial drug combinations that will be used in the near future. The results obtained could help us to predict the success of drug combination therapy. Experimental resistance of Leishmania donovani promastigotes to drug combinations was obtained after 10 weeks and remained in the intracellular amastigotes. We therefore conclude that L. donovani can easily develop resistance to drug combinations mainly miltefosine/paromomycin and SbIII/paromomycin. These results have been validated in intracellular amastigotes and are of considerable interest for future prediction of the success of drug combination therapy.
| The use of drug combinations, either in co-formulations or co-administrations, is an established approach for the treatment of several infectious diseases including malaria and tuberculosis [1]. This approach has also recently become a priority for other tropical parasitic diseases, such as visceral leishmaniasis [2]–[6]. Leishmaniasis, a neglected tropical parasitic disease that is prevalent in 98 countries spread across three continents, is caused by protozoan parasites belonging to the genus Leishmania [7]. The estimated incidence of leishmaniasis is 0.2–0.4 million cases of the visceral form (VL) and 0.7–1.2 million cases of the cutaneous form (CL) [7]. Although chemotherapy is the only current treatment option for leishmaniasis, its efficacy is increasingly limited by growing resistance to first-line drugs, especially antimonials, the frequent side-effects associated with their use, and the high cost of treatment [7], [8]. The recommended first-line therapies for VL include: i) pentavalent antimonials (meglumine antimoniate and sodium stibogluconate), except in some regions in the Indian subcontinent where there are significant areas of drug resistance [9]; ii) the polyene antibiotic amphotericin B (AmB); iii) the liposomal formulation AmBisome; iv) the aminoglycoside paromomycin (PMM); and v) the oral drug miltefosine (MLF). Recently, the WHO [7], [10], recommended to use either a single dose of AmBisome or combinations of anti-leishmanial drugs in order to reduce the duration and toxicity of treatment, prolong the therapeutic life span of existing drugs and delay the emergence of resistance.
Although recent clinical trials have highlighted the efficacy and safety of anti-leishmanial drug combinations [4], [5], [10]–[12], additional clinical studies are needed to investigate various other factors, such as the identification of an effective, well-tolerated and short treatment regimen, logistical aspects, and the potential risk of developing resistance considering that compliance in field conditions can be low [13].
Herein we describe the selection and characterization of experimental resistance to drug combinations in Leishmania parasites. Our findings clearly demonstrate the acquisition of resistance to different drug combinations in Leishmania donovani promastigotes using a step-wise adaptation process to increasing drug concentrations. Similarly, and perhaps importantly, we have obtained intracellular L. donovani amastigotes that are resistant to different drug combinations from promastigote forms resistant to these same combinations. These results indicate different patterns of resistance depending on the drug combinations used, with the combination MLF/PMM selecting resistant L. donovani more rapidly than the combination AmB/PMM. Significantly, we have also observed that promastigotes/amastigotes resistant to one drug combination show a marked cross-resistance profile to other anti-leishmanial drugs, a finding that could be of major clinical relevance. Additionally, our results indicate that the resistant lines remain protected against the drug-induced loss of ATP and mitochondrial membrane potential.
Trivalent antimony (SbIII), paromomycin (PMM), amphotericin B (AmB), paraformaldehyde, MTT [3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide], Rhodamine 123 (Rh123), buthionine sulfoximine (BSO), FCCP (carbonyl cyanide 4-trifluoromethoxyphenylhydrazone), and Triton X-100 were obtained from Sigma-Aldrich (St. Louis, MO). Miltefosine (MLF) was purchased from Zentaris GmbH (Frankfurt am Main, Germany), and CellTiter-Glo, CellTracker, and 4′, 6-diamidino-2-phenylindole dihydrochloride (DAPI) were purchased from Invitrogen. L-glutamine and penicillin/streptomycin were obtained from Gibco. All chemicals were of the highest quality available.
The L. donovani promastigotes (MHOM/ET/67/HU3) and derivative lines used in this study were grown at 28°C in RPMI 1640-modified medium (Invitrogen) supplemented with 20% or 10% heat-inactivated fetal bovine serum (HIFBS, Invitrogen). For thiol assays, they were grown in M-199 medium (Gibco) supplemented with 10% HIFBS.
The resistant lines were obtained following a previously described step-wise adaptation process [14], [15]. This process started with drug pressure in the wild-type (WT) L. donovani line at a concentration below the drug EC50 (the concentration of the drug required to inhibit parasite growth by 50%), gradually increasing the drug pressure over 10 weeks. After this period, the resistant lines were maintained for eight further weeks at the final drug concentration. The drug combination resistant lines generated, based on WHO recommendations [7], were AmB+MLF (AM), AmB+PMM (AP), AmB+SbIII (the antimonial active form; AS), MLF+PMM (MP) and SbIII+PMM (SP). Singly resistant lines named A, M, P, and S were obtained in a similar manner. All resistant lines were maintained in the continuous presence of drugs. Resistance stability was checked at one and four months after removal from drug pressure. The EC50, resistance index (EC50 ratio for resistant and WT parasites), and cross-resistance profile were determined for each line using an MTT colorimetric assay after incubation for 72 h at 28°C in the presence of increasing concentrations of the drug, as described previously [16].
Six-week-old male BALB/c mice were purchased from Charles River Breeding Laboratories and maintained in the Animal Facility Service of our Institute under pathogen-free conditions. They were fed a regular rodent diet and given drinking water ad libitum. These mice were used to collect primary peritoneal macrophages.
All experiments were performed according to National/EU guidelines regarding the care and use of laboratory animals in research. Approval for these studies was obtained from the Ethics Committee of the Spanish National Research Council (CSIC file CEA-213-1-11).
Mouse peritoneal macrophages were obtained as described previously [17] and plated at a density of 1×105 macrophages/well in RPMI 1640 medium supplemented with 10% HIFBS, 2 mM glutamate, penicillin (100 U/mL) and streptomycin (100 µg/mL) in 24-well tissue culture chamber slides. Late-stage promastigotes from WT and resistant lines were used to infect macrophages at a macrophage/parasite ratio of 1∶10. Eight hours after infection at 35°C in an atmosphere containing 5% CO2, extracellular parasites were removed by washing with serum-free medium. Infected macrophage cultures were maintained in RPMI 1640 medium plus 10% HIFBS at 37°C with 5% CO2 at different drug concentrations. After 72 h, macrophages were fixed for 30 min at 4°C with 2.5% paraformaldehyde in phosphate-buffered saline (PBS; 1.2 mM KH2PO4, 8.1 mM Na2HPO4, 130 mM NaCl, and 2.6 mM KCl adjusted to pH 7), and permeabilized with 0.1% Triton X-100 in PBS for 30 min. Intracellular parasites were detected by nuclear staining with Prolong Gold antifade reagent plus DAPI. Drug activity was determined from the percentage of infected cells and the number of amastigotes per cell in drug-treated versus non-treated cultures [17].
The levels of non-protein thiols were measured by flow cytometry using CellTracker, as described previously [18]. Parasites (107 promastigotes/mL), grown in M199 medium plus 10% HIFBS were washed twice with PBS and incubated with 2 µM CellTracker for 15 min at 37°C. They were then washed again with PBS and analysed by flow cytometry in a FACScan flow cytometer (Becton-Dickinson, San Jose, CA) equipped with an argon laser operating at 488 nm. Fluorescence emission between 515 and 545 nm was quantified using the Cell Quest software. Non-protein thiol-depleted parasites obtained after incubation with 3 mM BSO (a γ-glutamylcysteine synthetase inhibitor) for 48 h at 28°C were used as controls.
ATP was measured using a CellTiter-Glo luminescence assay, which generates a luminescent signal proportional to the amount of ATP present, as described previously [19]. Promastigotes (4×106/mL) were incubated at 28°C in RPMI plus 20% HIFBS containing 0.2 µM AmB or 25 µM MLF for 3 h, or 2 mM SbIII for 8 h. The drug concentration and incubation time were selected by monitoring parasite viability under a microscope. A 25-µL aliquot of parasites was then transferred to a 96-well plate, mixed with the same volume of CellTiter-Glo, and incubated in the dark for 10 min. The resulting bioluminescence was measured using an Infinite F200 microplate reader (Tecan Austria GmbH, Austria).
ΔΨm was measured by flow cytometry using Rh123 accumulation, as described previously [20]. The parasites (4×106 promastigotes/mL) were incubated with the drugs as described above. 0.5 µM Rh123 was then added and the parasites incubated for a further 10 min. They were then washed twice, resuspended in PBS and analysed by flow cytometry in a FACScan flow cytometer (Becton-Dickinson, San Jose, CA) equipped with an argon laser operating at 488 nm. Fluorescence emission between 515 and 545 nm was quantified using the Cell Quest software. Parasites fully depolarized by incubation in 10 µM FCCP for 10 min at 28°C were used as controls.
Statistical comparisons between groups were performed using Student's t-test. Differences were considered significant at a level of p<0.05.
The resistant lines were selected in vitro in L. donovani promastigotes by a stepwise adaptation process, with drug concentrations starting below the EC50 values and gradually increasing, over 10 weeks (equivalent to 90 generations), to a maximum concentration of 0.1, 8, 20 and 80 µM for AmB, MLF, PMM and SbIII, respectively. Resistance to single drugs and to double drug combinations was induced. The singly AmB-resistant line (A) and the lines resistant to the combination of AmB with MLF (AM), PMM (AP) or SbIII (AS) showed similar EC50 values for AmB of 0.14 µM, a value two-fold higher than for the WT line (Table 1). In contrast, the singly MLF-resistant line (M) and the lines resistant to the combination of MLF with AmB (AM) or PMM (MP) showed EC50 values for MLF 1.81, 3.10, and 4.43-fold higher than for the WT line, respectively (Table 1). Likewise, the singly PMM-resistant line (P) and the lines resistant to combinations with AmB (AP), MLF (MP) or SbIII (SP), showed EC50 values for PMM 11.02, 2.12, 14.25, and 18.35-fold higher than for the WT line, respectively (Table 1). Finally, the line resistant to SbIII alone (S) and the lines resistant to SbIII in combination with AmB (AS) or PMM (SP) showed EC50 values for SbIII that were 3.04, 2.05, and 2.18-fold higher than for the WT line, respectively (Table 1). All the resistant lines showed a similar growth rate, morphology, motility and macrophage infectivity to the WT line (data not shown).
We undertook additional experiments to determine whether the resistance to single drugs and drug combinations shown by the promastigote forms was maintained in intracellular amastigotes obtained after infection of mouse peritoneal macrophages (Table 2). The results indicated that the resistance indices to drugs in the different resistant intracellular amastigote lines were maintained and were very similar to those observed in their promastigote counterparts (Tables 1 and 2). The exception was the S-resistant line, which showed a significantly higher resistance index to SbIII (12-fold, Table 2) than that observed for the promastigote form (3-fold, Table 1). The higher resistance index for PMM in the singly P-resistant line (11-fold) and in the MP-(11- and 14-fold for intracellular amastigotes and promastigotes, respectively) and SP- resistant lines (16- and 18-fold for intracellular amastigotes and promastigotes, respectively) it worthy of note. Furthermore, a comparison of AP with P, and AS or SP with S, shows that the doubly resistant lines exhibit lower EC50 values for PMM or SbIII than their singly resistant counterparts. In contrast, the SP line exhibits a higher EC50 value for PMM than the P line (Table 2).
The resistant phenotypes were stable in a drug-free medium for 1 month in the singly A-resistant and AM, AP, and AS doubly-resistant lines for AmB. In contrast, the remaining resistances were unstable, although the EC50 values were higher than for the WT line, except for the M and AP lines, which lost resistance against MLF and PMM, respectively (Figure 1). After culture for four months in a drug-free medium, all lines lost their resistance levels either completely or partially, except the AP line, which maintained a similar initial resistance level for AmB (Figure 1). These findings suggest that the resistance phenotype of the induced drug combination resistance is unstable.
We investigated the cross-resistance profile of the promastigote and intracellular amastigote forms of each resistant line to different anti-leishmanial drugs (Tables 1 and 2). Both the A- and S-resistant lines showed a significant cross-resistance profile to PMM (Table 1), and the P-resistant line showed resistance to SbIII (Table 1). However, the M and P lines only showed resistance to SbIII in their intracellular amastigote forms (Table 2). In the case of drug combination resistant lines, we found that the AP-resistant line showed significant cross-resistance to MLF in both its promastigote and intracellular amastigote forms. Similarly, the AS-resistant promastigote line shows cross-resistance to PMM, and the AM-resistant promastigote line shows cross-resistance to PMM and SbIII (Tables 1 and 2). However, the MP and SP lines showed no cross-resistance to other anti-leishmanial drugs as either promastigotes or intracellular amastigotes (Tables 1 and 2). Our results concerning the stability of the cross-resistance in resistant lines maintained without drug pressure for one month showed that all lines maintained their resistance, except the AM line, which lost its cross-resistance to SbIII (Figure 2).
An increase in thiol levels has been considered to be one of the main detoxification mechanisms observed in lines selected for resistance to SbIII [21]. In light of this, we determined the total intracellular non-protein thiol content in the different resistant promastigote lines using CellTracker. The results of this study showed significantly higher thiol levels in the resistant lines than in the WT line, except for the M and S lines (Figure 3). The highest thiol values were found in the MP and SP lines. As expected, a drastic decrease in thiol content was observed in all lines after incubation with BSO (Figure 3).
We assessed the ATP levels in the presence of AmB, MLF and SbIII, which are known to induce an apoptotic-like process associated with ATP depletion in Leishmania [22]. PMM was not assessed as this drug kills the parasites by a different mechanism [22]–[24]. The WT parasites exhibited a significant decrease in total ATP levels after treatment with AmB, MLF or SbIII, with the former showing the highest decrease (Figure 4A). The lines resistant to single and drug combinations remained protected or were more tolerant to ATP loss (Figure 4A), with the AS-resistant line in particular showing a very small decrease in ATP levels after treatment with AmB or SbIII. A similarly small decrease in ATP levels was also observed after treatment of the M- and MP-resistant lines with MLF, and the S-resistant line with SbIII (Figure 4A). Additionally, we observed that the M- and S-resistant lines presented significantly higher basal ATP levels without drug pressure (Figure 4A), thus suggesting that these resistant lines have developed, amongst other resistance mechanisms, an increase in ATP levels. In contrast, the AM- and AP-resistant lines presented significantly lower basal ATP levels (Figure 4A).
ΔΨm is essential to mitochondrial ATP synthesis and changes to it are one of the markers for apoptosis induced by exposure to AmB, MLF and SbIII (but not exposure to PMM) [22]. Furthermore, mitochondrial oxidative phosphorylation in Leishmania accounts for most of the ATP expenditure of Leishmania parasites [25]. As a result, we tested the ΔΨm of WT and the various resistant lines by measuring Rh123 accumulation (Figure 4B). WT parasites incubated with AmB, MLF or SbIII showed a significant decrease in Rh123 accumulation (7.7-, 2.6- and 1.4-fold, respectively). These values (except for that for SbIII) were even lower than those obtained upon incubation of parasites with the control uncoupling reagent FCCP (Figure 4B). Except for the M-resistant line, the untreated resistant lines showed a significantly lower accumulation of Rh123; however, after treatment with the different anti-leishmanial drugs, the resistant lines showed a lower reduction ratio of Rh123 accumulation than the WT line (Figure 4B). Consequently, the resistant parasites remain protected against the oxidative stress induced by treatment with the anti-leishmanial drugs AmB, MLF and SbIII.
Drug combinations for the treatment of leishmaniasis represent a promising and challenging chemotherapeutic strategy that has recently been implemented in different endemic areas. This approach has several advantages over single-drug therapies, including shortening of the treatment period and reduction of the probability of selecting drug-resistant parasites. However, this approach must be used with care given to the possibility that, if not applied in a controlled and regulated way, resistance could be induced in Leishmania, thus resulting in a rapid loss of efficacy of not one but two therapeutic options [13]. It is therefore important to design relevant experimental studies in order to determine whether Leishmania parasites are able to develop resistance to the different potential anti-leishmanial drug combinations that are likely to be used in the near future. The results obtained from such experimental studies could help to predict the likely success of drug combination therapy.
There is still a great deal of debate concerning the clinical relevance of findings in promastigotes since this is not the stage that will eventually become exposed to the drug. It has recently been shown that differences can be obtained during the experimental induction of resistance to PMM using promastigotes and intracellular amastigotes depending on the resistance-selection protocol [26]. The methodology and technical difficulties required in inducing resistance to drug combinations justify the use of promastigote forms in this manuscript. However, experiments using intracellular amastigotes derived from resistant promastigotes could be useful when considering future recommendations for optimal drug combinations to combat different Leishmania species.
Studies in clinically resistant isolates, where the mechanisms of resistance involve multi-factorial events that contribute to the tolerance to chemotherapeutic agents in Leishmania, are somewhat more complex [27], [28]. In this paper, we have induced experimental resistance to the drug combinations AM, AP, AS, MP and SP in L. donovani (MHOM/ET/67/HU3, also known as LV9 or L82). The sensitivity values of the parental L. donovani strain to the different anti-leishmanial drugs used were similar to, or even lower than, the published data for this and other L. donovani strains [11], [26], [29]–[31]. It is important to point out that this is the first description of an experimental induction of resistance to a combination of different anti-leishmanial drugs. Consequently, the conditions and times required for the induction of resistance can not be compared to the previously described induction of resistance to single anti-leishmanial drugs. Additionally, the single-drug resistance studies in L. donovani described by different groups, obtained higher levels of resistance after increasing the drug pressure and exposure times. Thus, an approximately 14-fold resistance has been obtained for MLF resistance [32], an approximately 5- to11-fold resistance for PMM, depending on the L. donovani strain used [26], and an approximately 20-fold resistance for AmB [31]. Moreover, it is important to note that, in clinical isolates resistant to sodium stibogluconate, an up to 41-fold higher tolerance to SbIII has been observed with respect to the susceptible clones of promastigote forms in the stationary growth phase [27].
In summary, the results of this study show that L. donovani can easily develop resistance to the drug combinations MLF/PMM and SbIII/PMM, with higher resistance indices than those found for AmB/MLF, AmB/PMM or AmB/SbIII. These results have been validated in intracellular amastigotes and are of considerable interest for future applications. Experimental resistance of L. donovani to the drug combination MLF/PMM, a combination that could, in theory, have advantages over other drug combinations as regards future use, is easily achieved. Similarly, the experimental studies described herein confirm how the ease with which experimental resistance to SbIII/PMM is induced in Leishmania. These studies should therefore be taken into account when it comes to future recommendations for their use in endemic areas, especially as the SbIII/PMM combination appears to be effective against VL (in East Africa) and has the additional advantage of low cost [7]. Consequently, further research into combination regimens when given for a short period and at lower total dose are required.
We have also confirmed that Leishmania-resistant parasites develop an increase in cellular thiol redox metabolism as a drug-detoxification mechanism to protect against drug-induced loss of ATP and mitochondrial membrane potential. As described previously, the anti-leishmanial drugs AmB, MLF and SbIII (but not PMM) induce a significant decrease in the mitochondrial membrane potential, thus leading to a bioenergetic collapse of the parasite and drug-induced cell-death [22]. A link between the mode of killing of drugs against Leishmania infantum (such as AmB, MLF and Sb, which share a similar mode of killing), and the tolerance towards cell death induced by their respective anti-leishmanial drugs, has been described previously [22]. Although, in contrast to PMM, this was thought to facilitate the emergence of multidrug resistance, similar findings were not observed under our experimental conditions. Instead, our results show a significant cross-resistance profile to PMM in the AM- and AS-resistant lines and to MLF in the AP-resistant line. Conversely, the AP-resistant intracellular amastigote line acquired cross-resistance to MLF. A similar absence of a link between cross-resistance to drugs with similar mechanism-of-death pathways was observed in the singly A- and S-resistant promastigote lines, with a cross-resistance to PMM, and in the P-resistant line, with resistance to SbIII. The absence of any such correlation could be explained by taking into account that different Leishmania species present different drug susceptibilities and different abilities to respond to drug pressure.
In light of the characteristics of this infectious disease and the existence of different Leishmania species, with their different drug susceptibilities, it is possible that each Leishmania species will require a different drug combination. Suitable options for combination treatment must therefore be optimised in further experimental studies. In this respect, genome-sequencing and metabolomics experiments are currently underway to determine the specific resistance mechanisms developed by Leishmania parasites to different drug combinations. Finally, in view of the proven value of these results for the research community, and considering the debate as regards the use of promastigotes or intracellular amastigotes for induction of drug resistance to drug combinations, we are currently attempting to induce resistance to drug combinations in intracellular amastigotes, as the results obtained will be of greater significance in terms of the conditions found in the field.
|
10.1371/journal.pcbi.1006571 | powerTCR: A model-based approach to comparative analysis of the clone size distribution of the T cell receptor repertoire | Sequencing of the T cell receptor (TCR) repertoire is a powerful tool for deeper study of immune response, but the unique structure of this type of data makes its meaningful quantification challenging. We introduce a new method, the Gamma-GPD spliced threshold model, to address this difficulty. This biologically interpretable model captures the distribution of the TCR repertoire, demonstrates stability across varying sequencing depths, and permits comparative analysis across any number of sampled individuals. We apply our method to several datasets and obtain insights regarding the differentiating features in the T cell receptor repertoire among sampled individuals across conditions. We have implemented our method in the open-source R package powerTCR.
| A more detailed understanding of the immune response can unlock critical information concerning diagnosis and treatment of disease. Here, in particular, we study T cells through T cell receptor sequencing, as T cells play a vital role in immune response. One important feature of T cell receptor sequencing data is the frequencies of each receptor in a given sample. These frequencies harbor global information about the landscape of the immune response. We introduce a flexible method that extracts this information by modeling the distribution of these frequencies, and show that it can be used to quantify differences in samples from individuals of different biological conditions.
| Recent advances in high-throughput sequencing of the T cell receptor (TCR) repertoire provide a new, detailed characterization of the immune system. T cells, each displaying a unique TCR, are capable of responding to presented antigens and initiating an adaptive immune response. An immune response is described by rapid proliferation of T cell clonotypes whose TCRs are specific to the antigen. In humans, it is estimated that the body is capable of producing more than 1018 different TCRs [1, 2], where high diversity of the TCR repertoire implies a greater range of pathogens that can be fought off. A variety of studies have been published demonstrating the value in characterizing this immune response for purposes such as describing tumor cell origin [3] and predicting response to cancer therapy and infection [4]. The applications of TCR sequencing are many, but this type of data presents new needs for analysis techniques not met by existing tools for other kinds of genomic experiments.
Several groups have identified that the distribution of larger clone sizes in a sample can be approximated by a power law [5–8], which means that the number of clones of a given size decays approximately as a power of the clone size. This heavy-tailed distribution comes as a consequence of extensively proliferated clones actively participating in an ongoing immune response. More recent work has aimed to quantify statistically the diversity of the TCR repertoire, initially through the use of various estimators borrowed from ecology, such as species richness, Shannon entropy [9], and clonality. These estimators are known to be highly sensitive to sample size and missing observations. Given that the TCR repertoire is mostly populated by rare clonotypes, many of the clonotypes in the system are absent from any one sample. This presents a challenge to many of the ecological estimators. Model-based approaches to approximating the clone size distribution have also been proposed, with the goal of providing added stability and consequently more statistical power. Some examples are the Poisson-lognormal model [10], Poisson mixture models [11, 12], and a heuristic ensemble method [13]; however, these models lack a biologically meaningful interpretation, and further do not sufficiently account for the power law-like nature of the data. That is, power law distributions are heavier-tailed than the Poisson or even the lognormal distribution, leading to systematic bias in the model fit.
Previous research has also identified the imperfectness of the power law behavior for the clone size distribution below some clone size threshold [7, 8]. To handle the imperfectness, [7, 8] proposed to model large clones above the threshold using a type-I Pareto distribution, which is a member of the power-law distribution family, and omitted the clones with frequency below that threshold. The threshold is either user-specified or determined from the data based on a goodness-of-fit measure. Indeed, this model has certain biological basis. Through a stochastic differential equations setup that models the birth, death, selection, and antigen-recognition of cells active in the immune system, Desponds et al. [8] showed that the upper tail of the clone size distribution at equilibrium approximately follows a type-I Pareto distribution (Fig 1A). Unlike the Poisson and lognormal models, parameters in this model are related to relevant actors in the immune response, and can reveal certain biological insights into immune response, such as average T cell lifetime [8]. Yet, the resulting model excludes all clones below a certain frequency threshold.
However, even small clones may provide information; for example, Desponds et al. [8] indicated that the generation of new T cells affects the landscape of smaller clone sizes. Other studies have shown that low-frequency clones may support a diverse immune system and present a potential to mobilize against antigens, as in some cases having a clone size distribution highly dominated by a few clones has been correlated with unfavorable clinical outcome [14, 15]. With this in mind, we sought a means to exhaust all available data and consider modeling the complete clone size distribution.
To address this question, we propose a novel statistical tool, called powerTCR, to characterize the full distribution of the TCR repertoire. Our method models large clones that are above the threshold, where the power law begins, using the generalized Pareto distribution (GPD), which contains the type-I Pareto distribution as a special case, but provides a more flexible fit. It also models the small clones below the threshold using a truncated Gamma distribution. It determines the threshold in a data-driven manner simultaneously with the characterization of clone size distribution. Our final model contains parameters that are analogous to those found in the type-I Pareto model of Desponds et al. [8], relating our model to the biological interpretation of the dynamics of the immune system. Altogether, this allows our model to more accurately describe the shape of the clone size distribution for both large and small clones. Such a model is well suited for providing a global view of the state of the immune repertoire. It can also be employed to perform comparative analysis of healthy and compromised individuals to identify descriptors of strengths and deficiencies in the immune system.
Our goal is to model the clone size distribution of a sample immune repertoire. Fig 1A shows a typical distribution plotted using the repertoire of a Sarcoidosis patient in [16]. If the data are truly Pareto distributed, this plot would appear linear [17, 18]. However, noting the linear behavior is only true for the far upper tail of the data, this suggests that these data are a departure from the Pareto distribution. This imperfect power law implicates the use of a heavy-tailed distribution above some threshold and a lighter-tailed distribution below that threshold. Here, we model the tail part with a GPD. The GPD, introduced by [19], is a classical distribution typically used to model the values in the upper tail of a dataset. This formulation results in a distribution with density
f ( x ) = 1 σ ( 1 + ξ x - u σ )- ( 1 / ξ + 1 ) , (1)
where u ∈ (−∞, ∞) is a threshold that typically needs to be prespecified, σ ∈ (0, ∞) is a scale parameter, and ξ ∈ (−∞, ∞) is a shape parameter. The GPD has support x ≥ u when ξ ≥ 0 and u ≤ x ≤ u − σ/ξ when ξ < 0. We model the bulk part with a Gamma distribution with the upper tail truncated at the threshold. The Gamma distribution has a flexible shape and can fit many different clone size distributions. The threshold and the parameters in the two distributions are estimated from the data simultaneously. This setup, where data above and below an unknown threshold are drawn from the “bulk” and “tail” distributions respectively, falls into a class of models called spliced threshold models. The typical motivation for the model is the belief that the data above and below the threshold are driven by different underlying processes. We refer the interested reader to [20] for a thorough review of the general spliced threshold model, and its applications in fields such as insurance, hydrology, and finance.
Denote the proportion of data above the threshold u as ϕ. Let the bulk model distribution function be Hc(x|θb) and the tail model distribution function be Gc(x|θt), where subscripts b and t denote the bulk and tail model parameter vectors, respectively. Then the distribution function of the model is given by
F c ( x ) = { ( 1 - ϕ ) H c ( x | θ b ) H c ( u | θ b ) for x ≤ u 1 - ϕ + ϕ G c ( x | θ t , u ) for x > u (2)
with corresponding density
f c ( x ) = { ( 1 - ϕ ) h c ( x | θ b ) H c ( u | θ b ) for x ≤ u ϕ g c ( x | θ t , u ) for x > u . (3)
Because the clone size distribution is count data that typically exhibit numerous ties in the less frequently observed clonotypes, it is appropriate to treat this as a discrete problem. We modify the model in order to account for any quantized or censored data. Let ψ and Ψ be the density and distribution function of a continuous distribution, and let d be the interval length at which the data are censored. We obtain a quantized analog of ψ by letting
Pr ( X = x ) = Ψ ( x + d ) - Ψ ( x ) , x ∈ k · d , k ∈ Z .
This results in a discrete model with distribution function
F ( x ) = { ( 1 - ϕ ) H ( x | θ b ) H ( u - d | θ b ) for x ≤ u - d 1 - ϕ + ϕ G ( x | θ t , u ) for x ≥ u (4)
and corresponding probability mass function
f ( x ) = { ( 1 - ϕ ) h ( x | θ b ) H ( u - d | θ b ) for x ≤ u - d ϕ g ( x | θ t , u ) for x ≥ u . (5)
where h(x|θb) ∼ discrete Gamma(α, β), g(x|θt) ∼ discrete GPD(u, σ, ξ), and d = 1, which specifies that we model integer data (see Methods for the functional form of the discrete Gamma distribution and the discrete GPD). This discretization step turns out to be important for accurate estimation in our scenario. See S2 Text for a comparison between the performance of the discrete and continuous models in settings resembling true clone size distributions.
The relationship between the discrete Gamma-GPD spliced threshold model and the type-I Pareto model in Desponds et al. [8], hereafter referred to as the Desponds et al. model, allows us to draw connections between some of our parameters and the dynamics of immune response underpinning their approach. First, results from [8] show that the threshold at which the power law begins is indicative of the point over which a clone’s large size can be attributed to active immune response, as opposed to noise in the body that arises from processes such as self-recognition. The threshold fitted from the data provides an objective way to narrow down which clonotypes from a sample repertoire should be interrogated further. This notion is convenient for studying factors such as CDR3 (complementarity-determining region 3) amino acid motifs or specific V, D, and J genes important for combating certain antigens, which are typically determined based on a heuristic abundance cutoff. For example, [21] studies the 1,000 most abundant CDR3 amino acid motifs across all sampled peripheral blood mononuclear cell (PBMC) libraries, while [16] determines CDR3 amino acid motifs from clones that are present with 10 or more reads in a sampled repertoire. The threshold u estimated with our model, however, introduces a means to select motifs that does not rely on heuristics and automatically scales with sequencing depth.
Moreover, the shape parameter ξ of the GPD is inversely related to the shape parameter αd used in the Desponds et al. model (see Methods). As explained by Desponds et al., a small αd, i.e. a large ξ, implies increased average T cell lifetime and antigenic noise strength. They further show that antigenic noise strength grows as a consequence of a higher initial concentration of antigens and a higher rate at which new antigens are introduced. Interestingly, ξ also positively correlates with the familiar clonality estimator (1-Pielou’s evenness [22]). Indeed, as ξ increases, the clone size distribution becomes heavier-tailed—that is, more skewed towards dominating clones. This trend is in line with that of the clonality estimator, which favors a more uniform clone size distribution as clonality approaches 0 and a distribution dominated by expanded clones as clonality approaches 1. To numerically validate this relationship, we simulated the data from our model and computed the clonality (see Methods). We observed a high correlation between clonality and ξ (Spearman’s ρ ≈ 0.9), confirming that ξ reflects the skewness towards dominating clones (see S3 Text).
It is worth noting that our model acquires a theoretical gain via the threshold stability property of the GPD [23]. That is, for any generalized Pareto distributed data, the shape parameter ξ remains constant regardless of changes in u. In our context, this means that at decreasing sequencing depths, though the threshold u would decrease due to fewer cells being sampled, the shape parameter ξ in principle would be stable against the variation in sequencing depth. We will demonstrate this gain in stability on a murine tumor dataset. See Methods for our extension of the threshold stability property to the case of the discrete GPD.
In the following sections, we inspect four different datasets using our model. We compare our results to results from the Desponds et al. model to demonstrate the practical and theoretical benefits of our approach. We also make comparisons to results from the widely used richness, Shannon entropy, and clonality estimators. See Methods for information on computation of competing methods.
The expression of major histocompatability complex II (MHC-II) proteins in tumors correlates with boosted anti-tumor immunity. As part of a study of how MHC-II expression impacts tumor progression and functional plasticity of T cells [24], the CDR3 of TCRβ-chains of tumor infiltrating lymphocytes (TILs) were sequenced from breast cancer tumor tissue from six BALB/c mice [25]. Three of the mice were grafted with MHC-II expressing tumor cells and three control animals received parental MHC-II-negative cells. Samples were collected at 21 days after the date of treatment. Table A in S4 Text summarizes the number of unique clonotypes observed and the total number of reads in each sample.
Sarcoidosis is an inflammatory disease that typically is accompanied by an accumulation of activated CD4+ T cells in the lungs. A particularly acute form of Sarcoidosis, called Löfgren’s syndrom (LS), occurs with additional, more severe symptoms. A known signature of LS is the bombardment of the lungs with CD4+ T cells, which is expected to significantly alter the entire landscape of the TCR repertoire. We applied our method to TCR repertoire data of LS and non-LS Sarcoidosis patients [29], originally described in [16]. In this study, bronchoscopy with the bronchoalveolar lavage was performed on a cohort of 9 LS and 4 non-LS individuals and prepared for TCR α− and β–chain sequencing.
We compared the TCR distribution between LS and non-LS Sarcoidosis patients using our method and the competing methods. In order to visualize closeness of samples, we generated a distance matrix using JSD between fitted distributions using our method and the Desponds et al. model. The estimated parameters are in Tables E and F in S4 Text respectively. We then applied non-metric multidimensional scaling (MDS) to the distance matrix and plotted the first two coordinates. For the ecological estimators, we simply plotted centered and scaled estimates. As shown in Fig 4A, results from our model cluster LS patients into a tight group distinct from non-LS patients, bolstering the claim that LS patients exhibit a signature immune response. On the other hand, competing methods fail to uncover any pattern (Fig 4B and 4C).
We applied our method to data collected during a clinical trial of 13 glioblastoma patients receiving autologous tumor lysate-pulsed dendritic cell (DC) vaccine therapy [30], first detailed in [31]. Three intradermal injections were administered to patients at biweekly intervals. TCRβ-chains from PBMC samples were sequenced for the patients prior to vaccinations and two weeks following the final injection. Patients were followed up with and their time to progression (TTP) and overall survival (OS) were recorded. TTP was defined as the time from the first DC vaccination until MRI-confirmed tumor progression. OS was calculated as the time from the first DC vaccination until the patient’s death from any cause. We investigated whether current tools using TCRs sequenced only from blood samples indicate anything about patients’ survival time and time to progression.
We first fit our model to the pre- and post-treatment samples. In both cases, we classified the patients into two groups using the hierarchical clustering based on our model, the Desponds et al. model, and the richness, Shannon entropy, and clonality estimators. No clear grouping with respect to either TTP or OS could be observed from any clustering on the pre-treatment samples, whether by the model-based methods or the selected estimators (see S7 Text). However, among post-treatment samples, our method tends to cluster together patients with better clinical outcome (Fig 5A). This may indicate that the DC therapy alters the landscape of the TCR repertoire into a form that promotes favorable clinical outcome.
We do, however, cluster one patient (ID: 33296) with low TTP and OS in the group with overall higher TTP and OS. Interestingly, this misplaced patient had the lowest estimated TIL count and tumor/PBMC overlap of the entire cohort (S4 Text, Table G). Tumor/PBMC overlap was defined as the total number of reads of shared CDR3s normalized by total reads in the tumor and PBMC samples. Similarly, patient 17232 displayed among the best clinical outcome but clustered with lower-performing patients. Patient 17232 had the highest TIL count and level of tumor/PBMC overlap in the whole cohort. This information taken as a whole suggests that, while the clone size distribution found in blood may indicate something about a patient’s response to treatment, it still does not guarantee that T cells will infiltrate the tumor, an important factor for clinical benefit [32]. S8 Text highlights the clone size distributions of these two patients against all others.
Notably, inferred thresholds (minimum u = 4, maximum u = 6) on this dataset are much lower than on other datasets. This is likely because this dataset contains less deeply-sequenced samples than the others, which consequently reduces the threshold.
Noting that clones with size at or above the estimated threshold are considered active participators in the immune response, we sought to investigate whether any relationship existed between clinical outcome and the proportion of more highly stimulated cells. We defined the proportion of highly stimulated cells to be the total number of reads at or above the threshold, normalized by the total number of reads in the entire repertoire (S3 Text, Table I). We found correlations between this measure and both TTP (Spearman’s ρ = 0.54) and OS (Spearman’s ρ = 0.80). Rank scatterplots for these correlations are in Fig 5B. The positive correlation we uncovered suggests that this statistic could be a useful tool to quantify the antigen-specificity of the sample.
Risk factors for type 1 diabetes (T1D) are known to be heritable, yet genes alone are not sufficient explanation for drivers of the disease. Studies of monozygotic twins have revealed that, given one twin has T1D, the other will only have it at most half of the time [33]. The CD4+ T cell is viewed as the initiator of T1D as dysregulation of CD4+ antigen-recognition drives the autoimmune disease. Seeking out apparently non-heritable determinants of T1D, [34] conducted a deeper investigation of the CD4+ T cell. Briefly, the authors obtained PBMCs from 14 volunteer healthy donors (HDs) and 14 recently diagnosed patients with T1D. The cells were sorted using flow cytometry into distinct T cell subsets (true naïve; TN, central memory, CM; regulatory, Treg; and stem cell-like memory, Tscm) and TCRβ-chains were sequenced. The authors conducted a thorough analysis, finding shorter CDR3 sequence lengths and lower overall repertoire diversity among patients with T1D. However, on a per-individual basis, the authors were unable to uncover a relationship between repertoire diversity and disease status. Since the the spliced threshold model provides a new means to probe this complex data, we applied our approach to complement the original analyses.
We have developed a model, the discrete Gamma-GPD spliced threshold model, and demonstrated its utility on several datasets. As shown in our analyses, several biologically relevant descriptive features can be obtained from our model. One is the tail shape parameter ξ, a measure of the weight of the upper tail of the clone size distribution, where a heavier tail of the fitted model implies a more dominated distribution of expanded clones. Another is the proportion of total reads at or above the estimated threshold, a possible measure of intensity of the immune response. The third is the estimated threshold, which is a useful guide to objectively identify CDR3 motifs for downstream analysis. This could involve denoting motifs as only those CDR3s found in TCRs with frequencies at or above the estimated threshold for a given sample, or it could mean studying TCR gene usage among that same group of clonotypes. Though the dynamics driving our model form a compelling argument for this interpretation of the threshold, we acknowledge that further biological validation on more datasets is still needed to confirm this.
Similar to other estimators, our model requires that a repertoire be adequately sampled. Without adequate sampling, the differentiating features between TCR repertoires will be masked [8], and the estimated model parameters will not be reliable. Given the immense diversity of the TCR repertoire, one should in general be cautious about using any method to make inference about a sample TCR repertoire when few cells are sequenced. With sufficient samples, though, the spliced threshold model provides the user a meaningful high-level view of the TCR repertoire.
The diversity of the TCR repertoire and its responsiveness to stimuli provide a high-dimensional biomarker for monitoring the immune system and its adaptivity. Robust assessment of the clone size distribution through TCR sequencing is important for understanding this diversity. The discrete Gamma-GPD spliced threshold model is a flexible model that effectively captures the shape of the clone size distribution. It is especially appropriate since the heavy-tailed GPD is a good fit to model the highly expanded clones that dominate many TCR repertoire samples. The method also provides a means to comparatively analyze a collection of TCR repertoire samples while maintaining convenient theoretical properties and interpretations.
Compared with existing approaches, our method is more flexible, utilizes the full clone size distribution, is less sensitive to sequencing depth, and identifies the threshold in a data-driven manner. The parameters estimated from our method are biologically relevant and instructive to the dynamics of immune response. Our results on multiple datasets also show that the spliced threshold model is powerful in a range of scenarios for comparing TCR repertoires across samples, revealing potential trends in the landscapes of clone size distributions of affected immune systems.
We use maximum likelihood to estimate the parameters of our model. First, we more explicitly specify the form of our distribution. Letting x ∼ Gamma(α, β), we write the probability mass function of a discrete Gamma distribution as
h ( x ) = 1 Γ ( α ) [ γ ( α , β ( x + 1 ) ) - γ ( α , β x ) ]
for α > 0, β > 0, x ∈ Z, and where γ(α, βx) is the lower incomplete gamma function
γ ( α , β x ) = ∫ 0 β x t α - 1 e - t d t .
If x ∼ GPD(u, σ, ξ), we write the probability mass function of a discrete GPD as
g ( x ) = ( 1 + ξ x - u σ ) - 1 / ξ - ( 1 + ξ x + 1 - u σ ) - 1 / ξ
for u ∈ (−∞, ∞), σ ∈ (0, ∞), and ξ ∈ (−∞, ∞). The discrete GPD has support x ≥ u when ξ ≥ 0 and u ≤ x ≤ u − σ/ξ when ξ < 0, where x ∈ Z. In all analyses presented here, we make no assumptions on the sign of ξ, although empirically we tend to observe ξ > 0.
To proceed, we employ a profile likelihood approach. Let u be the threshold, θb be the bulk parameter vector {α, β}, θt be the tail parameter vector {σ, ξ}, and θ be the parameter vector {θb, θt}. Let also h and H be the density and distribution function of a discrete Gamma distribution, respectively, and let g be the density of a discrete GPD. Then the complete data likelihood is given by
L ( { θ , u } ∣ x ) = ∏ i = 1 n [ ( 1 - ϕ ) h ( x i ∣ θ b ) H ( u - 1 ∣ θ b ) 1 ( x i ≤ u - 1 ) + ϕ g ( x i ∣ θ t , u ) 1 ( x i ≥ u ) ]
and the profile likelihood of the model at u is denoted as
L p ( u ) = max θ L ( θ ∣ x , u ) .
A grid search over a suitable range of thresholds u⋆ = (u1, …, uk) may be implemented to maximize the profile likelihood. In this study, we adopted an approach similar to those of [19] and [40], searching for thresholds at or above the 75% quantile of the sample. The estimated parameters are then
u ^ = arg max u ⋆ ∈ u ⋆ L p ( u ⋆ ) , θ ^ = argmax θ L ( θ ∣ u = u ^ ) , and ϕ ^ = n u n ,
where n is the total number of clones and nu denotes the number of clones with size greater than or equal to the threshold.
The Desponds et al. model was fit as previously described [8]. Briefly, the model has density
f ( x ) = α d u α d x α d + 1 (6)
and distribution function
F ( x ) = 1 - ( u x ) α d (7)
where u > 0 is the threshold and αd > 0 is a shape parameter. For each sample TCR repertoire, a grid of potential thresholds u⋆ = (u1, …, uk) was constructed by considering every unique clone size in the repertoire. Then, for each ui, the shape parameter is estimated as
α ^ d = n i [ ∑ j = 1 n i ln x j u i ]- 1 (8)
where ni is the number of clones with size larger than the threshold ui. Once this value is computed for every threshold in u⋆, the threshold and corresponding α ^ were chosen to minimize the Kolmogorov-Smirnov statistic.
The ecological estimators [9, 22] were computed as follows. For a sample X, let S(X) be the sample richness, defined as the number of unique clonotypes in X, and let pi be the number of cells of clonotype i normalized by the total number of cells in the sample. Then, the Shannon entropy of X is
H ( X ) = - ∑ i = 1 S ( X ) p i ln p i (9)
and the clonality of X is
C ( X ) = 1 - H ( X ) ln S ( X ) . (10)
The Desponds et al. model, which is a type-I Pareto distribution, and the “tail” part of our model, which is a GPD, are closely related. In fact, the GPD contains the type-I Pareto distribution as a special case. We can write the distribution function of y, where y ∼ GPD(u, σ, ξ), as
F ( y ) = 1 - ( 1 + ξ y - u σ ) - 1 / ξ . (11)
Now, let x ∼ GPD ( u , u α d , 1 α d ). Then
F ( x ; u , u α d , 1 α d ) = 1 - [ 1 + 1 α d ( x - u u / α d ) ] - α d = 1 - [ 1 + ( x - u u ) ] - α d = 1 - ( u x ) α d
which is exactly the distribution function of a type-I Pareto distribution with threshold u and shape αd (Eq 7). Of course, this exact relationship only holds when σ = u α d. Nevertheless, αd and ξ perform the same function in their respective distributions, adjusting the weight of the tail. This relationship always holds—a larger ξ (smaller αd) implies a heavier-tailed distribution, while a smaller ξ (larger αd) implies a lighter-tailed distribution.
We conjecture that ξ, the shape parameter of the GPD, positively correlates with clonality. We numerically validated this claim using a simulated cohort of 48 clone size distributions. That is, we generated samples of n = 20, 000 clonotypes, where our 48 parameter settings were derived from every combination of α ∈ {3, 5, 10}, ξ ∈ {.25, .5, .75, 1.1}, and ϕ ∈ {0.1, 0.15, 0.2, 0.25}. We chose β = 0.15, σ = α β, and u = ⌊Qα,β(1 − ϕ)⌋ in each simulation, where Qα,β is the quantile function of the Gamma distribution with mean α β. To adjust for the effect of sample size on clonality, we downsampled the simulated data so that each sample contained the same number of reads (415,989 total reads per sample). We computed the clonality of each simulated TCR repertoire on these adjusted datasets.
The relationship between a pair of TCR repertoires can be elucidated by evaluating the distance between their fitted spliced threshold models. Several methods to compare densities are available. We propose measuring the distance between each pair of distributions using Jensen-Shannon distance (JSD) [41]. This metric is a symmetric and smoothed adaptation of the well-known Kullback-Leibler divergence that does not require the distributions under comparison to share the same support.
Given discrete distributions P and Q, the JSD between P and Q is
J S D ( P , Q ) = 1 2 [ ∑ i ( P i ln P i M i ) + ∑ i ( Q i ln Q i M i ) ] , (12)
where M i = 1 2 ( P i + Q i ). The resulting distances allow analysis and visualization via MDS or hierarchical clustering of the samples. Throughout our study, we use Ward’s method for hierarchical clustering.
The threshold stability property of the GPD is well-established [23]. Here, we show that the property also holds for the discrete GPD. Let X ∼ discrete GPD(u, σ, ξ) and denote its distribution function as F with Fc as its continuous analog. Then we can write
P ( X - u ≤ x + 1 | X ≥ u ) = P ( u ≤ X ≤ x + u + 1 ) P ( X ≥ u ) = F ( x + u + 1 ; u , σ , ξ ) - F ( u ; u , σ , ξ ) 1 - F ( u ; u , σ , ξ ) = F c ( x + u + 2 ; u , σ , ξ ) - F c ( u + 1 ; u , σ , ξ ) 1 - F c ( u + 1 ; u , σ , ξ ) = ( 1 + ξ σ ) - 1 / ξ - ( 1 + ξ x + 2 σ ) - 1 / ξ ( 1 + ξ σ ) - 1 / ξ = 1 - ( 1 + ξ x + 1 σ + ξ ) - 1 / ξ = F c ( x + 1 ; 0, σ + ξ , ξ ) = F ( x ; 0, σ + ξ , ξ ) .
This states that if X ∼ discrete GPD(u, σ, ξ), then X − u ∼ discrete GPD(0, σ + ξ, ξ). Or, for our application, consider a clone size distribution, where clones larger than some threshold u are distributed according to the discrete GPD. At decreasing sequencing depths, this estimated u decreases, implying naturally that the size a clone in the sample must achieve to be considered “expanded” decreases. Still, while u shrinks, the threshold stability property states that ξ remains constant.
|
10.1371/journal.pgen.1002232 | Frequent Beneficial Mutations during Single-Colony Serial Transfer of Streptococcus pneumoniae | The appearance of new mutations within a population provides the raw material for evolution. The consistent decline in fitness observed in classical mutation accumulation studies has provided support for the long-held view that deleterious mutations are more common than beneficial mutations. Here we present results of a study using a mutation accumulation design with the bacterium Streptococcus pneumoniae in which the fitness of the derived populations increased. This rise in fitness was associated specifically with adaptation to survival during brief stationary phase periods between single-colony population bottlenecks. To understand better the population dynamics behind this unanticipated adaptation, we developed a maximum likelihood model describing the processes of mutation and stationary-phase selection in the context of frequent population bottlenecks. Using this model, we estimate that the rate of beneficial mutations may be as high as 4.8×10−4 events per genome for each time interval corresponding to the pneumococcal generation time. This rate is several orders of magnitude higher than earlier estimates of beneficial mutation rates in bacteria but supports recent results obtained through the propagation of small populations of Escherichia coli. Our findings indicate that beneficial mutations may be relatively frequent in bacteria and suggest that in S. pneumoniae, which develops natural competence for transformation, a steady supply of such mutations may be available for sampling by recombination.
| Beneficial mutations have long been considered extremely rare events and were thought to occur with a frequency of approximately one out of a billion times that a bacterium replicates its genome. Rare beneficial mutations would then be amplified by natural selection from the more frequent background of harmful mutations. Mutation accumulation experiments probe the nature of these spontaneous mutations by monitoring changes in fitness of model organisms propagated in the laboratory through numerous generations under conditions where the effects of selection are minimal. Previous mutation accumulation experiments have shown that organisms under study have declined in fitness as random mutations accrue in their genomes, consistent with a predominance of deleterious mutations. We conducted a mutation accumulation study with the bacterial pathogen S. pneumoniae in which a broad measure of fitness instead rose. We demonstrate that this unexpected adaptation was due to frequent beneficial mutations that were further amplified by selection in stationary-phase bacterial colonies. Together with recent work using E. coli, these results demonstrate that beneficial mutations can be common in bacteria and may contribute to our understanding of the evolution of traits such as antibiotic resistance and virulence.
| Spontaneous mutations provide the underlying variability on which selection acts to drive evolution. Among newly-arising mutations that impact fitness, the balance between those that are beneficial and those that are deleterious has appeared to be heavily skewed toward maladaptive changes. This view has been supported by mutation accumulation (MA) studies, in which experimental lines of model organisms, including Drosophila [1], [2], Escherichia coli [3], [4], Caenorhabditis elegans [5], [6], Arabidopsis thaliana [7], Tetrahymena [8], Saccharomyces cerevisiae [9] and Salmonella typhimurium [10], have lost fitness when propagated under conditions of relaxed selection. Analysis of such MA lines in E. coli yielded an estimate for the rate of deleterious mutations as being at least 1.7×10−4 events per genome per generation [3]. In contrast, beneficial mutations were believed to occur in bacterial populations only at frequencies near 10−8 to 10−9 [11], [12].
More recent evidence, however, indicates that beneficial mutations are far more common in E. coli than previously recognized. In large populations of asexual microbes such as bacteria, clonal interference between competing mutations can prevent the fixation of beneficial mutations and limit the overall rate of adaptation [13]. Overcoming the experimental bias imposed by clonal interference by propagating populations of E. coli at a small effective population size (Ne), Perfeito et al. recently demonstrated in this model organism that beneficial mutations occurred as frequently as 2×10−5 per genome per generation [14].
We address here the frequency of beneficial mutations in the gram-positive bacterium, Streptococcus pneumoniae. This organism is a common pathogen of the human respiratory tract for which questions of the acquisition and spread of beneficial mutations have taken on additional medical importance as the bacterium has developed increasing resistance to antibiotics [15] and has shifted its population structure in response to the pressure of vaccination [16]. In contrast to E. coli, S. pneumoniae is a naturally-transformable bacterium in which recombination reduces the clonality of the population [17], [18] and may thereby facilitate more rapid adaptation. We present unexpected results from an MA experiment with S. pneumoniae in which beneficial mutations were sufficiently frequent to cause—in combination with selection in stationary-phase colonies—a net increase in fitness of the propagated populations. These findings support the emerging conclusion that beneficial mutations may be relatively frequent in bacteria.
Initially intending to characterize the rate of accumulation of deleterious mutations in S. pneumoniae, we designed an in vitro serial transfer experiment using a classical mutation accumulation design under conditions of weak selection and strong genetic drift. Forty bacterial lines were passaged in parallel over a period of 210 days by transferring single colonies each day to fresh THY agar plates. We anticipated a general decline in fitness among these lines due to accumulation of deleterious mutations in the setting of frequent population bottlenecks, as has been previously described in studies with Escherichia coli and Salmonella [3], [10]. Unanticipated results described below, however, suggested that adaptation had occurred despite these conditions and caused us to revise our plans so as to investigate this adaptation in detail. Evidence for adaptation emerged when, intending to exclude the possibility of specific adaptation to the conditions of growth on THY agar plates during passage, we tested a sample consisting of 24 consecutive lines out of the 40 that had been passaged. Fitness for growth on the agar surface was estimated by dissecting individual colonies from plates after 24 h of incubation and measuring the number of colony-forming units (CFU) in each mature colony. As described below, the 24 lines tested consisted of 12 lines each in wild-type and competence-deficient ΔcomAB backgrounds. Unexpectedly, the fitness of these bacterial lines—as measured by log2(CFU/colony) values—generally rose as the experiment progressed (Figure 1A and 1C). Because increasing fitness did not match the predictions of the mutation accumulation model under which the experiment had been planned, the scope of the project was revised at this point to focus on characterizing the apparent adaptation seen among these 24 serial transfer lines that had been initially tested.
Because pneumococcus is a naturally-transformable bacterium [19], [20] and we had initially been interested in later testing whether horizontal gene transfer between populations that have independently accumulated deleterious mutations can restore fitness, half of these lines had been initiated from a wild-type progenitor whereas the other half were derived from an isogenic, competence-deficient ΔcomAB mutant. In comparing the wild-type and ΔcomAB lines (12 lines of each genotype), we did not expect substantial differences between the groups because the THY medium used during this experiment is not permissive for development of spontaneous competence when bacteria are grown in broth culture and because daily serial transfer was expected to limit the time available for selection of recombinants after a transformation event. We could not, however, exclude the possibility that a fraction of the bacteria within a colony might develop competence at some point during the growth cycle and therefore examined our data for evidence of differences between wild-type and ΔcomAB lines. No significant differences in fitness were observed between the groups (Figure 1A), and data from these two sets of lines were pooled for subsequent analyses.
Among-line variance in fitness also increased over the course of serial transfer, consistent with mutational divergence of the passaged lines (Figure 1B). In contrast, within-line variance remained low throughout the experiment. Likewise, the among-line variance measured for the day 1 reference strains, which were analyzed in parallel with the evolved strains at each time point, did not increase as the experiment progressed.
To determine whether this increase in fitness was specific to the context of bacterial growth on the surface of agar plates, we next measured the growth of each isolate in THY broth cultures and estimated fitness in this setting based on the maximum rate of growth. In contrast to CFU/colony measurements (Figure 1C), growth rates in broth culture for many lines declined over the first 126 days of serial transfer (Figure 1D). The passaged isolate from one wild-type line even appeared to have lost the ability to grow in broth culture altogether. The adaptation of these pneumococcal lines therefore appeared to be specific for survival on agar plates and was further characterized by dissecting colonies off plates at intervals ranging from 7 to 26 h after inoculation. These time-series measurements of bacterial CFU/colony were performed for a sample consisting of the first 6 lines in the collection (representing 3 wild-type and 3 ΔcomAB lines, Figure 2) using isolates taken at days 1 and 126 of serial transfer. Developing colonies displayed an initial phase of active growth that extended through the first 11 h following inoculation. During this phase, growth of the colonies was approximately exponential. In this period, growth of the day 126 isolates lagged behind that of the day 1 isolates, consistent with the lower maximal rate of growth seen in the broth assays for the former samples.
After the initial phase of exponential growth, colonies entered what we refer to here broadly as stationary phase. Although changes in CFU/colony values observed during this period indicate that bacterial replication and death continue to at least some extent, this phase is characterized by birth and death rates being approximately equal. Within this stage, viable bacterial counts in maturing colonies briefly declined, followed by a slow rebound and plateau. It was during this stationary phase of colony development that the advantage of the passaged isolates became evident. When data were examined for each line individually, one line (shown as line B in Figure S1A and S1B) was found to have acquired a markedly greater stationary phase advantage than was seen in the other five lines. Even when this line was excluded, however, a pattern indicating adaptation to survival in stationary phase was seen among the remaining evolved isolates (Figure S1C). Notably, this advantage was particularly prominent from 22 to 24 h after inoculation, matching the time interval at which colonies were picked for serial transfer.
Although the definition of fitness is complicated in this serial transfer model by repeated exposure to both exponential and stationary phase growth conditions, it is important to note that the CFU/colony value at the end of a 24 h colony growth cycle reflects the net result of events during both the exponential and stationary phases of colony development. In this regard, CFU/colony measurements made at the end of the 24 h colony growth cycle reflect overall fitness of the passaged lines under the conditions of experimental propagation. Increases in CFU/colony observed over the course of the experiment—while driven by improved survival during stationary phase—therefore indicate increased overall fitness of the passaged lines. Because our serial transfer model did not permit seeding of individual colonies from a mixed inoculum, CFU/colony values provide a proxy for absolute fitness of strains in isolation but do not directly measure relative fitness. Potential implications of differences between relative and absolute fitness, and particularly the possibility of frequency-dependent variation in the strength of selection, are considered later.
Because mutations that are advantageous during stationary phase may be deleterious under conditions of rapid growth [21], we considered whether the decline in broth culture growth rates observed with passage might result in part from pleiotropic effects of mutations that were adaptive in the context of colonies grown on an agar plate. No correlation was observed between these two fitness measurements among the isolates that had been passaged for 126 days (Figure 3A). However, because some of these isolates may have accumulated multiple mutations and because the magnitude of an adaptive benefit may not correlate with the magnitude of a linked pleiotropic effect, the absence of such a correlation among the day 126 isolates cannot exclude the possibility that antagonistic pleiotropy contributes to the observed fitness changes.
To examine the relationship between these fitness changes more closely, a single bacterial line (shown in Figure S1 as line F, which had a ΔcomAB genotype) was chosen at random for further testing from among several lines for which additional samples had been frozen every 4 days during serial transfer. For this analysis, line B—which had shown distinctly stronger adaptation (Figure S1B)—was excluded from consideration as a potential outlier. Assuming that only rarely would more than one fitness-modifying mutation become fixed during a short interval, these samples served to isolate largely the effects of individual mutations. For each isolate in this series, the change in fitness as compared to that of the preceding isolate was measured in terms of both CFU/colony values and maximal broth growth rates (Figure 3B). Between most pairs of consecutive isolates, no significant change in either value was observed. At four time points, however, significant increases in fitness for growth on agar plates were observed, and at one of these time points (day 96) a significant decline in the growth rate in broth culture was seen to accompany this change. However, not all mutations that were beneficial during growth on plates were deleterious in broth, and one mutational event was observed that appeared to be deleterious for growth in broth without conferring a benefit for growth on agar plates (day 32). Occasional fixation of such a strictly deleterious mutation is in keeping with the original experimental design promoting mutation accumulation. Nonetheless, the observation of inversely correlated changes in broth and agar fitness measurements at day 96 also suggests that a component of the decrease in maximum growth rates seen in broth culture resulted from pleiotropic effects of mutations that were beneficial for growth on solid media rather than from the accumulation of strictly deleterious mutations. Consequently, the decline in growth rates of these samples in broth cannot be used here to estimate the rate of deleterious mutations in S. pneumoniae because it is likely to be confounded by the effects of beneficial mutations that are selected during growth on agar but are pleiotropically deleterious in broth.
Previous studies of the acquisition and selection of beneficial mutations in bacteria have typically been conducted using larger populations and without explicit consideration of the role of events during stationary phase [22]–[24]. While the development of adaptive mutations during single, prolonged episodes of stationary phase survival has been an issue of controversy [25]–[27], the evolution of bacterial populations experiencing repeated exposure to brief stationary phase conditions has not to our knowledge been explored systematically. Under these conditions in which the maximum population size within a colony remains under 105 and daily single-colony bottlenecks limit the duration of selection, multiple beneficial mutations should be unlikely to arise and become amplified to a level sufficient to compete with each other during a single colony growth cycle. The adaptation observed in this system therefore offers an opportunity to measure beneficial mutations largely free from clonal interference between competing mutations that can bias estimates of the mutation rate downwards (reviewed in [13]).
In order to understand the conditions that led to increased fitness for stationary phase survival during our serial transfer experiment, we developed a maximum likelihood (ML) model describing the processes of mutation and stationary phase selection in populations subjected to the strong drift imposed by serial transfer of single colonies. This model was used to estimate the rate of beneficial mutations and the distribution of effect strengths of such mutations. Parameters were evaluated by comparing output of the simulation model with fitness values—as measured by growth on agar plates—from our experimental serial transfer lines after 126 days of passage. This time point was selected because average fitness had shown a generally consistent and increasing trend over these first 126 days. The model assumed a constant beneficial mutation rate and effect distribution over time without epistasis, and these assumptions appeared more likely to be satisfied during this initial phase of the experiment than later when the gain in fitness was less steady.
Details of the model are presented in Materials and Methods. Because average fitness increased during the experiment and only 1 line showed decreased fitness, our data did not permit independent estimates of the deleterious mutation rate Ud or the change in fitness sd associated with such mutations. As these values have not been determined empirically for S. pneumoniae, parameters for the model were chosen based on a mutation accumulation study conducted in E. coli that had reported Ud of 1.7×10−4 per genome per generation and sd of 0.012 [3]. Under these conditions and allowing for selection throughout stationary phase, the ML estimate for the beneficial mutation rate Ub was 4.8×10−4 per genome per time interval (with intervals corresponding to the 40 min generation time of the bacterium during exponential growth, Figure 4B). The corresponding ML estimate for the scale parameter sb describing an exponential distribution of fitness effects of these mutations was 0.025. (Note that sb also represents the mean effect of beneficial mutations.) The predicted fitness distribution for simulated bacterial lines after 126 growth cycles using these parameters matched the values measured in our serial transfer experiment well (Figure 4A). Most notably, this analysis indicated that the beneficial mutation rate in S. pneumoniae may be particularly high.
We then tested the impact of the specific deleterious mutation parameters used in the model by generating new estimates for Ub and sb under the extreme case in which deleterious mutations were not permitted. This change did not substantially affect the output of the model (ML estimates for Ub = 3.8×10−4 and for sb = 0.027, Figure 4B). We cannot exclude the possibility that deleterious mutations may on the other hand be more frequent or more severe in S. pneumoniae as compared to E. coli. In that case, the beneficial mutation rate calculated here would likely underestimate the actual rate. One variation that was explored in considering the possibility of a higher deleterious mutation rate was a provisional estimate based on Bateman-Mukai [28], [29] analysis of the reduction in broth growth rates seen with our experimental lines. As noted earlier, however, this estimate appears to be biased by the pleiotropic effects of mutations selected during stationary phase and is not suggested to represent an accurate measurement of the deleterious mutation rate in S. pneumoniae. Nonetheless, such analysis would have generated an estimate for the deleterious mutation rate Umin of 2.5×10−4 events per genome per 40 min time interval. This represents a point estimate based on the magnitude and variance of fitness changes seen among our lines between days 1 and 126. If deleterious mutations occurred only during the 16 generations per growth cycle minimally required to generate colonies having the population size observed after exponential phase growth, the Bateman-Mukai estimate for our experiment would increase to 5.7×10−4 events per genome per generation. We therefore simulated the serial transfer process using this higher value for Ud of 5.7×10−4. As anticipated, the increase in Ud raised the ML estimate for the beneficial mutation rate (Ub = 7.5×10−4 and for sb = 0.021, Figure S2A).
The effect of our assumption regarding the extent of the colony growth cycle during which stationary phase selection is effective was also explored by varying this structural parameter of the model. ML estimates were generated under an alternative model in which selection was permitted only during the last 4 hours (6 time intervals) of the growth cycle. This final period, when competition for depleted nutrients may be most severe, was the stage at which our experimental passaged isolates showed the most consistent advantage (Figure 2 and Figure S1C). Under these conditions of limited selection, ML values for Ub (6.9×10−4) and sb (0.029) were slightly higher than when selection was allowed throughout stationary phase (Figure 4C). We also explored the limiting condition where stationary phase selection did not occur and mutations only accumulated neutrally (Figure 4D). As expected, this condition generated higher ML values for Ub (7.2×10−4) and sb (0.039). Finally, we tested the impact of the assumption that the mutation rate remained constant during fixed time intervals of both exponential growth and stationary phase survival. When mutations were restricted to occur only during exponential growth, ML estimates for Ub and sb were 8.4×10−4 and 0.024, respectively (Figure S2B). Despite small differences in output values, these model variations all support the presence of a high rate of beneficial mutation in S. pneumoniae in the range of 10−4 and possibly as high as 10−3 per genome per generation time interval.
Considering the potential limitations of the above model in which events within a growth cycle are described deterministically, a wholly stochastic model was also developed describing the replication, mutation and selection of individual bacteria within a single colony at each serial transfer step. Simulation of the serial transfer process by this model was more than 100,000 times slower than with the semi-deterministic model and was therefore not suitable for the comprehensive exploration of parameter space required for maximum likelihood estimation. Nonetheless, the distribution of fitness values after the simulated serial transfer process generated by the entirely stochastic model was similar to that of our experimental data when mutation and selection parameters matching the ML estimates derived from the faster model were used (Figure S3). This similarity demonstrates that the simplifying assumptions required by the semi-deterministic model do not strongly distort its behavior.
Prior mutation accumulation studies in bacteria and eukaryotes have uniformly shown that average fitness of the propagated populations decreases [1]–[3], [5]–[9], [29] even though increased fitness has been observed for individual lines in some studies [4], [30]–[32]. In contrast, the average fitness of S. pneumoniae lines in our experiment rose despite a serial transfer protocol of daily single-colony population bottlenecks. Because adaptation among these lines was specific for survival in stationary-phase bacterial colonies, we infer that selection during this period of the growth cycle contributed to the probability of fixation of beneficial mutations. Considering that microbial populations increase to large sizes even when subjected to frequent bottlenecks, assays with these organisms may be particularly prone to unintended selection during brief and recurrent episodes of stationary phase survival. Mutation accumulation systems have been designed to minimize selection, and therefore evidence that selection influences the outcome of a classically-designed microbial mutation accumulation experiment raises questions about the assumptions underlying these experiments.
Detailed examination of a pneumococcal line that was sampled at high frequency during the serial transfer process demonstrated that on at least one occasion the fixation of a mutation adaptive for growth on agar was accompanied by a large decrease in the rate of growth in broth culture. This observation suggests that antagonistic pleiotropy may be responsible for a portion of the overall loss of fitness for growth in broth that was observed with the larger set of evolved strains as they adapted to stationary phase conditions. Such an inverse relationship between fitness for survival during stationary phase and maximal exponential growth rate has been reported previously in E. coli [21] and is consistent with the hypothesis of a trade-off between the optimization of self-preservation and nutritional competence (referred to as the SPANC balance) [33], [34].
Although we initially set out to assess the rate of deleterious mutations in S. pneumoniae, this antagonistic pleiotropy appears to confound attempts to estimate deleterious mutation parameters based on the declining growth rates observed in broth cultures. It is not known whether the stationary-phase adaptation that appears to be the source of this bias is unique to S. pneumoniae or might extend to other bacteria. This bias may be limited if selection during stationary phase is weak. It is noteworthy, however, that the mutation accumulation experiment of Kibota and Lynch in E. coli that reported the anticipated decline in fitness exclusively employed fitness assays measuring maximal rates of growth in broth cultures [3]. For comparison with this earlier study, we determined that application of a similar Bateman-Mukai analysis to our data would have generated an estimate for the deleterious mutation rate Umin of 2.5×10−4 events per genome per 40 min time interval. Although this is only a point estimate based on data for two time points, the result is similar to the value of 1.7×10−4 per genome per generation derived from the work in E. coli [3]. Due to the limitations noted above, however, the actual rate of deleterious mutations in S. pneumoniae remains to be determined.
Attempting to understand the process of selection in fluctuating populations, we developed a ML model of mutation and selection in the context of frequent single-colony bottlenecks. This model analyzed the fitness values measured for our experimental lines after 126 days of passage to estimate parameters for beneficial mutations. These fitness values, which were derived from assays of CFU/colony on agar plates, suggested that the lines had attained relatively modest increases in fitness and were consistent with weak selection in the system. (The possibility of stronger, frequency-dependent variation in the strength of selection is considered later.) Our model demonstrated that weak selection even in the setting of frequent single-colony bottlenecks would be sufficient to explain the adaptation that was observed if the beneficial mutation rate, Ub, were in the range of 4.8×10−4 per genome per 40 min time interval. Consistent with weak selection in the model, similar ML estimates of Ub were obtained even when selection was restricted or eliminated (Figure 4C and 4D). The model furthermore predicted that selection extending throughout stationary phase (as in Figure 4A, in which newly arising beneficial mutations have an average selection coefficient of 0.025) would increase the average selection coefficient only to 0.038 for beneficial mutations that become fixed. This estimate is in good agreement with the average selective advantage of 0.037 measured for the 4 beneficial mutational events (seen at days 8,16, 92 and 96 of serial transfer in Figure 3B) that were observed during our high-frequency analysis of one experimental line.
If selection in the serial transfer system were weak, a high beneficial mutation rate would therefore be required as the major force contributing to adaptation among these lines of S. pneumoniae. Although oxidative stress from strong endogenous production of peroxide is a major source of mutations in this organism [35], mutation rates in S. pneumoniae do not appear to be unusually high compared to other bacteria. The rate of point mutations conferring resistance to the antibiotic optochin, for instance, was estimated by fluctuation analysis to be 1.4×10−8 per cell division for the D39 strain used in our study [36]. It is possible, however, that bacteria experience more intense oxidative stress, and consequently a higher mutation rate, at a high density within stationary-phase colonies than during growth in broth. A higher mutation rate has also been suggested to occur in aging colonies of E. coli (over a period of 7 days) associated with the induction of an RpoS stress response [37], and a high deleterious mutation rate has been reported for E. coli during prolonged episodes of stationary phase (100 days) [38]. While both of these observations have been subject to controversy [26], [39], their relevance to the current study may be reduced by the consideration that S. pneumoniae lacks a homologue of the RpoS stationary-phase sigma factor.
Regardless of the overall mutation rate, our results indicate that beneficial mutations in particular may be relatively common for this organism when single colonies are propagated in vitro (as compared to deleterious mutations, which may yet be more frequent but are not amplified by selection). The balance of mutational effects between those that are beneficial and those that are deleterious may shift as the fitness of an organism changes [40], [41]. For phage in which fitness has been substantially degraded through previous mutations, additional random mutations, for instance, have been shown even to cause a net restoration of fitness [42]. In contrast to such phage, however, the D39 isolate of S. pneumoniae used for this experiment is a laboratory strain that has not been subjected to mutagenesis and is fit enough that it retains virulence in animal models. To the extent, however, that this organism is better adapted to exponential growth in broth culture than to survival within a stationary-phase colony, this consideration may have contributed to the net increase in fitness we observed for colony growth.
An alternative to the above scenario of weak selection is the possibility of strong frequency-dependent variation in the strength of selection within developing bacterial colonies. If beneficial mutations generate a GASP (growth advantage in stationary phase) phenotype [43] that allows mutants to continue growing after wild-type bacteria in a colony have stopped replicating, the initial gain in relative fitness associated with such mutations may potentially be large. A newly arisen mutant might increase from a single individual to a sizeable percentage of the total colony population within a single growth cycle. Yet once a GASP mutant becomes common in the population, its relative fitness advantage would decline because the higher carrying capacity associated with the mutation would be reached after fewer generations of bacterial growth. When fitness of an evolved strain is measured in isolation, such a GASP mutation might confer only a small fitness benefit—as measured by CFU/colony—reflecting a modest increase in the carrying capacity of the colony. If such frequency-dependent selection of GASP mutants was a source of strong selection in our serial transfer lines, the rate of beneficial mutations may be lower than we have estimated. Because we are unable to initiate individual pneumococcal colonies from a mixed inoculum for competitive fitness assays, we have not been able to evaluate this possibility directly.
The potential effects of recurrent exposure to stationary phase and frequency-dependent variation in the strength of selection on other microbial systems of experimental evolution may also need to be considered. Although different in many regards from the current study, the work of Perfeito et al. [14], for instance, showing a high rate of beneficial mutations in E. coli also utilized populations that appear to have experienced daily periods of stationary phase survival between controlled population bottlenecks. If GASP mutants were to arise and reach measurable frequencies (>1%) during a single growth cycle, the probability of adaptive mutations escaping stochastic loss may have been substantially higher than was estimated from fitness values based on changes in microsatellite allele frequencies that were observed once mutants had already risen to measurable levels. Furthermore, for the initial stages of selection when a GASP mutant is at low frequency and may be strongly selected within a single growth cycle, Ne may approach the maximum size of the stationary phase population. For events within a single growth cycle, therefore, Ne may be similar even for populations subjected to very different bottlenecks. The bottleneck size, however, may still affect the outcome of stationary phase selection because a smaller bottleneck would also correspond to a greater dilution factor and a shorter duration of stationary phase accompanying each growth cycle. In contrast to the rapid initial enrichment of a GASP mutant, the final stages of selection—when a mutant is at higher frequency and less strongly selected—should extend over multiple growth cycles and thereby operate under a lower Ne as driven by the recurrent bottlenecks. These considerations may impact the apparent rate of beneficial mutation in studies of adaptive evolution.
Our data imply that either the rate of beneficial mutations in S. pneumoniae is quite high (approximately 4.8×10−4 per genome per 40 min time interval) or that strong selection is occurring in this system despite a classical mutation accumulation experimental design. The apparent rate of beneficial mutations calculated under our model incorporating weak selection is several orders of magnitude higher than early estimates for bacteria and viruses that clustered near rates of 10−8 to 10−9 [11], [12], [44]. This estimate, however, more closely agrees with the recent work of Perfeito et al. [14] that obtained a beneficial mutation rate for E. coli of 2×10−5 and for which Monte Carlo simulations of the same data indicated that the actual rate could be as high as 10−4. Our findings may therefore provide additional support for the presence of high rates of beneficial mutations among bacteria. The frequency with which such beneficial mutations occur is likely to contribute to the ability of these bacteria to evolve rapidly as conditions change. In the case of S. pneumoniae, the spread of these new beneficial mutations may be facilitated by the natural competence of the organism for transformation and may be an important factor in the virulence of this pathogen and in its development of resistance to antibiotics.
Experiments were performed using the encapsulated serotype 2 strain D39 [19] of S. pneumoniae. A competence-deficient and kanamycin-resistant mutant, TMP1, was constructed by allelic replacement mutagenesis inserting the aph3 gene into the comAB locus. DNA from a region on the 5′ side of comAB was amplified by PCR using the primers 5′-TTTTGTTTAGTGATTGGGGTAAG-3′ and 5′-ACGAGGATCCGAGAGCAGACCATTTTTTTGTTC-3′, and DNA from a 3′ region was amplified with primers 5′-AGCAGGGCCCAATACCAAGAAGGGGCAGAGGG-3′ and 5′-TAGCGAACAGAATCACCGAC-3′. A PCR product encoding aph3 was amplified from CP1296 DNA [45] using primers 5′-GATCGGATCCGTTTGATTTTTAATGGATAATG-3′ and 5′- ACCTGGGCCCCTCTCCTGTGTTTTTTTATTTTTGG-3′. These products were linked using PCR ligation mutagenesis [46] and introduced into D39 by transformation as previously described [47]. The resulting mutation in comAB was confirmed by PCR and sequencing.
Serial transfer lines of S. pneumoniae were derived from one colony each of strains D39 and TMP1. These colonies were streaked onto fresh THY agar plates (Todd-Hewitt agar supplemented with 0.2% yeast extract), and 20 of the resulting colonies from each (designated growth cycle 0) were picked to initiate the experimental lines. A single colony from each line was chosen daily at random for transfer to a fresh THY plate by selecting the most well-separated colony present on an initial field of view under a stereomicroscope. The interval between inoculation and the next transfer step was maintained between 22 to 26 h and was generally close to 24 h. Between transfer steps, plates were kept at 37°C in an incubator containing ambient air. Samples were preserved at fixed intervals by suspending patches of colonies from the plates in a 4∶1 mixture of THY broth:80% glycerol and stored at −75°C for later testing. During serial transfer, THY plates were divided into 6 radial sectors, which were used for propagation of separate bacterial lines in an alternating pattern such that D39 (kanamycin-sensitive) and TMP1 (kanamycin-resistant) lines were present on adjacent sectors. To check for contamination between lines, we verified that each line had maintained its initial pattern of kanamycin susceptibility or resistance at the end of the experiment.
The growth of isolates on THY agar plates was assessed by dissecting individual colonies under a stereomicroscope after intervals ranging from 7 to 26 h following inoculation. Because we were not able to initiate a single bacterial colony from a controlled mixture of experimental and reference strains, growth was measured by examining colonies composed of a single strain type rather than through direct competition. Six sectors per plate were inoculated with isolates in a randomized order in which passaged and control samples were intermixed. Both the passaged and control isolates had been frozen prior to testing and were grown for 24 h on an initial agar plate before inoculating the colonies to be dissected. The most well-separated colony on a sector was dissected and the number of CFU within the colony was determined by dilution and colony counting. The log2(CFU/colony) was used as a measure of the net growth rate of a colony. The ratio of this value for a given isolate to the average value for control isolates from growth cycle 1 (processed in parallel for each assay) was taken as a proxy for fitness. Where a fitness value is given to represent the average of multiple passaged lines, the fitness of each individual line was first estimated by replicate measurements. The average fitness across the lines was then determined as the mean of these estimates for the independent lines, and standard deviation was calculated using a sample size equal to the number of lines.
Growth in liquid culture was measured by suspending individual bacterial colonies in 1 ml aliquots of THY broth. Aliquots were placed in a randomized order into wells of a 48-well plate, which was sealed with optically-transparent film to limit evaporation and incubated at 37°C in a Synergy2 plate reader (BioTek; Winooski, VT). Optical density (OD) readings at 620 nm were taken every 10 min, and the maximum growth rate was estimated by determining the maximum change in log10(OD) over any 30 min window. Fitness for growth in this context was calculated as the ratio of the maximum growth rate for a given isolate to the average maximum growth rate for control isolates from growth cycle 1 that were processed in parallel for each assay. For the time series analysis of samples archived at 4-day intervals (Figure 3B), each isolate was compared directly to samples from the preceding time point that were assayed in parallel in order to estimate the change in fitness over the interval (for both broth and agar plate assays) rather than comparing each to the growth cycle 1 reference.
A computational model was developed to simulate changes in fitness due to mutation and selection, as well as the drift imposed by single-colony bottlenecks, over repeated growth cycles of the serial transfer process. A description of the model is here followed by a more detailed consideration of the assumptions made in the design of the model. The fitness of the population was modeled as consisting of a discrete distribution of fitness classes. Each simulation was initiated with a uniform population having fitness of 1. The model divided each growth cycle into discrete time intervals corresponding to the generation time of S. pneumoniae during exponential growth (approximately 40 min in THY medium either on plates or in broth). For simplicity, the first half of the growth cycle was considered to constitute the period of exponential growth while the second half was considered stationary phase.
Beneficial and deleterious mutations were allowed to occur at each time step at rates Ub and Ud, respectively. The fitness effects of beneficial mutations were assumed to be exponentially distributed with a scale parameter sb. Deleterious mutations, however, were considered to have a constant fitness effect sd. Fitness was assumed to be multiplicative without epistatic effects. For simplicity, the model did not consider the effects of selection during the exponential phase of growth because our experimental data had indicated that isolates resulting from serial transfer did not have an advantage during this period. Furthermore, the smaller populations present during much of exponential growth would have limited the efficacy of selection during this phase. Mutations therefore were allowed to accumulate neutrally during exponential phase.
Selection during stationary phase operated at the end of each specified time interval and changed the frequency, , of the ith fitness class at time t in proportion to its fitness, wi, relative to the average fitness of the overall population, , such that:To limit complexity, the variable fluctuations in population size seen in our experimental samples during stationary phase survival were not incorporated into the model.
At the end of each simulated growth cycle, the change in fitness of the bacterium chosen to continue the serial transfer process was determined by a random draw from the fitness distribution after the last time interval. The change in fitness of a bacterial line over the serial transfer experiment was simulated by repeated iterations of the growth cycle model. Because fitness effects were assumed not to show epistasis, the same distribution of fitness changes was used at the end of each growth cycle. Simulation of serial transfer for 105 independent lines was used to generate a distribution of expected fitness values for passaged isolates. This distribution was modified by application of a Gaussian smoothing function with a variance matching the average within-line variance of our experimental data to account for the effects of environmental noise in laboratory measurements of fitness. The log likelihood of the experimental data under the parameters Ub and sb of a given simulation was calculated as:where F(126,wj) is the simulated frequency after 126 growth cycles of the fitness class wj matching the average measured fitness of the jth experimental line of S. pneumoniae. Maximum likelihood estimates were then generated for Ub and sb by exploring the parameter space between Ub of 1×10−6 and 8×10−3 and between sb of 0.005 and 0.08. Boundaries of 95% confidence regions were calculated as contours for which the likelihood estimatewhere UbMLE and sbMLE represent the ML estimates of Ub and sb, respectively [48].
Although controversy exists regarding mutation rates during stationary phase and the extent to which mutations depend on DNA replication, we assumed initially that mutation rates remained constant as a function of time throughout the growth cycle. The impact of this assumption was tested through a variation in model parameters that eliminated mutations during stationary phase. The potential for selection to amplify preferentially beneficial mutations that have larger fitness effects than average made it important to consider the distribution of effect strengths for these mutations. The actual distribution of selection coefficients for newly arising mutations is uncertain, but the exponential distribution employed here has been proposed on theoretical grounds [48], [49], and the distributions of new beneficial mutations escaping stochastic loss [14] or reaching fixation [48] in E. coli show a good match to predictions based on this theory. Recent results also support an exponential distribution for newly arising beneficial mutations in Pseudomonas fluorescens [50]. Because the frequency of deleterious mutations causing larger fitness effects than average would be reduced rather than amplified by selection, however, deleterious mutations were considered to have a constant fitness effect sd.
Within each growth cycle, the evolution of the bacterial population was modeled deterministically for the reasons outlined below. Because there are approximately 3×104 CFU in each mature colony and there are generally thousands of colonies on a plate at each growth cycle from which a founder may be selected for the next stage, the population from which this founder is chosen is large (at least 107). The overall serial transfer experiment, however, is still characterized by a low effective population size (Ne = 18 for the model population) due to the single colony bottlenecks.
In order to estimate the fitness distribution at the end of each growth cycle, rather than modeling the thousands of colonies present during each growth cycle individually we considered these bacteria to represent a single, large population. This assumption reflected the experimental design in which a serial transfer line is propagated through selection of a single CFU from within the collection of colonies present at each growth cycle in a two-step, random process. One colony among many is first randomly chosen for transfer to the next agar plate and spread so as to isolate individual colony-forming units. The random selection of which of these daughter colonies will serve to continue the line is delayed until the end of the next growth cycle, but the evolutionary trajectories of these daughter colonies become independent once they are separated. The model collapses this process into a single random selection event at the end of each growth cycle in which the probability of founding the next population is given by the frequency of a variant within the entire population but is not affected by the distribution of variants among colonies. This strategy reduced the computational complexity of the model by requiring calculation of the fitness distributions only across a single, large, deterministic population rather than in many small populations with stochastic behavior but effectively assumed that bacteria were able to compete equally well across the plate as within a colony. While not strictly realistic, the limited period of stationary-phase selection available before the next growth cycle prevented even highly beneficial mutations from amplifying to a level that would correspond to fixation within a single colony, much less the entire population. (For instance, after 12 hours [i.e., 18 rounds] of stationary-phase selection a new mutation with fitness of 1.5 might be amplified roughly 1500-fold, which would represent less than 5% of a colony having a total size of 3×104 CFU.) The assumption of a well-mixed population did not therefore permit amplification by selection to continue beyond limits compatible with the physical constraint of each variant within a colony.
To simplify the model during the period of neutral mutation accumulation, the frequency of new mutations at the start of stationary-phase selection was given by , where g1 represents the number of time intervals before the onset of selection. Mutations were modeled as if they all occurred in separate individuals (i.e., a Poisson distribution of mutations in the population was not assumed). This simplification was reasonable because the contribution of the exponential distribution of mutational effects toward generating individuals in higher fitness classes is much stronger than the contribution of rare individuals with multiple beneficial mutations. Calculation of the distribution of mutations at the start of stationary phase did not require incorporation of Luria-Delbrück distributions [51] because mutational jackpots within colonies are homogenized among colonies. Because the relative enrichment of mutations within jackpot colonies would have reduced the difference between the fitness of a new variant and that of the colony as a whole, the assumption of a well-mixed population causes the strength of selection to be overestimated in our model. As the frequency of each new variant remains low during a growth cycle (see above), this effect however should be small. Nonetheless, the rate of beneficial mutations inferred using the model may be considered to be a minimum estimate in this regard.
An alternative and entirely stochastic model was constructed to simulate mutation, selection and replication of individual bacteria within a single propagated colony throughout the serial transfer process. The model was initiated with a single bacterium with fitness of 1. The colony growth cycle was divided into 36 time intervals of equal length, corresponding to the 40 min pneumococcal generation time during exponential growth. At each time interval, individuals acquired beneficial and deleterious mutations with probabilities Ub and Ud, respectively. Selection coefficients for beneficial mutations were chosen at random from an exponential distribution with scale parameter sb. Deleterious mutations were assigned a fixed selection coefficient of sd. New mutations had a multiplicative effect on fitness without epistasis. During the first half of the colony growth cycle, each individual replicated at every time interval (i.e., no selection during exponential phase), contributing 2 progeny of identical fitness to the next generation. During the second half of the growth cycle (i.e., stationary phase), acquisition of new mutations continued in the same manner with each time interval. Selection was introduced by stipulating that individuals with fitness less than the population average failed to contribute to the subsequent population with probabilitywhere wi is the fitness of the ith individual and is the mean population fitness. Likewise, individuals with fitness greater than the population average succeeded in replicating during stationary phase (i.e., contributing 2 rather than 1 individuals to the subsequent population) with a probabilityTo assess the sensitivity of the model to the choice of fitness function, a variation of the model was also analyzed in which the probability of an individual with below average fitness failing to contribute to the subsequent population was given byand the probability of an individual with above average fitness reproducing in stationary phase was given byAt the end of the growth cycle, a single individual was chosen at random to initiate the next colony of the serial transfer. After iterative modeling of 126 growth cycles, the final fitness for the simulated line was determined by choosing an individual at random at the end of the final growth cycle. Both the stochastic and semi-deterministic models were implemented in Python version 2.6.4 and are available as Text S1.
|
10.1371/journal.pbio.2002458 | Identifying genetic variants that affect viability in large cohorts | A number of open questions in human evolutionary genetics would become tractable if we were able to directly measure evolutionary fitness. As a step towards this goal, we developed a method to examine whether individual genetic variants, or sets of genetic variants, currently influence viability. The approach consists in testing whether the frequency of an allele varies across ages, accounting for variation in ancestry. We applied it to the Genetic Epidemiology Research on Adult Health and Aging (GERA) cohort and to the parents of participants in the UK Biobank. Across the genome, we found only a few common variants with large effects on age-specific mortality: tagging the APOE ε4 allele and near CHRNA3. These results suggest that when large, even late-onset effects are kept at low frequency by purifying selection. Testing viability effects of sets of genetic variants that jointly influence 1 of 42 traits, we detected a number of strong signals. In participants of the UK Biobank of British ancestry, we found that variants that delay puberty timing are associated with a longer parental life span (P~6.2 × 10−6 for fathers and P~2.0 × 10−3 for mothers), consistent with epidemiological studies. Similarly, variants associated with later age at first birth are associated with a longer maternal life span (P~1.4 × 10−3). Signals are also observed for variants influencing cholesterol levels, risk of coronary artery disease (CAD), body mass index, as well as risk of asthma. These signals exhibit consistent effects in the GERA cohort and among participants of the UK Biobank of non-British ancestry. We also found marked differences between males and females, most notably at the CHRNA3 locus, and variants associated with risk of CAD and cholesterol levels. Beyond our findings, the analysis serves as a proof of principle for how upcoming biomedical data sets can be used to learn about selection effects in contemporary humans.
| Our global understanding of adaptation in humans is limited to indirect statistical inferences from patterns of genetic variation, which are sensitive to past selection pressures. We introduced a method that allowed us to directly observe ongoing selection in humans by identifying genetic variants that affect survival to a given age (i.e., viability selection). We applied our approach to the GERA cohort and parents of the UK Biobank participants. We found viability effects of variants near the APOE and CHRNA3 genes, which are associated with the risk of Alzheimer disease and smoking behavior, respectively. We also tested for the joint effect of sets of genetic variants that influence quantitative traits. We uncovered an association between longer life span and genetic variants that delay puberty timing and age at first birth. We also detected detrimental effects of higher genetically predicted cholesterol levels, body mass index, risk of coronary artery disease (CAD), and risk of asthma on survival. Some of the observed effects differ between males and females, most notably those at the CHRNA3 gene and variants associated with risk of CAD and cholesterol levels. Beyond this application, our analysis shows how large biomedical data sets can be used to study natural selection in humans.
| A number of central questions in evolutionary genetics remain open, in particular for humans. Which types of variants affect fitness? Which components of fitness do they affect? What is the relative importance of directional and balancing selection in shaping genetic variation? Part of the difficulty is that our understanding of selection pressures acting on the human genome is based either on experiments in fairly distantly related species or cell lines or on indirect statistical inferences from patterns of genetic variation [1–3].
The statistical inferences rely on patterns of genetic variation in present-day samples (or, very recently, in ancient samples [4]) to identify regions of the genome that appear to carry the footprint of positive selection [2]. For example, a commonly used class of methods asks whether rates of nonsynonymous substitutions between humans and other species are higher than expected from putatively neutral sites in order to detect recurrent changes to the same protein [5]. Another class instead relies on polymorphism data and looks for various footprints of adaptation involving single changes of large effect [6]. These approaches detect adaptation over different timescales and, likely as a result, suggest quite distinct pictures of human adaptation [1]. For example, approaches that are sensitive to selective pressures acting over millions of years have identified individual chemosensory and immune-related genes (e.g., [7]). In contrast, approaches that are most sensitive to selective pressures active over thousands or tens of thousands of years have revealed strong selective pressures on individual genes that influence human pigmentation (e.g., [8–10]), diet [11–13], as well as sets of variants that shape height [14–16]. Even more recent still, studies of contemporary populations have suggested that natural selection has influenced life-history traits like age at first childbirth as well as educational attainment over the course of the last century [17–23].
Because these approaches are designed (either explicitly or implicitly) to be sensitive to a particular mode of adaptation, they provide a partial and potentially biased picture of what variants in the genome are under selection. In particular, most have much higher power to adaptations that involve strongly beneficial alleles that were rare in the population when first favored and will tend to miss selection on standing variation or adaptation involving many loci with small beneficial effects (e.g., [24–27]). Moreover, even where these methods identify a beneficial allele, they are not informative about the components of fitness that are affected or about possible fitness trade-offs between sexes or across ages.
In line with Lewontin’s proposal to track age-specific mortality and fertility of hundreds of thousands of individuals [28], we sought a more direct and, in principle, comprehensive way to study adaptation in humans, focusing on current viability selection. Similar to the approach that Allison took in comparing frequencies of the sickle cell allele in newborns and adults living in malarial environments [29], we aimed to directly observe the effects of genotypes on survival by taking advantage of the recent availability of genotypes from large cohorts of individuals of different ages. Specifically, we tested for differences in the frequency of an allele across individuals of different ages, controlling for changes in ancestry and possible batch effects. This approach resembles a genome-wide association study (GWAS) for longevity yet does not focus on an end point (e.g., survival to an old age) but on any shift in allele frequencies with age. Thus, it allows the identification of possible nonmonotonic effects at different ages or sex differences. Any genetic variant that affects survival by definition has a fitness cost, even if the cost is too small to be effectively selected against (depending on the effective population size, the age structure of the population, and the age at which the variant exerts its effects [30]). Of course, a genetic variant can influence fitness without influencing survival through effects on reproduction or inclusive fitness. Our approach is therefore considering only 1 of the components of fitness that are likely important for human adaptation.
As a proof of principle, we applied our approach to 2 recent data sets: to 57,696 individuals of European ancestry from the Genetic Epidemiology Research on Adult Health and Aging (GERA) cohort [31,32] and, by proxy [33–35], to the parents of 117,648 individuals of British ancestry surveyed as part of the UK Biobank [36]. We did so for individual genetic variants then jointly for sets of variants previously found to influence 1 of 42 polygenic traits [37–40].
If a genetic variant does not influence viability, its frequency should be the same in individuals of all ages. We therefore tested for changes in allele frequency across individuals of different ages, while accounting for systematic differences in the ancestry of individuals of different ages (for example, due to migration patterns over decades) and genotyping batch effects. We used a logistic regression model in which we regressed each individual’s genotype on their age bin, their ancestry as determined by principal component analysis (PCA) (S1 Fig), and the batch in which they were genotyped (see Materials and methods for details). In this model, we treated age bin as a categorical variable; this approach allowed us to test for a relationship between age and the frequency of an allele regardless of the functional form of this relationship. We also tested a model with an interaction between age and sex to assess whether a variant affects survival differently in the 2 sexes.
We first evaluated the power of this method using simulations. We considered 3 possible trends in allele frequency with age: (i) a constant frequency up to a given age followed by a steady decrease, i.e., a variant that affects survival after a given age (e.g., variants contributing to late-onset disorders), (ii) a steady decrease across all ages for a variant with detrimental effect throughout life, and (iii) a U-shaped pattern in which the allele frequency decreases to a given age but then increases, reflecting trade-offs in the effects at young and old ages, as hypothesized by the antagonistic pleiotropy theory of aging [41] or as may be seen if there are protective alleles that buffer the effect of risk alleles late in life [42] (Fig 1). In all simulations, we used sample sizes and age distributions that matched the GERA cohort (S2 Fig). For simplicity, we also assumed no population structure or batch effects across age bins (see Materials and methods). For all trends, we set a maximum of 20% change in the allele frequency from the value in the first age bin (Fig 1).
Because of the age distribution of individuals in the GERA cohort (S2 Fig), our power to detect the trend is greater when most of the change in allele frequency occurs in middle age (Fig 1). For example, for an allele with an initial allele frequency of 0.15 that begins to decrease in frequency among individuals at age 20, age 50, or age 70 years, there is around 20%, 90%, and 60% power, respectively, to detect the trend at P < 5 × 10−8, the commonly used criterion for genome-wide significance [44]. We also experimented with a version of the model in which the age bin is treated as an ordinal variable; as expected, this model is more powerful if there is in fact a linear relationship between age and allele frequency. Because we do not know the functional form of the relationship between age and allele frequency a priori in most cases, we used the categorical model for all analyses unless otherwise noted.
In the UK Biobank, all individuals were 45–69 years old at enrollment, so the age range of the participants is restricted and our method has low power. However, the UK Biobank participants reported the survival status of their parents: age of the parents if alive or age at which their parents died; following recent studies [33–35], we therefore used these values (when reported) instead in our model. In this situation, we are testing for correlations between an allele frequency and father’s or mother’s age (if alive) or age at death (if deceased). This approach obviously comes with the caveat that children inherit only half of their genome from each parent and so power is reduced (e.g., [45]). Furthermore, the patterns expected when considering individuals who have died differ subtly from those generated among surviving individuals. Notably, when an allele begins to decline in frequency starting at a given age (Fig 1A), there should be an increase in the allele frequency among individuals who died at that age followed by a decline in frequency, rather than the steady decrease expected among surviving individuals (S3 Fig, see Materials and methods for details). In a first analysis, we therefore focused on the majority of participants who reported father’s or mother’s age at death, 88,595 and 71,783 individuals, respectively. We compared the results of this approach with the results of a Cox proportional hazards model [46], which allowed us to include individuals who reported their parents to be alive but has the disadvantage of assuming fixed effects across all ages.
We further adapted this model to allow us to test for changes in frequency at sets of genetic variants jointly. Many phenotypes of interest, from complex disease risk to anthropomorphic and life-history traits such as age at menarche, are polygenic [47,48]. If a polygenic trait has an effect on fitness, either directly or indirectly (i.e., through pleiotropic effects), the individual loci that influence the trait may be too subtle in their survival effects to be detectable with current sample sizes. We therefore investigated whether there is a shift across ages in sets of genetic variants that were identified as influencing a trait in GWASs (S1 Table). Specifically, for a given trait, we calculated a polygenic score for each individual based on trait effect sizes of single variants previously estimated in GWASs and then tested whether the scores vary significantly across 5-year age bins (see Materials and methods for details). These scores are calculated under an additive model, which appears to provide a good fit to GWAS data [49].
If a polygenic trait is under stabilizing selection (e.g., human birth weight [50]), i.e., an intermediate polygenic score is optimal, no change in the mean value of polygenic scores across different ages is expected. However, if extreme values of a trait are associated with lower chance of survival, the spread of the polygenic scores should decrease with age. To consider this possibility, we tested whether the squared difference of the polygenic scores from the mean varies significantly with age (see Materials and methods for details).
We first applied the method to the GERA cohort using 8,868,517 filtered genotyped and imputed autosomal biallelic single-nucleotide polymorphisms (SNPs) and indels. We focused on a subset of 57,696 filtered individuals who we confirmed to be of European ancestry by PCA (see Materials and methods, S4 and S5 Figs). The ages of these individuals were reported in bins of 5-year intervals (distribution shown in S2 Fig). We tested for significant changes in allele frequencies across these bins. For each variant, we obtained a P value comparing a model in which the allele frequency changes with age to a null model. No inflation was detected in the quantile-quantile plot (S6A Fig), indicating that, for common variants at least, our control for population structure (and other potential confounders) is sufficient. To illustrate this point, we looked at the lactose intolerance-linked SNP rs4988235 within the LCT locus, which is among the most differentiated variants across European populations [11]; the trend in the expected allele frequency based on the null model (i.e., accounting for confounding batch effects and changes in ancestry) tracks the observed trend quite well (S7 Fig).
By our approach, all variants that reached genome-wide significance (P < 5 × 10−8) reside on chromosome 19 near the APOE gene (Fig 2A and S8 Fig). This locus has previously been associated with longevity in multiple studies [51,52]. The APOE ε4 allele is known to increase the risk of late-onset Alzheimer disease (AD) as well as of cardiovascular diseases [53,54]. We observe a monotonic decrease in the frequency of the T allele of the ε4 tag SNP rs6857 (C, protective allele; T, risk allele) beyond the age of 70 years old (Fig 2B). This trend is observed for both the heterozygous and homozygous risk variants (S9 Fig) and for both males and females (S10 Fig). No variant reaches genome-wide significance testing for age by sex interactions (quantile-quantile plot shown in S6B Fig).
We further investigated the trends in frequency with age for the other 2 major APOE alleles defined by rs7412 and rs429358 SNPs: ε2 (rs7412-T, rs429358-T) and ε3 (rs7412-C, rs429358-T), while ε4 is known by rs7412-C and rs429358-C alleles. Unlike the ε4 allele, ε2 allele carriers are suggested to be at lower risk of AD, cardiovascular disease, and mortality relative to the ε3 carriers [51,55]. We focused on a subset of 38,703 individuals with unambiguous counts of each APOE allele. There is a significant change in the frequency of the ε4 allele with age in this subset (P~6.0 × 10−12), similar to the trend observed for the tag SNP rs6857 (S11 Fig). The ε3 allele shows the reverse trend, with a significant, monotonic increase in frequency beyond the age of 70 years old (P~1.7 × 10−8) (S11 Fig). The enrichment of the ε3 allele in elderly individuals can be explained by the corresponding depletion of the ε4 allele and does not necessarily imply an independent, protective effect of the ε3 allele. The frequency of the ε2 allele does not change significantly with age (P~0.21), possibly reflecting low power given its allele frequency of approximately 0.06 (S11 Fig).
We considered the possibility that some unobserved confounding variable was driving the strength of this signal at APOE. Since there are 2 genotyped SNPs with signals similar to rs6857 within the locus, genotyping error seems unlikely to be driving the pattern (S8 Fig). Another concern might be a form of ascertainment bias, in which individuals with AD are underrepresented in the Kaiser Permanente Medical Care Plan. However, there is no correlation in these data between the amount of time that an individual has been enrolled in this plan and the individual’s APOE genotype (S12 Fig). These observations, along with previously reported associations at this locus, argue that the allele frequency trends in Fig 2B are driven by effects of APOE genotype on mortality (or severe disability). Moreover, the effects that we identified are concordant with epidemiological data on the mean age at onset of AD, given 0 to 2 copies of APOE ε4 allele [53]. This case not only serves as a positive control for our approach, it illustrates the resolution that it provides about age effects of genetic variants.
We estimated that we have about 93% power to detect the trend in allele frequency with age as observed for rs6857 (at a genome-wide significance level, see Materials and methods). Using both versions of the model treating age bin as a categorical or an ordinal variable, we have similar power to detect other potential trends considered in Fig 1 for variants as common as rs6857 and with similar magnitude of effect on survival. Yet across the genome, only APOE variants show a significant change in allele frequency with age for both versions of the model (Fig 2 and S13 Fig). Thus, our finding only APOE ε4 allele indicates that there are few or no other common variants in the genome with an effect on survival as strong as is seen in the APOE region.
We then turned to the UK Biobank data set. We applied our method to individuals of British ancestry whose data passed our filters; of these, 88,595 had death information available for their father and 71,783 for their mother. We analyzed 590,437 genotyped autosomal variants, applying similar quality control (QC) measures as with the GERA data set (see Materials and methods). We tested for significant changes in allele frequencies with father’s age at death and mother’s age at death stratified in eight 5-year interval bins. As in the GERA data set, no inflation was detected in the quantile-quantile plots (S14 Fig).
Consistent with recent studies [33,34], the variants showing a genome-wide significant change in allele frequency with father’s age at death (P < 5 × 10−8) reside within a locus containing the nicotine receptor gene CHRNA3 (Fig 3A). The A allele of the CHRNA3 SNP rs1051730 (G, major allele; A, minor allele) has been shown to be associated with increased smoking quantity among individuals who smoke [56]. We observe a linear decrease in the frequency of the A allele of rs1051730 throughout almost all age ranges (Fig 3B) (P~1.3 × 10−7 and P~2.7 × 10−10, treating paternal age at death as a categorical or an ordinal variable, respectively). Although it does not reach genome-wide significance, this allele shows a similar trend with age in GERA (P~8.6 × 10−3, S15 Fig). We note that 30,819 of the UK Biobank individuals included in the above analysis were genotyped on the UK BiLEVE Axiom array (see Materials and methods), selected based on lung function and smoking behavior (while the remaining 57,776 samples were genotyped on the UK Biobank Axiom array) [57]. Expectedly, the frequency of the A allele is significantly higher among UK BiLEVE subjects (P~2.3 × 10−10), but the age effects are similar across both arrays (P~0.72, see Materials and methods).
For mother’s age at death, a SNP in a locus containing the MEOX2 gene reached genome-wide significance (Fig 3C). The C allele of rs4721453 (T, major allele; C, minor allele) increases in frequency in the age bin centered at 76 years old (S16 Fig), i.e., there is an enrichment among individuals who died at 74 to 78 years of age, which corresponds to a deleterious effect of the C allele in this period. The trend is similar and nominally significant for other genotyped common SNPs in moderate linkage disequilibrium with rs4721453 (S16 Fig). Also, the signal for rs4721453 remains nominally significant when using subsets of individuals genotyped on the same genotyping array: 44,552 individuals on the UK Biobank Axiom array (P~6.6 × 10−5) and 25,231 individuals on the UK BiLEVE Axiom array (P~1.1 × 10−4). These observations suggest that the result is not due to genotyping errors, but it is not reproduced in GERA (P~0.023, S17 Fig) and so it remains to be replicated. Variants within the APOE locus are among the top nominally significant variants (Fig 3C). At the APOE SNP rs769449 (G, major allele; A, minor allele), there is an increase in the frequency of A allele at around 70 years old before subsequent decrease (Fig 3D, P~1.2 × 10−7). This pattern is consistent with our finding in GERA (of a monotonic decrease beyond 70 years of age), considering the difference in patterns expected between allele frequency trends with age among survivors versus individuals who died (S3 Fig).
We note that by considering parental age at death of the UK Biobank participants—as done also in [33–35]—we introduce a bias towards older participants, who are more likely to have deceased parents (S18 Fig). We confirmed that our top signals are not significantly affected after adjusting for age of the participants (among other potential confounders, including participants’ sex, year of birth, and socioeconomic status, as measured by the Townsend deprivation index): results remain similar for the SNP rs4721453 near MEOX2 (P~2.1 × 10−9), APOE SNP rs769449 (P~1.5 × 10−6), and CHRNA3 SNP rs1051730 (P~1.8 × 10−6 and P~4.3 × 10−9, treating paternal age at death as a categorical or an ordinal variable, respectively).
We further tested for trends in allele frequency with parental age at death that differ between fathers and mothers, focusing on 62,719 individuals with age at death information for both parents. No variant reached genome-wide significance level (S19A Fig). The rs4721453 near the MEOX2 gene and APOE variant rs769449 show nominally significant sex effects (P~7.2 × 10−8 and P~2.2 × 10−3, respectively), with stronger effects in females (S19B Fig). Variants near the CHRNA3 locus are nominally significant when using the model with parental ages at death treated as ordinal variables (rs11858836, P~5.7 × 10−4), with stronger effects in males (S19B Fig).
We next turned to sets of genetic variants that have been associated with polygenic traits rather than individual genetic variants. We focused on 42 polygenic traits, including disease risk and traits of evolutionary importance such as puberty timing, for which a large number of common variants have been mapped in GWASs (see S1 Table for the list of traits and number of loci) [37–40]. For each individual and each trait, we calculated a polygenic score based on the genetic variants that reached genome-wide significance level for association and then tested whether this polygenic score, or its squared difference from the mean in the case of stabilizing selection, is associated with survival (after controlling for covariates, see Materials and methods).
We first applied the Cox proportional hazards model in the UK Biobank for parental survival, focusing on the participants whose genetic ancestry is British and who reported their father’s or mother’s age or age at death (114,122 and 116,323 individuals, respectively). We then compared the results with our approach of testing for changes in the polygenic score across parental ages at death. We further analyzed 2 data sets for replication purposes: participants of the UK Biobank of non-British ancestry (29,511 and 30,372 individuals reporting father’s or mother’s age information, respectively) and the GERA cohort.
Using the Cox model, the scores for several traits show significant associations with father’s survival after accounting for multiple testing (Fig 4A, Table 1): total cholesterol (TC, P~4.3 × 10−11), low-density lipoproteins (LDL, P~8.1 × 10−9), body mass index (BMI, P~1.8 × 10−8), and coronary artery disease (CAD, P~9.0 × 10−6), consistent with 2 recent studies [34,35]. In addition, we uncovered significant association for the polygenic score for puberty timing (P~6.2 × 10−6); in this analysis, we used age at menarche-associated variants in females, motivated by the high genetic correlation between the timing of puberty in males and females [58]). A higher score for puberty timing is associated with longer paternal survival (per year hazard ratio of 0.96) (Table 1), indicating that variants delaying puberty timing are associated with a higher chance of survival, consistent with epidemiological studies suggesting early puberty timing to be associated with adverse health outcomes [59]. For all other traits, a higher score is negatively associated with paternal survival: 1 unit polygenic score hazard ratio of 1.09 for TC, 1.08 for LDL, 1.08 for CAD, and 1.22 for BMI (Table 1). With the exception of lipid traits, the effects on survival are not significantly changed after accounting for the effect of the polygenic score of another trait (S20 Fig). This is especially relevant to BMI and puberty timing, for which there is substantial genetic overlap [38]; the per year hazard ratio is 0.97 for the puberty timing score (P~4.8 × 10−4) after adjusting for the BMI score.
Using our approach instead, i.e., considering the father’s age at death, led to very similar results. Specifically, all traits significantly associated with paternal survival show a significant change in polygenic score with father’s age at death using the model with parental ages at death treated as ordinal variables (S21 Fig): TC (P~8.8 × 10−9), CAD (P~3.3 × 10−8), puberty timing (P~1.6 × 10−7), LDL (P~8.6 × 10−7), and BMI (P~3.4 × 10−6). In addition, we uncovered significant changes in polygenic score with father’s age at death for asthma (ATH, P~9.4 × 10−5) and triglycerides (TG, P~4.4 × 10−4, the effect of which does not seem to be distinct from other lipid traits, S20 Fig). The score for puberty timing increases monotonically with the father’s age at death (Fig 4B), indicative of a protective effect of later predicted puberty timing, whereas all other traits with significant signal show a monotonic decline in score with age (Fig 4C–4F).
In a Cox survival model, for mothers as with for fathers, scores for TC, CAD, and LDL are significantly associated with survival, with similar hazard ratios (Fig 5A and Table 1): 1 unit polygenic score hazard ratio of 1.09 for LDL (P~5.2 × 10−8), 1.09 for CAD (P~5.2 × 10−6), and 1.07 for TC (P~7.8 × 10−6). In addition, the high-density lipoproteins (HDL) score is associated with maternal survival (1 standard deviation (SD) hazard ratio of 0.94, P~8.9 × 10−5). Also, suggestive evidence was detected for protective effects of increased predicted age at first birth (AFB) (per year hazard ratio of 0.94, P~1.4 × 10−3) as well as predicted puberty timing (per year hazard ratio of 0.97, P~2.0 × 10−3) (Fig 5A and Table 1). Other than the LDL and TC, all signals seem to be distinct (S20 Fig), including for puberty timing and AFB, despite the genetic correlation between the 2 phenotypes [39].
In turn, applying our approach to maternal age at death, puberty timing and AFB are the top signals (P~2.2 × 10−4 and P~3.1 × 10−3, respectively, S21 Fig). Higher polygenic scores for puberty timing are enriched among longer-lived mothers (Fig 5B), as seen for fathers. Similarly, the score for AFB increases with mother’s age at death (Fig 5C), indicating an association between variants that delay AFB and longer life span. Scores for CAD, LDL, and HDL did not show significant monotonic change across mother’s age at death bins (P~7.7 × 10−3, P~0.058, and P~0.35, respectively); however, the trends are suggestive of subtle age-dependent effects, with an effect of CAD score in middle age and late-onset effects of LDL and HDL scores (Fig 5D–5F). Testing for age by sex interactions, the TC and CAD score trends with parental ages at death are significantly different between fathers and mothers (P~4.0 × 10−4 and P~7.4 × 10−4, respectively, S22 Fig).
To further investigate the age dependency of the effects, we plotted polygenic scores among parents who had survived up to a given age as compared to the trends with parental ages at death (S23 and S24 Figs). All traits associated with paternal survival seemingly show more pronounced effects in middle age (S23 Fig). Similar patterns were observed for maternal survival-associated traits except for LDL and HDL, which had more pronounced late-age effects (S24 Fig). We also compared the hazard ratios for ages at death of ≤ 75 and > 75 years (Materials and methods), similar to a recent study [33]. Consistent with trends in scores with parental age, among the traits associated with paternal survival, almost all traits have seemingly stronger effects among younger fathers, particularly for CAD (S3 Table): 1 unit log-odds hazard ratio of 1.14 for younger fathers (P~2.6 × 10−9) and 0.99 for older fathers (P~0.70). Unlike in fathers, in mothers, TC, LDL, and HDL scores had more pronounced late-age effects (S3 Table): for TC, 1 SD hazard ratio of 1.03 for younger mothers (P~0.15) and 1.11 for older mothers (P~1.4 × 10−6) and for LDL, 1 SD hazard ratio of 1.05 for younger mothers (P~0.03) and 1.12 for older mothers (P~3.3 × 10−8).
Next, we sought to replicate the top associations observed among the UK Biobank participants of British ancestry (discovery cohort) in 2 other data sets: participants of the UK Biobank of non-British ancestry and the GERA cohort. Applying the Cox model using parental survival for UK Biobank participants of non-British ancestry, the direction of hazard ratios for all traits (as well as the estimated values for most traits) are consistent with the discovery cohort for both fathers and mothers (S4 Table). The congruence of results in 2 cohorts with different ancestries suggests that our top signals are not false positives caused by poor control for population structure. In the GERA cohort, we tested whether polygenic scores change with the age of the participant, similar to our approach for individual genetic variants in this cohort. All top signals except AFB have directionally consistent effects with the discovery cohort (S5 Table). Of particular interest, the strongest signal is an increase in the polygenic score for puberty timing with age of the participants (P~6.7 × 10−3, S25 Fig).
In the discovery cohort, we further investigated if there are significant changes in the squared difference of polygenic scores with parental ages at death, as might be expected if the mean value of the trait leads to the highest chance of survival. No trait shows evidence of such stabilizing selection (S26 Fig).
We introduced a new approach to identify genetic variants that affect survival to a given age and thus to directly observe viability selection ongoing in humans. Attractive features of the approach include that we do not need to make a decision a priori about which loci or traits matter to viability and focus not on an end point (e.g., survival to an old age) but on any shift in allele frequencies with age, thereby learning about the ages at which effects are manifest and possible differences between sexes.
To illustrate the potential of our approach, we performed a scan for genetic variants that impact age-specific mortality in the GERA and the UK Biobank cohorts. We only found a few individual genetic variants, almost all of which were identified in previous studies. This result is in some ways expected: available data only provide high power to detect effects of common variants (>0.15–0.2) on survival (Fig 1), yet if these variants were under viability selection, we would not expect them to be common, short of strong balancing selection due to trade-offs between sexes, ages, or environments. As sample sizes increase, however, the approach introduced here should provide a comprehensive picture of viability selection in humans. To illustrate this point, we repeated our power simulation with 500,000 samples and found that we should have high power to detect the trends for alleles at a couple percent frequency in the sample (S27 Fig).
Already, however, this application raises a number of interesting questions about the nature of viability selection in humans. Notably, we discovered only a few individual variants influencing viability in the 2 cohorts, most of which exert their effect late in life. On first thought, this finding may suggest such variants to be neutrally evolving. We would argue that if anything, our findings of only a few common variants with large effects on survival late in life suggest the opposite: that even variants with late-onset effects have been weeded out by purifying selection. Indeed, unless the number of loci in the genome that could give rise to such variants (i.e., the mutational target size) is tiny, other variants such as the APOE ε4 allele must often arise. That they are not observed when we have very high power to detect them suggests they are kept at lower frequency by purifying selection. Why might they be selected despite affecting survival only at old ages? Possible explanations include that they decrease the direct fitness of males sufficiently to be effectively selected (notably given the large, recent effective population size of humans [60]) or that they impact the inclusive fitness of males or females. If this explanation is correct, it raises the question of why the APOE ε4 allele has not been weeded out. We speculate that the environment has changed recently, making this allele more deleterious. For example, it has been proposed that the evolution of this allele has been influenced by changes in physical activity [61] and parasite burden [62].
Considering 42 traits that have been investigated by GWASs, we found a number of cases in which the mean polygenic score changes with age. Of course, detecting an effect of age on the traits does not imply that these are the phenotypes under viability selection, as the variants that contribute likely have pleiotropic effects on other traits [37]. Nonetheless, it is perhaps not surprising that we found detrimental effects of higher genetically predicted TC, LDL, BMI, and risk of CAD on survival, as these phenotypes are studied in GWASs precisely because of their adverse health effects. Intriguingly, however, we also found associations for fertility traits, notably, protective effects of later predicted puberty timing and AFB. If these findings reflect life-history trade-offs (e.g., longer life span at the cost of delayed reproduction), they may help to explain the persistence of extensive variation in such fitness-correlated traits [63,64]. Intriguingly, we saw a negative correlation between genetically predicted AFB and number of siblings of the UK Biobank participants, a proxy for the fertility of their parents (P~4.2 × 10−8, S28 Fig), consistent with previous reports of a genetic correlation between AFB and the number of children ever born [21,39]. These findings underscore that consideration of survival or fertility effects alone does not allow one to infer whether the net effect of a variant or set of variants is beneficial. Instead, to convert effects on viability such as those detected here or effects on fertility reported elsewhere [22,23] into an understanding of how natural selection acts on an allele requires a characterization of its effects on all components of fitness (including potentially inclusive fitness).
In this regard, it is also worth noting that while our method is designed to detect changes in allele frequencies (and in polygenic scores) caused by genetic effects on age-specific mortality, such changes could in principle also arise from effects on other components of fitness. For example, if the frequency of a genetic variant in a population decreases over decades due to an effect on fertility, its frequency would increase with the age of surviving individuals sampled at a given time (as in the GERA cohort). This confounding is less of an issue when considering effects on the age at death (what we measured in the UK Biobank). Nonetheless, even in the UK Biobank, fertility effects may manifest as effects on age at death; for example, when sampling a cohort of children, parents with later ages at death are possibly born earlier (S29 Fig). To this end, in the UK Biobank, we accounted for changes in allele frequencies with year of birth of the participants themselves (ideally, we would want to condition on parents born at similar times, which we cannot do; instead, we used year of birth of the participants as an estimator for year of birth of the parents). Thus, we believe our results in the UK Biobank not to be confounded by fertility effects. Moreover, a number of our findings in this study are consistent with prior knowledge of effects on survival, such as those for disease risk variants like the APOE ε4 allele. Nonetheless, some caution is required in interpreting trends with age as strictly reflecting viability effects.
Also of interest are the marked differences between males and females in our analysis of mothers and fathers of individuals in the UK Biobank. The differences between sexes are most notable at the CHRNA3 locus, which shows a strong effect only in fathers, and sets of genetic variants associated with risk of CAD and cholesterol levels, which exhibit different age-dependent effects between fathers and mothers. Results for the CHRNA3 locus, in which variants are associated with the amount of smoking among smokers, may reflect a gene-by-environment interaction rather than a sex effect per se. Consistent with a more pronounced effect on male than female age at death, smoking prevalence in men has been consistently higher than women over the past few decades in the United Kingdom: from 1970 to 2000, smoking prevalence decreased from around 70% to 36% in middle-aged men, compared to from around 50% to 28% in middle-aged women [65].
Moving forward, the application of approaches such as ours to the millions of samples in the pipeline (such as the UK Biobank [66], the Precision Medicine Initiative program [67], and the BioVU biobank at Vanderbilt University [68]) will allow viability effects of rare as well as common alleles to be examined. These analyses will provide a comprehensive answer to the question of which loci affect survival, helping to address long-standing open questions such as the relative importance of viability selection in shaping genetic variation and the extent to which genetic variation is maintained by fitness trade-offs between sexes or across ages.
This study used data sets from the UK Biobank (application number 11138), as approved by the UK Biobank Board, and the Genetic Epidemiology Research on Adult Health and Aging (GERA), obtained through dbGaP (request numbers 28113–4 and 57119–2) and approved by Columbia University Institutional Review Board, protocols AAAQ2700 and AAAN4411.
We performed PCA using the EIGENSOFT v6.0.1 package with the fastpca algorithm [76,77] for 2 purposes: (i) as a QC on individuals to validate self-reported European ancestries (only in GERA data set) and (ii) to correct for population structure in our statistical model (for individuals in the UK Biobank of non-British ancestry, we used the PCs provided with the data).
We downloaded the list of variants contributing to 39 traits (all traits but age at menarche, AFB, and age at natural menopause) and their effect sizes recently described in Pickrell et al. [37] from: https://github.com/PickrellLab/gwas-pw-paper/tree/master/all_single. For age at menarche, we used the variants and effect sizes recently identified by Day et al. [38]. We used variants associated with AFB from Barban et al. identified in either sex-specific analyses or analyses of both sexes and used the effect sizes estimated in the combined analysis [39]. We used age at natural menopause-associated variants and their effect sizes from Day et al. [40]. For all traits, we used variants that were genotyped/imputed with high quality in our data (see S1 Table).
We ran simulations to determine the power of our statistical model to detect deviation of allele frequency trends with age across 14 age categories mimicking the GERA cohort’s age structure (57,696 individuals with age distribution as in S2 Fig) from a null model, which for simplicity was no change in frequency with age, i.e., no changes as a result of age-dependent variation in population structure and batch effects. For a given trend in frequency of an allele with age, we generated 1,000 simulated trends in which the distribution of the number of the alleles in age bin i is Bin(2Ni, fi), where Ni and fi are the sample size and the sample allele frequency in bin i. We then estimated the power to detect the trend as the fraction of cases in which P < 5 × 10−8, by a chi-squared test.
We ran simulations to investigate the relationship between allele frequency with age of the surviving individuals and the age of the individuals who died in a cohort. We simulated 2 × 106 individuals going forward in time in 1-year increments. For each time step forward, we tuned the chance of survival of the individuals based on their count of a risk allele for a given variant such that the number of individuals dying in the increment complies with: (i) a normal distribution of ages at death with mean of 70 years and standard deviation of 13 years, roughly as is observed for parental ages at death in the UK Biobank and (ii) a given frequency of the risk allele among those who survive. Specifically, we modeled the survival rate of the population, S, as the weighted mean for 2-alleles carriers, S2, 1-allele carriers, S1, and noncarriers, S0:
S(x)=∑i=02fiSi(x)
where f denotes the frequency of genotypes in the population and x denotes the age. Si and S are related: Si(x) = S(x) fi(x)/fi, where fi(x) is the genotype frequency among individuals survived up to age x. Given a trend in allele frequency with age, we calculated genotype frequencies with age assuming Hardy-Weinberg equilibrium and then estimated genotype-dependent chance of survival, Si(x), taking S(x) as the survival function for N(70, 132).
|
10.1371/journal.ppat.1000745 | Protection of Mice against Lethal Challenge with 2009 H1N1 Influenza A Virus by 1918-Like and Classical Swine H1N1 Based Vaccines | The recent 2009 pandemic H1N1 virus infection in humans has resulted in nearly 5,000 deaths worldwide. Early epidemiological findings indicated a low level of infection in the older population (>65 years) with the pandemic virus, and a greater susceptibility in people younger than 35 years of age, a phenomenon correlated with the presence of cross-reactive immunity in the older population. It is unclear what virus(es) might be responsible for this apparent cross-protection against the 2009 pandemic H1N1 virus. We describe a mouse lethal challenge model for the 2009 pandemic H1N1 strain, used together with a panel of inactivated H1N1 virus vaccines and hemagglutinin (HA) monoclonal antibodies to dissect the possible humoral antigenic determinants of pre-existing immunity against this virus in the human population. By hemagglutinination inhibition (HI) assays and vaccination/challenge studies, we demonstrate that the 2009 pandemic H1N1 virus is antigenically similar to human H1N1 viruses that circulated from 1918–1943 and to classical swine H1N1 viruses. Antibodies elicited against 1918-like or classical swine H1N1 vaccines completely protect C57B/6 mice from lethal challenge with the influenza A/Netherlands/602/2009 virus isolate. In contrast, contemporary H1N1 vaccines afforded only partial protection. Passive immunization with cross-reactive monoclonal antibodies (mAbs) raised against either 1918 or A/California/04/2009 HA proteins offered full protection from death. Analysis of mAb antibody escape mutants, generated by selection of 2009 H1N1 virus with these mAbs, indicate that antigenic site Sa is one of the conserved cross-protective epitopes. Our findings in mice agree with serological data showing high prevalence of 2009 H1N1 cross-reactive antibodies only in the older population, indicating that prior infection with 1918-like viruses or vaccination against the 1976 swine H1N1 virus in the USA are likely to provide protection against the 2009 pandemic H1N1 virus. This data provides a mechanistic basis for the protection seen in the older population, and emphasizes a rationale for including vaccination of the younger, naïve population. Our results also support the notion that pigs can act as an animal reservoir where influenza virus HAs become antigenically frozen for long periods of time, facilitating the generation of human pandemic viruses.
| Influenza A viruses generally infect individuals of all ages and cause severe respiratory disease in very young children and elderly people (>65 years). However, the 2009 pandemic H1N1 virus infection is predominantly seen in children and adults (<35 years of age), but rarely in people older than 65 years of age. Recent serological studies indicate that older people carry antibodies that recognize the 2009 H1N1 virus. This suggests that they may have been exposed to or vaccinated with an influenza virus similar to 2009 H1N1 virus. In this study, we wanted to identify the older H1N1 virus(es) that may confer protection to the elderly population. Using 11 different inactivated influenza A viruses that have circulated between 1918 to 2007, we immunized mice and challenged them with a lethal dose of the 2009 novel H1N1 virus. We find that mice vaccinated with human H1N1 viruses that circulated in 1918 and in 1943 were protected from the 2009 H1N1 virus. Also, the 1976 swine origin H1N1 virus, against which nearly 40 million people were immunized in 1976 in the United States, protects mice from death by the 2009 H1N1 virus. This indicates that people carrying antibodies against H1N1 viruses that circulated between 1918–1943 and to the 1976 swine origin H1N1 virus are likely to be protected against the 2009 pandemic H1N1. Importantly, our data underscores the significance of vaccinating people under 35 year of age, since the majority of them do not have protective antibodies against the 2009 H1N1, and provide a possible mechanism by which pandemic viruses could arise from antigenically frozen influenza viruses harbored in the swine population.
| Influenza A viruses (IAV), members of the Orthomyxoviridae family, cause severe respiratory diseases in humans with an average mortality rate of 36,000/year in the United States alone [1]. Apart from yearly seasonal outbreaks, IAV can cause frequent epidemics and occasional pandemics in humans [2],[3]. Vaccination has been one of the most effective means of protection against IAV. Vaccine induced production of antibodies against the viral surface glycoprotein hemagglutinin (HA) is crucial for immune protection [4]. The HA plays a critical role in the virus life cycle by mediating virus binding to sialic acid containing receptors on the cell surface and fusion of viral and endosomal membranes, leading to viral entry into the host cell [2],[5]. HA-specific antibodies have been demonstrated to block the IAV infection by preventing receptor binding and/or fusion. However, the HA protein, due to antibody mediated immune selection pressure, undergoes rapid antigenic evolution by accumulation of mutations (“antigenic drift”) and through genetic reassortments of segments (“antigenic shift”). In the 20th century, influenza virus caused three pandemics in humans: 1918 “Spanish influenza” (H1N1), 1957 “Asian influenza” (H2N2) and 1968 “Hong Kong influenza” (H3N2) [6],[7]. In April 2009 the Centers for Disease Control and Prevention (CDC) of the United States of America announced the detection of a novel strain of influenza virus in humans. Further investigation revealed that this novel virus derived its genes from viruses circulating in the pig population. Due to sustained human-to-human transmission of this novel virus throughout the world on June 11th, the World Health Organization (WHO) raised the worldwide pandemic alert level to phase 6 (e.g. ongoing global spread and community level outbreaks in multiple parts of world). All of the past pandemics and the recent 2009 swine-origin IAV H1N1 pandemic have been caused by IAV strains carrying an antigenically novel HA segment in populations immunologically naïve to that particular HA.
Apart from humans, IAV can infect a variety of species including poultry, aquatic birds, horses, pigs, dogs and seals [8],[9]. Aquatic wild birds are considered to be natural reservoirs of different subtypes of IAV, and pigs, in addition to harbor swine influenza virus strains, are thought to serve as “mixing vessels” for genetic reassortments of viral segments between avian and human influenza virus strains [8],[10]. Sometime during the 1918–20 pandemic, IAV H1N1 resembling the 1918 pandemic virus was introduced into the domestic swine population [11],[12]. Until today, these H1N1 viruses, referred as classical swine H1N1, have circulated in the swine population with relatively modest changes in HA antigenicity [13]. In contrast to HA antigenic stasis in swine, descendants of the 1918 virus circulated in humans with considerable antigenic drift in HA, until they were replaced by the 1957 H2N2 pandemic virus [4]. In 1978, an H1N1 virus resembling a late 1950's human H1N1 virus reemerged in the human population [14],[15]. From 1977 to the present day, both H3N2 and H1N1 viruses co-circulate in humans. In addition, in 1976, a swine H1N1 IAV (classical H1N1) outbreak was reported among soldiers at the Fort Dix Army base in New Jersey [16],[17]. Because no H1N1 viruses had circulated in humans since 1957, the Fort Dix outbreak raised fears of an H1N1 pandemic. Therefore, a vaccine based upon an inactivated A/NJ/76 (H1N1) virus was developed and administered to nearly 40 million people in the United States. However, the 1976 swine H1N1 virus did not cause a pandemic and no infections were reported outside of Fort Dix.
Although several infections with swine IAV have been reported in humans, such infections have been rare [18],[19],[20],[21]. Surprisingly, the current 2009 pandemic H1N1 has proved to be very efficient in human-to-human transmission compared to previous swine influenza viruses. Apart from lack of pre-existing immunity against this virus in most humans, it is still unclear what viral genetic factors contribute to this higher transmission rate. It is interesting to note that recent epidemiological data indicate a higher rate of confirmed 2009 H1N1 infection in individuals younger than 18 years of age as compared to older individuals [22]. It is speculated that some of the older population may have been exposed to a virus antigenically similar to 2009 H1N1. Nevertheless, it is currently unknown which specific virus or viruses that circulated in the past and during what years, might be responsible for this apparent serum cross-reactivity reported in the older population (above 65 years of age) [23],[24],[25].
In this study, using a lethal mouse model system of the current 2009 H1N1 virus, we tested the efficacy of a panel of 11 different virus vaccines spanning from 1918 to 2009. We demonstrate that mice immunized with inactivated vaccines based on human H1N1 virus from 1918 to 1943 or based on classical swine H1N1 viruses confer complete protection from death against a lethal challenge with the 2009 H1N1 virus. In contrast, vaccination of mice with inactivated human H1N1 viruses isolated after 1950, including contemporary H1N1 viruses, offered only partial protection and resulted in greater morbidity due to 2009 H1N1 virus infection. In agreement with vaccination/challenge studies, pre-challenge sera from mice vaccinated with 1918, and classical swine antigens show neutralization of hemagglutination of 2009 H1N1 virus. In addition, we isolated anti-HA specific monoclonal antibodies (mAb) that have cross-reactivity against 1918 and 2009 H1N1 HA's. In challenge experiments, these mAbs offer complete protection against death by the 2009 H1N1 virus following passive immunization. Analysis of escape mutant viruses selected in the presence of these mAbs mapped to antigenic site Sa. Together, these data suggest that the novel 2009 pandemic H1N1 virus shares a great level of antigenic similarity to the 1918 virus and that site Sa is highly conserved in these two viruses. These results agree well with epidemiological data which indicate that the older population (age >65) are less susceptible to the 2009 H1N1 [24],[25],[26] and with data arising from two recent studies that demonstrated high prevalence of cross-neutralizing antibodies in people born before 1940 [27],[28]. Taken together, our observations provide a rationale for the protection observed in the older population to the 2009 H1N1 pandemic virus and the greater susceptibility seen in younger individuals. Thus, our data indicate that individuals that have been previously exposed to and contain antibodies against 1918-like H1N1 viruses or classical swine H1N1 are likely to be protected against the novel swine-origin 2009 H1N1 virus.
All animal procedures performed in this study are in accordance with Institutional Animal Care and Use Committee (IACUC) guidelines, and have been approved by the IACUC of Mount Sinai School of Medicine.
Human embryonic kidney (293T) cells were maintained in DMEM supplemented with 10% FBS and 1000u/ml penicillin/streptomycin. Madin–Darby canine kidney (MDCK) cells were maintained in MEM supplemented with 10% FBS and penicillin/streptomycin. Reagents for cell culture were purchased from Gibco Life Technologies.
H1N1 viruses used in this study are as follows: A/Swine/Iowa/30 (Sw/30), A/Puerto Rico/8/34 -MSSM (PR8), A/Weiss/43 (Wei/43), A/New Jersey/8/1976 (NJ/76), A/USSR/92/77 (USSR/77), A/Houston/20593/84 (Hou/84), A/Texas/36/1991 (Tx/91), A/Brisbane/59/2007 (Bris/59/07), A/California/04/2009 (Cal/09) and A/Netherlands/602/2009 (Neth/09). A/Northern Territory/60/1968 (NT/68) and A/Brisbane/10/2007 (Bris/10/07) were used as H3N2 controls. All experiments involving 2009 H1N1 viruses were conducted under Biosafety level 3 (BSL3) conditions for animal work and Biosafety level 2 with BSL3 practices laboratory conditions for in vitro work, in accordance with guidelines of the Centers for Disease Control and Prevention. Cal/09 and Neth/09 virus stocks used for mice experiments were all grown in MDCK cells. All other viruses were grown in 10d old eggs at 37°C for 2–3 days.
The recombinant influenza A virus carrying HA/NA from Cal/09 and the internal six genes from PR8, referred to as Cal/09 6:2 in the reminder of the study, was generated as described previously [29]. Briefly, 293T cells were transfected with eight ambisene pDZ vectors encoding the 8 viral genes and proteins and 24 h post-transfection, the supernatant was inoculated into 8-day-old embryonated chicken eggs. The allantoic fluid was harvested after 3 days of incubation at 35°C. The rescued virus was plaque purified in MDCK cells and re-grown in 10-day-old embryonated chicken eggs.
All viruses used to prepare inactivated vaccines were grown in 10d old eggs at 37°C for 2 days. After the clarification of debris in the allantoic fluid by low speed centrifugation, the viruses were pelleted on a 30% sucrose cushion by centrifugation at 25,000 rpm for 2hr. The viral pellet was resuspended in calcium borate buffer pH = 7.0 (143mM Sodium chloride, 10mM Calcium chloride, 20mM Boric acid 2.5mM sodium borate) at 1mg/ml concentration calculated by the Bradford method (Bio-Rad laboratories, Hercules, CA) and the virus was inactivated by formaldehyde (0.9% final conc.). For the Cal/09 vaccination group, in place of wild type virus, a Cal/09 6:2 was used (described above). For the 1918 virus-based vaccine, virus-like particles (VLP) were used for vaccination. VLPs were produced by co-transfecting HA (A/South Carolina/1/18) and NA (A/Brevig Mission/1/18) into 293T cells as previously described [30]. Released VLPs were purified on a 30% sucrose cushion. Vaccine doses used were based on the amount of total protein concentration measured by the Bradford method.
Pandemic H1N1 2009 and 1918 HA-specific mouse mAbs were generated by the hybridoma shared research facility at Mount Sinai School of Medicine, New York, NY. For the generation of 2009 H1N1 HA-specific antibodies, 6-week-old BALB/C mice were infected intranasally with 5×104 pfu of influenza virus Cal/09. Four weeks later, the mice were given 2×107 pfu of Cal/09 6:2 intravenously to boost Cal/09 HA-specific B cells. Three days after the boost the spleen was harvested and B cells from the spleen were fused with SP2/0 myeloma cells by the addition of polyethylene glycol. The supernatants of the resulting hybridomas were tested by immunofluorescence staining on 293T cells transfected with pDZ-HA (Cal/09). Positive hybridomas were subcloned and re-tested. To generate 1918 virus HA-specific antibodies, mice were immunized by DNA vaccination with pCAGGS-HA of the influenza A/South Carolina/1/18 H1N1 virus and then boosted with whole inactivated virus as above. Hemagglutination inhibition (HI) assays were used to select cross-reactive antibodies against homologous HA using 1918 VLPs. The 1918 HA-specific mAbs 6B9 and 39E4 correspond to isotype IgG2a and the Cal04/09 mAb 29E3 correspond to an IgG2b isotype. The mAbs were isotyped using the IsoStrip kit (Roche, Indianapolis, IN) and purified using a Protein A sepharose column.
Mouse sera were inactivated using a trypsin-heat-periodate treatment as previously described [31]. Briefly, one volume of sera was mixed with half a volume of trypsin 8 mg/ml (Sigma-Aldrich, St. Louis, MO) in 0.1 M phosphate buffer, pH 8.2 and then incubated for 30 min at 56°C. The samples were allowed to cool to room temperature (RT) and were mixed with 3 volumes of 0.11 M metapotassium periodate and incubated at RT for 15 min. Three volumes of 1% glycerol saline were then added and mixed with the samples and further incubated for 15 min at RT. The samples were diluted to a final 1∶10 dilution by adding and mixing 2.5 volumes of 85% saline. HI assays of mAbs and sera were conducted following standard protocols [30]. Two-fold serial dilutions of sera or mAbs were mixed and pre-incubated in 96-well plates for 30 min at 4°C with 8 HA units of virus per well. Turkey red blood cells were added to a final concentration of 0.25%, and the plate was incubated on ice for 30 min. Hemagglutination inhibition (HI) titers of sera were determined as the highest dilution that displayed hemagglutinating activity. Specific HI activity of mAbs was calculated as the lowest concentration of mAb that displayed hemagglutinating activity.
mAb escape mutants were generated as previously described [32]. Briefly, Cal/09 (6:2) virus was incubated with excess monoclonal antibody for 1hr at room temperature. The virus-mAb mix was inoculated into 10 day old embryonated eggs and incubated at 37°C for 48 hrs. The virus (allantoic fluid) grown under this conditions was harvested and re-tested for loss of HI activity against the same mAb used for selection. Mutations in the HA responsible for the Ab escape were identified by direct sequencing of vRNA obtained by RT-PCR.
The structural models for HA of Cal//09 and Bris/59/07 were generated using structural automated protein structure homology-modeling prediction server (Swiss-Model) with best fitting templates (PDB: 1ruy for Cal/09 and 1rvx for Bris/59/07) [34]. The structure of 1918 HA was obtained from PDB (PDB ID: 2wrg). The final images of the HA structures were generated using Pymol (Delano Scientific) [35].
To evaluate the virulence and infection kinetics of novel pandemic H1N1 viruses in mice, six-week old BALB/c and C57B/6 mice were infected with two virus isolates of 2009 H1N1 and monitored for signs of body weight loss and survival. Inoculation of BALB/c mice with Cal/09 isolate produced a marked decrease in body weight even at 103 pfu (∼13%, Fig. 1A), a phenomenon not commonly observed with non-mouse adapted human H1N1 viruses [36]. When infected with 5×104 pfu 50% of the mice reached the experimental end point of >25% weight loss by day 9 (Fig. 1C). The remaining mice showed substantial weight loss by day 8 p.i., with an average weight of 77.5% (Fig. 1A). Viral lung titers in these mice at both day 2 and 5 were notably high with no significant difference at both time points (Fig. 1D). Similar results were observed in C57B/6 mice after inoculation with Cal/09 virus. At doses of 103 and 5×104 pfu mice showed body weight losses comparable to those of BALB/c mice (Fig. 1B). The viral titers in the lungs of C57B/6 mice were almost ten-fold higher on day 6 p.i. than on day 3 p.i. (Fig. 1D), indicating active replication of virus. As with BALB/c mice, infection with 5×104 pfu resulted in 50% of mice reaching the experimental end point by day 8 p.i. (Fig. 1C), and an average weight loss of 21.8% on day 7 p.i. (Fig. 1B).
To evaluate whether similar virulence is observed with other isolates of 2009 pandemic H1N1, we infected mice with the Neth/09 isolate. Interestingly, Neth/09 showed a considerably higher virulence in C57B/6 mice. At a dose of 5×104 pfu/ml, the mice exhibited rapid and substantial weight loss by day 4 (Fig. 1B). All mice succumbed to infection or reached experimental end point by day 6 p.i. (Fig. 1C), and had higher lung viral titers than those observed with the Cal/09 isolate at this time point (Fig. 1E).
To further evaluate the pathogenesis of Neth/09, we next determined the LD50 of Neth/09 isolate in mice. Nine-week old female C57B/6 mice were infected with different doses of virus (5×102, 5×103, 5×104, 5×105 or 5×106 pfu) and monitored for weight loss and survival over a period of 14 days p.i. (Fig. 2A and B). All mice inoculated with 5×104 pfu or higher succumbed to infection or reached the experimental end point between days 4–8 p.i. (Fig. 2A). Mice inoculated with 5×103 pfu showed a 12.2% decrease in body weight, and mice infected with 5×102 pfu showed no significant drop in weight (Fig. 2B). The LD50 for Neth/09 was determined to be 1.58×104 pfu for 9-week-old C57B/6 mice (Reed and Muench method, [33]). Overall, these data indicates that the Neth/09 isolate is more pathogenic in mice than Cal/09 and thus, in the challenge experiments described herein, the Neth/09 isolate and C57B/6 mice were used as a lethal model for 2009 pandemic H1N1.
Next, to evaluate the level of antigenic resemblance between 2009 H1N1 viruses and other H1N1 viruses, we performed hemaglutination inhibiton (HI) assays and challenge experiments. Analysis of sera from mice infected with Cal/09 revealed similar levels of HI activity against the Cal/09 and Neth/09 isolates, suggesting that they are antigenically very similar, despite significant differences in pathogenicity in mice (data not shown).
As a proof of principle and to establish if prior infection with Cal/09 is sufficient to confer protection against a lethal challenge with an antigenically similar Neth/09 isolate, C57B/6 mice were first infected with Cal/09 and allowed to seroconvert for 21 days, followed by lethal challenge with 106 pfu of Neth/09 (>50 LD50). As expected, all mice previously infected with Cal/09 survived (Fig. 2D). Mice previously inoculated with 103 pfu of Cal/09 showed a decrease of 10% body weight. However, all the mice regained the weight by day 6 (Fig. 2C). No significant weight change was observed in the 5×104 pfu group. In contrast, mice previously mock infected succumbed to Neth/09 challenge by day 8 (Fig. 2D). Together, this demonstrates that Cal/09 and Neth/09 are antigenically very similar and there is cross-protection between different 2009 pandemic H1N1 isolates.
Recent serological studies have shown that the human population under the age of 30 has little or no level of neutralizing antibodies against the 2009 pandemic H1N1 virus [22],[27],[28],[37]. However, sera from adults older than 35 years of age showed varied levels of neutralizing activity to 2009 pandemic H1N1, with people born before 1940's showing highest degree of neutralizing activity [27]. To investigate if antibodies against older H1N1 viruses can cross-react with and therefore offer protection against the 2009 pandemic H1N1 virus, C57B/6 mice were immunized with 1918 VLP or different inactivated H1N1 virus vaccines, spanning from 1918–2009. Prior to lethal challenge, all mice were tested for seroconversion against the homologous virus and were found to have HI antibody titers equivalent to ≥40 (Table 1). To test if vaccination with any of the other H1N1 viruses offered protection against novel 2009 pandemic H1N1, immunized mice were challenged with a lethal dose of Neth/09 (50 LD50), and the vaccine efficacy was evaluated by assessment of weight loss and survival over a 14 day period p.c., and also by measuring the virus titer in the lower respiratory track (Fig. 3). In the no vaccination group (control), mice succumbed to infection and reached the experimental endpoint by day 5 (Fig. 3B). This group of mice showed the highest viral titers in the lungs at days 3 and 6 p.c. (Fig. 3C). As anticipated immunization with Cal/09 6:2 offered full protection from death to mice (Fig. 3B). Modest levels of disease were observed in this group as evidenced by a decrease in body weight (Fig. 3A). However, we did not detect infectious virus in the lungs of these mice on days 3 or 6 p.c. (Fig. 3C; limit ≥10 pfu). Interestingly, inactivated classical swine H1N1 viruses (Sw/30 or NJ/76), 1918 VLPs and an inactivated human H1N1 virus isolated in 1943 (Wei/43) offered 100% protection against death from 2009 pandemic H1N1 challenge. The 1918 VLP-vaccinated mice all survived and showed weight loss comparable to that of the Cal/09 6:2 vaccinated animals (Fig. 3B and 3A). We detected 100-fold lower levels of virus in the lungs of challenged mice on day 3 p.c. and no virus on day 6 p.c. Similarly, in mice immunized with Sw/30 or NJ/76, Neth/09 virus was only detected in the lungs on day 3 p.c., indicating that these two vaccines had an efficacy in reducing viral replication in lungs comparable to that of 1918 VLPs (Fig. 3C). However, a slightly higher weight loss (∼18.2%) was observed in these groups (Fig. 3A). Although vaccination with Wei/43 resulted in 100% survival, these mice showed a much greater and sustained weight loss than the 1918 VLP, SW/30 and NJ/76 vaccinated mice, and the viral titers in the lungs were high on both day 3 and day 6 p.c., suggesting only modest levels of inhibition of Neth/09 virus replication (Fig. 3A, B and C). In contrast, immunization with human H1N1 viruses that circulated from 1977–2007 showed only partial protection against lethal Neth/09 challenge (Fig. 3E). This was evidenced by extensive weight loss observed in these mice (Fig. 3D). In mice vaccinated with USSR/77, Hou/84, Tx/91 or Bris/59/07 only 20–60% survival was observed (Fig. 3D and E). The viral titers on day 3 p.c. were similar to controls. Nevertheless, on day 6, viral titers were lower, particularly in the Tx/91 vaccinated mice that had around 1000-fold lower viral titers as compared to day 3 p.c. (Fig. 3F). In the groups vaccinated with the NT/68 and Bris/10/07 strains, viruses belonging to the H3N2 subtype, lethal challenge resulted in either a 25% survival or all mice succumbing to infection, respectively. In agreement with the increased morbidity seen in Bris/10/07 vaccinated mice, weight loss and viral titers were also high. The reasons for a lower viral titer in the NT/68 vaccination group on day 6 p.c. are unclear (Fig. 3F). Nevertheless, these results indicate that inactivated vaccines based on classical swine H1N1 viruses and on human 1918 and 1943 H1N1 viruses protect against death in mice infected with the 2009 pandemic H1N1 virus and suggest that they likely share significant antigenic similarity.
Next, to test whether the protection observed in 1918–1943 and NJ/76 vaccinated correlates with cross-reactive neutralizing antibodies in the sera, we tested the pre-challenge sera from vaccinated mice for HI activity against different H1N1 viruses (Table 2). All the pre-challenge sera showed HI titer between 160–2560 against the homologous virus. As expected, sera from 1918 VLP, SW/30 and NJ/76 vaccinated animals showed HI activity against both 2009 H1N1 pandemic viruses (HI titer ≥40), suggesting that protection is due to cross-reactive antibodies. However, despite complete protection against death from Neth/09 lethal challenge, the sera from Wei/43 showed no HI activity against 2009 H1N1 viruses. This might indicate that the HI assay is less sensitive than the challenge assay in vivo to evaluate the presence of protective Abs. Taken together, our results demonstrate that 1918, SW/30 and NJ/76 share antigenically similar epitope(s) that might be responsible for the cross-protection.
To further examine the common possible epitopes shared between 1918 and 2009 H1N1 viruses, we examined cross-reactivity of 1918 HA-specific monoclonal antibodies with Cal/09 by HI assay. Two 1918 HA mAb's, 6B9 and 39E4, showed HI activity against Cal/09. Also, one of the Cal/09 HA specific antibodies (29E3) showed HI activity against Cal/09 and against 1918 VLPs (Table 3). To evaluate if these mAbs could confer protection (in vivo) against lethal Neth/09 challenge, mice were given 150 µg of mAb as a prophylactic treatment 24h prior to challenge (50 LD50). To assess the level of HI titers in the blood prior to challenge, the HI activity in passively immunized mice sera was measured (Table 4). Sera from mice immunized with an Ab raised against an unrelated viral protein (Nipah virus W) did not show HI activity. Mice immunized with HA-specific monoclonal antibodies showed considerable levels of HI titers (HI = 40–320). Mice immunized with polyclonal sera showed a lower HI titer (HI = 20–80; Table 4). As expected, administration of the control antibody resulted in no protection from lethal challenge and all mice had high virus titers in the lungs equivalent to those seen in untreated mice (Fig. 4). Polyclonal sera from mice previously infected with Cal/09 virus were used as a positive control prophylactic treatment to establish the baseline of protective antibodies in vivo (Poly sera). Mice in this group were completely protected from lethal challenge and showed a decrease in lung virus titers on day 3 and 6 compared to untreated controls. Nonetheless, these mice showed a significant weight loss during the first 4 days (Fig. 4A, B and C). Treatment with 150 µg of 1918 specific anti HA monoclonal antibodies (6B9 and 39E4) also conferred full protection from challenge and showed reduced weight loss as compared to the polyclonal sera treatment group. The Cal/09 specific monoclonal antibody (29E3) treatment resulted in 100% survival and no weight loss, indicative of little or no morbidity in these mice. Assessment of infectious virus in the lower respiratory track of mice treated with all monoclonal antibodies revealed decreased viral titers as compared to controls especially on day 3 p.c. (Fig. 4C). Taken together, these data suggest that cross-reactive monoclonal antibodies protect against lethal Neth/09 challenge by blocking a conserved antigenic site between 1918 and 2009 pandemic H1N1 viruses.
To identify the shared cross-protective epitope(s) in HA, we generated mAb escape mutants of Cal/09 6:2 virus. Escape mutants generated by pre-incubation of Cal/09 virus (6:2) with excess of each of the cross-reactive mAbs (6B9, 39E4 and 29E3) resulted in a loss of HI activity against the same mAb. Of interest, all the escape mutant viruses generated with either 1918- or 2009 HA-specific mAbs carried mutations in the antigenic site Sa (Table 5; Fig. 5). The escape mutants that arose in the presence of the 1918 specific mAb 6B9 and 39E4 contained mutations G172E (G158E by H3 numbering) and K171E (K157E by H3 numbering), respectively. The escape mutants generated by selection with the 2009 H1N1 specific mAb (29E3) carried either a K171E or K171Q or K180N mutation (residues K171 and K180 correspond to K157 and K166 in H3 numbering). These results indicate that all the mAbs bind to the conserved antigenic site Sa (Fig. 5 and 6).
Seasonal influenza viruses predominantly cause severe disease in very young children and in the older population. However, infections with the 2009 pandemic H1N1 are considerably lower in people 65 years or older, likely due to immunity from prior exposure and/or vaccination to an antigenically similar influenza virus [24],[25],[26]. Using a panel of 11 different influenza viruses from 1918–2009, we tested the ability of inactivated vaccines based on these viruses to protect against a mouse-lethal 2009 pandemic H1N1. Here, we show that vaccination of mice with human 1918 influenza virus VLPs, and inactivated human Wei/43 and classical swine H1N1 viruses completely protects from death by lethal challenge with the 2009 pandemic H1N1. Also, analysis of pre-challenge sera from mice vaccinated with these viruses, except Wei/43, show cross-reactivity against the 2009 H1N1 virus (HI titer ≥40). In contrast, vaccination with more contemporary H1N1 viruses offered only partial protection. In addition, passive immunization with 1918 HA-specific monoclonal antibodies protects against the 2009 H1N1 virus, suggesting antigenic similarity between these viruses. Cross-protective epitope mapping shows that 1918 HA-specific mAbs protect by binding to the conserved antigenic site Sa in 2009 H1N1 virus. Based on our HI data (Table 2) and since inactivated vaccines are known to induce protective humoral and minimal cellular immunity, this protection was likely mediated by antibodies, as also evidenced by the lack of complete protection by H3N2-based vaccines. Our results indicate that prior exposure to human influenza A viruses from 1918–1943 or vaccination with classical swine H1N1 virus (NJ/76) offers significant levels of cross-protection against the novel 2009 H1N1 pandemic virus and thus provides a mechanistic explanation for the lower incidence of disease and/or infection seen in people aged 65 or older.
Usually human influenza viruses do not show substantial replication in mice without prior adaptation to this host [36]. Only a few viruses including 1918 and highly pathogenic H5N1 viruses have been shown to replicate efficiently and cause severe pathogenicity [38]. The current 2009 pandemic H1N1 shows active replication and pathogenesis in mice. However, there appear to be differences in the pathogenicity and transmission of 2009 H1N1 isolates in different animal models [39],[40],[41]. In this study we have established a lethal mouse model for the 2009 pandemic H1N1 virus and demonstrate differences between the Cal/09 and Neth/09 viruses in terms of mouse virulence. Mice infected with Neth/09 showed an early and dramatic loss in body weight and succumbed to infection or reached the experimental end point by day 6 (Fig. 1B and C; Fig. 2A and B). Cal/09 and Neth/09 viruses were both isolated from humans showing mild upper respiratory symptoms and their genomes only differ at eight amino acid positions [39],[40]. The reasons for the differences in mouse virulence between these two isolates remains unclear. Despite these differences, antigenicity based on HI cross-reactivity between the two virus isolates is similar. This is in agreement with recent reports from the World Health Organization stating that all currently circulating 2009 H1N1 pandemic isolates are antigenically similar (www.who.int).
Although at this moment we can not exclude the possibility that the NA protein contributed at least in part to the protection seen with the inactivated vaccines, our results from the passive immunization and vaccination studies suggest that human H1N1 viruses from or prior to 1943, including the 1918 virus, share common protective epitope(s) in the HA protein. The antigenic sites of influenza virus HA's has been extensively characterized by monoclonal antibody epitope mapping studies [32],[42]. Interestingly, sequence alignment of the HAs from 1918, Sw/30, NJ/76 and 2009 H1N1 viruses showed a high degree of similarity in the known antigenic sites (Fig. 5, bottom). To further understand the antigenic relatedness between 1918 and 2009 H1N1 viruses, we compared the structure of the 1918 HA with the predicted HA structures of Cal/09 and Bris/59/07 (Swiss-Model; Fig. 6A). It is apparent that the globular head regions of HA's, that harbor the known antigenic sites, share close similarities between the Cal/09 and 1918 viruses (Fig. 6A and B). However, the antigenic sites in Bris/59/07 HA contain several amino acid changes as compared to the 1918 HA (evidenced from the higher number of red colored residues, Fig. 6B, top panel). This is supported by analysis of the pre-challenge sera from mice vaccinated with 1918 VLP that show high cross-reactivity against 2009 H1N1 virus (Table 2). Although the Wei/43 vaccine conferred complete protection from death, Wei/43 HA shows less sequence similarity to the 2009 H1N1 HA in the antigenic sites as compared to 1918 and the sera from Wei/43 vaccinated animals showed no cross-reactivity with 2009 H1N1 viruses by HI assay (Fig. 5; Table 2). In agreement with our findings, Wei/43 vaccinated animals showed more morbidity and higher viral titers after Neth/09 challenge than Cal/09 and 1918 vaccinated mice. Thus, Wei/43 appears to have a slightly decreased protective effect, suggesting that circulation of H1N1 viruses in humans for approximately 2 decades (after the 1918 pandemic) resulted in sufficient drift leading to considerable loss of antigenic cross-reactivity. One would therefore predict that H1N1 viruses that have circulated in the human population beyond the 1950's would offer little or no protection against the 2009 pandemic H1N1. Although, there is a correlation between sequence identity in the antigenic sites and protection, it is possible that some of the changes in the antigenic sites may not contribute to the overall alterations in antigenicity. In contrast to 1918 and classical swine H1N1 viruses, alignment of 2009 H1N1 with H1N1 HAs from viruses isolated in humans after 1977 showed significant variations in the antigenic sites especially in sites Sa and Sb (Fig. 6A, top). This is in agreement with our results showing that contemporary H1N1 virus vaccines only offered partial protection in mice against Neth/09 challenge and that surviving mice showed greater morbidity as evidenced by a more dramatic and sustained weight loss (Fig. 3D). Of interest, two of the more recent viruses, Tx91 and Hou/84, showed nearly 60% protection from death in mice to Neth/09 challenge. However, the reasons for this better partial protection offered by these two virus vaccines are unclear. Nevertheless, similar partial protection from lethal challenge has been previously described for antigenically different H1N1 viruses [43]. The mechanism for this partial protection is unclear and has been speculated to be largely due to anti-HA immunity [43]. Also, we cannot rule out completely the possible roles of T-cell mediated immunity or anti-NA-specific antibodies [44],[45],[46],[47],[48]. This warrants further investigation.
To characterize the antigenic similarity, we generated mAbs against the HA of 1918 and 2009 H1N1 viruses. Interestingly, two 1918 HA-specific mAbs, 6B9 and 39E4, and one Cal/09 HA specific mAb, 29E3, showed HI activity against both 1918 VLP and Cal/09 virus (Table 4). Passive immunization of mice with both 1918 HA-monoclonal antibodies afforded complete protection against lethal Neth/09 challenge and a 100–1000 fold reduction in viral titers in the lower respiratory track of mice. However, we did notice a loss in body weight starting on day 4–6 and an increase in lung titers on day 6, likely due to waning of neutralizing antibody levels in the sera. In the case of polyclonal sera treatment, the HI titers in the blood prior to challenge were 4-fold lower than monoclonal antibody treatment. In agreement with the lower HI titer, we observed an initial decrease in body weight up to 4 days following Neth/09 challenge. All these mice started recovering weight by day 5. Despite lower HI titers in the sera prior to challenge, the polyclonal sera treated mice had lung titers 1000 fold lower than control mice. This could be due to the action of NA specific antibodies which have been demonstrated to play a protective role [46], and to other HA specific antibodies without HI activity. The ability of 1918 HA specific antibodies to neutralize the 2009 pandemic H1N1 further confirms the close antigenic relationship of these viruses. Indeed, by generating antibody escape mutant viruses of Cal/09, we mapped antigenic site Sa as the cross-reactive epitope (Table 5 and Fig. 6B). Overall, this highlights the apparent slower pace of antigenic drift of the classical swine H1N1 viruses since their introduction into the pig population over 90 years ago, most likely due to the absence of pre-existing immunity in the majority of this population.
In 1976, an H1N1 virus closely related to classical swine H1N1 isolates from the 1930's caused an influenza outbreak at an American military facility, Fort Dix, in the state of New Jersey. After the 1976 swine H1N1 outbreak, nearly 40 million people in the United States were immunized with an NJ/76 vaccine. Our study in mice, in agreement with previous serology done with ferret sera [49], shows that immunization with NJ/76 vaccine protected from 2009 pandemic H1N1 lethal challenge, and therefore it is likely that people carrying antibodies against the NJ/76 will be protected against the 2009 H1N1. Additionally, sera from adults immunized with NJ/76 vaccine cross-neutralized the 2009 H1N1 [27].
Two recent human serology studies showed a high prevalence of neutralizing antibodies against 2009 pandemic H1N1 in people born before 1930 [28],[39]. Our data reveals that immunization with human H1N1 viruses that circulated before 1945 (e.g. specific antibodies against 1918 and Wei/43) is sufficient for immune protection from the 2009 pandemic H1N1. These findings provide a rationale for the epidemiological data arising from different parts of the world that have indicated a consistently higher rate of infection with this novel virus in the younger population (<35 of age), due to the lack of previous exposure to any of the antigenically-related viruses described above. As such, these results greatly emphasize that vaccination efforts and resources should also be directed at this susceptible population.
Our data are consistent with a picture in which domestic pigs have served as a reservoir for an “antigenically frozen” H1 hemagglutinin derived from the 1918 influenza virus. A significant drift in humans of that same 1918 H1N1 virus has resulted in a lack of significant pre-existing immunity against the 2009 pandemic H1N1 virus in humans born after the 1940–50s. Of concern, swine influenza viruses containing HA genes derived from more modern human H3 and H1 viruses have also established lineages in North American pigs during 1997–1998 and 2003–2005, respectively [13]. Also, human H3 viruses have established stable lineages in European swine populations after the 1968 “Hong Kong” pandemic, and are still circulating in the form of human-avian H3N2 reassortant viruses [50]. It is possible that these new swine viruses would also remain “antigenically frozen”, leading to potential human pandemic H3 and H1 viruses in the future [51],[52]. As such, surveillance and containment of swine influenza viruses is desirable for the prevention of future pandemic episodes.
|
10.1371/journal.pntd.0004401 | Yeast-Based High-Throughput Screens to Identify Novel Compounds Active against Brugia malayi | Lymphatic filariasis is caused by the parasitic worms Wuchereria bancrofti, Brugia malayi or B. timori, which are transmitted via the bites from infected mosquitoes. Once in the human body, the parasites develop into adult worms in the lymphatic vessels, causing severe damage and swelling of the affected tissues. According to the World Health Organization, over 1.2 billion people in 58 countries are at risk of contracting lymphatic filariasis. Very few drugs are available to treat patients infected with these parasites, and these have low efficacy against the adult stages of the worms, which can live for 7–15 years in the human body. The requirement for annual treatment increases the risk of drug-resistant worms emerging, making it imperative to develop new drugs against these devastating diseases.
We have developed a yeast-based, high-throughput screening system whereby essential yeast genes are replaced with their filarial or human counterparts. These strains are labeled with different fluorescent proteins to allow the simultaneous monitoring of strains with parasite or human genes in competition, and hence the identification of compounds that inhibit the parasite target without affecting its human ortholog. We constructed yeast strains expressing eight different Brugia malayi drug targets (as well as seven of their human counterparts), and performed medium-throughput drug screens for compounds that specifically inhibit the parasite enzymes. Using the Malaria Box collection (400 compounds), we identified nine filarial specific inhibitors and confirmed the antifilarial activity of five of these using in vitro assays against Brugia pahangi.
We were able to functionally complement yeast deletions with eight different Brugia malayi enzymes that represent potential drug targets. We demonstrated that our yeast-based screening platform is efficient in identifying compounds that can discriminate between human and filarial enzymes. Hence, we are confident that we can extend our efforts to the construction of strains with further filarial targets (in particular for those species that cannot be cultivated in the laboratory), and perform high-throughput drug screens to identify specific inhibitors of the parasite enzymes. By establishing synergistic collaborations with researchers working directly on different parasitic worms, we aim to aid antihelmintic drug development for both human and veterinary infections.
| We have developed and validated a yeast-based high-throughput screening assay for the identification of specific inhibitors of filarial targets. We engineered yeast strains to functionally express parasite and human enzymes, labeling these with fluorescent proteins and growing them in competition in the presence of test compounds. These strains express different target proteins from Brugia malayi (as well as their human orthologs) and our results demonstrate that it is possible to identify compounds that can discriminate between filarial and human enzymes. Accordingly, we are confident that we can extend our assay to novel targets from Brugia malayi and other worms of medical and veterinary importance, and perform high-throughput screens to identify new drugs against different parasitic worms.
| Lymphatic filariasis is a neglected tropical disease caused primarily by the parasitic nematodes Wuchereria bancrofti and Brugia malayi. The painful and disfiguring manifestations of this disease, also known as elephantiasis, can lead to permanent disability, causing an annual loss of approximately 5.5 million disability adjusted life years, affecting the poorest populations in Africa, Asia, and Latin America [1]. Current antifilarial therapies aim to eliminate filariasis through mass drug administration. However, in standard doses, the drugs used for this purpose (diethylcarbamazine, ivermectin and albendazole) are not effective against adult nematodes. As the adult worms can live in the human body for ca. 15 years [2], patients need to undergo multiple rounds of treatment, increasing not only the cost of therapy, but also the risk of drug-resistant worms emerging [3–5].
Filarial worms are difficult to cultivate in vitro, so adult worms for laboratory studies have to be obtained from animal models. Marcellino et al. [6] successfully developed a whole-plate, motion-based screen for monitoring drug activity against macroscopic parasites (WormAssay). This method was subsequently employed in screens against B. malayi [2], leading to the identification of the antifilarial activity of the FDA-approved drug auranofin. Unfortunately, there is no small animal model for other filarial worms, such as W. bancrofti [7] or Onchocerca volvulus; hence, there is a requirement for novel assays in the search for better treatments targeting filariasis cell-based assays also require extensive optimization). An alternative to parasite-based assays is to use in vitro drug screens based on protein targets. However, in vitro target-based assays require careful (and costly) optimization of the screening platform for each individual target protein to be tested, and provide no information on whether the drug is likely to be taken up by cells or whether it has general cytotoxicity.
To address these problems, we have developed and successfully validated a novel approach to high-throughput screens (HTS) for antiparasitic compounds using yeast [8,9]. Yeast cultures, which can be grown rapidly and at low cost, are ideal for use in automated screens [8–11]. Yeast cells are suitable hosts for the expression of nematode proteins [12–18], including enzymes essential for different life-cycle stages of the parasites, many of which cannot be propagated in vitro [17]. We engineered Saccharomyces cerevisiae strains to express either different parasite drug targets [9], or their equivalent human proteins, such that the growth of the yeast is dependent on the functioning of these heterologous proteins. We then transformed the engineered strains with plasmids expressing either CFP (cyan fluorescent protein), Venus (yellow fluorescent protein), Sapphire (blue fluorescent protein) or mCherry (red fluorescent protein), to enable their in vivo labeling.
Our engineered yeast strains are genetically identical, apart from expressing different heterologous drug targets and fluorescent labels that allow the growth of multiple strains to be followed in a single culture. These mixed cultures can be treated with chemical libraries to identify compounds capable of specifically inhibiting strains with the parasite targets but not their human counterparts. By these means, the drug sensitivity observed in a particular strain can be directly linked to the in vivo inhibition of the heterologous target protein. This approach has a number of significant advantages over conventional screens: it is very easy to set up for different drug targets; it is cheap, as the volumes used are very small and the yeast growth medium is inexpensive; it discriminates between compounds affecting parasite enzymes and human enzymes, and, by definition, active compounds must be able to enter living cells.
In this work, we evaluated the potential of such yeast-based drug screens in the identification of novel antifilarial compounds. We constructed yeast strains expressing different B. malayi target proteins, and used them to screen for novel inhibitors of filarial enzymes. We utilized a publicly available small-chemical library (400 Malarial Box compounds; http://www.mmv.org/malariabox) and identified compounds with significant inhibitory activity against the B. malayi enzymes, but little or no detectable activity against the equivalent human enzymes expressed in yeast. These first hit compounds were then validated in vitro against the closely related species, Brugia pahangi (continuously cultivated in our laboratory) with encouraging results, providing a proof of principle for this approach.
All animal protocols were carried out in accordance with the guidelines of the UK Home Office, under the Animal (Scientific Procedures) Act 1986, following approval by the University of Glasgow Ethical Review Panel. Experiments were performed under the authority of the UK Home Office, project numbers 60/4448 and 60/3792.
The filarial enzymes selected for testing are listed in Table 1. The coding regions of Bm1_22900 (BmNMT), Bm1_01925 (BmPGK), Bm1_29130 (BmTPI), Bm1_ 49000 (BmPIS), Bm1_48165 (BmSAH), Bm1_38705 (BmSEC53), Bm1_11585 (BmADE13) and Bm1_23075 (BmCDC8) were PCR-amplified from an adult Brugia malayi cDNA library, kindly donated by the Filariasis Research Reagent Resource Center (University of Georgia). These were cloned into the BamHI-PstI sites of pCM188 [19] to produce pCMBmNMT, pCMBmPGK, pCMBmTPI, pCMBmPIS, pCMBmSAH, pCMBmSEC53, pCMBmADE13 and pCMBmCDC8. These constructs placed the heterologous genes under the control of the TetO2 promoter. Synthetic genes encoding Brugia malayi Bm1_33465 (BmCDC21), Bm1_16500 (BmKRS), Bm1_42945 (BmMVD), Bm1_57600 (BmRKI) and Bm1_16300 (BmDYS) were synthesized by Geneart and sub-cloned into the BamHI-PstI sites of pCM188 [19] to produce pCMBmCDC21, pCMBmKRS, pCMBmMVD, pCMBmRK1, and pCMBmDYS.
The coding regions of human TPI1 (HsTPI), CDIPT (HsPIS), AHCYL1 (HsSAHa), AHCY/SAHH (HsSAHb), PMM2 (HsSEC53), PUR8 (HsADE13), CDC8/DTYMK (HsCDC8) and PPA1 (HsIPP1a) were PCR amplified from a cerebellum cDNA library, kindly donated by Dr. Nianshu Zhang (University of Cambridge). These were cloned into the BamHI-PstI sites of pCM188 [19] to produce pCMHsTPI, pCMHsPIS, pCMHsSAHa, pCMHsSAHb, pCMHsSEC53, pCMHsADE13, pCMHsCDC8 and pCMHsIPP1a. Synthetic genes encoding Homo sapiens TYMS (HsCDC21), LysRS (HsKRS), MVD (HsMVD), and RPIA (HsRKI) were synthesized by Geneart and sub-cloned into the BamHI-PstI sites of pCM188 [19] to produce pCMHsCDC21, pCMHsKRS, pCMHsMVD, and pCMHsRK1. Plasmid maps for new constructs developed in this work are available in Figs A to AF in S1 Text. Plasmids encoding human NMT2 and PGK1 (pCMHsNMT and pCMHsPGK) are described in Bilsland et al [9].
PDR5 encodes the major drug export pump of Saccharomyces cerevisiae, hence we deleted PDR5 in all of our yeast strains to increase their susceptibility to the test compounds. Deletion of the yeast PDR5 coding sequence was performed as described previously [20].
pCMBmNMT constructs were transformed into nmt1Δ::KanMX/NMT1 pdr5Δ::HisMX/PDR5 strains (BY4743 background [21]). pCMBmPGK constructs were transformed into pgk1Δ::KanMX/PGK1 pdr5Δ::HisMX/PDR5 strains (BY4743 background). The same approach (transformation of the heterologous construct into a yeast strain heterozygous for a deletion mutant of the orthologous yeast gene and heterozygous for pdr5) was employed for all subsequent constructs. Heterozygous strains harbouring the heterologous constructs were sporulated and derived haploids were selected for growth assays and drug screens. Strain genotypes are described in Table A in S1 Text.
Saccharomyces cerevisiae, Brugia malayi and Homo sapiens orthologues were selected based on data available at: http://inparanoid.sbc.su.se/cgi-bin/index.cgi. Multiple protein sequence alignments were performed using https://www.ebi.ac.uk/Tools/msa/, creating a pairwise identity matrix between each protein orthologue.
Yeast strains were grown in 2.5 mL YPD (1% yeast extract, 2% peptone, 2% glucose) cultures overnight at 30°C. Cultures were diluted 100 times in fresh YPD with 0, 5 or 10 mg/L doxycycline to tune-down the expression of the target protein from the TetO2 promoter. Clear 384-well plates (Corning) were then filled with 70 μL of each diluted culture (in triplicate). Plates were incubated in a BMG Optima plate reader and OD595 measures for each culture were acquired every 15 minutes. Growth scores were obtained by calculating the maximum exponential growth rate for each culture, and then multiplying this value by the yield of the culture (yield = maximum OD595—minimum OD595). Growth scores of the yeast strains dependent for growth on either the B. malayi or human coding sequence for the target enzyme were divided by the score for the corresponding wild-type strains (BY4741 or BY4742), to estimate their relative growth.
Yeast strains were transformed with fluorescent plasmids expressing one of the following fluorescent proteins mCherry (yEp_Cherry_LEU2), Venus (yEp_venus_LEU2), Sapphire (yEp_sapphire_LEU2) or CFP (yEp_CFP_LEU2) [8]. This allowed us to monitor, in real time, the growth of 2–4 different yeast strains growing in competition. Fluorescently labelled strains were grown in 2.5 mL YNB (1, 7 g/l yeast nitrogen base, 5 g/L ammonium sulphate, 2% glucose, and amino acid supplements) overnight at 30°C. Cultures were pooled and diluted 50 times in fresh YNB. An aliquot (35 μL) of the diluted pooled-culture was added to each well of black 384-well plates, together with 35 μL of YNB with 20 μM of each test compound (final 70 μL of 1:100 cultures with 10 μM of the test compound). Each experiment was run in quadruplicate, with two replicates using one combination of fluorescent markers, and two replicates using the alternate combination, in order to minimize false-positive results. Plates were incubated in a BMG Optima plate reader and fluorescence measured every 25 minutes, for 48 hours, in 4 different channels: Venus (excitation 500 nm/emission 540 nm), CFP (440 nm/490 nm), Sapphire (405 nm/510 nm) and mCherry (580 nm/612 nm). We confirmed each of our hits by generating dose:response curves with 0, 2, 10 and 50 μM of each hit compound.
Time-series fluorescence data from each channel was analyzed using a fitting procedure written in R, based on the model-free spline method of Kahm et al. [22]. The first 26 fluorescence measurements were used for analysis (corresponding to 26.66 hours of incubation); beyond this time, fluorescence decayed due to photo-bleaching. Specifically, a smoothed spline was fitted to the raw data for each channel and the first derivative through each point of the fit calculated. The highest derivative was taken as the maximum exponential growth rate (μ). This gradient was extrapolated back towards the start of the growth curve by fitting a straight line of the form y = mx + c. The lag phase duration (λ) was estimated by finding the intercept of this line and the baseline fluorescence from the start of the experiment. The fluorescence yield was calculated by subtracting the baseline fluorescence from the maximum value of the fit. These parameters are summarized in Fig AG in S1 Text.
Each compound was scored according to:
Score = rate × yieldlag
The average score for each compound was normalized to the average score for the DMSO control from the corresponding row in the plate. Compounds were scored for each pool (marker swap) independently. A compound was considered a hit when the relative growth score, in both marker combinations, of the recombinant strains expressing the parasite target protein was significantly lower than the growth score of the strain expressing the human enzyme.
Female worms of B. pahangi were recovered from infected jirds (the rodent, Meriones unguiculatus) approximately four months post-infection, exactly as described previously [23]. Worms were washed once in Hank’s Balanced Salt Solution and then cultured individually in 24-well tissue culture plates containing 2.0 ml of RMPI-1640 (Invitrogen, Cat No: 52400) supplemented with 5% heat-inactivated fetal bovine serum, 1% glucose and 100 units per mL of penicillin/streptomycin (all Invitrogen) for 24 hours to monitor viability and microfilarial output. Healthy worms were then selected for drug testing. Initially, each compound was tested in triplicate at concentrations of 50 μM and 10 μM. Control wells contained the appropriate concentration of the DMSO solvent. Parasites were also exposed to geldanamycin, a known Hsp90 inhibitor, at the same concentrations, as a positive control. All cultures were maintained at 37°C in 5% v/v CO2 in air and were monitored daily for viability. In a second experiment, adult worms were exposed to a wider concentration range of three selected drugs that had proved active in the first experiment. These were MMV396794, MMV665941 and MMV666022. Each drug was tested in triplicate at 25, 10, 5 and 2.5 μM.
We initially selected 15 enzymes (Table 1) to test as anti-filarial targets according to the following criteria: highly expressed in the adult stages of the parasite; likely to be essential for the viability of the nematode in the human host; and have an essential yeast ortholog. As drugs such as DEC and ivermectin have potent microfilaricidal activity, we chose to focus on adult parasites to control the stages not affected by current therapies. Furthermore, adult worms are the cause of much of the pathology associated with lymphatic filariasis [24].
We obtained Wuchereria bancrofti and Brugia malayi cDNA libraries, prepared from mRNA extracted from adult stages of the nematodes, from the Filariasis Research Reagent Resource Center (The University of Georgia). These cDNA libraries were used as templates for amplification by polymerase chain reaction (PCR), and cloning into yeast plasmids.
From the outset of our work, we noticed that the publicly available genome sequence of W. bancrofti contains multiple gaps, which frequently overlapped with our target genes. This made the design of primers for coding sequence (cds) amplification problematic. On the other hand, the genome sequence of B. malayi is very well annotated, facilitating the cloning procedure. Hence, we focused most of our cloning efforts on B. malayi targets. In spite of this, after multiple attempts and primer designs, we failed to amplify seven of the intended B. malayi targets (Table 2). This was probably due to either the absence of that particular cDNA in the libraries or to errors in the annotated sequence. The problem due to the low coverage of our cDNA library was overcome by synthesizing DNA encoding each of the selected Brugia targets to allow their expression in yeast. We identified two instances of errors in the published Brugia genome sequence by either sequencing our clones derived from cDNA (BmSAH), or by phenotyping strains expressing a synthetic cds based on the predicted sequence of the BmMVD gene product.
We observed that the synthetic human MVD could complement the yeast deletion very successfully (88% of wild-type growth), whereas the synthetic B.malayi MVD could not. Alignment of the publicly available B. malayi, Loa loa (eye worm), human, and yeast sequences, demonstrated that the B. malayi sequence diverged from those of the other three species, strongly suggesting either a non-conserved insertion in the Brugia malayi protein (which leads to the loss of function of the heterologous protein in yeast) or a problem with genome assembly of the published B. malayi sequence (Fig 1).
Similarly, sequencing of the cloned B. malayi adenosylhomocysteinase (Bm1_48165, BmSAH), demonstrated an insertion of 27 amino-acids (aa) between aa 88 and 89 of Bm1_48165. Eight independent BmSAH plasmids were constructed and sequenced and the same insertion was always detected. As the same insertion is present in both yeast and human SAH (Fig 2), it suggests that there may be an error in the B. malayi genome sequence, or that B. malayi encodes more than one splice variant of the enzyme. Furthermore, the encoded protein from our clone is enzymatically functional in yeast, complementing the deletion of the orthologous yeast gene.
We constructed seven additional strains where B. malayi gene products (Bm1_22900/BmNMT, Bm1_01925/BmPGK, Bm1_29130/BmTPI, Bm1_49000/BmPIS, Bm1_33465/BmCDC21, Bm1_57600/BmRKI, Bm1_16300/BmDYS) were able to complement the essential functions of the yeast orthologous gene. The heterologous genes were cloned into the yeast expression vectors under the regulation of the TetO2 promoter [25]. This promoter is constitutively on; however, by adding doxycycline to the growth medium, it is possible to tune-down the expression of the construct. Decreasing the expression of the target enzyme facilitates the growth inhibition by the test compounds (less target = more efficient inhibition). We have previously tested the inhibition of expression from the TetO2 promoter by addition of 1, 2, 5, 10, 20, 50 or 100 mg/L of doxycycline [9] and found that 5 and 10 mg/L would be ideal for the current work. Hence, we performed growth assays with each of the heterologous strains in medium with 0, 5 or 10 mg/L of antibiotic (Fig 3). However, we noticed that most of our strains require the full expression of the heterologous targets for optimum growth. Hence, drug screens employing the strains expressing B. malayi targets were performed in the absence of doxycycline.
In addition to BmMVD, four other B. malayi clones constructed in our studies failed to complement the essential functions of their cognate yeast genes; these were: Bm1_38705 (BmSEC53), Bm1_16500 (BmKRS1), Bm1_11585 (BmADE13) and Bm1_23075 (BmCDC8). With a view to establishing a sequence-similarity cut-off on which to base the selection of new targets, we calculated the percentage identity between yeast and filarial proteins that could successfully replace each other and those that did not. We found no clear correlation between sequence similarity and functional complementation; therefore, with the data collected so far, we cannot predict which filarial enzymes will be functional in yeast (Fig 4).
The screening method developed in our laboratory allows the simultaneous screening of up to three parasite targets and of the human counterpart in a single assay [8,9]. Hence, to make best use of the screening efforts, we included yeast strains expressing Schistosoma mansoni PGK [9], S. mansoni NMT [9], Homo sapiens NMT [9], H. sapiens PGK [9], H. sapiens TPI, H. sapiens PIS and H. sapiens SAHa and SAHb (S1 Text). These eight strains, as well as the five strains expressing the corresponding Brugia malayi targets (Table 2), were labeled by expression of fluorescent proteins and screened against each of the 400 compounds from the Malaria Box collection at a concentration of 10 μM. Comparing the growth of yeast strains expressing either parasite or human enzymes, we identified a number of compounds that specifically inhibited the growth of strains dependent on the parasite enzymes (Table 3). We then performed dose:response experiments with 0, 2, 10 or 50 μM of each hit compound to confirm our NMT, PGK and PIS hits, and confirmed the specificity of the compounds at the indicated concentrations (Table 3, last column).
We performed drug-sensitivity assays of our nine Brugia malayi hits using adult Brugia pahangi worms. Two of the compounds were not soluble in the test conditions, but four of the soluble compounds killed the worms at concentrations of 10 or 50 μM. Furthermore, the three remaining compounds compromised the motility of B. pahangi (Fig 5). Of the four compounds that had pronounced effects on adult worms at 24h incubation, two had immediate effects on adult viability. For MMV396794 and MMV665941, adult worms were dead within 3h of incubation at 50 μM concentration. These two compounds and MMV666022 were re-tested over a wider range of concentrations starting at 25 μM. The results were very similar, with worms incubated in MMV MMV396794 and MMV665941 at 25 μM dying within 3h of exposure. At 10 μM, all worms were dead in MMV665941 by 6h of exposure and at 5.0 and 2.5 μM were tightly coiled, although still motile at 6h, but dead by 24h. Compound MMV396794 killed all worms at 10 μM by 24h, with worms at lower concentrations being sluggish at 24h, although still alive. MMV666022 was the least active compound—although, after 48h, worms were dying at 25 and 10 μM, they were largely unaffected at 5 or 2.5 μM. Considering that the compounds used in our screens (Malaria Box) have low toxicity to human cells, this initial validation was extremely encouraging.
We previously developed and validated an efficient high-throughput screening method for the identification of compounds that selectively inhibit the activity of parasite proteins that are candidate drug targets, but have little effect on the orthologous human proteins [8,9,11]. Most of our initial screens were performed with targets from protozoan parasites (Plasmodium, Trypanosoma, Leishmania) that are evolutionarily very distant from the human host. In the present work, we evaluated the suitability of our approach for identifying inhibitors of enzymes from a metazoan animal, the nematode B. malayi.
Our assay is based on replacing essential yeast genes with their parasite or human counterparts, to allow the identification of compounds that inhibit the growth of yeast cells expressing the parasite target but not the growth of yeast strains expressing their human counterpart. This discrimination is very efficient for targets from parasites that are only distantly related to humans. Therefore, at the outset of our studies, we were unsure as to whether we would identify any compound capable of inhibiting metazoan targets such as Brugia, without also inhibiting the human ortholog.
Through the construction of yeast strains expressing the heterologous drug targets, we identified two instances where the publicly available B. malayi protein sequences are incorrect (Bm1_48165 and Bm1_42945), and have made a small contribution towards improving the annotation of the parasite genome. We also created tools for further studies of the functions of eight different Brugia enzymes.
It was gratifying to note that even working with a small library of compounds (400 Malaria Box drugs), it was possible to find compounds that could discriminate between human and metazoan pathogen enzymes. Most importantly, the compounds identified in our yeast-based screens can kill Brugia parasites in vitro, or inhibit their motility. Further studies will be necessary to follow up on these findings and titrate minimal effective concentrations of drug with activity against adult filarial worms. Interestingly, inhibition of filarial NMT proved to be very effective in our screen, confirming previous data of Galvin et al [27]. In that study, NMT was shown to be essential for viability in both B. malayi and in the free-living model nematode Caenorhabditis elegans. Hence, we are confident of the potential of our approach in pre-selecting novel compounds against metazoan parasites, and will be able to extend our screens to targets from other parasitic nematode species. Recent studies on the systematic humanization of yeast strains [28] will be particularly helpfully in suggesting proteins that can successfully complement essential yeast deletions.
In conclusion, our yeast-based screen offers many advantages over organism-based screens as a first pass for identifying compounds with activity against key filarial targets. Future studies will screen for inhibitors of additional essential targets, and extend the screens to larger chemical libraries in the search for much-needed novel nematicides.
Bm1_22900, Bm1_01925, Bm1_29130, Bm1_48165, Bm1_16955, Bm1_38705, Bm1_33465, Bm1_16500, Bm1_11585, Bm1_42945, Bm1_57600, Bm1_23075, Bm1_16300, Bm1_32340, Bm1_49000.
|
10.1371/journal.pbio.2001220 | Smek promotes corticogenesis through regulating Mbd3’s stability and Mbd3/NuRD complex recruitment to genes associated with neurogenesis | The fate of neural progenitor cells (NPCs) during corticogenesis is determined by a complex interplay of genetic or epigenetic components, but the underlying mechanism is incompletely understood. Here, we demonstrate that Suppressor of Mek null (Smek) interact with methyl-CpG–binding domain 3 (Mbd3) and the complex plays a critical role in self-renewal and neuronal differentiation of NPCs. We found that Smek promotes Mbd3 polyubiquitylation and degradation, blocking recruitment of the repressive Mbd3/nucleosome remodeling and deacetylase (NuRD) complex at the neurogenesis-associated gene loci, and, as a consequence, increasing acetyl histone H3 activity and cortical neurogenesis. Furthermore, overexpression of Mbd3 significantly blocked neuronal differentiation of NPCs, and Mbd3 depletion rescued neurogenesis defects seen in Smek1/2 knockout mice. These results reveal a novel molecular mechanism underlying Smek/Mbd3/NuRD axis-mediated control of NPCs’ self-renewal and neuronal differentiation during mammalian corticogenesis.
| Neural progenitor cells are self-renewing, multipotent cells that generate major neural cell types, including neurons and glia. Their fate during development of the cerebral cortex is determined by a complex interplay of genetic and epigenetic components. It has been shown that Suppressor of Mek null (Smek) proteins—which are evolutionarily conserved—play a role during the asymmetric cell division of neuroblasts in invertebrates. Methyl-CpG–binding domain 3 (Mbd3) protein, a core component of the repressive nucleosome remodeling and deacetylase (NuRD) complex, is an important epigenetic regulator that plays an essential role in mammalian development. In this study, we discovered that Smek interacts with Mbd3 and promotes its degradation via a posttranslational modification called polyubiquitylation. Degradation of Mb3, in turn, blocks recruitment of Mbd3/NuRD complex on target gene promoters, leading to an increase in neuronal differentiation during cortical development. This study not only elucidates a distinct mechanism for Smek-mediated neuronal differentiation but also identifies Smek as a negative regulator of the Mbd3 protein during cortical brain development.
| Neural stem cells (NSCs) are self-renewing, multipotent cells that generate major neural cell types, including neurons and glia, in the developing central nervous system (CNS) [1,2]. During neurogenesis, NSCs are derived from neuroepithelial cells (NECs), which first divide symmetrically to expand the population and then undergo a series of asymmetric cell divisions to produce neural progenitor cells (NPCs), lineage-restricted precursor cells (RPCs), and mature neural cells [3]. NSC fate determination is tightly regulated by intrinsic and extrinsic factors [4–6]. Recent findings suggest that neurodevelopmental and neurological anomalies, such as schizophrenia, autism, and depression, can emerge from abnormal specification, growth, and differentiation of NSCs [6–8].
Suppressor of Mek null (Smek), an evolutionarily conserved protein family, consists of two isoforms, Smek1 (PP4R3A) and Smek2 (PP4R3B), first reported as playing a role in the formation of a functional phosphatase group with PP4c, PP4R1, and PP4R2 complex [9]. Smek was initially identified in Dictyostelium discoideum as a playing a role in cell polarity, chemotaxis, and gene expression [10]. Smek also has several functions in lower eukaryotes, such as Caenorhabditis elegans, including roles in longevity by modulating DAF-16/FOXO3a transcriptional activity [11], DNA repair through dephosphorylation of phosphorylated H2AX (g-H2AX) during DNA replication [12], and glucose metabolism by controlling cAMP-response element binding protein (CREB)-regulated, transcriptional coactivator 2 (CRTC2)-dependent gene expression [13]. Notably, Smek also plays a critical role in cell-fate determination in higher eukaryotes. In Drosophila neuroblasts, PP4R3/Falafel (Flfl), which is an orthologous of Smek and is conserved throughout eukaryotic evolution, regulates asymmetric cell division by controlling localization of Miranda [14–16]. In mice, which express orthologous Smek 1 and 2, both Smek proteins suppress brachyury expression in embryonic stem cells (ESCs), and Smek1, especially, promotes NSC neuronal differentiation by negatively regulating Par3 [14–16]. Although we have shown that the Smek isoform Smek1 promotes NSC neuronal differentiation, signaling pathways required for that activity remain unclear [15].
Methyl-CpG–binding domain protein 3 (Mbd3), a core component of the repressive nucleosome remodeling and deacetylase (NuRD) complex, possesses a conserved methyl-CpG–binding domain (Mbd) [17,18]. Unlike other family members, which recognize 5′-methyl-cytosine (5′-mC)-modified DNA, Mbd3 specifically recognizes 5′-hydroxymethyl-cytosine (5′-hmC), an epigenetic marker highly enriched in NSCs [19,20]. Mbd3 plays an important role in brain development. Mbd3 expression is reported to be predominant in cortical NECs of the embryonic forebrain [21]. Mice lacking Mbd3 die in utero before neurogenesis is completed [22]. Conditional knockout of Mbd3 in neural progenitor cells leads to defects of differentiation of appropriate cell types during neurogenesis [23]. Despite emerging evidence that Mbd3 has a critical function in the CNS, little is known about its regulatory mechanism in NSCs.
To understand Smek protein function during mammalian CNS neurogenesis, we screened for novel Smek-binding proteins that regulate NPC neuronal differentiation and identified Mbd3, a potent epigenetic regulator, as a Smek-interacting protein. We found that Mbd3 is highly expressed in NPC populations in the ventricular zone, and it was predominantly expressed in the nucleus. Smek interacted directly with the Mbd3’s Mbd domain, destabilizing Mbd3 protein and its interaction with NuRD components, and sequentially, preventing accumulation of the Mbd3/NuRD complex on target gene loci functioning in neurogenesis. Such dissociation of Mbd3/NuRD complex promotes NPC neuronal differentiation. Moreover, overexpression of Mbd3 significantly inhibited neuronal differentiation of wild-type NPCs, while Mbd3 depletion rescued neurogenesis defects seen in Smek knockout mice.
This work identifies a novel pathway of Smek and Mbd3/NuRD complex in brain development and could encourage discovery of novel epigenetic regulators governing neuronal differentiation.
Recently, we reported that Smek1 promotes neurogenesis during mouse cortical development [15]. To further characterize Smek function in neurogenesis, we generated Smek1 and Smek2 double knockout (dKO) mice and set out to analyze cortical development in Smek1 knockout (KO) and Smek1 and Smek2 dKO embryo brains (S1A and S1B Fig). To do so, we undertook immunohistochemical analysis of the embryonic cortex derived from WT and Smek1/2 dKO mice and observed a decrease in the number of cells positive for Tuj1, an early neuronal marker (~15% and ~25% fewer Tuj1+ cells at E12.5 and E14.5, respectively) (Fig 1A). We observed similar decreases in postmitotic cortical neuron marker Tbr1-positive cells (~15% and ~20% reductions at E12.5 and E14.5, respectively) and Tuj1 and Tbr1 double-positive cells (near 20% reduction in each stage) (Fig 1A). The number of mature microtubule-associated protein 2 (MAP2)-positive neurons also significantly decreased by ~12% in the E12.5 cortex (Fig 1A, right and S1C Fig). In Smek1/2 mice, the number of Pax6-positive NPCs increased significantly (by ~23% at both E12.5 and E14.5), as did Nestin/Ki67 double-positive cells (~24% increase at E12.5) compared with wild-type (WT) mice (Fig 1B and S1D–S1F Fig). As expected, neurogenesis defects in Smek1/2 dKO embryonic brains were greater than those seen in Smek1 KO mice (S2A and S2B Fig, S1 Table). These results demonstrated that in the Smek1/2 dKO mice, the number of neurons is reduced while the number of neural stem cells is increased.
To assess direct effects of Smek loss on NPC differentiation capacity, we cultured Smek1/2 dKO NPCs derived from E11.5 mouse embryo brains in the presence of basic fibroblast growth factor (bFGF) and then withdrew bFGF to induce neural differentiation. Consistent with in vivo results, the number of Tuj1-positive cells decreased in Smek1/2 dKO cultures while the number of Nestin positive cells increased slightly in Smek1/2 dKO cultures over the course of differentiation (Fig 1C and S3 Fig). In addition, we assessed the role of Smek in maintenance of self-renewal activity using the single-cell clonal neural sphere formation assay. Smek1/2 dKO NPCs showed higher sphere-forming ability than those derived from WT NPCs in both primary and secondary sphere-forming assays (Fig 1D). Further analysis using quantitative PCR (qPCR) showed that Smek1/2 dKO cells exhibited decreased expression of Dlx1, Dlx2, Tuj1, Gad67, NeuN, and NeuroD1, as well as other neural differentiation genes such as Gfap and Mbp (the latter an oligodendrocyte marker), and increased expression of Nestin (Fig 1E and S3 Fig). Severe differentiation defects seen in cortical NPCs lacking both Smek 1 and 2 suggest that these factors compensate for each other during cortical development. All these experiments demonstrated that Smek plays a role in self-renewal and neural differentiation of NPCs in vivo and in vitro.
In order to determine a detailed molecular mechanism modulating the differentiation of NPCs by Smek, we sought to identify proteins interacting with Smek protein. To identify how Smek mediates neuronal differentiation of NPCs, a yeast two-hybrid (Y2H) screening assay was performed and revealed that Smek interacts with full-length Mbd3 (Fig 2A and S2 Table). An immunoprecipitation (IP) assay confirmed the interaction of Smek1 and Smek2 with Mbd3 in 293T cells (Fig 2B, S4A and S4B Fig). As shown in Fig 2B and S4B Fig, Smek1 or Smek2 coimmunoprecipitate with Mbd3, while no signals were detected in cells transfected with a negative control vector. Furthermore, colocalization of endogenous Mbd3 and Smek2 protein was also observed in the nucleus of in vitro cultured NPCs (Fig 2C). Both Smek1 and Mbd3 proteins were expressed in NPCs of the ventricular zone (VZ) and subventricular zone (SVZ) in vivo (Fig 2D, S4C and S4D Fig), and merged images reveal that Smek and Mbd3 staining was nuclear (Fig 2C and 2D). Mbd3 is highly enriched in the nucleus of the VZ progenitor cells and its expression in the nucleus showed a gradually decreasing pattern in the direction of the intermediate zone (IZ) and cortical plate (CP). Smek1 is expressed in the nuclear or perinuclear region of cortical progenitor cells in the VZ, but its expression and nuclear localization is significantly increased in the differentiated neurons in the IZ and CP (Fig 2D, S4C and S4D Fig). The interaction of endogenous Smek and Mbd3 were further confirmed by co-IP of Smek and Mbd3 using NPC lysates. Smek/Mbd3 interaction, however, was apparently disrupted during neuronal differentiation of wild-type NPCs, suggesting that protein complexes may function in NPC differentiation (Fig 2E and S4C Fig).
To map Smek and Mbd3 domain(s) required for interaction, we generated Smek1 and Mbd3 mutants and assessed their interaction (Fig 2F–2H). A Smek N-terminal deletion mutant (ΔRanBD) did not interact with Mbd3 protein, shown by co-IP of Smek and Mbd3 in HEK293T cells, whereas interaction of other Smek deletion mutants with Mbd3 was comparable to the full-length protein, suggesting the RanBD domain of Smek mediates Smek’s interaction with Mbd3 (Fig 2F). To map to the domains on Mbd3 required for Smek interaction, we generated glutathione-S-transferase (GST) fusion Mbd3 mutant proteins in bacteria and incubated these proteins with HEK293T cell lysates expressing Smek2. GST pull-down assays revealed that the Smek-Mbd3 interaction was disrupted in Mbd3 deletion mutants lacking the first 92 N-terminal amino acids (ΔN92), a region encompassing the Mbd domain (Fig 2G, upper panel). This finding was confirmed by IP experiments in HEK293T cells transfected with full-length or ΔN92 forms of Mbd3 and Smek (Fig 2H). Those results indicated that the Mbd domain of Mbd3 is required for Smek interaction.
To characterize Smek and Mbd3 expression during neural differentiation, we examined protein and mRNA levels after withdrawal of bFGF from NPC culture media. In NPCs, Mbd3 protein levels gradually decreased by day 1 of differentiation, while Mbd3 protein levels did not decrease in Smek1 and Smek2 single KO or dKO NPCs (Fig 3A–3C and S5A, S5B Fig). However, Mbd3 mRNA levels did not change during differentiation in both WT and Smek1, Smek2, or Smek1/2 dKO NPCs, suggesting that changes in Mbd3 protein levels that occur during differentiation require Smek (Fig 3D and 3E and S5A Fig, lower panel). These observations led us to further analyze Mbd3 stability. We found that Mbd3’s half-life was ~6 h in NPCs and 293T cells in which new protein synthesis was inhibited and significantly prolonged in the presence of the proteasomal inhibitor MG132 (Fig 3F). Endogenous Mbd3 protein levels in NPCs were increased by MG132 treatment (Fig 3G). Overexpressed Mbd3 protein was polyubiquitylated, and polyubiquitylated proteins accumulated in cells treated with MG132 or MG-101, respectively (Fig 3H and 3I and S5C, S5D Fig). Furthermore, endogenous Mbd3 in NPCs was found to be polyubiquitylated as well (Fig 3J).
To determine whether Smek regulates Mbd3 protein stability, we monitored endogenous Mbd3 protein levels in wild-type NPCs, in HEK293T cell lines stably overexpressing Smek1 or Smek2 (S5E and S5F Fig), or in NPCs derived from wild-type and Smek1/2 dKO embryonic mouse brains (Fig 4A and 4B). Mbd3 protein turnover rate was increased by overexpression of either Smek1 or Smek2 and decreased upon Smek loss (Fig 4A and 4B). Consistent with Mbd3 degradation, Smek1 or Smek2 overexpression in HEK293T cells significantly promoted Mbd3 polyubiquitylation (Fig 4C and S6A Fig). Furthermore, Mbd3 was ubiquitylated in wild-type NPCs but not in Smek1/2 dKO NPCs (Fig 4D). To further assess the effects of Smek expression on Mbd3 degradation, we examined Mbd3 ubiquitylation following expression of the Mbd3 ΔN92 mutant, which cannot interact with Smek. Polyubiquitylation of this mutant was not significantly changed by overexpression of either Smek1 or Smek2 (Fig 4E and S6B Fig). These results suggest that interaction with Smek destabilizes Mbd3.
Smek has a nuclear localization signal (NLS) and is nuclear localized [15], suggesting a potential role in regulating transcription. To determine whether the Smek protein is associated with chromatin—and if so, whether it is enriched on chromatin loci of neurogenesis-associated genes—we performed chromatin immunoprecipitation sequencing (ChIP-seq) in NPCs using a Smek1 antibody. Genome-wide binding profiling demonstrated that Smek proteins bind to chromatin loci of genes related to organ morphogenesis, cell-fate determination, and CNS development and differentiation (Fig 5A and 5B and S7A, S7B Fig). Furthermore, Smek specifically bound proximal promoter regions and gene bodies of neuronal genes such as Dlx1, Dlx2, Tlx3, NeuroD1, Ascl1, and Lbx1, which were known to highly express in neuron or proneuronal cells (Fig 5A–5C and S7C, S7D Fig). Unlike Smek, which lacks a known DNA-binding motif, Mbd3 exhibits the Mbd domain, which reportedly binds 5′-hydroxymethyl cytosine (5′-hmC) regions [20]. We therefore asked whether Smek and Mbd3 share similar genomic regions in NPCs, initially by determining whether Mbd3 binds neuronal gene promoters that Smek binds to. To do so, we undertook ChIP-qPCR with a Mbd3 antibody in wild-type and Smek1/2 dKO NPCs cultured in undifferentiation or differentiation conditions. This analysis confirmed enrichment of Mbd3 on the Smek-bound loci of genes including Dlx1, Dlx2, Tlx3, NeuroD1, Ascl1, and Lbx1 in undifferentiated conditions (Fig 5D and S7C Fig). Moreover, Mbd3 enrichment on these gene loci significantly decreased under differentiation conditions in wild-type NPCs but was unchanged in Smek1/2 dKO NPCs under the same conditions (Fig 5D and S7C Fig). Then we asked whether enrichment of the Smek1 protein on chromatin loci of neurogenesis-associated genes is dependent on Mbd3 protein, and we performed ChIP-qPCR with Smek1 and Mbd3 antibodies in shScramble or shMbd3 knockdown (KD) NPCs cultured in undifferentiation or differentiation conditions. These results demonstrated that occupancy of Smek1 on the promoters of Dlx1, Dlx1as, Tlx3, NeuroD1, Ascl1, and Lbx1 genes, but not Gfap genes, is dependent on Mbd3 protein (Fig 5E and S7F Fig). Mbd3 has been reported to represses transcription by recruiting the NuRD complex to target gene loci [20]. Co-IP experiments showed that Mbd3 interacted with MTA1, RbAP46, HDAC1, and HDAC2, which are components of NuRD complex (S7G Fig). We then undertook ChIP-qPCR with HDAC1, HDAC2, MTA1, and acetyl histone H3 antibodies using wild-type and Smek1/2 dKO NPCs cultured in differentiation conditions. ChIP analysis revealed that enrichments of NuRD components HDAC1, HDAC2, and MTA1 to target gene loci were significantly increased in Smek1/2 dKO NPCs when compared with those of the wild type (Fig 5F). Inversely, the amount of acetyl histone H3 was decreased (Fig 5F). These findings suggest that Smek inhibits enrichment of Mbd3/NuRD complex to neurogenesis-associated gene loci and increases acetyl histone H3 activity for their gene transcription during neurogenesis. These data also suggest that NuRD activity is dependent on Mbd3 ability, which can bind to target DNA, and Smek, as an upstream regulator of Mbd3/NuRD complex, promotes Mbd3 degradation, potentially allowing transcription of neuronal differentiation–associated genes by disrupting association and enrichment of Mbd3/NuRD complex on target gene loci.
Our findings suggest that Mbd3 regulates expression of neuronal target genes in NPCs, an activity modulated by Smek. To assess whether Mbd3 represses neurogenesis-associated target genes, we overexpressed full-length or mutant (ΔN92) Mbd3 in NPCs and then induced differentiation over 2 d. Full-length Mbd3 (but not mutant form) overexpression attenuated Dlx1, Tlx3, NeuroD1, Tuj1, Gad67, and NeuN gene expression, all neuronal lineage markers, but had no effect on glial cell differentiation or gene expression (Fig 6A and S8A–S8D Fig). As noted above, Mbd3 bound specifically to the Dlx1, Tlx3, NeuroD1, Ascl1, and Lbx1 gene loci, and this association decreased upon induction of differentiation conditions (Fig 5D). Thus, we asked whether decreased neuronal gene expression following Mbd3 overexpression paralleled increased occupancy of Mbd3 on target promoters (Fig 6B). As expected, the amount of overexpressed Mbd3 bound to gene promoters in NPCs in differentiation conditions over 2 d was similar to that seen in nondifferentiation conditions, leading to attenuate Dlx1, Tlx3, NeuroD1, Tuj1, Gad67, and NeuN gene expression, all markers of neuronal lineage (Fig 6A and S8B Fig). In contrast, there was little or no accumulation of Mbd3 on the Gfap promoter under the differentiation condition over 2 d for NPCs after Mbd3 overexpression, suggesting that Mbd3 blocks neuronal rather than glial cell differentiation (Fig 6B and S8B Fig). Immunocytochemistry analysis also showed that Mbd3 overexpression prevented NPC neuronal differentiation but did not affect astrocyte differentiation (Fig 6C, 6D and S8C, S8D Fig). Moreover, in both control and Mbd3-overexpressing NPCs, Nestin staining was comparable in staining intensity (S8C Fig). These data suggest that Mbd3 is a novel regulator for neuronal cell-fate determination of NPCs.
To further investigate whether the Smek-Mbd3 axis regulates neurogenesis, we knocked down endogenous Mbd3 using an shMbd3 lentiviral vector in cultured Smek1/2 dKO NPCs and then induced differentiation for 2 d. Mbd3 KD significantly rescued effects of Smek loss on Dlx1, Tlx3, NeuroD1, Tuj1, Gad67, and NeuN expression but had no effect on astrocyte differentiation or gene expression (Fig 7A). Increased neuronal gene expression seen following Mbd3 KD was accompanied by decreased occupancy of target gene promoters by Mbd3 (Fig 7B). The epistatic relationship of Mbd3 and Smek in neurogenesis was analyzed by Mbd3 knockdown in Smek1/2 dKO NPCs. Mbd3 shRNA were expressed from the same vector that coexpressed enhanced green fluorescent protein (EGFP). The percentage of EGFP and Tuj1 double-positive cells among EGFP-positive cells was increased in cultures expressing Mbd3 shRNA but not in cells expressing Scramble shRNA (Fig 7C). We next assessed Mbd3 function in neurogenesis using an in utero electroporation system. Electroporated embryos were readily identifiable by EGFP expression (Fig 8A). About 74% of total EGFP-positive Mbd3 knockdown cells migrated toward the IZ or CP, while only ~39% of EGFP-positive control cells showed a similar migration pattern (Fig 8B and 8C). Quantitative analyses showed that the number of Tuj1-positive cells significantly increased in the VZ, SVZ, and IZ regions in Mbd3 KD EGFP-positive cells relative to control EGFP-positive cells (Fig 8D). These results strongly suggest that Mbd3 regulates NPC neuronal differentiation in the VZ or SVZ during cortical development.
Here, we have analyzed mouse embryos lacking functional Smek1 and Smek2 genes as well as cultured NPCs derived from those animals to understand Smek function during cortical development. We discovered that Smek1/2 dKO NPCs exhibit significantly reduced capacity for neuronal differentiation and increased self-renewal activity. Furthermore, we employed a Y2H screen to search for Smek binding partners and identified Mbd3 as a novel Smek-interacting protein (Fig 2A and S2 Table). Importantly, we observed that Smek promotes Mbd3 protein degradation and reduces Mbd3 occupancy of neural differentiation–associated gene promoters, likely increasing transcription of those genes via inhibiting recruitment of the repressive NuRD complex.
Interestingly, in the developing CNS, increased Mbd3 instability had an effect only on neuronal differentiation, with little or no effect on glial cell fate (Figs 6, 7 and S7F and S8B–S8D Figs). We could not determine the molecular mechanism by which the Smek-Mbd3 axis specifically regulates neuronal cell-fate determination but not glial cell fate. In our previous study, we found that protein phosphatase PP4c interacts with Smek and this complex suppressed Par3 activity for differentiation of NPCs [15]. PP4c is known to regulate neuronal cell-fate determination and organization of early cortical progenitors in the ventricular zone of the embryo brain by modulating spindle orientation during mitosis [24], and we could confirm the role of PP4c in neuronal differentiation of NPC by a PP4c loss-of-function study (S9A Fig). Interestingly, knockdown of PP4c significantly abolished neuronal cell as well as glial cell differentiation of NPCs, similar to Smek loss of function, and this finding suggests that Smek/PP4c/Par3 might have a different biological function from Smek/Mbd3 in at least regulating glial cell gene expression of NPCs (Fig 1E and S9A Fig). Moreover, Par3 regulation of Smek/PP4c during neurogenesis exclusively occurs in a cytosolic fraction but not in the nucleus of NPCs [15]. However, Smek and Mbd3 expression and transcriptional repression of Mbd3/NuRD complex mainly occurs in the nucleus of NPCs (Fig 2C and 2D). Our preliminary investigation of the relationship between Smek-PP4c complex and Mbd3 protein stability also reveals that loss of PP4c could not affect Smek-mediated Mbd3 polyubiquitylation (S9B Fig). In addition, ChIP-seq and ChIP-qPCR data show that Smek and Mbd3 are not significantly enriched at Gfap gene loci (S7F Fig). Thus, overall data suggest that the Smek-Mbd3 axis likely functions independently of the Smek-PP4c-Par3 axis, at least in regulation of Gfap gene expression, and that the Smek-Mbd3 interaction plays a crucial role in neuronal cell-fate determination in NPCs.
So far, five vertebrate MBD proteins have been identified as members of the MBD protein family: Mbd1, Mbd2, Mbd3, Mbd4, and MECP2 [25, 26], and these members are more highly expressed in the brain than in other tissues, leading investigators to hypothesize that they may play a critical role in normal brain development and in behavior [27, 28]. Our data indicate that Mbd3 represses neurogenesis and likely functions differently from other family members. For example, Mbd2 and Mbd3 are closely related and share a highly conserved methyl-CpG–binding domain, but mouse studies indicate that the two proteins are not functionally redundant [24], possibly because Mbd3 specifically recognizes methylated DNA, especially, 5′-hydroxymethylcytosine (5′-hmC) [19]. Deletion of Mbd3 gene in neural progenitor cells leads to generation of neurons expressing both deep- and upper-layer markers [23], suggesting that Mbd3 is required to maintain appropriate transcription in progenitor and neurons during neural development. A recent study suggests that Mbd3 may fine-tune expression of both active and silent genes [29]. Other studies suggest that conversion of 5′-mC to 5′-hmC coincides with increased transcriptional activity by excluding Mbd proteins from target genes [30]. Consistent with these findings, we found that Mbd3 is specifically bound to neuronal gene loci, and our findings suggest that it is likely released from these loci by Smek during NPC differentiation (Fig 5D).
Mbd3 is a subunit of the NuRD complex, which has nucleosome remodeling and histone deacetylase activities [17,18] and thus regulates gene expression. The molecular function of this complex has been extensively studied in the context of tumorigenesis, stem cell pluripotency, and brain development [23, 31–34]. Mbd3 mutation or abnormal expression may function in tumorigenesis by perturbing gene expression. Like Mbd3, other Mbd proteins, especially Mbd2 and Mbd4, are associated with progression of cancer such as colorectal cancer, albeit by different mechanisms [34–36]. Furthermore, Mbd3 knockdown during somatic cell reprogramming significantly increases reprogramming efficiency [31–33]. Although Mbd3 activity is likely relevant to pathologies seen in cancer, neurological disease, and developmental defects, mechanisms underlying its regulation remain unclear. We propose a novel function in which Mbd3 protein levels, depending on Smek activity, decrease during neurogenesis (Figs 2D, 3A–3C and S4C and S4D, S5A and S5B Figs). Smek1/2 dKO NPCs or the embryonic cortex show aberrantly high Mbd3 levels that may repress neuronal gene expression and underlie developmental defects seen in the latter. In accordance, we report that Smek promotes ubiquitylation and degradation of Mbd3 (Figs 3A–3C, 4A–4D and S5A–S5B, S6A Figs). Our data also indicate that Smek regulation of Mbd3 is not transcriptional, based on the lack of significant change in Mbd3 mRNA levels over NPC differentiation. Conversely, Mbd3 protein levels decreased during neuronal differentiation in the embryonic cortex starting at E12.5 in mice. In addition, decreased Mbd3 levels seen in cultured NPCs are blocked by MG132 treatment concomitant with accumulation of polyubiquitylated Mbd3. These results overall indicate that Mbd3 activity is regulated at the level of protein stability and that Smek likely governs this process. Changes in protein stability often constitute a more rapid means of regulating protein activity than does modulation of transcription. Therefore, regulation of Mbd3 protein stability might function epigenetically to recruit the NuRD complex to 5′-hmC–modified gene promoters. To our knowledge, this is the first report of regulation of an Mbd family protein by stability changes.
We also examined potential factors or complexes that might function in Smek-dependent Mbd3 degradation. To do so, we sought potential E3 ligase proteins that might catalyze Mbd3 ubiquitylation by using Biograph software and identified the E3 ligase TRIpartite Motif protein 33 (TRIM33) protein, which has an N-terminal Really Interesting New Gene (RING)-domain (S10 Fig). TRIM33 specifically targets phosphorylated nuclear proteins for degradation [37]. Interestingly, we have previously identified protein kinase C (PKC) lambda/iota (λ/ι), a serine/threonine kinase, as a binding partner of Smek1 from a mass spectrometry analysis [15]. PKC isoforms contain an NLS and contribute diverse cellular physiology [38–40]. Smek1/2 also have NLS sequences and are localized exclusively in the nucleus in interphase [15]. To further investigate the involvement of PKCλ/ι in the molecular mechanism for Mbd3 protein stability, we performed prediction of putative kinases for phosphorylation of Mbd3 protein by GPS (group-based prediction system) software 3.0 (http://gps.biocuckoo.org/). (S3 Table). Interestingly, we predicted PKCλ/ι as putative kinases for Mbd3 phosphorylation. Although it still remains unclear how Smek1/2 promotes ubiquitylation and stability of Mbd3, accumulating data and predictions suggest that nuclear-localized Smek1/2-PKCλ/ι complex with TRIM33 may function in Mbd3 ubiquitylation and degradation. Alternatively, Aurora-A protein, a serine/threonine kinase, reportedly physically associates with Mbd3 at centrosomes in early M phase in vivo and phosphorylates Mbd3 protein in vitro [41]. These findings suggest that Aurora-A may also be involved in the regulation of Mbd3 protein stability as a different mechanism from Smek-PKCλ/ι complex. This topic will be addressed in future studies.
Smek orthologues in Drosophila play a critical role in neuroblast mitosis [16]. In Smek-deficit neuroblasts, cell-fate determinants, such as Prospero and Miranda, are no longer localized to the cell cortex; instead, they are distributed in the cytoplasm of dividing neuroblasts [16]. As a result, asymmetric cell division and neurogenesis are defective. In the Drosophila system, another class of asymmetric cell division regulators are epigenetic modulators. However, it is not clear if Smek functions through epigenetic modulators. Our studies suggest that Smek and Mbd3 have the opposite function in the NPCs’ differentiation in vertebrate systems. Although these studies do not address asymmetric cell division, our research may shed light on asymmetric cell division and neurogenesis in Drosophila and mammals.
Our findings also highlight the importance of Smek/Mbd3 interaction in regulating NPC differentiation. Other studies suggest that Smek and Mbd3 may have overlapping functions or activities in brain development, stem cell activity, and regulation of transcription [15,16,20,21,42]. Our studies support a functional relationship of Smek and Mbd3 in NPCs. Our mapping analysis shows that Smek/Mbd3 interaction is mediated by the Mbd domain of Mbd3. Immunofluorescence analysis confirmed close proximity of these proteins in the NPC nucleus, and we have observed coincident expression of Smek and Mbd3 in the mouse embryonic brain [16,21]. Finally, we found that Smek and Mbd3 target the same neuronal gene loci for regulating transcription (Fig 5D). Further analysis suggests a model in which Smek regulates target gene transcription by regulating Mbd3 protein stability, interaction with NuRD components, and recruitment of Mbd3/NuRD complex to the promoters of target genes. Smek1/2 dKO exhibits reduced neuronal differentiation and decreased expression of Dlx1, Dlx2, NeuroD1, Tuj1, Gad67, NeuN, and stabilizing Mbd3 protein, while Mbd3 overexpression attenuated Smek-mediated neuronal differentiation (Figs 1E, 3A–3C, 6A, 6C and 6D). Thus, this study is significant not only for demonstrating Smek-mediated Mbd3 protein degradation but also in providing evidence that Smek/Mbd3 interaction regulates neuronal gene expression and neuronal differentiation during cortical development. In conclusion, we report functional interaction of Smek with Mbd3 in neuronal differentiation of NPCs.
All animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) and the National Institutes of Health (IACUC protocol number: 11489). Mouse embryos and primary neural progenitor cells were obtained from a deceased pregnant mouse following CO2 asphyxiation. For in utero electroporation experiments, timed-mated pregnant mice had been anesthetized with Avertin (2.5%) (Sigma, St. Louis, MO) following IACUC instruction. At the experimental endpoints, mice were euthanized by CO2 asphyxiation.
Smek1/2 dKO mice were generated using gene trap mutant ES cells obtained from the Gene Trapping Consortium. Gene trap vectors were targeted between exons 3 and 4 of the Smek1 gene and between exons 10 and 11 of the Smek2 gene, respectively. Smek1 and Smek2 mutant ES cells (E14) were injected into mouse blastocysts and chimeric mice were backcrossed with C57BL/6 mice. Smek1/2 dKO mice were generated by crossing C57BL6J-Smek1+/- with Smek2+/- mice. After six generations, mice were used for analysis. Although Smek1/2 dKO mice can die at later stages of embryonic development, we were able to obtain dKO embryos as late as E14.5 with a normal Mendelian distribution. Thus, we have conducted functional analysis of Smek1/2 dKO embryos at E11.5, E12.5, and E14.5. Embryos and pups of wild-type and heterozygous KO mice were collected from timed-mated pregnant females.
Antibodies used in this study were anti-Smek1 (rabbit polyclonal 1:500 dilution) anti-Smek2 (rabbit polyclonal 1:500 dilution), anti-Flag (mouse monoclonal 1:2,000 dilution) (Sigma), anti-Mbd3 (rabbit polyclonal 1:500 dilution), anti-CDH3 (rabbit polyclonal 1:500 dilution), anti-RbAP46 (rabbit polyclonal 1:2,000), anti-GFAP (rabbit polyclonal 1:200 dilution), anti-MTA1 (rabbit polyclonal 1:2,000 dilution), anti-RbAp46 (rabbit polyclonal 1:2,000 dilution), anti-HDAC1/HDAC2 (mouse monoclonal 1:2,000 dilution) (Cell Signaling Technology, Beverly, MA), anti-HA (rabbit polyclonal 1:500 dilution), anti-GST (rabbit polyclonal 1:500 dilution), anti–α-tubulin (mouse monoclonal 1:5,000 dilution), anti-HA (mouse monoclonal 1:3,000 dilution) (Santa Cruz Biotechnology, Santa Cruz, CA), anti-MAP2ab (rabbit polyclonal 1:200 dilution) (Chemicon, Temecula, CA), anti-NeuN (rabbit polyclonal 1:200 dilution) (EMD Millipore, Billerica MA), anti-Nestin (mouse monoclonal 1:350 dilution) (BD Biosciences, San Jose, CA), anti-Tuj1 (mouse polyclonal 1:200 dilution) (Covance, Princeton, NJ), anti-Tbr1 (rabbit polyclonal 1:200 dilution), and anti-Pax6 (rabbit polyclonal 1:200 dilution) (Abcam Ltd, Cambridge, MA). Secondary antibodies were anti-rabbit Alexa Fluor 488-, anti-mouse Alexa Fluor 488-, anti-rabbit Alexa Fluor 555-, or anti-mouse Alexa Fluor 555-conjugated IgG (1:200 dilution) (Molecular Probes, Eugene, OR). bFGF was purchased from PeproTech (Rocky Hill, NJ). The protease inhibitor cocktail was from Roche Applied Science (Indianapolis, IN). To HDAC inhibition, Trichostatin A (TSA) and 4-(dimethylamino)-N-[6-(hydroxyamino)-6-oxohexyl]-benzamide (DHOB) were purchased from Santa Cruz Biotechnology. TRIzol, Protein A/G agarose beads, and DAPI were from Sigma. The ECL Kit and KOD Hot Start DNA polymerase were from EMD Millipore. Glutathione magnetic beads, phenol:chloroform:isoamyl alcohol, and the First Strand cDNA Synthesis Kit were from Thermo Fisher Scientific (Rockford, IL).
Cells were gently lysed with IP buffer (50 mM Tris-HCl, pH 7.4, 130 mM NaCl, 10mM NaF, 2 mM EGTA, 2 mM EDTA, 0.5% Triton X-100, 0.5% NP-40, 5% glycerol, 1 mM dithiothreitol [DTT], and a protease inhibitor cocktail) for 1 h on ice and then centrifuged at 14,000 rpm at 4°C for 15 min. The supernatant was collected and precleared with 30 μl of Protein A/G beads (Santa Cruz Biotechnology) for 2 h, and then precleared lysates were incubated with 4 μg of each specific antibody overnight at 4°C. Lysates were then incubated with 30 μl of Protein A/G beads for 4 h at 4°C. After immune complexes were washed six times with IP buffer, they were eluted by boiling for 3 min at 95°C in SDS sample buffer and separated on 10% SDS-PAGE. After blocking, membranes were incubated with primary antibody and then with a peroxidase-conjugated secondary antibody. Bound secondary antibody (anti-mouse or anti-rabbit 1:10,000) (Santa Cruz Biotechnology) was detected using the enhanced chemiluminescence (ECL) reagent (Santa Cruz Biotechnology).
For immunohistochemistry, embryonic brains were dissected and fixed in 4% parafomaldehyde (PFA) at 4°C, cryoprotected in 30% sucrose, embedded and frozen in Tissue Tek OCT compound, and sectioned at 30 μm on a cryostat. Sections were incubated with primary antibody at 4°C for 18 h. For immunocytochemistry, cells cultured on coverslips were fixed with 4% PFA/PBS for 30 min and immunostained after permeabilizing with 0.2% Triton X-100. Tissues and cells were incubated with secondary antibodies at room temperature for 1 h and counterstained in 4'-6-diamidino-2-phenylindole (DAPI) (Boehringer Mannheim, Mannheim, Germany), and images were visualized using confocal microscopy (LSM5 PASCAL; Zeiss, Jena, Germany). Values obtained from at least three independent experiments were averaged and reported as means ± SD. The two-tailed Student’s t test was used to compare two experimental groups.
DH5 bacteria were transformed with GST-tagged plasmids (Mbd3, ΔN36, ΔN92, ΔC249, ΔC221, ΔC174, and ΔC93) and protein expression was induced by addition of 0.5 mM isopropyl 1-thio-β-D-galactopyranoside (IPTG) at 25°C at mid-log phase. Cells were lysed with B-PER Bacterial Protein Extraction Reagent (Thermo Fisher Scientific), lysates were purified, and proteins were captured using with Glutathione magnetic beads. HEK293T cells were transfected with Flag-tagged Smek2 plasmid and lysed with lysis buffer for 1 h on ice. Cell lysates were centrifuged at 14,000 rpm at 4°C for 15 min, and collected supernatants were incubated with Glutathione magnetic beads bound to GST or GST proteins. Bound proteins were eluted by boiling for 3 min at 95°C in SDS sample buffer, followed by immunoblotting.
NPCs were prepared from E11.5 cortex of Wild-type, Smek1-/-, Smek2-/-, and Smek1/2 dKO mice in Hank’s balanced salt solution (HBSS) (Invitrogen) and cultured as described [43]. To maintain stem cell characteristics, NPCs were cultured in N2 medium containing bFGF for 4 d. Stemness of cultured NPCs was confirmed by Nestin and Sox2 expression. To induce NPC differentiation, cells were seeded and further cultured in the absence of bFGF2.
NPCs derived from Wild-type or Smek1/2 dKO E11.5 forebrain were transfected with 4 μg pUltra-hot-Mbd3-flag or a Myc-tag vector for Mbd3 gain-of-function experiments and with pLKO3G-shMbd3 for Mbd3 loss-of-function experiments using Lipofectamine LTX and Plus Reagent (Invitrogen) or electroporation with an AMAXA nucleofector (Ronza AG, Basel, Switzerland). pLKO3G-shcontrol vector was used for negative control of pLKO3G-shMbd3 vector. After 12 h, transfection efficiency was confirmed to be >90% by monitoring mcherry or EGFP signaling. After two more days in differentiation conditions, cells were analyzed by qPCR or ChIP.
Mbd3 expression vectors were constructed by subcloning full-length mouse Mbd3 from a lentiviral FUIGW-Mbd3-Flag vector we previously created into XbaI/EcoRI sites of pUltra-hot (Addgene plasmid # 24130). For Mbd3-myc, the reverse primer included the full Myc sequence. PCR was carried out using the KOD Hot Start Polymerase Kit (EMD Millipore) with corresponding primer pairs. PCR products were ligated into double-digested pUltra-Hot vector and inserted ligations were confirmed by PCR and DNA sequencing (Genewiz, Inc). For the shRNA vector, the Mbd3 target sequence was designed based on the RNAi Consortium library top hits for mouse Mbd3. Details for pLKO3G shMbd3 and shcontrol construction are listed in S4 Table.
Cells were harvested and total RNA was isolated using TRIzol reagent (Invitrogen). The SuperScript III qRT-PCR kit (Invitrogen) was used to synthesize cDNA from total RNA. Quantitative PCR was carried out using the ABI PRISM 7900 Sequence Detection System with SYBR Green Master Mix (iTaq) with conditions of 95°C for 10 min followed by 50 cycles at 95°C for 15 sec and 60°C for 3 sec. Samples were run in triplicate and Dlx1, Dlx2, Tlx3, NeuroD1, Tuj1, Gad67, NeuN, Mbp, Gfap, Ascl1, and Id1 transcript quantitation was undertaken by comparing Cycle Threshold (Ct) values for each reaction with the Gapdh reference. Primer sets for quantitative PCR are listed in S5 Table.
For the ChIP assay, NPCs derived from wild-type or Smek1/2 dKO E11.5 forebrain or transfected with 4 μg pUltra-hot-Mbd3-flag or a pUltra-hot-empty vector for Mbd3 gain-of-function experiments and with pLKO3G-shMbd3 or pLKO3G-shScramble for Mbd3 loss-of-function were treated with 1% formaldehyde for 10 min at room temperature and quenched with 0.125 M glycine for ten more minutes at room temperature. Cross-linked chromatin was sonicated to fragment DNA to 200–1,000 base pairs, and then immunoprecipitation was performed with rabbit anti-IgG, anti-Smek1 (Sigma), anti-Mbd3 (Cell Signaling), anti-HDAC1, anti-HDAC2, anti-MTA1, and acetyl histone H3 (Santa Cruz Biotechnology) antibodies overnight at 4°C, followed by incubation with 50 μl of magnetic Protein A/G Dynabeads (EMD Millipore). Abundance of sequences in immunoprecipitates was determined by PCR and normalized as a fold-value relative to input chromatin. Smek ChIP-seq data were analyzed with the MACS online tool, and cis-regulatory sequences were analyzed using the Genomic Regions Enrichment of Annotations Tool (GREAT) interface (http://bejerano.stanford.edu/great/public/html/). We also utilized the Intergrative Genomics Viewer (IGV v2.3) to visualize distribution of ChIP-seq–identified peaks in different genomic regions. Primer sets for ChIP-qPCR are listed in S6 Table.
All procedures followed guidelines of the Institutional Animal Care and Use Committee (IACUC) and the National Institutes of Health. A total of 2.5% (w/v) avertin (1 g/ml solution of 2, 2, 2-Tribromoethanol, 97% in tert-amylalcohol [99%]; Aldrich, catalog numbers T4,840–2 and 24,048–6, respectively) in 0.9% saline was injected i.p. (15 μl/g of body weight) to anesthetize pregnant mice (E13.5). A laparotomy was performed, and the uterus with embryos was exposed. A total of ~2–5 μl of plasmid DNA (approximately 2 μg/ μl, dissolved in water) was injected into the lateral ventricle using a fine-glass microcapillary and a PV830 pneumatic PicoPump. Electroporation was performed using a Nepagene CUY21SC electroporator (amplitude, 50 V [E13.5]; duration, 50 ms; intervals, 150 ms). To deliver electrical pulses, tweezer-type circular electrodes (7-mm diameter) were used with the positive side directed to the medial wall of the ventricle into which DNA was injected. Uterine horns were repositioned in the abdominal cavity, and the abdominal wall and skin were sewed with surgical sutures. Mice were kept on a warm plate (37°C) for recovery. Two to three days later, embryos were taken from mothers and fixed with 4% (w/v) PFA (Sigma) in PBS (pH 7.4). After a 24 h fixation at 4°C, embryo brains were transferred to a 30% (w/v) sucrose solution in 4% PFA. Tissues were sectioned at 30 μm using a cryotome (Leica) and analyzed by immunohistochemistry.
Smek2 DNA was fused in frame to the LexA bait vectors pBTM116 for use as bait in the yeast two-hybrid screen. Preys were expressed as fusions to the activation domain of GAL4 in pACT2 (BD Biosciences Clontech, Palo Alto, CA, US). Transformed bait strains with or without transformed prey strains were mated and analyzed by using β-galactosidase activity. The following preys were used: mbd3. The Saccharomyces cerevisiae strain L40 (MATa trp1 leu2 his3 LYS2:: lexA-HIS3 URA3::lexA-lacZ) was cotransformed with bait and prey plasmids using the PEI method and selected for histidine prototrophy on minimal medium, containing 2% glucose; 6.7% yeast nitrogen base (BD Diagnostic Systems, Sparks, MD, US); complete amino acid mixture lacking histidine, leucine, and tryptophan (Qbiogene, Carlsbad, CA, US); and 2% bacto agar (BD Biosciences, Franklin Lakes, NJ, US). Yeast transformants were grown for 3 d at 30°C.
Statistical differences among groups were analyzed using Student’s t test and are indicated in each Fig as follows: *p < .05, **p < .005, and ***p < .0005. *p < .05 was considered statistically significant.
|
10.1371/journal.ppat.1006951 | Host shifts result in parallel genetic changes when viruses evolve in closely related species | Host shifts, where a pathogen invades and establishes in a new host species, are a major source of emerging infectious diseases. They frequently occur between related host species and often rely on the pathogen evolving adaptations that increase their fitness in the novel host species. To investigate genetic changes in novel hosts, we experimentally evolved replicate lineages of an RNA virus (Drosophila C Virus) in 19 different species of Drosophilidae and deep sequenced the viral genomes. We found a strong pattern of parallel evolution, where viral lineages from the same host were genetically more similar to each other than to lineages from other host species. When we compared viruses that had evolved in different host species, we found that parallel genetic changes were more likely to occur if the two host species were closely related. This suggests that when a virus adapts to one host it might also become better adapted to closely related host species. This may explain in part why host shifts tend to occur between related species, and may mean that when a new pathogen appears in a given species, closely related species may become vulnerable to the new disease.
| Host shifts, where a pathogen jumps from one host species to another, are a major source of infectious disease. Hosts shifts are more likely to occur between related host species and often rely on the pathogen evolving adaptations that increase their fitness in the novel host. Here we have investigated how viruses evolve in different host species, by experimentally evolving replicate lineages of an RNA virus in 19 different host species that shared a common ancestor 40 million years ago. We then deep sequenced the genomes of these viruses to examine the genetic changes that have occurred in different host species that vary in their relatedness. We found that parallel mutations–that are indicative of selection–were significantly more likely to occur within viral lineages from the same host, and between viruses evolved in closely related species. This suggests that a mutation that may adapt a virus to a given host, may also adapt it to closely related host species.
| Host shifts–where a pathogen jumps into and establishes in a new host species–are a major source of emerging infectious diseases. RNA viruses seem particularly prone to host shift [1–4], with HIV, Ebola virus and SARS coronavirus all having been acquired by humans from other host species [5–7]. Whilst some pathogens may be pre-adapted to a novel host, there are increasing numbers of examples demonstrating that adaptation to the new host occurs following a host shift [8, 9]. These adaptations may allow a pathogen to enter host cells, increase replication rates, avoid or suppress the host immune response, or optimise virulence or transmission [10, 11]. For example, in the 2013–2016 Ebola virus epidemic in West Africa, a mutation in the viral glycoprotein gene that arose early in the outbreak and rose to high frequency was found to increase infectivity in human cells and decrease infectivity in bats, which are thought to be the source of Ebola virus [12, 13]. Likewise, a switch of a parvovirus from cats to dogs resulted in mutations in the virus capsid that allowed the virus to bind to cell receptors in dogs, but resulted in the virus losing its ability to infect cats [14, 15].
In some instances adaptation to a novel host relies on specific mutations that arise repeatedly whenever a pathogen switches to a given host. For example, in the jump of HIV-1 from chimps to humans, codon 30 of the gag gene has undergone a change that increases virus replication in humans, and this has occurred independently in all three HIV-1 lineages [5, 16]. Similarly, five parallel mutations have been observed in the two independent epidemics of SARS coronavirus following its jump from palm civets into humans [17]. Similar patterns have been seen in experimental evolution studies, where parallel genetic changes occur repeatedly when replicate viral lineages adapt to a new host species in the lab. For example, when Vesicular Stomatitis Virus was passaged in human or dog cells, the virus evolved parallel mutations when evolved on the same cell type [18]. Likewise, a study passaging Tobacco Etch Potyvirus on four plant species found parallel mutations occurred only when the virus infected the same host species [19]. These parallel mutations provide compelling evidence that these genetic changes are adaptive, with the same mutations evolving independently in response to natural selection [20]. These studies have only used a limited number of hosts, and so do not provide information on how viral evolution occurs across a wide phylogenetic breadth of host species.
The host phylogeny is important for determining a pathogens ability to infect a novel host, with pathogens tending to replicate most efficiently when they infect a novel host that is closely related to their original host [2, 21–34]. Here, we asked whether viruses acquire the same genetic changes when evolving in the same and closely related host species. We experimentally evolved replicate lineages of an RNA virus called Drosophila C Virus (DCV; Discistroviridae) in 19 species of Drosophilidae that vary in their relatedness and shared a common ancestor approximately 40 million years ago [35, 36]. We then sequenced the genomes of the evolved viral lineages and tested whether the same genetic changes arose when the virus was evolved in closely related host species.
To examine how viruses evolve in different host species we serially passaged DCV in 19 species of Drosophilidae. In total we infected 22,095 adult flies and generated 173 independent replicate lineages (6–10 per host species). We deep sequenced the evolved virus genomes to generate over 740,000 300bp sequence reads from each viral lineage. Out of 8989 sites, 584 contained a SNP with a derived allele frequency >0.05 in at least one viral lineage, and 84 of these were tri-allelic. None of these variants were found at an appreciable frequency in five sequencing libraries produced from the ancestral virus, indicating that they had spread though populations during the experiment (Fig 1). In multiple cases these variants had nearly reached fixation (Fig 1).
We next examined whether the same genetic changes occur in parallel when different populations encounter the same host species. Of the 584 SNPs, 102 had derived allele frequencies >0.05 in at least two viral lineages, and some had risen to high frequencies in multiple lineages (Fig 1). We estimated the genetic differentiation between viral lineages by calculating FST. We found that viral lineages that had evolved within the same host were genetically more similar to each other than to lineages from other host species (Fig 2; P<0.001). Furthermore, we found no evidence of differences in substitution biases in the different host species (Fisher Exact Test: p = 0.14; see methods), suggesting that this pattern is not driven by changes in the types of mutations in different host species.
To examine the genetic basis of parallel evolution, we individually tested whether each SNP in the DCV genome showed a signature of parallel evolution among viral lineages passaged in the same host species (i.e. we repeated the analysis in Fig 2 for each SNP). We identified 56 polymorphic sites with a significant signal of parallel evolution within the same host species (P<0.05; significantly parallel sites are shown with a red asterisk in Fig 1; the false discovery rate is estimated to be 17% [37]).
We investigated if viruses passaged through closely related hosts showed evidence of parallel genetic changes. We calculated FST between all possible pairs of viral lineages that had evolved in different host species. We found that viral lineages from closely related hosts were more similar to each other than viral lineages from more distantly related hosts (Fig 3A). This is reflected in a significant positive relationship between virus FST and host genetic distance (Fig 3B, Permutation test: r = 0.15, P = 0.002). We lacked the statistical power to identify the specific SNPs that are causing the signature of parallel evolution in Fig 3 (false discovery rate >0.49 for all SNPs).
Two of the most striking examples of parallel evolution in related species are in Scaptodrosophila pattersoni and S. lebanonensis, which show two high frequency parallel mutations. These are a synonymous mutation in the 2C replicase protein at position 1901 and a triallelic non-synonymous mutation in a viral capsid protein at position 8072. However, the wider pattern of parallel evolution is not driven by these two examples, as the results remained significant after viruses that had evolved in these two species were removed from the dataset (within species parallelism: P<0.001; between species parallelism: P = 0.013).
When a pathogen infects a novel host species, it finds itself in a new environment to which it must adapt [4, 8, 10, 44]. When DCV was passaged through different species of Drosophilidae, we found the same genetic changes arose repeatedly in replicate viral lineages in the same host species. Such repeatable parallel genetic changes to the same host environment are compelling evidence that these changes are adaptive [20]. We then examined whether these same genetic changes might occur in closely related host species, as these are likely to present a similar environment for the virus. We found that viruses evolved in closely related hosts were more similar to each other than viruses that evolved in more distantly related species. Therefore, mutations that evolve in one host species frequently arise when the virus infects closely related hosts. This finding of parallel genetic changes in closely related host species suggests that when a virus adapts to one host it might also become better adapted to closely related host species.
Phylogenetic patterns of host adaptation may in part explain why pathogens tend to be more likely to jump between closely related host species. This pattern is seen in nature, where host shifts tend to occur most frequently between closely related hosts, and in laboratory cross-infection studies, where viruses tend to replicate more rapidly when the new host is related to the pathogens natural host [2, 21–34]. For example, in a large cross-infection experiment involving Drosophila sigma viruses (Rhabdoviridae) isolated from different species of Drosophila, the viruses tended to replicate most efficiently in species closely related to their natural hosts [34]. This suggests that these viruses had acquired adaptations to their host species that benefitted them when they infected closely related species. Our results demonstrate that this pattern is apparent at the level of specific nucleotides, and can arise very shortly after a host shift.
While the susceptibility of a novel host is correlated to its relatedness to the pathogens’ original host, it is also common to find exceptions to this pattern. This is seen both in nature when pathogens shift between very distant hosts [45, 46], and in laboratory cross-infection experiments [33, 34]. This pattern is also seen in our data where we also observe parallel genetic changes occurring between more distantly related hosts. For example, a mutation at position 8072 was not only near fixation in most of the lineages infecting two closely related species, but also occurred at a high frequency in replicate lineages in a phylogenetically distant host (Fig 1).
The function of these mutations is unknown, but in other systems adaptations after host shifts have been found to enhance the ability of the virus to bind to host receptors [11], increase replication rates [16] or avoid the host immune response [8, 10, 47]. Of the high frequency significant SNPs (shown in Fig 1 with red asterisk) nine occur in the non-structural proteins (ORF1: RNA dependant RNA polymerase, the putative protease and helicase proteins), and eleven occur in the capsid proteins (ORF2). Interestingly, none were in DCV-1A, which suppresses the host antiviral RNAi defences [48]. It will be of interest to examine the functions of the parallel mutations we detected, and characterise phenotypically how they affect viral infectivity and replication.
One mutation rose to a high frequency across all the host species (Fig 1, position 214 in the 3’ un-translated region). This is unlikely to be an error in the genome sequencing, as it did not occur when we sequenced the ancestral virus. This may have been due to natural selection favouring this change in all species, perhaps because there was a strongly deleterious mutation at this site in the virus we cloned or due to the virus going from cell culture to being passaged in vivo.
Previous studies have elegantly demonstrated parallel evolution following host shifts (eg [18]). However, these are often in cell culture, and so do not reflect the heterogeneity of tissue and cell types in whole animals that occur in studies in vivo (although see [49] that suggests otherwise). The complex nature of different tissue types in vivo coupled with a limited number of generations may explain why some parallel SNPs have remained at a low frequency in this study. Following a host shift, viruses must sometimes acquire specific mutations that allow them to be transmitted in their new host [9, 10]. As we artificially inoculated the virus, this aspect of adaptation to a new host is missing from our study.
In conclusion, we have found that host relatedness can be important in determining how viruses evolve when they find themselves in a new host. This study suggests that while some genetic changes will be found only in specific hosts, we frequently see the same changes occurring in closely related host species. These phylogenetic patterns suggest that mutations that adapt a virus to one host may also adapt it to closely related host species. Therefore, there may be a knock-on effect, where a host shift leaves closely related species vulnerable to the new disease.
DCV is a positive sense RNA virus in the family Discistroviridae that was isolated from D. melanogaster, which it naturally infects in the wild [50, 51]. To minimise the amount of genetic variation in the DCV isolate we used to initiate the experimental evolution study, we aimed to isolate single infectious clones of DCV using a serial dilution procedure. DCV was produced in Schneider’s Drosophila line 2 (DL2) cells [52] as described in [53]. Cells were cultured at 25°C in Schneider’s Drosophila Medium with 10% Fetal Bovine Serum, 100 U/ml penicillin and 100 μg/ml streptomycin (all Invitrogen, UK). The DCV strain used was isolated from D. melanogaster collected in Charolles, France [54]. DL2 cells were seeded into two 96-well tissue culture plates at approximately 104 cells in 100 μl of media per well. Cells were allowed to adhere to the plates by incubating at 25°C for five hours or over-night. Serial 1:1 dilutions of DCV were made in complete Schneider’s media, giving a range of final dilutions from 1:108–1:4x1014. 100 μl of these dilutions were then added to the cells and incubated for 7 days, 8 replicates were made for each DCV dilution. Each well was then examined for DCV infection of the DL2 cells, and a well was scored as positive for DCV infection if clear cytopathic effects were present in the majority of the cells. The media was taken from the wells with the greatest dilution factor that were scored as infected with DCV and stored at -80°C. This processes was then repeated using the DCV samples from the first dilution series. One clone, B6A, was selected for amplification and grown in cell culture as described above. Media containing DCV was removed and centrifuged at 3000 x g for 5 minutes at 4°C to pellet any remaining cell debris, before being aliquoted and stored at -80°C. The Tissue Culture Infective Dose 50 (TCID50) of the DCV was 6.32 x 109 infectious particles per ml using the Reed-Muench end-point method [55].
We passaged the virus through 19 species of Drosophilidae, with 6–10 independent replicate passages for each species. We selected species from across the phylogeny (that shared a common ancestor approximately 40 million years ago [35, 36]), but included clades of closely related species that recently shared common ancestors less than 5 million years ago (Fig 1). All fly stocks were reared at 22°C. Stocks of each fly species were kept in 250ml bottles at staggered ages. Flies were collected and sexed, and males were placed on cornmeal medium for 4 days before inoculation. Details of the fly stocks used can be found in the supplementary materials.
4–11 day old males were infected with DCV using a 0.0125 mm diameter stainless steel needle (26002–10, Fine Science Tools, CA, USA) dipped in DCV solution. For the first passage this was the cloned DCV isolate in cell culture supernatant (described above), and then subsequently was the virus extracted from the previous passage (described below). The needle was pricked into the pleural suture on the thorax of flies, towards the midcoxa. Each replicate was infected using a new needle and strict general cleaning procedures were used to minimise any risk of cross-contamination between replicates. Species were collected and inoculated in a randomised order each passage. Flies were then placed into vials of cornmeal medium and kept at 22°C and 70% relative humidity. Flies were snap frozen in liquid nitrogen 3 days post-infection, homogenised in Ringer’s solution (2.5μl per fly) and then centrifuged at 12,000g for 10 mins at 4°C. The resulting supernatant was removed and frozen at -80°C to be used for infecting flies in the subsequent passage. The remaining homogenate was preserved in Trizol reagent (Invitrogen) and stored at -80°C for RNA extraction. The 3 day viral incubation period was chosen based on time course and pilot data showing that viral load reaches a maximum at approximately 3 days post-infection. This process was repeated for 10 passages for all species, except D. montana where only 8 passages were carried out due to the fly stocks failing to reproduce. Each lineage was injected into a mean of 11 flies at each passage (range 4–18). Experimental evolution studies in different tissue types have seen clear signals of adaptation in 100 virus generations [18]. Based on log2 change in RNA viral load we estimate that we have passaged DCV for approximately 100–200 generations.
After passaging the virus, we sequenced evolved viral lineages from 19 host species, with a mean of 9 independent replicate lineages of the virus per species (range 6–10 replicates). cDNA was synthesised using Invitrogen Superscript III reverse-transcriptase with random hexamer primers (25°C 5mins, 50°C 50mins, 70°C 15mins). The genome of the evolved viruses, along with the initial DCV ancestor (x5) were then amplified using Q5 high fidelity polymerase (NEB) in nine overlapping PCR reactions (see supplementary Table S2 for PCR primers and cycle conditions). Primers covered position 62-9050bp (8989bp) of the Genbank refseq (NC_001834.1) giving 97% coverage of the genome. PCRs of individual genomes were pooled and purified with Ampure XP beads (Agencourt). Individual Nextera XT libraries (Illumina) were prepared for each viral lineage. In total we sequenced 173 DCV pooled amplicon libraries on an Illumina MiSeq (Cambridge Genomic Service) v3 for 600 cycles to give 300bp paired-end reads.
FastQC, version 0.11.2 [56] was used to assess read quality and primer contamination. Trimmomatic, version 0.32 [57] was used to removed low quality bases and adaptor sequences, using the following options: MINLEN = 30 (Drop the read if it is below 30 base pairs), TRAILING = 15 (cut bases of the end of the read if below a threshold quality of 15), SLIDINGWINDOW = 4:20 (perform a sliding window trimming, cutting once the average quality within a 4bp window falls below a threshold of 20), and ILLUMINACLIP = TruSeq3-PE.fa:2:20:10:1:true (remove adapter contamination; the values correspond in order to: input fasta file with adapter sequences to be matched, seed mismatches, palindrome clip threshold, simple clip threshold, minimum adapter length and logical value to keep both reads in case of read-through being detected in paired reads by palindrome mode).
To generate a reference ancestral Drosophila C Virus sequence we amplified the ancestral starting virus by PCR as above. PCR products were treated with exonuclease 1 and Antarctic phosphatase to remove unused PCR primers and dNTPs and then sequenced directly using BigDye reagents (ABI) on an ABI 3730 capillary sequencer in both directions (Source Bioscience, Cambridge, UK). Sequences were edited in Sequencher (version 4.8; Gene Codes), and were manually checked for errors. Fastq reads were independently aligned to this reference sequence (Genbank accession: MG570143) using BWA-MEM, version 0.7.10 {Li, 2009 #1605} with default options with exception of the parameter–M, which marks shorter split hits as secondary. 99.5% of reads had mapping phred quality scores of >60. The generated SAM files were converted to their binary format (BAM) and sorted by their leftmost coordinates with SAMtools, version 0.1.19 (website: http://samtools.sourceforge.net/) [58]. Read Group information (RG) was added to the BAM files using the module AddOrReplaceReadGroups from Picard Tools, version 1.126 (https://broadinstitute.github.io/picard).
The variant calling was then performed for each individual BAM using UnifiedGenotyper tool from GATK, version 3.3.0. As we were interested in calling low frequency variants in our viruses, we assumed a ploidy level of 100 (-sample_ploidy:100). The other parameters were set to their defaults except—stand_call_conf:30 (minimum phred-scaled confidence threshold at which variants should be called) and—downsample_to_coverage:1000 (down-sample each sample to 1000X coverage)
We used a trimmed version of a phylogeny produced previously [33]. This time-based tree (where the distance from the root to the tip is equal for all taxa) was inferred using seven genes with a relaxed molecular clock model in BEAST (v1.8.0) [43, 59]. The tree was pruned to the 19 species used using the Ape package in R [60, 61].
We examined the frequency of alternate alleles (single nucleotide polymorphisms: SNPs) in five ancestral virus replicates (aliquots of the same virus stock that was used to found the evolved lineages). SNPs in these ancestral viruses may represent pre-standing genetic variation, or may be sequencing errors. We found the mean SNP frequency was 0.000923 and the highest frequency of any SNP was 0.043 across the ancestral viruses. We therefore included a SNP in our analyses if its frequency was >0.05 in any of the evolved viral lineages. For all analyses we included all three alleles at triallelic sites.
As a measure of genetic differentiation we estimated FST between all the virus lineages based on the heterozygosity (H) of the SNPs we called [62]:
FST=Hb−HwHb
(Eq 1)
where Hb is the mean number of differences between pairs of sequence reads sampled from the two different lineages. Hw is mean number of differences between sequence reads sampled from within each lineage. Hb and Hw were calculated separately for each polymorphic site, and the mean across sites used in Eq (1). Hw was calculated separately for the two lineages being compared, and the unweighted mean used in Eq (1).
To examine whether there had been parallel evolution among viral lineages that had evolved within the same fly species, we calculated the mean FST between lineages that had evolved in the same fly species, and compared this to the mean FST between lineages that had evolved in different fly species. We tested whether this difference was statistically significant using a permutation test. The fly species labels were randomly reassigned to the viral lineages, and we calculated the mean FST between lineages that had evolved in the same fly species. This was repeated 1000 times to generate a null distribution of the test statistic, and this was then compared to the observed value.
To identify individual SNPs with a signature of parallel evolution within species, we repeated this procedure separately for each SNP.
We next examined whether viral lineages that had evolved in different fly species tended to be more similar if the fly species were more closely related. Considering all pairs of viral lineages from different host species, we correlated pairwise FST with the genetic distance between the fly species. To test the significance of this correlation, we permuted the fly species over the Drosophila phylogeny and recalculated the Pearson correlation coefficient. This was repeated 1000 times to generate a null distribution of the test statistic, and this was then compared to the observed value. To identify individual SNPs whose frequencies were correlated with the genetic distance between hosts we repeated this procedure separately for each SNP.
We confirmed there was no relationship between rates of molecular evolution (SNP frequency) and either genetic distance from the host DCV was isolated from (D. melanogaster) or estimated viral population size (see supplementary S1 and S2 Figs) using generalised linear mixed models that include the phylogeny as a random effect in the MCMCglmm package in R [63] as described previously [34]. We also examined the distribution of SNPs and whether they were synonymous or non-synonymous (see supplementary results).
To test whether there were systematic differences in the types of mutations occurring in the different host species, we classified all the SNPs into the six possible types (A/G, A/T, A/C, G/T, G/C and C/T). We then counted the number of times each type of SNP arose in each host species at a frequency above 5% and in at least one biological replicate (SNPs in multiple biological replicates were only counted once). This resulted in a contingency table with 6 columns and 19 rows. We tested for differences between the species in the relative frequency of the 6 SNP types by simulation [64].
Sequence data (fastq files) are available in the NCBI SRA (Accession: SRP119720). BAM files, data and R scripts for analysis in the main text are available from the NERC data repository (https://doi.org/10.5285/4434a27d-5288-4f2e-88ac-4b1372e4d073).
|
10.1371/journal.pbio.1000365 | Sex and the Single Cell. II. There Is a Time and Place for Sex | The Drosophila melanogaster sex hierarchy controls sexual differentiation of somatic cells via the activities of the terminal genes in the hierarchy, doublesex (dsx) and fruitless (fru). We have targeted an insertion of GAL4 into the dsx gene, allowing us to visualize dsx-expressing cells in both sexes. Developmentally and as adults, we find that both XX and XY individuals are fine mosaics of cells and tissues that express dsx and/or fruitless (fruM), and hence have the potential to sexually differentiate, and those that don't. Evolutionary considerations suggest such a mosaic expression of sexuality is likely to be a property of other animal species having two sexes. These results have also led to a major revision of our view of how sex-specific functions are regulated by the sex hierarchy in flies. Rather than there being a single regulatory event that governs the activities of all downstream sex determination regulatory genes—turning on Sex lethal (Sxl) RNA splicing activity in females while leaving it turned off in males—there are, in addition, elaborate temporal and spatial transcriptional controls on the expression of the terminal regulatory genes, dsx and fru. Thus tissue-specific aspects of sexual development are jointly specified by post-transcriptional control by Sxl and by the transcriptional controls of dsx and fru expression.
| Morphologically, fruit flies are either male or female. The specification of sex is a multi-step process that depends on whether the fertilized egg has only one X chromosome (will develop as male) or two X chromosomes (will develop as female). This initial assessment of sex activates a cascade of regulatory genes that ultimately results in expression of either the male or female version of the protein encoded by the doublesex gene (dsx). These sex-specific proteins from the dsx gene direct most aspects of somatic sexual development, including the development of all of the secondary sexual characteristics that visibly distinguish males and females. In flies, as in most animal species, only some tissues are obviously different between the two sexes, so we asked the question of whether all cells in the animal nevertheless know which sex they are. That is, do all cells express dsx? We have developed a genetic tool that lets us visualize the cells in which the dsx is expressed. Strikingly, dsx is only expressed in a subset of tissues. Thus, adult flies of both sexes appear to be mosaics of cells that do know their sex and cells that do not know their sex.
| A single, multi-branched regulatory hierarchy specifies all somatic sexual differences in Drosophila melanogaster [1]–[5]. The fly sex hierarchy is of great intrinsic interest as a model developmental system to dissect both how information is passed through a molecular network and how the actions of the terminal regulatory genes in that network are coordinated with the actions of other patterning hierarchies to orchestrate sex-specific aspects of development, morphogenesis, differentiation, and adult functions.
One outstanding issue with respect to the functioning of the sex hierarchy in flies is whether all cells are sexually differentiated. While the different roles of the two sexes in courtship and reproduction have led to the evolution of the arrays of female- and male-specific features that characterize the two sexes, in flies, as in most animal species, these overt sexual differences are limited to subsets of tissues. Nonetheless, we tend to think of “femaleness” and “maleness” as two distinct states of being that pervade individuals (certainly with respect to humans, and we would submit this viewpoint carries over to color our thinking of sexuality in other species as well). However, from a developmental point of view, there is a question as to whether tissues and organs that are not discernibly different between males and females “know” their sex (i.e., express and utilize the final regulatory genes in the hierarchy) and as a consequence differ sexually in subtle ways between males and females, or alternatively whether males and females are really sexual mosaics in which some cells know their sex and differentiate sex-appropriately, while other cells do not know their sex and thus differentiate identically in males and females.
The three branches of the fly sex hierarchy (Figure 1A) govern: (1) X chromosome dosage compensation (reviews [1],[6]), (2) male sexual and aggressive behaviors via the action of the fruitless (fru) gene in the nervous system (reviews [4],[5],[7]–[9]), and (3) all other somatic sexual differences via the action of the doublesex (dsx) gene (which also functions in the nervous system; see below) (reviews [1],[2],[5],[10]). The initial steps in somatic sex determination assess the X chromosome∶autosome ratio and establish sex by setting the RNA splicing activity encoded by Sex lethal (Sxl) to “ON” in females (XX) and “OFF” in males (XY). Once turned ON, SXL activity in females is maintained by a positive autoregulatory feedback loop. In females, SXL also blocks translation of msl-2 mRNA and thus prevents dosage compensation. In addition, SXL directs the female-specific splicing of transcripts from the transformer (tra) gene. The female-specific TRA protein together with the TRA-2 protein directs female-specific splicing of pre-mRNAs arising from the dsx and the P1 fruitless (fruM) promoters to generate female-specific dsx and fru mRNAs. In males, there is no SXL activity, and so (1) dosage compensation occurs and (2) tra transcripts containing premature stop codons are produced by the default-splicing pathway. The absence of TRA protein in males leads to the default splicing of dsx and P1 promoter-derived fru transcripts to produce male-specific mRNAs. The female- and male-specific dsx transcripts both encode proteins (DSXF and DSXM, respectively), whereas only the P1 promoter-derived fru male-specific transcripts are translated and they encode FRUM protein.
The widely accepted view has been that regulatory genes of the Drosophila sex hierarchy are expressed ubiquitously in the soma and that all cells “know” their sex. In particular, it was shown that Sxl is expressed in all somatic cells [11], which made sense since Sxl is needed to make dosage compensation sex-specific throughout the soma (reviews [12]–). Surprisingly, tra-2 and ix are expressed in both males and females, although they function in sexual development only in females, leading to the proposals that these genes were also ubiquitously expressed [16]–[18]. Finally, as the sex-specific splicing of dsx and tra pre-mRNAs was sufficient to account for all known roles of these genes, it was inferred that they too were ubiquitously expressed.
However, emerging evidence indicated that the terminal genes in the hierarchy, fru and dsx, are not expressed ubiquitously. Expression of fruM is restricted to subsets of neurons in the central and peripheral nervous systems [19]–[22]. Evidence suggesting that expression of dsx is also spatially and temporally restricted came first from the finding that expression of dsx mRNA is restricted to the gonad in embryos (www.fruitfly.org/cgi-bin/ex/insitu.pl) [23]. It was subsequently shown by immunolocalization that DSX expression in the central nervous system is restricted to a small subset of neurons in adults [24]–[26] and that DSXM expression in male gonads is restricted to somatic cells in both embryos and adults [27]. However, characterization of the spatial and temporal expression pattern of dsx has been quite limited. Moreover, the implications of a restricted pattern of dsx expression have been minimally considered.
To better understand the function of dsx in sexual development and the sexuality of adults, we generated a tool with which we could both visualize dsx expression throughout development and also manipulate dsx-expressing cells. Specifically, we used homologous recombination to insert the GAL4 transcriptional activator coding sequence into the dsx locus immediately following the start codon to generate dsxGAL4. In only those cells that express it, dsxGAL4 can be used to drive the expression of any transgene of interest under control of the GAL4-responsive upstream activating sequence (UAS). dsxGAL4 faithfully reproduces the known features of dsx expression. In addition, we observed dsxGAL4 expression in many tissues where no phenotypic effects of dsx mutants have been reported. Strikingly, dsx, like fru, is only expressed in a subset of tissues and is highly restricted within those domains; in many cells and tissues, it is not expressed at all. Thus in both XY and XX Drosophila, only a subset of cells appear to have the potential to sexually differentiate, and hence both males and females are mosaics of sexually differentiated and sexually undifferentiated cells. These findings have led to a significant revision in our understanding of how sex is specified in Drosophila. These findings also have significant implications for evolutionary considerations of sex.
Alternatively spliced mRNAs derived from a common promoter and common 5′ exons encode the DSXM and DSXF proteins, which have the same N-terminus but different C-termini [28]. We used homologous recombination to insert the GAL4 coding sequence immediately after the translational start codon of the dsx gene in exon 2, which is common to male and female dsx mRNAs, to generate dsxGAL4 (Figure 1B) [29]. The GAL4 coding sequence is terminated by a single stop codon. The two codons 3′ of the GAL4 stop codon encode different amino acids than those found in the DSX proteins, but otherwise the dsx sequences both 5′ and 3′ of the GAL4 insertion site are unaltered in the targeted chromosome. Thus, the dsxGAL4 gene is anticipated, barring internal reinitiation of translation, to be a null allele of dsx.
The proper targeting of the GAL4 coding sequence into the dsx gene was confirmed by genomic PCR. Additionally, we found that the insertion creates a mutant allele of dsx that produces classical dsx morphological phenotypes when heterozygous with the null alleles dsx1 or In(3LR)dsxM+R13 (unpublished data). However, the dsxGAL4 chromosome unexpectedly causes reduced fertility in both sexes.
Both molecular and genetic characterizations of the dsxGAL4 expression pattern indicate that it faithfully recapitulates the endogenous dsx expression pattern.
dsx transcripts and DSXM protein are only detected in cells of the developing gonads in stage 13–17 embryos (www.fruitfly.org/cgi-bin/ex/insitu.pl) [23],[27], and we saw dsxGAL4 driven UAS-mCD8::GFP [30] expression that precisely recapitulated this expression pattern (Figure 2A). This pattern of reporter gene expression was completely dependent on the presence of dsxGAL4, as were all other patterns of dsxGAL4 expression described below.
In third instar larvae, dsxGAL4 was expressed in the gonads of both sexes (Figure 2B,D), consistent with detection of DSXM at these stages [27]. Further, dsxGAL4 expression is coincident with the protein EYES ABSENT (EYA), which marks all somatic cell nuclei of the gonad (unpublished data) [31], but dsxGAL4 was not expressed in germ cells, which are marked by the cytoplasmic protein VASA (Figure 2B,D) [32]. Somatic cells in which dsxGAL4 was evident include the apical hub cells that form the germline stem cell niche (Figure 2E), cyst cells that envelop the developing germ cells throughout spermatogenesis (Figure 2B,C), and cells at the basal end of the testis that likely correspond to male-specific gonadal precursor cells (Figure 2B,C) [33]. This pattern of expression, like that of DSXM, continued into the adult testis. Thus, dsxGAL4 expression recapitulates the DSXM pattern in the gonad and reveals expression in the female gonad.
DSX antibodies have also been used to describe dsx expression in the CNS [24]–[26]. As presented below, dsxGAL4 expression in the CNS recapitulates all major features of the reported temporal and spatial patterns of DSX expression and in addition reveals new features of dsx expression in the CNS.
To further assess dsxGAL4's accuracy, we asked if it was expressed in cells that give rise to the sexually dimorphic external parts of the fly. The major external sexual dimorphisms in Drosophila occur on the forelegs, tergites (dorsal aspects of abdominal segments), sternites (ventral aspects of abdominal segments), and the genitalia. In all of these regions, males and females differ in the number, location, morphology, and/or pigmentation of specific cuticular elements. To ascertain whether dsxGAL4 was expressed in the cells producing these sexually dimorphic cuticular elements, we asked whether dsxGAL4-directed expression of an inhibitory RNA (UAS-dsxIR) that targets both the wild-type male and female dsx transcripts would produce dsx mutant cuticular phenotypes. Intersexual differentiation was seen in all of these cuticular elements in dsxGAL4/dsx+ individuals with two copies of UAS-dsxIR reared at 29°C, although to a lesser degree in some cases than what was displayed by dsx null control individuals, possibly indicating that UAS-dsxIR did not fully suppress dsx+ expression and/or also targeted the fused GAL4-dsx transcripts. For example, on the first tarsal segment of forelegs of wild-type individuals, there are ca. 10 bristles that develop as the thick, blunt sex comb teeth in the male (Figure 3A, XY) and ca. 5 tapered, pointed homologous bristles in the female (Figure 3A, XX) [34],[35]. In both XY and XX dsx mutant individuals, these bristles are intermediate in both their number and morphology between those of wild-type males and females. In both XY and XX individuals in which dsxGAL4/+ drove expression of UAS-dsxIR, these bristles had an intermediate morphology (Figure 3A). However, the sexual dimorphism in bristle number was not eliminated by UAS-dsxIR expression, although the numbers of these bristles were significantly decreased in males and increased in females (Figure 3B). Differentiation of non-sexually dimorphic regions of the cuticle was normal in both sexes (unpublished data). Taken together, the above findings are all consistent with dsxGAL4 accurately reporting the dsx expression pattern.
Below, we extended these studies by examining dsxGAL4-driven expression of UAS-fluorescent protein reporters at many developmental stages. Except as noted for sensory organs of the foreleg and genitalia, and the CNS, we saw no difference between males and females in the dsxGAL4 expression patterns and our descriptions apply to both sexes. The results reveal a very dynamic and elaborate pattern of dsxGAL4 expression in a wide variety of cell types reflecting the transcriptional regulation of dsx across development.
In embryos, dsxGAL4 expression was only detected in the gonad, as described above. We did not see expression in the genital imaginal disc precursor cells [36]. In first instar larvae, expression was seen in ca. 4 cells that are likely part of the genital disc based on their location (unpublished data). The number of genital disc cells expressing dsxGAL4 increased in the second instar to ca. 8–10 cells, which is substantially below the number of cells in the genital disc in young larvae (ca. 60) as determined by direct counts [37].
In intact third instar larvae, fluorescence from dsxGAL4-driven reporters in the gonads and genital disc was easily detectable through the body wall. Fluorescence from other discs, some of which do express dsxGAL4 (see below), was not routinely detectable in intact larvae. In very late stage third instar larvae (wandering stage), dsxGAL4 expression became detectable in most, if not all, of the larval fat body cells of both sexes (unpublished data). Larval fat body expression of dsxGAL4 was detected with three different UAS-fluorescent protein reporters and was never seen in the absence of dsxGAL4 and represents the first case we know of in D. melanogaster where a purely larval tissue may be sexually dimorphic.
Since dsx is known from genetic data to function during development in several imaginal discs [34],[35],[38],[39], we examined dsxGAL4 expression in dissected imaginal discs of mature third instar larvae using the nuclear reporter UAS-RedStinger. No dsxGAL4 expression was seen in clypeolabral, labial, or humeral discs. In the second leg, third leg, wing, and haltere discs there were only a few cells expressing dsxGAL4, mostly in the region near the stalks of these discs, and the number of dsxGAL4-expressing cells increased with age so that at the time of puparium formation the number of labeled cells in the wing disc approached 30, the haltere 20, and the second and third legs 5–15 each (Figure S1; see Figure 4E). The nature of these cells and what parts of the adult they correspond to is unknown. There are no known anatomic sexual dimorphisms in the adult derivatives of these discs. In the genital, foreleg, and eye-antennal discs, some of whose adult derivatives are sexually dimorphic, there was significant dsxGAL4 expression.
In the genital disc of both sexes, dsxGAL4 was broadly expressed, consistent with the fact that all major regions of this disc give rise to sexually dimorphic structures of the adult genitalia and analia [2],[40]. The only regions of the genital disc not expressing GFP were the lateral edges of the disc (Figure S1).
In contrast, foreleg discs from early wandering third instar larvae exhibited dsxGAL4 expression in a thin, crescent-shaped band of epithelial cells plus several scattered cells in both sexes (Figure 4A). The crescent of foreleg disc cells expressing the highest levels of fluorescence is outside of the engrailed domain of the disc, indicating that these cells reside in the anterior compartment (Figure 4A). However, a few cells with lower levels of fluorescence could be seen scattered within the engrailed domain. At this stage, a total of 15–30 cells expressed dsxGAL4 in the foreleg disc.
In mature wandering third instar larvae, the crescent of dsxGAL4-expressing cells in the foreleg disc increased markedly. dsxGAL4 expression largely overlapped with Sex combs reduced protein (SCR) expression, consistent with the known requirement for both Scr and dsx in sex comb formation (Figure 3B) [35],[41]. The relative levels of dsxGAL4 and SCR expression varied significantly between cells. Further examination of foreleg discs from less mature third instar larvae showed that a low level of SCR could be detected in all tarsal regions of the disc at a time when dsxGAL4 was expressed in only a few cells (unpublished data), suggesting that the specification of tarsal segment identities by Scr precedes expression of dsxGAL4. The crescent of dsxGAL4 expression was in tarsal segment 1 (T1) based on its location relative to that of bric-a-brac-lacZ (bab-lacZ) (Figure 4C), which labels the tarsal segment boundaries [42].
In contrast to the foreleg disc, dsxGAL4 was expressed in only a few cells within the tarsal segments of the other leg discs in mature third instar larvae (Figure 4D). Thus, the pattern of dsxGAL4 expression in the foreleg disc epithelium is dynamic during late larval life.
dsxGAL4 was expressed in subsets of cells within both portions of the eye-antennal disc (Figure 3E). In the eye domain, there were ca. 32 such cells with the majority residing in the ventral region of the disc. Although a number of these cells were within or immediately below the plane of the photoreceptor cells, only a few of them expressed the proneural marker neuralized-lacZ (neur-lacZ) [43]. An additional ca. 10 dsxGAL4-expressing cells were in the region of the frons primordia, while additional cells having very faint expression were found in both eye disc regions. In the antennal domain of the eye-antennal disc of mature wandering third instar larvae, there were ca. 71 cells expressing dsxGAL4. Approximately 10 of these were found in segment 1, located near the stalk opposite the palpus, and the remaining ca. 61 were found in segment 3, which contains the olfactory sensory organ precursors, and segment 4. Of the cells expressing dsxGAL4 in segments 3 and 4, a small number co-expressed neur-lacZ, suggesting they may be sensory organ precursors. No dsxGAL4 expression was seen in segment 2, the Johnston's organ primordia; segment 5, the basal cylinder primordia; or segment 6, the arista primordia. Like the thoracic discs, the number of cells expressing dsx in the eye-antennal disc undergoes a dramatic increase over the course of late larval life.
We examined other imaginal tissues of mature wandering third instar larvae for dsxGAL4 expression. In the gut, expression was detected in a number of small cells in a region posterior to the slight tapering of the anterior midgut (unpublished data). Each of these cells localized to clusters of cells that constitute the gut imaginal nests, and they were thus imaginal cells fated to contribute to the adult midgut. The number of such cells expressing dsxGAL4 ranged between 12 and 17 in males and 4 and 7 in females, but this difference may reflect variability in the degree of maturity of the small number of larvae examined. Notably, in those imaginal nests containing at least one dsxGAL4-expressing cell, the majority of other cells in the cluster did not express dsxGAL4. Further, other gut imaginal nests in the anterior midgut and those in posterior regions of the gut did not express dsxGAL4 at the stage examined. Nor was expression observed in the tracheal nests, abdominal histoblast nests, gut imaginal rings, or salivary gland imaginal rings.
dsxGAL4 was expressed in the cells of most (but not all) adult tissues in which dsx is known to have a developmental role. Expression in adult adipose cells (or fat body) and oenocytes was anticipated from genetic studies [2],[44], and we observed dsxGAL4 expression in the large sheets of adipose cells associated with the dorsal abdominal body wall, as well as in oenocytes (Figure S2). dsxGAL4 was also expressed in all of the internal derivatives of the genital disc in both sexes with the notable exception of the male accessory gland, which showed no expression (unpublished data). With respect to the lack of adult expression of dsxGAL4 in the male accessory gland, it is worth noting that dsx function is required during the late larval period, but not subsequently, for its specification [45].
Muscles that attach to sex-specific elements of the internal and external genitalia, such as the ejaculatory bulb and penis apparatus in males and the gonopod in females, also express dsxGAL4 (Figure S5 and unpublished data), as do epidermal cells underlying the cuticular elements of the external male genitalia and analia (Figure 6C,D). Consistent with dsx's regulation of cuticle pigmentation in abdominal tergites 5 and 6 [2], dsxGAL4 was expressed in many epidermal cells underlying the abdominal cuticle (unpublished data). In the five tarsal segments of the foreleg, dsxGAL4 was expressed in a complex pattern as outlined below.
Intriguingly, dsxGAL4 was also expressed in a number of locations and cell types for which there are no reported sexual dimorphisms, including thoracic and leg muscles, various head structures, epidermal cells, and subsets of cells in the digestive system. dsxGAL4 was also expressed in a subset of tracheal cells of the abdomen, although consecutive cells along the same tracheole varied between expressing or not expressing (unpublished data). Expression was also seen in trachea associated with the internal genitalia and the gonads. dsxGAL4 was also expressed in cells of the Malpighian tubules and salivary glands (Figure S3 and unpublished data). While we did not examine all muscles, we observed expression in muscle groups of the proboscis, the tergosternal muscles underlying the ventral abdominal cuticle, many of the muscles that overlay the proventriculus and midgut (unpublished data), as well as muscles associated with parts of the reproductive structures mentioned above. In addition, dsxGAL4 was expressed in muscles of the mesothorax, as well as the trochanter, coxa, femur, and tibia of all three pairs of legs (Figure S4 and unpublished data).
dsxGAL4 expression in the digestive system illustrates the complex transcriptional regulation governing dsx expression. dsxGAL4 was expressed in the proventriculus, midgut, rectum, and crop of the adult digestive system (Figure S3). In each of these components, expression is spatially restricted to subsets of cells, as revealed by both nuclear and membrane reporters (UAS-RedStinger and UAS-mCD8::GFP, respectively). In the proventriculus, dsxGAL4 was expressed in subsets of epithelial cells distributed across the epithelial folds of this organ and was expressed in much of the epithelium of the stomodaeal valve (Figure S3) [46]. Great variability in dsxGAL4 expression was seen in the populations of large enterocytes residing in different regions along the length of the midgut (Figure 5A,B). Smaller nuclei likely corresponding to both the basally located intestinal stem cells and the enteroendocrine cells [47],[48] did not express dsxGAL4 in the regions examined. In contrast to the midgut and rectum, no expression was detected in cells of the hindgut that connects them (Figure S3 and unpublished data). While dsxGAL4 was expressed in many cells covering the rounded surface of the rectum, it was not expressed in the large epithelial cells forming the rectal papillae (epithelial folds) that project into the lumen of the rectum (unpublished data). Expression was also absent in cells forming the tracheoles that extend into each rectal papilla.
While dsx is clearly implicated in distinct aspects of sexual behaviors [26],[49] and has been shown to have roles generating sexual dimorphisms in both peripheral [50] and central neurons [24],[26],[51],[52], it has been difficult in most cases to link dsx-dependent behavioral deficits to particular aspects of the nervous system (but see [26],[51]). Part of the difficulty has stemmed from the fact that most courtship defects detected in dsx null males manifest as general decrements in courtship. These general decrements in courtship could reflect requirements for dsx in the CNS, peripheral sensory neurons, or even non-neuronal cells.
As a first step towards distinguishing between these possibilities, we have characterized in detail the patterns of dsxGAL4 expression in the second and third antennal segments, tarsal segments of the foreleg, proboscis, maxillary palp, and external genital structures, which collectively contain primary sensory neurons of the olfactory, gustatory, auditory, and mechanosensory systems that are known to be important for courtship. In those instances where dsxGAL4 was expressed in peripheral sensory structures known to contain fru-expressing neurons, we examined whether their expression overlaps at the cellular level. To do this, we simultaneously imaged expression of dsxGAL4 and fruP1.LexA [50] using UAS-nuclear GFP (UAS-Stinger) and lexA operator-nuclear tdTomato (lexAop-tdTomato::nls) red fluorescent protein [50] in males and females at 72 h after puparium formation (APF) and as 0–12 h adults. The patterns of dsxGAL4 expression seen at these two time points were virtually the same.
In the head, dsxGAL4 was not expressed in any of the chemosensory or auditory neurons of the antennae, maxillary palps, or proboscis; accordingly, there is no overlap with fruP1.LexA. We did observe expression in other, non-neuronal cells in these tissues. For example, dsxGAL4-expressing cells were associated with the bases of large mechanosensory bristles on the second antennal segment and maxillary palps, and dsxGAL4 was expressed in epithelial cells along the lateral aspect of the proboscis (Figure 6A,B). Several days after eclosion, we also observed expression of dsxGAL4 in cells of the basal cylinder, the small antennal segment from which the arista projects (unpublished data).
At the late pupal/young adult stage, dsxGAL4 was expressed in a complex pattern in the five tarsal segments of the foreleg. In all five tarsal segments, dsxGAL4 was expressed in a subset of gustatory sense organs (GSOs), as evidenced by its pattern of expression with respect to that of fruP1.LexA, which is expressed in all foreleg GSOs with the exception of two GSOs in tarsal segment 5 (Figure 6E) [50]. Each GSO of the foreleg is composed of four gustatory neurons, one mechanosensory neuron, and several non-neuronal cells [53]. In those GSOs in which it was expressed, dsxGAL4 was expressed in both neuronal and non-neuronal cells (unpublished data). In tarsal segments 1–4 dsxGAL4 was expressed in fewer GSOs in females than in males (unpublished data), likely because males have more foreleg GSOs than do females [53]. In the first tarsal segment, dsxGAL4 was expressed in cells associated with the sex comb bristles, as well as pericuticular cells around the sex comb. Less pericuticular expression was seen in other tarsal segments.
dsxGAL4 was co-expressed with fruP1.LexA in the neurons of the clasper bristles of the male genitalia, but not in fruP1.LexA-expressing neurons of the mechanosensory bristles of the lateral plate and anal plate (Figure 6C,D).
We visualized dsxGAL4 expression in the CNS using the UAS-mCD8::GFP membrane-bound GFP reporter and observed excellent correspondence with the dsx neuron clusters previously identified by immunolocalization (Figure 7) [26]. We observed prominent expression of dsxGAL4 in neurons of the posterior brain, including the pC1 and pC2 clusters. Both were sexually dimorphic, with females showing ∼85% fewer cells in the medial pC1 cluster and ∼80% fewer cells in the pC2 cluster relative to males (Figure 7). The pC2 cluster also appeared to consist of at least two distinct clusters, which were each associated with distinct fasciculations from the medial (pC2m) and lateral (pC2l) clusters.
We also noted a small cluster of dsxGAL4-expressing neurons in the dorsal aspect of the posterior brain that had not been reported previously, which we here identify as posterior Cells, dorsal (pCd) (Figure 7). This cluster is apparent at mid-pupal stages in previous work, but it was not named (Figure 4 in [26]; Figure 1 in [25]; Figure S1 in [26]). Females had ∼50% fewer neurons in this cluster than did males. We also confirmed expression in a small number of isolated neurons in the male brain, including two anterior-dorsal neurons (aDNs), and a single medial suboesophageal neuron (SN) within each hemibrain; these neurons were absent in females (Figure 7). In addition to these clusters, we noted several neurons that had been previously identified as being part of the pC1 and pC2 clusters, but their distinct fasciculations and large cell bodies suggested that they are unlikely to be associated with these clusters. We rename these neurons in Figure 7.
Although previous authors have reported a neuronal morphology for all dsxGAL4-expressing cells in the anterior brain, we could not confirm this for the suboesophageal lateral neurons (SLNs), located in the anteroventral optic cleft [25],[26]. In this area, we instead observed a novel pattern of expression in gliaform cells (Figure 8A,B). These putative glia had a cortical morphology and were seen in both sexes beginning at about 24 h APF (unpublished data). Although they first appeared in the ventrolateral optic cleft, we observed a scattered distribution of these cells at later time points across the anterolateral cortex of the suboesophageal ganglion (SOG), consistent with the migratory behavior of glia during CNS development (review [54]). We confirmed these cells to be glia by examining overlap with the glial marker REPO (Figure 8B,D), and we refer to them as Suboesophageal Lateral Glia (SLG). All other cells expressing dsxGAL4 in the CNS appeared to show overlap with the neuronal marker ELAV (Figure 8B,C).
In the male VNC, dsxGAL4 expression was observed in the prominent TN1 neuronal cluster, which is derived from the prothoracic ganglion (Jim Truman, personal communication), and in the densely packed neurons of the abdominal ganglion (AbN) (Figure 7). Expression was also seen in single or paired TN2 neurons within each thoracic hemisegment. In contrast, dsxGAL4 was not expressed in the thoracic ganglia of females, while AbN expression was significantly reduced in females through pupal stages before increasing towards eclosion (Figures 7 and 9 and unpublished data). These data are consistent with the reported patterns of DSX immunoreactivity [24],[25].
While previous works reported the position of neuronal cell bodies, the projections of these neurons could not be described by immunolocalization because DSX is a nuclear protein [27]. To determine which areas of the CNS might be influenced by the activity of neurons that express dsx, we visualized the projections of dsxGAL4-expressing neurons throughout the brain and VNC (Figure 9). In the VNC of males, projections from foreleg gustatory neurons were visible in the prothoracic ganglion and showed male-specific crossing of the VNC midline [50]. TN1 neurons were seen to innervate the dorsal pro/mesothoracic neuropile, which may subserve wingsong [26],[55],[56]. The various TN2 neurons also innervated dorsal neuropile regions, while the AbNs innervated the posterior abdominal neuropile and projected through the abdominal trunk nerve (Figure 9).
In the brain, projections from dsxGAL4-expressing neurons were conspicuously absent from regions of sensory processing or multimodal integration such as the antennal lobes, lateral horn, and mushroom body (Figure 9), many of which are innervated by fru-expressing neurons [21],[22],[57]. Projections from dsxGAL4-expressing neurons were similarly absent from areas subserving memory (e.g., mushroom body, see [58]) and motor patterning (e.g., central complex, see [59]). Instead, dsxGAL4-expressing neurons projected to and arborized in core neuropile of the central brain, encircling the peduncle and innervating the superior ventrolateral protocerebrum and dorsofrontal protocerebrum (Figure 9) [60]. All four of the posterior dsxGAL4-expressing clusters contributed to this pattern of innervation in both male and female brains. However, we also observed a male-specific projection through the anterior dorsal commissure, in accord with the morphology reported for fru-expressing P1 neurons within the dsx-expressing pC1 cluster [51]. In females, no dsxGAL4-expressing neurons projected across the anterior dorsal commissure, consistent with dsxF-mediated cell death of the P1 neurons [51], but a subset of the remaining pC1 neurons contributed to non-commissural projections in the protocerebrum. Taken together, the projection patterns of dsxGAL4 neurons in the brain suggests that dsx may regulate sexual dimorphism in neurons that are not simply serving sensory processing or motor output but instead define or modulate a core circuit underlying sexual behavior.
Findings from our work and that of others suggest that dsx has several roles within the CNS. First, dsx sculpts sexually dimorphic CNS development in several neuronal structures, including the P1 cluster of the posterior brain [51], the TN1 cluster of the mesothoracic ganglion [26] (but see [24]), neurons of the abdominal ganglion [52],[61], and projections of foreleg gustatory neurons in the VNC [50]. In some of these contexts, dsx has been shown to sex-specifically regulate neuron number by either prolonging neuronal proliferation [61] or preventing apoptosis [24]. However, we also note that dsxGAL4 is expressed in single neurons that are not obviously sexually dimorphic in number (e.g., pMN1, pMN2, and aDN). Since dsxGAL4 continues to be expressed in both classes of these neurons into adulthood, this implies that dsx may play an additional role in regulating sex-specific connectivity or function in these neurons. Lastly, the expression of dsxGAL4 in glia raises the possibility that dsx acts to influence brain function through a previously unanticipated mechanism.
Here we report the generation of a targeted insertion of the GAL4 coding sequence into doublesex to generate dsxGAL4. dsxGAL4 has allowed us to broadly characterize the expression patterns of dsx across development in many tissues and provides a tool to investigate the role of dsx in particular aspects of sexual development.
Perhaps the most striking feature of our findings is that dsx transcription is regulated in a very dynamic and precise temporal and spatial pattern. Temporally, expression of dsx begins as early as mid-embryogenesis in the somatic cells of the gonad, whereas in some imaginal tissues it is not expressed until the mid- to late pupal period and persists into adults. Spatially, dsx is expressed in nearly all cells of some tissues, but in only a few, or no cells in other tissues. Moreover, even within one imaginal disc there can be very dynamic changes in dsx expression. There are several important implications of this diverse and dynamic expression pattern.
First, our findings provide significant insight in to how dsx functions through other regulatory genes to orchestrate sexual development. Studies have shown that dsx brings about many aspects of sexual development by sex-specifically modulating, in specific cells and tissues, the activities of generic transcription factors and cell–cell signaling molecules that are deployed sex-nonspecifically in other cells and tissues. For example, DSXF negatively regulates the expression of bnl, the Drosophila FGF gene, in a subset of cells of the genital disc so that it is not transcribed in females, but is transcribed in males, where DSXF is absent [62]. Male-specific expression of FGF in turn induces genital-disc-associated mesodermal cells that express the fly FGF receptor to migrate into the disc, dedifferentiate into ectoderm, and ultimately produce part of the male internal genitalia [62]. FGF and its receptor also function sex-nonspecifically in other aspects of Drosophila development [54],[63]. Findings such as these raised the question: How is the specificity that allows dsx to regulate genes like bnl in one tissue, but not another, encoded? Our findings here suggest that some of the specificity of dsx function likely comes from the fact that the DSX proteins are deployed in precise patterns across development.
Second, our findings also provide insight into the results of studies with temperature-sensitive sex determination mutants, which revealed that dsx likely acts at different times in various cell lineages to determine different aspects of sex [38]. Indeed, even within one cell lineage, different aspects of sex require the functioning of the sex hierarchy at different times [38]. The finding that dsx has a highly dynamic temporal expression pattern offers a basis for such observations.
Third, until now, sexual differentiation in flies has been considered to be an adult characteristic. No sexual dimorphisms in purely larval cells have been reported. Thus we were surprised to see that dsx is expressed throughout larval fat bodies of very late third instar larvae. If the expression of dsx in very late third instar larval fat bodies is reflective of sexual differentiation, we believe this may be related to the fact that larval fat bodies persist until shortly after adult eclosion. Perhaps the remnants of larval fat bodies are utilized during the pupal period or in young adults and the optimal compositions of these materials differ between males and females.
Fourth, basic evolutionary considerations suggest that the highly refined patterns of dsx and fru transcriptional regulation arose through the piecemeal modification of their cis-regulatory regions. While one focus of evolutionary studies on the Drosophila sex hierarchy has very profitably centered on cis-regulatory regions of genes that are immediately downstream of dsx [64],[65], our findings suggest that another rich and complementary evolutionary history with respect to sex in flies resides in the cis-regulatory regions of dsx and fru that are integral to the elaborate temporal and spatial regulation of these genes.
Finally, and perhaps most importantly, not all cells express dsx. This finding has significant implications with respect to the nature of sexuality. Our reasoning is as follows. Most aspects of sexual differentiation in flies are determined cell autonomously, i.e. the functioning of dsx and/or fru within a cell directs that cell's pattern of sexual differentiation (reviews [1],[66]). Thus, the findings that many cells express neither dsx nor fru suggest, most simply, that such cells do not differentiate sexually. In that case, Drosophila males and females are mosaics composed of some cells that know their sex (express dsx and/or fru) and sexually differentiate, and other cells that do not express dsx and/or fru and thus don't sexually differentiate. A caveat to this simple reasoning comes from more recent findings that in certain cells in Drosophila, sexual differentiation is specified non-autonomously (reviews [2],[67]). In these cases, cells expressing dsx direct neighboring cells, via the modulation of cell–cell signaling, to undergo sex-specific differentiation. The neighboring cells that are thus directed may or may not express dsx. We look on these cases where local cell–cell interactions specify sexual differentiation as “exceptions that prove the rule.” We believe it is very unlikely that the existence of such local cell–cell interactions offer an alternative explanation to our general conclusion that most cells that don't express dsx and/or fru don't know their sex (either directly by dsx/fru expression, or indirectly by a cell–cell interaction of the type just noted). Most simply, many cells that don't express dsx are found in large contiguous domains (e.g., all of the embryo except the somatic gonad, large parts of and even entire imaginal discs and their adult derivatives, etc.). Thus we suggest that in flies, both males and females are, in fact, mosaics in which some cells are competent to sexually differentiate and other cells are not.
These findings have also contributed substantially to crystallizing a major revision of our understanding of how sex is specified in Drosophila. In the canonical view of how the sex hierarchy functions, Sxl plays a unique, key role as the master regulatory gene, as its activity state (ON in females, OFF in males) is the sole factor dictating whether male or female sexual differentiation occurs in each cell throughout the soma (review [1]). Once the state of Sxl activity was determined in response to the embryonic assessment of the X chromosome∶autosome ratio, the autoregulatory function of Sxl fixed the splicing milieu in all somatic cells as either male or female throughout development. Downstream of Sxl in the sex hierarchy, the tra and dsx genes were believed to be constitutively expressed, and their expression governed solely by sex-specific alternative splicing. Thus, in this canonical view, there is only one decision point in the hierarchy—setting Sxl's activity state. Further, under this view, the findings that dsx controlled many different processes in cell-type specific ways were attributed to there being different arrays of other regulatory molecules with which DSX worked in each of these cell types.
The findings that the expression of the pre-mRNAs encoding fru's sex-specific functions are under precise transcriptional regulation [20], as is the expression of dsx in the embryo and CNS (www.fruitfly.org/cgi-bin/ex/insitu.pl) [23]–[27] and indeed in all somatic tissues (this report), have shown that the canonical view of how the fly sex hierarchy functions must be modified. There are, in fact, two levels of regulation governing the expression of sex by a cell. First, the fundamental role of Sxl in setting up the competency of all somatic cells to sex-specifically splice the transcripts of the downstream sex regulatory genes is unchanged. However, there is another, previously unrecognized layer of regulation of the hierarchy at the transcriptional level: temporally and spatially regulated transcription of the terminal sex determination regulatory genes, dsx and fru, results in only some cells having the competency to respond to Sxl's action and produce the sex-specific DSXF, DSXM, and FRUM transcription factors that confer the potential for sexual differentiation. One particularly satisfying aspect of this revised perspective on the fly sex hierarchy are the striking parallels between the dsx and fru branches of the hierarchy, which indicate that both branches are likely governed by the same developmental and evolutionary logics.
There is an intriguing caveat to our suggestion that in flies both sexes are mosaics of cells that express dsx and/or fru and thus have the potential for sexual differentiation, and other cells that don't express dsx and/or fru and thus cannot undergo sexual differentiation. If the dosage compensation branch of the hierarchy is included in these considerations, then it appears that two different types of sex-regulated functions (universal and tissue-specific) are governed in distinct ways. There are two features of somatic sexual differentiation in Drosophila that probably involve all somatic cells—dosage compensation and body size (females are larger than males). Dosage compensation is controlled directly by Sxl via a universal mechanism. How exactly body size is specified is not clear, but it is known that its regulation resides above tra in the sex hierarchy [1] and may thus also be controlled via a universal mechanism. In contrast, the tissue-specific aspects of sexual differentiation are jointly specified by Sxl splicing regulation and by the transcriptional controls of dsx and fru expression.
Our revised perspective on sexual development also offers insight into certain aspects of the evolution of sexuality. First, the highly refined temporal and spatial patterns of dsx and fru expression indicate that these genes have likely evolved by the progressive addition or subtraction of elements from their expression patterns. Second, we know that the genes that respond to dsx, either directly or indirectly, seem to be different in each cell type in which dsx functions (reviews [2],[64],[65]. Thus evolution has likely acted both (1) to allow the array of genes that are regulated by dsx to shift as such genes lose or acquire DSX binding sites and (2) to shift the cells and tissues in which dsx functions by altering the cis-regulatory regions of dsx that specify the temporal and spatial patterns of its expression. Significantly, evolutionary changes that lead to dsx expression in a new cell population open the possibility of dsx acquiring novel regulatory targets in those cells. While there is currently much less information on the targets of fru, we suggest that the situation is likely the same. Given the fundamental role of dsx in fly sexuality, coupled with our finding of the enormous spatial and temporal diversity in dsx's deployment across development, we hope that these findings will stimulate analogous studies of dsx homologues in other species.
When we look at how sexual differentiation is controlled in Drosophila today, we are seeing a sum of the evolutionary decisions that led to novel groups of cells or tissues, and particular sets of genes therein, becoming sex-specifically regulated. Strikingly, nearly all the decisions as to how to implement such sexual differentiation were the same—to deploy dsx or fru. It is reasonable then to wonder why it has consistently been dsx or fru, rather than some other transcription factor, which was selected to govern a new aspect of sexual development. We think part of the explanation for this observation is that dsx and fru are already structured for sex-specific alternative splicing, and thus evolution has only to select for either a novel place and/or a new target to come under their regulation. For another transcription factor to fulfill this role, evolution would have to select not only for the same events as just enumerated for dsx and fru but in addition for the new transcription factor to evolve to be sex-specifically regulated in response to the activity of Sxl.
Our current perspective on sexual development in flies is likely to be of significant heuristic value in understanding the processes of sexual differentiation in other species. First, that males and females are both sexual mosaics suggests that there is an evolutionary advantage to only some cells having the competency to sexually differentiate (see also [67]). This may simply indicate that particular cells and tissues acquire ad hoc the potential for sexual differentiation (expression of dsx and/or fru) when there is a selective advantage to the individual. That evolution appears to have taken this route repeatedly in Drosophila suggests that giving all cells the ability to sexually differentiate may negatively impact fitness. In this regard, it is of interest to note that pan-neural expression of FRUM or ubiquitous expression of DSXM is highly detrimental to survival (our unpublished results) [68],[69]. Second, the varied deployment of these factors in otherwise sex-neutral tissues may mediate the diversification of species. Finally, given the extensive commonality across animal species in (1) the genes they possess and (2) how they regulate most fundamental cellular and developmental processes, it seems likely that the mosaic nature in one species of a process as ancient and fundamental as sexual differentiation will likely be representative of sexuality in many other animal species.
Lines dsxGAL4 and UAS-dsxIR (P{WIZ-dsxIR}4 P{WIZ-dsxIR}10) were generated as described below. w; P{w[+mC] = UAS-Stinger}2, P{w[+mC] = lexAop-FRT-tdTomato(nls)}6.2; fruP1.LexA/TM6B is described in [50]. dsx1 and In(3LR)dsxM+R13 are from our lab stocks. Lines from the Bloomington Drosophila Stock Center were: P{ry+t7.2 = 70FLP}11 P{v[+t1.8] = 70I-SceI}2B nocSco/CyO, S2, P{w[+mC] = UAS-mCD8::GFP.L}LL5P, {w[+mC] = UAS-RedStinger}4/CyO, P{ ry+t7.2 = lArB}neurA101 ry506/TM3, and ryRK Sb1 Ser1. Lines received as gifts were: P{w[+mC] = UAS-Stinger}2 (Scott Barolo), P{lArB}bab1A128.1F3 (Frank Laski), and JFRC-IVSA1 (Barret Pfeiffer). Crosses were performed at 25°C except for inhibitory RNA crosses, which were performed at 29°C.
dsx sequences were PCR-amplified using AccuPrime Supermix (Invitrogen) from genomic DNA prepared with the DNeasy Tissue Kit (Qiagen), and sequenced prior to use. 721-bp dsx exon 2 fragment was amplified with primers AAGTCACTTACCCAAGGGCACATTG and CCTCCTGAGTCATCACCATCATGTC and used to make pP{WIZ-dsxIR} as per [70]. The dsx 2.8-kb 5′ and 2.7-kb 3′ homology arms, extending between genomic sequences ATGTACTAGTCCGTCCGTTTGTCTG and CATGATTCCAGCTTCTGATATCCTA, and GAATTCAATTTGCCTCGCTTTAAAT and GGTTTCGGAGGAGAACTGGAATAGC, respectively, were cloned flanking the GAL4 coding sequence and transferred into pP{WhiteOut2} (gift of Jeff Sekelsky) to make pP{WO2-dsx-GAL4}. Details available upon request.
pP{UAS-dsxIR} transgenics were made by P element-mediated germline transformation using standard methods. Line UAS-dsxIR4/10 contains two copies of the transgene. pP{WO2-dsx-GAL4} transgenics (Rainbow Transgenic Flies, Inc.) were made as described above and four independent integrant lines were isolated to serve as donors of the dsxGAL4 DNA substrate for homologous recombination [29]. Donors were crossed to the line containing heat-shock-inducible FLP recombinase and I-SceI endonuclease transgenes [29] and larvae were heat shocked for 1 h at 37°C on days 3 and 4 of development. ∼5,000 female F1 progeny containing all three elements were crossed to UAS-mCD8::GFP, and the F2 progeny screened for candidates with changes in the GFP expression pattern relative to the donors alone. Candidate lines producing intersexual progeny when crossed to dsx1 or In(3LR)dsxM+R13 were PCR-tested using the 5′ genomic and Gal4 primers, GTGTGTGAGGCTGCCTATGTACTAG and ATGCTTGTTCGATAGAAGACAGTAG, and the 3′ genomic and GAL4 primers CCCATGGTGTCGGTATCTCAAAG and TCACTACAGGGATGTTTAATACCAC, respectively. Using these primer pairs, insert-specific PCR products were generated for the 5′ and 3′ ends of the inserted GAL4. White-eyed, w;dsxGAL4 lines were established over the balancer TM6B.
Males from a w+;dsxGAL4 line containing a wild-type X chromosome (red-eyed) were crossed to w;UAS-dsxIR females to determine the chromosomal sex of intersexual progeny (red-eyed are XX, white-eyed are XY).
Native fluorescence of the UAS-induced reporter proteins was imaged in embryos and imaginal discs, with the exception of the second instar genital disc, which was immunostained with anti-GFP. CNS expression of GFP reporters was also detected with anti-GFP. Larval instars were staged by spiracle morphology [46]. Dissected gonads and imaginal discs were fixed with 4% paraformaldehyde (Electron Microscopy Sciences) in PBS for 15–25 min at 22°C. Blocking and antibody incubations were done in PBS with 0.1% Triton X-100 and 5% normal goat serum (Vector Laboratories) overnight at 4°C or 4–8 h at 22°C. Discs and gonads were mounted in Vectashield mounting media with DAPI (Vector Laboratories). CNS tissues were immunostained largely as described in [71]. Briefly, CNS tissues were fixed for 30 min with 4% paraformaldehyde in PBS, blocked in 4% normal goat serum in TNT, incubated at 4°C overnight with primary or secondary antibodies in TNT alone, and mounted in Fluoromount (Electron Microscopy Sciences). Primary antibodies from the Developmental Studies Hybridoma Bank were used at the indicated dilution: mouse anti-SCR 6H4.1 (1∶20), anti-EN 4D9 (1∶20), anti-Cut 2B10 (1∶20), and anti-EYA 10H6 (1∶25); rat anti-DN-Cadherin Ex#8 (1∶40), anti-Neuroglian BP104 (1∶40), anti-REPO 8D12 (1∶20), and anti-ELAV 7E8A10 (1∶10). Additional antibodies were: rabbit anti-VASA (1∶1000; gift of R. Lehmann) and rabbit anti-GFP (1∶1000 for discs or 1∶800 for CNS; Molecular Probes/Invitrogen); lacZ reporter expression was visualized with anti-β-galactosidase (1∶1000, Promega). Alexa Fluor fluorescently conjugated goat secondary antibodies (Molecular Probes/Invitrogen) were used at 1∶500 or 1∶800: 488 anti-mouse, 546 anti-mouse, 568 anti-rabbit, 647 anti-rabbit, and 647 anti-rat. Fluorescence imaging of live embryos and brightfield imaging of adult legs were done on an Axio Imager M1 (Zeiss). All other imaging was done on an LSM510 Meta or LSM710 laser scanning confocal microscope (Zeiss) using 20× air or 40× oil immersion objectives. Confocal slices were manipulated using Image J. Photoshop CS3 software was used to adjust brightness and contrast, as well as to crop images.
|
10.1371/journal.ppat.1005210 | Perivascular Arrest of CD8+ T Cells Is a Signature of Experimental Cerebral Malaria | There is significant evidence that brain-infiltrating CD8+ T cells play a central role in the development of experimental cerebral malaria (ECM) during Plasmodium berghei ANKA infection of C57BL/6 mice. However, the mechanisms through which they mediate their pathogenic activity during malaria infection remain poorly understood. Utilizing intravital two-photon microscopy combined with detailed ex vivo flow cytometric analysis, we show that brain-infiltrating T cells accumulate within the perivascular spaces of brains of mice infected with both ECM-inducing (P. berghei ANKA) and non-inducing (P. berghei NK65) infections. However, perivascular T cells displayed an arrested behavior specifically during P. berghei ANKA infection, despite the brain-accumulating CD8+ T cells exhibiting comparable activation phenotypes during both infections. We observed T cells forming long-term cognate interactions with CX3CR1-bearing antigen presenting cells within the brains during P. berghei ANKA infection, but abrogation of this interaction by targeted depletion of the APC cells failed to prevent ECM development. Pathogenic CD8+ T cells were found to colocalize with rare apoptotic cells expressing CD31, a marker of endothelial cells, within the brain during ECM. However, cellular apoptosis was a rare event and did not result in loss of cerebral vasculature or correspond with the extensive disruption to its integrity observed during ECM. In summary, our data show that the arrest of T cells in the perivascular compartments of the brain is a unique signature of ECM-inducing malaria infection and implies an important role for this event in the development of the ECM-syndrome.
| Cerebral malaria is the most severe complication of Plasmodium falciparum infection. Utilizing the murine experimental model of cerebral malaria (ECM), it has been found that CD8+ T cells are a key immune cell type responsible for development of cerebral pathology during malaria infection. To identify how CD8+ T cells cause cerebral pathology during malaria infection, in this study we have performed detailed in vivo analysis (two photon imaging) of CD8+ T cells within the brains of mice infected with strains of malaria parasites that cause or do not cause ECM. We found that CD8+ T cells appear to accumulate in similar numbers and in comparable locations within the brains of mice infected with parasites that do or do not cause ECM. Importantly, however, brain accumulating CD8+ T cells displayed significantly different movement characteristics during the different infections. CD8+ T cells interacted with myeloid cells within the brain during infection with parasites causing ECM, but this association was not required for development of cerebral complications. Furthermore, our results suggest that CD8+ T cells do not cause ECM through the widespread killing of brain microvessel cells. The results in this study significantly improve our understanding of the ways through which CD8+ T cells can mediate cerebral pathology during malaria infection.
| Malaria remains a significant global health problem with 207 million cases, resulting in 584,000–1,238,000 deaths, annually [1, 2]. A high proportion of these deaths are due to cerebral malaria (CM), a neuropathology induced primarily by the species Plasmodium falciparum [2]. Current treatment of cerebral malaria is limited to parasiticidal chemotherapies, typically administered late in the course of infection. These traditional and narrowly targeted interventions are ineffective in many cases, and the mortality rate of CM, even after treatment, remains at 10–20% [3–5]. A greater understanding of the parasitological and immunological events leading to the development of CM would aid the development of improved therapeutic options to treat the condition.
Infection of susceptible strains of mice with Plasmodium berghei ANKA (Pb ANKA) results in the development of a serious neurological syndrome, termed experimental cerebral malaria (ECM), which recapitulates many of the clinical and pathological features of CM [6–10]. Susceptible mice typically develop neurological signs of disease including ataxia, convulsions, paralysis and coma between 6 and 8 days post infection [7, 11]. Histologically visible hemorrhages, widespread disruption of the vascular integrity and accumulation of leukocyte subsets are observed within the brain concomitant with the onset of signs of disease, [12–14]. The reason why Pb ANKA causes ECM while other strains of P. berghei, such as P. berghei NK65, do not is an area of active investigation. However, the differing virulence of P. berghei parasites does not appear to be due to extensive genetic polymorphisms between strains [15, 16].
Multiple cell types, including monocytes, macrophages, NK cells and CD8+ T cells accumulate within the brain at the onset of ECM [17–20]. However, to date, only CD8+ T cells have been identified as playing an unequivocal role in the development of cerebral pathology; protection from ECM is afforded by their depletion as late as one day prior to the development of neurological signs [10, 12, 19, 21]. The pathogenic parasite-specific CD8+ T cells are primed in the spleen by CD8α+ dendritic cells (DCs) [22] before migrating to the brain through homing dependent upon IFNγ-stimulated CXCL10 production in the CNS [23]. Monocytes play a role in recruitment of the pathogenic CD8+ T cells to the brain during ECM; however, the relative importance of this event in development of cerebral pathology remains undefined [19]. It has previously been shown that parasite-specific CD8+ T cells mediate ECM development through perforin- and granzyme B-dependent mechanisms [11, 24, 25], yet where the CD8+ T cells localize within the brain to cause ECM has remained unclear.
Parasite-specific CD8+ T cells appear to require in situ antigen-dependent stimulation within the brain to program their pathogenic activity necessary for ECM development [11]. To date, however, the identity of the putative antigen cross-presenting cells that interact with pathogenic CD8+ T cells during ECM is unknown. Recently, it has been shown that parasite specific CD8+ T cells can specifically interact with antigen cross-presenting microvessel cells obtained from mice experiencing ECM [26], but the relevance of this interaction for development of ECM in vivo is undefined.
In other models of neuroinflammatory diseases, such as experimental autoimmune encephalomyelitis (EAE), it has been demonstrated that professional antigen presenting cells (APCs) within the subarachnoid (SA) and perivascular spaces of the central nervous system (CNS) present antigen to T cells, instructing their pathogenic function [27–31]. Whether interaction of CD8+ T cells with brain-resident or infiltrating APC types is a canonical event in ECM development is largely unexplored and may represent a hitherto unexplored mechanism in the development of ECM.
In this study, we have attempted to reveal, in vivo, the mechanisms through which brain-infiltrating CD8+ T cells cause ECM. Using transcranial intravital two-photon microscopy, we report that T cells are recruited to, and accumulate perivascularly within, the SA and perivascular spaces of mice infected with both ECM-inducing and non-ECM-inducing Plasmodium berghei strains. However, a high proportion of perivascular T cells exhibited arrested behavior, consistent with immunological synapse formation [32–34], in the meninges specifically during ECM-inducing malaria infection. These arrested perivascular T cells formed cognate interactions with cells expressing CX3CR1, comprising inflammatory monocytes, macrophages and dendritic cells, but this event was redundant for ECM development. Pathogenic CD8+ T cells co-localized with apoptotic CD31+ cells in brains of mice with ECM, but apoptosis was a rare event in relation to the extensive vascular leakage observed during ECM. Combined, our results support a model where CD8+ T cells mediate ECM via direct recognition of cognate antigen on target cells without the need for additional in situ secondary activation in the brain by professional APCs and without causing apoptosis.
To investigate the immunopathological events that contribute to the development of ECM, we used the well-characterized Pb ANKA infection of C57BL/6 mice. Infected mice developed fatal neurological symptoms of ECM on day 6–7 post infection (p.i.) (Fig 1A), with a peak peripheral parasitaemia of around 15% (Fig 1B). The brains of symptomatic mice (day 6 p.i) displayed extensive vascular leakage, as assessed by Evans blue leakage, with diffuse blue coloration throughout the brain along with a few intense blue foci, which identify sites of petechial hemorrhage (Fig 1C). In contrast, brains from uninfected mice showed no discoloration (Fig 1C). Spectrophotometric quantification of Evans blue extravasation due to disruption of the cerebral vascular integrity revealed this to be a late occurring phenomenon, coinciding with the onset of ECM. (Fig 1D).
A pathological hallmark of human CM is sequestration, or cytoadhesion, of parasitized RBCs (pRBCs) within the cerebral blood vessels [35, 36]. In agreement, utilizing static immunofluorescence detection methods, we observed low numbers of Pb ANKA pRBCs adhering to vascular endothelial cells in mice with advanced symptoms of ECM (Fig 2A and 2C). Importantly, no parasite accumulation was observed in the brains of mice during Pb NK65 infection (Fig 2B), a strain of malaria that causes similar peripheral parasite burdens but does not cause signs of cerebral dysfunction (S1 Fig). We subsequently performed intravital imaging through the thinned skull to study the nature of pRBC adhesion to vascular endothelial cells under physiological flow conditions during ECM. Comparable with results recently obtained by Nacer et al. [13], we observed low frequencies of Pb ANKA pRBC adhering to vascular endothelial cells (Fig 2D and 2E). This interaction was weak, and the pRBCs were quickly removed by the sheer force of blood flow (S1 Video). Surprisingly, by intravital imaging we also observed an increase in the number of pRBCs located in the perivascular space in mice with advanced symptoms of ECM (Fig 2F and 2G, S2 Video). In fact, on day 6 p.i., when the majority of mice developed ECM (9/14), perivascular pRBCs (∼3 pRBC/mm2) were found more often than adherent luminal pRBCs (∼2 pRBC/mm2). In those mice that developed ECM on day 7 p.i., there was a further significant increase in numbers of perivascular pRBCs (∼22 pRBC/mm2), which represents a 10-fold increase that is substantially more than that of peripheral parasitaemia (1.5-fold increase), or adherent luminal pRBCs (no increase), observed between days 6 and 7 p.i. In contrast to LCMV encephalitis, another CD8+ T cell-dependent immunopathological model, we did not observe any evidence of petechial hemorrhages during intravital imaging of meninges of mice with ECM. Thus, it is unlikely that the accumulation of pRBCs within the perivascular space during ECM was simply due to formation of and carriage within hemorrhages. (S2 Fig and S3 Video). These results show that parasite accumulation within the brain is a specific event associated with ECM. However, at the point of ECM development, perivascularly located pRBCs, determined as such by their location relative to the fluorescent endothelium, are at least as common as briefly adherent intravascular pRBCs (Fig 2H). The interaction of pRBCs or parasite-derived material with cells behind the blood vessel endothelial wall within the perivascular space may, therefore, be an important event in the development of ECM.
CD8+ T cells are known to play a central role in the development of ECM [11, 17], yet the mechanisms through which they promote cerebral pathology during malaria infection remains poorly understood. It has been shown that despite the divergent infection outcomes, CD8+ T cells accumulate at similar levels in the brains of mice infected with Pb ANKA and Pb NK65 parasites [37]. We therefore hypothesized that brain accumulating CD8+ T cells exhibited intrinsic differences in phenotype activation status during Pb ANKA and Pb NK65 infections, explaining their pathogenic activity specifically during ECM-inducing infections. To test this hypothesis, we isolated leukocytes from the brains of infected and uninfected mice and characterized them via flow cytometry. As reported, similar frequencies and numbers of CD8+ T cells accumulated within the brains of mice infected with Pb ANKA and Pb NK65 on day 7 p.i., when mice infected with Pb ANKA developed ECM (Fig 3A and 3B). Moreover, comparable frequencies of brain accumulating CD8+ T cells expressed high levels of CD11a during Pb ANKA and Pb NK65 infections (Fig 3C), a marker of antigen experience [38, 39]. The intracerebral CD8+CD11ahigh parasite-specific T cells also displayed comparable activation in Pb ANKA and Pb NK65 infections, as evidenced by effector status (CD44+CD62L-) and increased expression of CD69, ICOS, KLRG1, CXCR3, and granzyme B (Fig 3D). Combined, these results show that CD8+ T cells recruited to the brains of mice infected with either ECM-causing or non-ECM causing parasites are similarly activated to cause cerebral pathology and that, although necessary, their presence alone is not sufficient for ECM development.
The comparable activation status of CD8+CD11ahigh parasite specific T cells in the brains of mice infected with Pb ANKA and Pb NK65 parasites on day 7 p.i. suggested that the disparate pathogenic activity of the cells during the two infections may be driven by CD8+ T cell extrinsic factors, which were not quantifiable through traditional flow cytometric analysis. To investigate this, we performed intravital imaging to study the compartmentalization and dynamics of T cells within the brain of hCD2-DsRed transgenic B6 mice during Pb ANKA and Pb NK65 infections. Few DsRed+ T cells were identified within the brains of naïve mice (Fig 4A) or on day 5 p.i. with Pb ANKA (S3 Fig), confirming the late accumulation of T cells in the brain during infection. Very few NK cells or B cells, which may express hCD2, were observed in the brains of infected mice on day 7 p.i. (S4 Fig). Characterization of T cells from isolated meningeal vessels of infected mice on day 7 p.i. showed them to be mainly CD8+ (>70%) (S5 Fig). The number of T cells/mm2 of vessel was increased within the brains of Pb ANKA and Pb NK65 infected mice (day 7 p.i.) compared with uninfected mice (Fig 4A and 4B). The relatively high numbers of T cells quantified by two photon microscopy compared with those given by whole brain flow cytometric analysis may reflect preferential meningeal T cell localization, diluted out by whole brain homogenization. Alternatively it may represent the failure to recover for flow cytometric analysis the Pb specific T cells that are tightly bound to target cells during ECM, a problem recently highlighted for T cell isolation from non-lymphoid tissues [40].
Surprisingly, the majority of T cells were compartmentalized to the perivascular side of the blood vessels during both Pb ANKA and Pb NK65 infections (70.5±10.3% vs 59.2±21.9%, respectively) (Fig 4C). Perivascularly located T cells were closely associated with the abluminal surface of blood vessels in mice infected with Pb ANKA and Pb NK65 parasites. Average distance from the vessel surface was, however, greater in mice infected with Pb NK65 than Pb ANKA parasites (5.9±7.4 μm v 3.1±4.8 μm, respectively) (Fig 4D). Interestingly, the distribution of T cells was not homogenous around the vessels imaged, with clusters of T cells observed preferentially around particular vessels (Fig 4E), which are most likely post-capillary venules [20].
It is becoming evident that the behavior of effector T cells, such as interaction with antigen presenting cells or responsiveness to chemokines and adhesion molecules, determines the local activity and function of the cells within the tissue [41, 42]. Thus, as T cells, of which greater than 70% are CD8+, were observed in comparable compartments of the brain during Pb ANKA and Pb NK65 infections, we next examined whether the cells displayed disparate behavior in the two infections, underlying their pathogenic activity specifically during Pb ANKA infection. Tracking DsRed+ cells revealed that perivascular T cells within brains of mice infected with Pb ANKA (S4 Video and S5 Video) exhibited a more arrested phenotype than those in mice infected with Pb NK65 (S6 Video and S7 Video). These differences were reflected in a lower mean track speed (8.7±4.4 v 15.7±5.6 μm/min), higher arrest coefficient (0.18±0.21 v 0.05±0.1) and lower confinement ratio (0.35±0.23 v 0.46±0.24), reflecting greater confinement (Fig 4F). The more constrained movement of perivascular T cells in brains of mice infected with Pb ANKA compared with Pb NK65 parasites is graphically illustrated by plotting 2D projections of tracks over 17 minutes that were then fixed to a common origin (Fig 4G). Thus, a major correlate of ECM is the perivascular arrest of CD8+ T cells.
The arrest of perivascular T cells within brains of mice infected with Pb ANKA was consistent with immune synapse formation with antigen-expressing cells. The interaction of perivascular T cells with brain APCs has previously been shown to instruct T cell pathogenic functions in other models of T cell mediated cerebral pathology [43, 44]. Consequently, we next assessed the interaction of T cells with professional APC populations in the brains of mice infected with Pb ANKA and Pb NK65 utilizing hCD2-DsRed X CX3CR1+/GFP dual reporter mice, where GFP is expressed by subsets of monocytes, macrophages and DCs, and all microglia [45]. We found that perivascular T cells within brains of mice infected with either Pb ANKA or Pb NK65 were closely associated with CX3CR1+/GFP cells and made frequent interactions with their cellular processes (S8 Video and S9 Video). Consistent with results from the single-reporter hCD2-DsRed mice, arrested perivascular T cells were more numerous in mice infected with Pb ANKA than Pb NK65, and many were stably bound to CX3CR1+/GFP cells (Fig 5A–5C, S10 Video). Formation of longer lasting stable interactions between CX3CR1+/GFP cells and perivascular T cells was reflected by longer average contact times for individual perivascular T cells (Fig 5D) and the formation of fewer contacts with new CX3CR1+/GFP cells in mice infected with Pb ANKA than in mice infected with Pb NK65 (Fig 5E) over the course of the 17 minute imaging period. Combined with CD8+ T cell recruitment results, this raised the possibility that differences within the CX3CR1+/GFP population may determine the behavior, and associated pathogenicity, of perivascularly located T cells that enter the brain subsequent to their full activation in the spleen during Pb infection.
We hypothesized that the different nature of cognate interaction between T cells and CX3CR1+/GFP cells in the brains during Pb ANKA and Pb NK65 infections was due to alterations in the composition and/or activation of the brain CX3CR1+/GFP population during infection with Pb ANKA and Pb NK65. Thus, to more specifically identify and characterize the CX3CR1+/GFP cells present within the brain during the two infections, we isolated GFP+ leukocytes from the brains of infected and uninfected mice and characterized them for expression of phenotypic and functional markers. The frequencies of GFP+ cells (out of total leukocytes) in the brain increased comparably during infection with both Pb ANKA and Pb NK65 (Fig 6A). We subsequently sub-gated GFP+ cells into three subsets based on expression of CD11b and CD45; R1 = CD45intCD11bhi microglia, R2 = CD45hiCD11bhi meningeal and perivascular macrophages and inflammatory monocytes and R3 = a mixed population of CD45hiCD11bint leukocytes [46, 47] (Fig 6B). The majority of GFP+ cells in brains from uninfected mice were microglial cells (73.7±7.9%). During both infections, the proportion of microglia within the GFP+ population decreased (52.3±12.2% for Pb ANKA, 47.2±14.1% for Pb NK65), likely due to other GFP+ cells infiltrating the brain (Fig 6C). CD45hiCD11bint leukocytes were mainly CD11c+ DCs (Pb ANKA 82.8±10.3%, Pb NK65 74.5±8.6%) and CD45hiCD11bhi leukocytes were mainly Ly6Chi inflammatory monocytes (Pb ANKA 84.4±2.5%, Pb NK65 83.8±2.5%). Lack of CD11c and Ly6C expression on the CD45int CD11bhi population confirmed their identity as microglia (Fig 6D). These results demonstrate that the composition of the CX3CR1+/GFP population changed within the brain during malaria infection, and that it changed comparably during ECM-causing and non-ECM-causing malaria infection. Furthermore, in contrast to our hypothesis, infection with both Pb ANKA and Pb NK65 caused largely comparable activation of all three CX3CR1+/GFP populations with the most striking up-regulation of co-stimulatory (CD40 and CD80) and antigen presenting molecules (MHC-I) occurring in the CD45hiCD11bhi population (Fig 6E). Changes in T cell motility correlating with ECM could not be attributed to different T cell or myeloid cell surface phenotypes and, thus, in vivo motility represents a distinct parameter of value in assessing T cell function.
The CD45hiCD11bhi monocyte and macrophage containing population (R2) displayed the most activated phenotype during infection and were also found to be enriched within the meninges, the site of imaging, compared with the whole brain (S6 Fig). Previous studies investigating the role of monocytes and macrophages in ECM pathogenesis have employed systemic administration of clodronate liposomes, CCR2 and Gr1 depleting antibody or AP20187 drug administration to MAFIA mice [19, 26, 48, 49]. None of these interventions prevented ECM development when given late in the course of infection. Crucially, however, these methods did not deplete populations located behind the endothelial barrier, including meningeal and perivascular macrophages or microglia. Accordingly, we addressed the contribution of these cells, and by association the importance of T cell interactions with them, in the development of ECM. Depletion of systemic and meningeal and perivascular macrophages through combined intraperitoneal (i.p.) and intracerebroventricular (i.c.v.) injection of clodronate liposomes (C.L.) using a protocol adapted from Galea et al. [50], from day 5 p.i. failed to protect mice against development of ECM (Fig 7A). Furthermore, specific depletion of CX3CR1 expressing cells, including perivascular and meningeal macrophages and microglia, from day 3 of infection utilizing CX3CR1-iDTR mice [51] also failed to inhibit development of ECM (Fig 7B). This was despite large scale depletion of brain resident macrophages and microglia, as shown by reduced Iba1 staining (Fig 7C). Our results, therefore, indicate that although stable interaction of T cells with CX3CR1+/GFP APCs is a frequent event, specifically during Pb ANKA infection, this activity is redundant for the development of ECM.
To further analyze the mechanisms through which CD8+ T cells mediate ECM, we developed a tractable Ag-specific model, where CD8+ T cells with bona fide pathogenic activity can be visualized and tracked. Adoptive transfer of 106 SIINFEKL-specific OT-I CD8+ T cells into otherwise ECM-resistant P14 TCR transgenic mice, in which most CD8+ T cells express the receptor specific for the gp33 epitope of LCMV [52], led to the robust development of ECM when mice were subsequently infected with a GFP-SIINFEKL-expressing strain of Pb ANKA parasite (Pb-TG) (Fig 8A and 8B). Consistent with results obtained when examining the polyclonal T cell response (Fig 4), DsRed-expressing OT-I CD8+ T cells were recruited to the brain of P14 hosts in response to infection with Pb-TG (Fig 8C). These brain recruited OT-I CD8+ T cells were predominantly perivascular, as determined by their location relative to the fluorescent endothelium (Fig 8D). Perivascular OT-I CD8+ T cells were, however, more highly arrested than the polyclonal T cells with a mean speed of 2.36±2.32 μm/min and a mean arrest coefficient of 0.829±0.256 (Fig 8E, S11 Video). Differences in Ag-specific and polyclonal T cell behavior are likely due to the varied specificity of the complex polyclonal CD8+ T cell population for different parasite molecules, heterogeneously expressed by cells within the brain, [53]. Transferred OT-I CD8+ T cells behaved comparably in infected wild type and P14 hosts, indicating that their highly arrested behavior was not an artifact of the transgenic recipient (S7 Fig). As our results argue against the requirement for secondary activation by brain APCs to endow brain-infiltrating CD8+ T cells with pathogenic activity during Pb ANKA infection (Fig 3), it is likely that the arrested T cells are attached directly to antigen-presenting target cells, and that such interaction contributes to ECM pathogenesis. These results further emphasize that perivascular arrest of antigen-specific CD8+ T cells is a consistent signature of ECM.
It has recently been proposed that CD8+ T cells can directly interact with antigen expressing endothelial cells, and that this interaction may be a proximal event in causing endothelial cell apoptosis and ECM development [24, 26]. We, therefore, assessed the level of apoptosis within the brains of mice with ECM. Detection of apoptotic cells during intravital imaging was attempted by i.v. injection of an active-caspase3/7-reporter molecule, CellEvent. However, no CellEvent-positive cells were observed by transcranial microscopy, suggesting that cellular apoptosis is a rare event in the brain during ECM. This was confirmed by ex vivo examination of thick coronal sections from mice injected with CellEvent, allowing for a larger volume of the brain to be surveyed (Fig 9A), with only 11.2 cells/mm3 apoptotic cells found in the vasculature of the cortex of brains of mice with ECM, compared with 1.36 cells/mm3 in naïve brains (Fig 9B). Although rare, apoptotic cells could be found in immediate contact with parasite-specific CD8+ T cells (Fig 9C), suggesting that the CD8+ T cells were mediating the cellular apoptosis. The low level of cellular apoptosis in the brains of mice with ECM was further confirmed by immunofluorescent staining of thick cortical sections with antibodies against activated caspase 3 (Fig 9D–9F). Brains from mice subject to middle cerebral artery occlusion induced stroke were used to confirm successful activated caspase 3 staining (S8 Fig). Co-staining of activated caspase 3 and CD31 demonstrated that the rare apoptotic cells present during ECM were associated with the vasculature. By examination of morphology, apoptotic cells were identified as both endothelial cells (Fig 9G) and vasculature associated leukocytes (Fig 9H). Apoptotic cells may be rapidly cleared, particularly under physiological blood flow conditions, potentially explaining the low levels of apoptosis observed in the brains from mice with ECM. However, no differences were observed in either total number of vessels (Fig 9I and S9 Fig) or vessel area (Fig 9J and S9 Fig) in brains from mice with ECM compared with uninfected mice. Despite this, as previously reported, extensive vascular leakage was observed in ECM (Fig 9K). Thus, our results indicate that CD8+ T cells do not cause extensive vascular leakage via endothelial cell apoptosis; rather, our results support an alternative model in which perforin and granzyme B release, by perivascular parasite-specific CD8+ T cells, induces opening of intercellular junctions of the endothelium.
In this study, we have shown that CD8+ T cells accumulate at high levels in the perivascular space of the brain during malaria infection. The perivascular compartment of the brain is a hitherto understudied location in the study of ECM pathogenesis, even though, consistent with our results, leukocytes have previously been observed in this space during ECM using electron microscopy [54, 55]. Although, our intravital studies examined the superficial regions of the brain, perivascularly located CD8+ T cells were also observed around vessels deeper in the brain during ECM (S10 Fig), suggesting that this phenomenon is not unique to pial vasculature. Our results extend the study of Nacer et al, which used acute labeling with intravenous fluorescent antibodies to detect intravascular CD8+ T cells during ECM [20]. We have utilized a genetically encoded fluorescent protein, expressed in all T cells, coupled with the phenotypic analysis of T cells in the CNS to reveal a large population of extravasated, but still perivascular, CD8+ T cells that were previously unappreciated.
Interestingly, we observed accumulation of CD8+ T cells in the perivascular spaces of the brain during both ECM-inducing and non-inducing infections, indicating that the presence of perivascular T cells per se is insufficient to cause ECM. Instead, dynamic time-lapse movies revealed distinct behaviors of these cells in the perivascular compartment during the different infections, suggesting that it is T cell motility in the brain that correlates with their pathogenic activity during malaria infection. Infection with ECM-causing parasites resulted in a higher proportion of perivascular T cells, including Ag-specific CD8+ T cells that directly contribute to ECM, exhibiting behavior consistent with immunological synapse formation, including lower mean speed, higher mean arrest coefficient and lower mean confinement ratio. Whilst, further work is required to definitively resolve whether T cells interact with other cells within the brain during ECM through classical or atypical synaptic associations, these findings demonstrate the utility of two-photon microscopy in revealing information on important dynamic cellular behaviors of leukocytes that contribute to disease.
Perivascular CD8+ T cells were observed forming stable interactions with CX3CR1+/GFP cells, specifically during infection with ECM-causing Pb ANKA parasites. Such interactions between perivascular APCs and T cells, leading to secondary in situ reactivation, are known to be essential for the development of pathology in other neuro-inflammatory models such as EAE [43, 44]. However, perturbation studies using clodronate liposomes and the CX3CR1/iDTR system [51] to deplete CX3CR1+ cells demonstrated that this interaction is not critical for development of ECM. Thus, CD8+ T cells mediate ECM via interaction with other non-myeloid cell populations within the brain. These results extend previously published data that peripheral myeloid cells are not required for terminal ECM development [17, 19, 48, 49] to also show that resident or recruited myeloid cells positioned behind the tight endothelial barrier of the cerebral vasculature do not non-redundantly contribute to ECM pathogenesis. This difference between EAE and ECM is perhaps unsurprising when the location of disease-associated tissue damage is considered. Whilst in EAE, perivascular T cells must receive further stimulation to cross the glia limitans into the parenchyma, where the disease-associated pathology occurs [27, 56], in malaria the ECM-associated pathology is focused primarily to the cerebral vasculature [13], which is accessible to both luminal and perivascular CD8+ T cells. However, in contrast to luminal cells, it is also possible that arrested perivascular CD8+ T cells may exert a secondary activity during ECM by interacting with the glia limitans. This may induce signals into the parenchyma of the brain, contributing to the activation and injury of brain-resident cells including astrocytes, microglia and neurons, which is observed during ECM [57–59].
Thus, the critical questions are 1) why do T cells behave differently in the brains of ECM-inducing and non-inducing malaria infections, and 2) why can they specifically cause cerebral pathology only during Pb ANKA infection? We found that intracerebral CD8+ T cells possess an equally activated phenotype in mice with ECM-causing and non-ECM causing infections, indicating that upon migration to the brain, CD8+ T cells possess the intrinsic ability to mediate cerebral pathology, as long as they receive the necessary tissue signal. This result is in agreement with Howland et al. who reported that parasite specific CD8+ T cells with cytolytic potential [26] are found within the brains of mice infected with non-ECM causing parasites. Indeed, it is likely that T cells recruited to the brain during Pb ANKA and NK65 infections exhibit comparable antigen specificities, as a number of the most immunogenic antigens are conserved between Pb ANKA and other murine Plasmodium parasites [60, 61]. Moreover, CD8+ T cells specific for EIYIFTNI (F4: replication protein A1), IITDFENL (Pb2: bergheilysin) and SQLLNAKYL (Pb1: PbGAP50) epitopes are observed at similar frequencies in the brains of Pb NK65 and Pb ANKA infected mice [26, 61]. Thus, Howland et al. suggested that the ability of CD8+ T cells to mediate pathology specifically during Pb ANKA infection is the presentation of parasite antigen by microvessel endothelial cells in the brain during this infection. This hypothesis is consistent with Haque et al.’s report that antigen-specific CD8+ T cells migrate to the brain, but do not induce ECM until a critical antigen threshold is reached within the brain [11].
The source of malarial antigen and its mode of presentation to pathogenic T cells specifically during ECM-inducing malaria infections has been a subject of intense investigation. A number of studies have shown that, despite similar peripheral parasitemia in the blood, parasite accumulation within the brain is higher during infection with ECM-causing parasites than non-ECM causing parasites [26, 37]. Similarly, greater parasite accumulation is found in the brains of ECM-susceptible strains of mice than resistant strains [37]. In agreement with these studies, we observed significantly higher accumulation of Pb ANKA parasites than Pb NK65 parasites within the brain on day 7 of infection, when Pb ANKA infected mice developed ECM. Whilst these data support a major role for parasite accumulation in the brain in ECM-pathogenesis, Pb ANKA parasite accumulation was infrequent compared with the level of P. falciparum sequestration observed during CM. Moreover, consistent with Nacer et al [13], we failed to observe high levels of adhesion of luminal parasites to brain vascular endothelium during Pb ANKA infections by intravital imaging. Thus, although Pb ANKA sequestration does not lead to vessel occlusion during ECM, accumulation of pRBCs within the brain, and specifically within the perivascular space, may provide a localized source of antigen for cross-presentation by endothelial or associated cells during Pb ANKA infection, enabling perivascular CD8+ T cells to mediate their pathogenic activity. Notably very recently, Howland et al. [53] have also shown that brain endothelial cells are significantly more efficient at cross presenting Pb ANKA parasites than Pb NK65 or P. yoelii NL parasites, providing a further mechanism for increased antigen expression with the brain during Pb ANKA infection than in other non-ECM inducing infections. Detailing the properties and features of Pb ANKA parasites that cause them to reproducibly promote ECM, unlike other Pb isolates, require further investigation, especially as there appears to be little genetic polymorphism between them [15, 16].
Irrespective of how CD8+ T cells encounter antigen within the brain during Pb ANKA infection, the question remains as to the mechanism through which they mediate ECM pathology. The requirement for perforin and granzyme B in CD8+ T cell-mediated disruption of cerebral vascular integrity, leading to ECM, has previously led to the hypothesis that CD8+ T cells directly induce apoptosis of endothelial cells within the brain [11, 24, 25]. In support of this, perivascularly located T cells were closely associated with the abluminal surface of blood vessels in both infections but more so in mice infected with Pb ANKA than Pb NK65 parasites. However, although increased above levels seen in brains from uninfected mice, the number of apoptotic endothelial cells within the brains of mice with ECM appeared insufficient to account for the extensive vascular leakage, which occurs during ECM. Importantly, the low level of endothelial cell apoptosis during ECM, detected by two independent methods, was not due to rapid clearance of dying cells, as there was no widespread loss of vascular endothelial cells in brains of mice with ECM. The surprisingly low level of cellular apoptosis in the brain during ECM is potentially due to the rapid progression of the ECM syndrome and the late recruitment of relatively low numbers of activated CD8+ T cells. Thus, as CTLs appear to locate target cells through a stochastic search strategy, and CTL immunological synapses last between 30 minutes and 6 hours [62, 63], each individual recruited CTL may target and kill a limited number of endothelial cells before the animal succumbs to infection. Consequently, our results support a model whereby vascular leakage during ECM predominantly occurs at the level of the interendothelial tight junctions without, or long before, causing endothelial cell death [13]. Supporting this hypothesis, Suidan et al. have demonstrated the ability of intracerebral antigen-specific CD8+ T to initiate central nervous system vascular leakage through a perforin-dependent mechanism involving VEGF, that down-regulates tight junction proteins 12–24 hours before activation of the apoptotic caspase cascade [64]. Notably, VEGF is increased in the brains of mice with ECM [65] and humans with CM [66, 67] and has been shown to inhibit endothelial cell apoptosis [68–70]. This scenario provides a potential explanation for the rapid recovery from ECM and reestablishment of the vascular integrity that can occurs after administration of anti-malarial drugs, which would not be possible if vascular dysfunction was mediated through extensive loss of endothelial cells.
In summary, we have presented data that highlights the perivascular space as a site of significant CD8+ T cell accumulation during murine malaria infections. This site was additionally found to be a site for pRBC accumulation. Together, our results suggest that pathogenically relevant interactions between CD8+ T cells and their target cells, most likely cross presenting endothelial cells, may occur at the perivascular aspect of affected vessels, shifting the focus from intraluminal events to include this previously underappreciated location. We, therefore, propose that infection with malaria parasites results in accumulation of perivascular antigen-specific CD8+ T cells and antigen-dependent in-situ active engagement of the T cell receptor, only during Pb ANKA infection, leading to alteration of tight junction proteins, increased vascular permeability and death before the widespread occurrence of endothelial cell apoptosis (Fig 10).
Animal work in New York was carried out in strict accordance with the recommendation in the Guide for the Care for the Care and Use of Laboratory Animals of the Public Health Service (National Institutes of Health) and was approved by New York University School of Medicine Institutional Animal Care and Use Committee (IACUC). Animal work in the U.K. was approved following local ethical review by the Universities of Manchester and Glasgow Animal Procedures and Ethics Committees and was performed in strict accordance with the U. K Home Office Animals (Scientific Procedures) Act 1986 (approved H.O Project Licenses 70/6995 and 70/7293). All surgery was performed under anesthesia: ketamine (50 mg/kg), xylazine (10 mg/kg), acepromazine (1.7 mg/kg); or isoflurane (2% in O2 at 0.2 L/min).
The following mice were used in the study: At Skirball Institute Animal facility, C57BL/6 (H-2b) from the National Cancer Institute or Taconic Labs, OT-I [71], P14 [52], CFP [72], DsRed [73], CX3CR1CreER X Rosa26iDTR mice [51]. Both CFP and DsRed were expressed under the control of a chicken β-actin promoter and CMV enhancer cassette. CX3CR1CreER X Rosa26iDTR (referred to as CX3CR1-iDTR) mice express Cre in CX3CR1+ cells upon treatment with tamoxifen, inducing expression of the diphtheria toxin receptor (DTR) and, thus, sensitivity to diphtheria toxin (DT) in those cells. At the Universities of Glasgow and Manchester, C57BL/6 mice from Harlan and Charles River, UK, hCD2-DsRed [74], CX3CR1GFP/GFP [45], CX3CR1GFP/GFP X C57BL/6 (F1) and CX3CR1GFP/GFP X hCD2-DsRed+/+ mice (F1). In all cases, transgenic mice were fully backcrossed to a C57BL/6 background and were used between 6 and 12 weeks of age. Mice were maintained in specific pathogen-free conditions.
GFP-expressing P.berghei ANKA parasites (GFP expressed under control of elogation factor 1a [eEF1a] promoter) were a kind gift from Chris Janse (Leiden University Medical Center) [75]. GFP/SIINFEKL-expressing P.berghei ANKA (Pb-TG) parasites were a kind gift from William Heath (University of Melbourne) [22]. P.berghei NK65 parasites expressed GFP under control of the circumsporozoite promoter [76]. Parasites were maintained in liquid nitrogen and passaged through naive mice prior to being used to infect experimental animals. Experimental infections were initiated by i.v. inoculation with 104 or 106 pRBCs, depending upon the experiment, and infected mice were monitored for neurological symptoms (paralysis, ataxia, convulsions, and coma occurring between day 6 and 10 post-infection). Parasitemia was measured daily from day 3 p.i. by either examination of Giemsa-stained thin blood smears or by flow cytometric detection of DAPI stained GFP+ parasites (S11 Fig).
The induction and severity of ECM was assessed using the following well-defined grading system [77] 1: no signs; 2: ruffled fur/and or abnormal posture; 3: lethargy; 4: reduced responsiveness to stimulation and/or ataxia and/or respiratory distress/hyperventilation; 5: prostration and/or paralysis and/or convulsions. Stages 2/3 were classified as prodromal signs of ECM and stages 4/5 were classified as ECM.
Splenic DsRed+/+OT-I CD8+ T cells were purified (>95% purity) from a naïve DsRed+/+OT-I TCR Tg mouse using a Dynal Mouse CD8+ Negative Isolation Kit (Invitrogen). 104 and 106 DsRed+/+OT-I CD8+ T cells were transferred i.v. into C57BL/6 and P14 TCR Tg recipients, respectively, one day prior to infection with 106 Pb-TG-pRBCs, as described above.
Vascular leakage was detected as described previously, with some modifications [78]. Briefly, 50 μL 3% Evans blue/PBS (w/v) was injected i.v. into anesthetized mice and allowed to circulate for 5–6 hours. 100–150 μL blood was collected just before perfusion for serum sample. Mice were exsanguinated under KXA anesthesia by intracardial perfusion with 20 mL ice cold PBS. Each brain was removed, weighed, deposited in 500 μL formamide and incubated in darkness at 37°C for 48 hours to extract Evans blue. Formamide was then aspirated and Evans blue absorbance at 610 nm was measured for serum and brain samples with EnVision 2104 Multilabel Reader plate reader (PerkinElmer). Leakage was calculated as Evans blue concentration(brain) multiplied by extraction volume (500 μL), divided by Evans blue concentration(serum), divided by brain mass, divided by hours of circulation, yielding μL serum/g brain/hr.
Non-recoverable intravital transcranial imaging was performed utilizing adapted published protocols [79–81]. Infected mice were imaged on days 5, 6 or 7 post infection. For the imaging of mice with ECM, mice were selected only when they scored 3 or above using the grading system described above. Mice imaged on day 5 post infection showed no signs of ECM. The following numbers of mice were imaged to assess the polyclonal T cell response within the subarachnoid and perivascular spaces:—2 mice infected with Pb ANKA on day 5 p.i., before ECM development, 4 mice with ECM (score > 3) on day 7 p.i., 4 mice infected with Pb NK65 on day 7 p.i. and 4 uninfected mice. To visualize blood vessels, mice were injected (i.v.) with either 20 ng Evans blue in PBS or 10 μL Qtracker 705 non-targeted quantum dots (Invitrogen) in PBS prior to imaging. After exposure of the skull by removal of the scalp and periosteum, mice were immobilized in a stereotaxic apparatus. An imaging window (∼3-mm diameter) was created on the right parietal bone 2–3 mm lateral and posterior to bregma by thinning the bone with a micro-drill, under a dissecting microscope.
Two-photon transcranial microscopy was performed with either a LSM-7 MP or LSM-710 system (Zeiss), using a 20x W Plan-Apochromat water immersion objective (NA 1.0, Zeiss). Excitation wavelengths between 910 and 940 nm were generated by a tunable Ti-sapphire femtosecond pulsed laser. Fluorescent emission signals were detected using a combination of non-descanned fluorescence detectors. Emission signals were sequentially separated by dichroic mirrors and bandpass filters arranged in two configurations: 1) 740 and 625 dichroic mirrors (Qtracker 705), 490-nm dichroic mirror with a 485-nm shortpass (SHG), and 593-nm dichroic mirror in combination with 525/25 (GFP) and 585/22 (DsRed) bandpass filters (Semrock). 2) 442/45 (SHG) filter, 465-nm dichroic mirror with 483/35 (CFP/GFP) filter, 505-nm dichroic mirror with 538/36 (GFP/CellEvent) filter, 555-nm dichroic mirror with 610/70 (DsRed) filter and 660-nm dichroic mirror with 710/100 (Evans blue-albumin) filter (Chroma or Semrock). CFP and GFP signals were distinguished by assessing the ratio of 483/35 signal to 538/36 signal. Mice were maintained under anaesthesia throughout surgical and transcranial imaging procedures. Core body temperature was maintained at 37°C, thermostatically controlled by a rectal temperature probe. Perfusion of the cranial window with an isotonic solution was maintained throughout the imaging and provided a meniscus for the dipping lens objective. Preparation and imaging of an individual animal lasted up to 3 hours. Individual movies lasted between 15 and 35 minutes for uninfected and Plasmodium infected mice. For LCMV infected mice, movies lasted up to 1 hour. Imaging sessions started with the generation of a tile scan (1416x1416 μm) of the area in the center of the imaging window. This ‘map’ of the visible area included an average of 8.8±2.1 large vessels (30–100 μm) along with their associated smaller ancillary vessels. Further image acquisition was centered on these larger vessels, in non-overlapping regions of the tile scan. Vessels with T cells visible in the initial tile scan were selected for further image acquisition for motility analyses.
Two-photon time-lapse sequences acquisition (283x283x30 μm, 2 μm Z step) (Zen software, Zeiss) was performed with low laser power and short pixel dwell time. As a result, significant dark current shot noise was present in some of our images, which had a detrimental impact on our chosen automated tracking software. In order to eliminate this high frequency speckled noise, a multiscale, -undecimated "A Trous" wavelet transform [82] based on a 3x3x3 linear kernel was applied on each volume of interest. Each channel and each time-point were decomposed separately into 7 additive layers plus final residual layer. Noise was mostly contained in the first layer, and cells were best detected in the 4th layer (S12 Fig). By selecting the layers containing only cells, we produced time-lapse sequences which were devoid of noise and therefore suitable for cell tracking. Similarly, the blood vessel stacks were filtered by summing layers 3 to 7 of the wavelet decomposition. Values below 0 were ignored and a volumetric distance map was obtained by successive greyscale dilations and summations of this new image stack using a 3x3x3 structuring element. In the distance map, pixel values are proportional to the distance from the blood vessels and 0 inside the blood vessels. The volumetric distance map was used both to discard any cells inside the blood vessels, and for the remaining cells, to measure their distance from the blood vessels and perform motility analyses (S13 Fig). Imaris (Bitplane) or Volocity (Improvision) software packages were used track cells using a combination of automated and manual processing. Motility analyses were subsequently exported and processed in Excel. Mean speed was calculated as path length/time (μm/min). Mean instantaneous speed was calculated by determining speed of a tracked cell at each consecutive time-point. Arrest coefficient of a cell was defined as the percentage of time points when its instantaneous speed was <2 μm/min. The confinement ratio was calculated as track displacement (distance between start and end point)/track length. Cellular contacts were determined manually in 3D to determine a lack of space between interacting cells at each time point. Analyses were performed for cells with tracks of at least 5 time-points.
Brain sequestered leukocytes were isolated from PBS perfused mice as previously described [83]. Isolated brain leukocytes were surface stained with α-mCD8 (53–6.7), CD45 (30-F11), CD3 (17A2), CD11a (M17/4), CD11b (M1/70), CD69 (H1.2F3), CXCR3 (CXCR3-173), KLRG1 (2F1), CD44 (IM7), CD62L (MEL-14), ICOS (15F9), CD40 (3/23), CD80 (16-10A1), MHC I (28-8-6), CD19 (1D3), CD49b (DX5). Intracellular staining for granzyme B (GB11) was performed for 1 hr, after treatment with fix-perm (eBioscience). Dead cells were excluded using forward scatter and side scatter properties and/or LIVE/DEAD Fixable Blue Dead Cell Stain Kit (Life Technologies). Fluoresce minus one controls were used to set gates. Cells were analysed with a BD LSR II (Becton Dickinson) or MACSQuant (Miltenyi) using BD FACSDiva software (Becton Dickinson), and data was analyzed with FlowJo (Tree Star Inc.). All antibodies were from eBioscience and Biolegend
Macrophages were depleted during infection by the administration of clodronate-liposomes. Systemic depletion was performed from day 5 p.i. by injection (i.p.) of 300 μL clodronate liposomes. Depletion of perivascular macrophages, which are present behind the blood vessel wall and are not depleted by systemic administration of clodronate liposomes [84], was performed on day 5 p.i. by i.c.v. injection of 8 μL clodronate liposomes into the right lateral ventricle, adapted from Galea et al. [50]. Briefly, animals were anesthetized with isoflurane (2%) in O2 (0.2 L/min) and N2O (0.4 L/min) and craniectomy performed. I.c.v. injection was performed using a glass microneedle (co-ordinates from bregma: anterior–posterior -0.22 mm, lateral −1.0 mm, ventral −2 mm. Sub-cutaneous buprenorphine was prophylactically administered at 0.1 mg/kg. The i.c.v. protocol was optimized and the injection placement within the ventricle confirmed by assessing the circulation of trypan blue throughout the perivascular compartments.
CX3CR1+ cells were depleted from Pb ANKA infected mice as follows: CX3CR1-iDTR mice were initially infected with 106 Pb ANKA pRBCs i.v. On day 1 and 2 p.i., infected mice were treated daily with 5 mg tamoxifen dissolved in corn oil via oral gavage. On day 3, 4 and 5 p.i., mice were treated daily with 1 μg diphtheria toxin i.p. to thoroughly deplete DTR-expressing CX3CR1+ cells. CX3CR1+ cell depletion was confirmed by brain histology of CX3CR1-expressing microglia (Fig 7C). Mice were monitored daily for ECM symptoms and survival, as described.
For detection of GFP+ parasites, CD31 (MEC 13.3, BD Pharmingen), activated caspase 3 (Asp175, Cell Signalling), CD8 (YTS105.18, AbD Serotec) and Lectin (biotinylated Lycopersicon esculentum, Sigma) in 30 μm thick free floating brain sections: mice were sacrificed by exposure to a rising concentration of CO2. Spleens were removed before sequential transcardial perfusion with PBS and 4% PFA. Brains were removed and sequentially incubated at 4°C overnight in 20% sucrose/4% PFA and 20% sucrose/PBS. Brains were snap frozen with dry ice and sectioned on a freezing sledge microtome (Bright Instruments, Cambridge, UK) at a thickness of 30μm. Brain sections were stored in cyroprotectant solution (30% ethylene glycol, 20% glycerol in PBS) at -20° C until used. Free-floating sections were washed in PBS and then blocked in 10% goat serum (Sigma Aldrich, UK) in primary diluent (PBS and 0.3% Triton X-100) for 1.5 hrs. Combinations of primary antibodies in primary diluent were applied for 12–48 hrs at 4°C. Combinations of secondary antibodies in primary diluent (goat anti-rabbit 488 and 546, Life Technologies, goat anti-rat 647, Life Technologies and Streptavidin conjugated 546, Life Technologies) were applied for 1.5 hrs at RT. Cell nuclei were counterstained with DAPI (Life Technologies, UK). Sections were mounted onto gelatine coated slides and coverslipped in Prolong Diamond anti-fade mountant (Life Technologies, UK). Images were collected on an Olympus BX51 upright microscope using 20x/ 0.50 and 40x/ 0.75 Plan Fln objective and captured using a Coolsnap ES camera (Photometrics) through MetaVue Software (Molecular Devices). Images were processed using ImageJ (http://rsb.info.nih.gov/ij). Quantitation of parasite load, vasculature and apoptosis was performed in a semi-automated fashion using Image-Pro Premier (MediaCybernetics). ImagePro’s smart segmentation technology was utilized to ascribe areas of interest (i.e parasite/apoptotic cell/vessel) and background from an image. The software then automatically determined all regions of interest from the total image; providing data regarding number of regions of interest and total area these regions occupied in the image. These settings were globally applied to all samples ensuring a non-biased analysis of all samples.
For detection of CD31 and Iba1 in fresh frozen sections: mice were exsanguinated under KXA anesthesia by intracardial perfusion with 20 mL ice cold PBS supplemented with 10 U/mL heparin. For immunohistofluorescent staining, brains were embedded in Optimum Cutting Temperature compound (O.C.T., Tissue-Tek) and frozen in a dry ice/isopropanol bath. 6–8 μm thick coronal sections were cut with a Leica CM3050 S cryotome sectioning system and mounted onto SuperFrost slides. Mounted sections were fixed in acetone at -20°C for 5 min, allowed to air dry for at least 5 hours, followed by rehydration, in TBS (20 mM Tris, 150 mM NaCl, pH 7.5) with orbital shaking for 10 min at RT. Sections were permeablized by soaking in TBST (0.05% Tween-20/TBS) with orbital shaking for 15 min at RT, then blocked by addition of 5% BSA, 10% mouse serum, TBS for 30 min at RT. Following washing in TBST sections were stained in the dark for 1 hour at RT with rat α-CD31-AlexaFluor647 (MEC13.3, Invitrogen) or rabbit α-Iba1 (Wako) antibodies diluted in TBST, and washed with TBST. The Iba1-stained tissue was further probed with goat α-rabbit-AlexaFluor647. After washing probes away, cryosections were mounted with glycerol under coverglass, and sealed with nail polish. For unstained histology, each brain was removed and fixed in 4% PFA/PBS (w/v) at 4°C overnight. Brains were washed with PBS 3 times for 5 min each. 500-μm thick coronal sections were cut with a Pelco100 vibratome sectioning system (Ted Pella Inc.) and mounted onto Superfrost slides with glycerol and sealed with nail polish. Samples were imaged with a 25x LD LCI Plan-Apochromat objective (NA 0.8, Zeiss) and analysed using Photoshop CS5 (Adobe).
10 μL 2 mM CellEvent Caspase-3/7 Green Detection Reagent (Invitrogen) was injected i.v. into anesthetized mice and allowed to circulate for 30–60 minutes. Z stacks were captured by two-photon microscopy spanning a surface area of 4–5 mm2 and a depth of ∼150 μm. CellEvent-positive cells were counted and their number divided by the volume of brain tissue imaged.
All statistical analyses were performed using GraphPad PRISM (GraphPad Software, USA). Comparison between two groups was made using unpaired t tests, with Welch’s correction where needed. Comparison between multiple groups was made using a one-way ANOVA with Tukey’s test for multiple comparisons. Differences in survival were analysed using the Mantel-Cox log-rank test.
|
10.1371/journal.pbio.1001839 | A Novel Protein, CHRONO, Functions as a Core Component of the Mammalian Circadian Clock | Circadian rhythms are controlled by a system of negative and positive genetic feedback loops composed of clock genes. Although many genes have been implicated in these feedback loops, it is unclear whether our current list of clock genes is exhaustive. We have recently identified Chrono as a robustly cycling transcript through genome-wide profiling of BMAL1 binding on the E-box. Here, we explore the role of Chrono in cellular timekeeping. Remarkably, endogenous CHRONO occupancy around E-boxes shows a circadian oscillation antiphasic to BMAL1. Overexpression of Chrono leads to suppression of BMAL1–CLOCK activity in a histone deacetylase (HDAC) –dependent manner. In vivo loss-of-function studies of Chrono including Avp neuron-specific knockout (KO) mice display a longer circadian period of locomotor activity. Chrono KO also alters the expression of core clock genes and impairs the response of the circadian clock to stress. CHRONO forms a complex with the glucocorticoid receptor and mediates glucocorticoid response. Our comprehensive study spotlights a previously unrecognized clock component of an unsuspected negative circadian feedback loop that is independent of another negative regulator, Cry2, and that integrates behavioral stress and epigenetic control for efficient metabolic integration of the clock.
| The circadian clock has a fundamental role in regulating biological temporal rhythms in organisms, and it is tightly controlled by a molecular circuit consisting of positive and negative regulatory feedback loops. Although many of the clock genes comprising this circuit have been identified, there are still some critical components missing. Here, we characterize a circadian gene renamed Chrono (Gm129) and show that it functions as a transcriptional repressor of the negative feedback loop in the mammalian clock. Chrono binds to the regulatory region of clock genes and its occupancy oscillates in a circadian manner. Chrono knockout and Avp-neuron-specific knockout mice display longer circadian periods and altered expression of core clock genes. We show that Chrono-mediated repression involves the suppression of BMAL1–CLOCK activity via an epigenetic mechanism and that it regulates metabolic pathways triggered by behavioral stress. Our study suggests that Chrono functions as a clock repressor and reveals the molecular mechanisms underlying its function.
| Circadian rhythms with a period of approximately 24 h endow organisms with the ability to adapt to changes of solar light following earth's rotation. The mammalian circadian clock system consists of inputs from light and feeding, a core pacemaker located in a paired nuclei, called suprachiasmatic nucleus (SCN), and outputs including, but not limited to, cycles of locomotor activity, sleep–awake, and hormonal secretion. Disturbance of the biological clock causes not only sleep rhythm disruptions but also various pathological conditions such as cancer, metabolic, and psychiatric disorders [1]–[5].
The clock gene, period, was first identified in fly [6]–[8] and later in various organisms [9]. The molecular mechanism of circadian transcription was then found to be based on interconnected transcription–translation feedback loops (TTFL), conserved from prokaryotes to humans [10]–[12]. In mammals, the complex of positive elements BMAL–CLOCK (NPAS2) activates PER and CRY that repress their own transcription to form a negative feedback loop. An accessory feedback loop involves ROR and REV–ERBα, which regulate BMAL1 transcription positively and negatively, respectively, whereas BMAL1 activates REV–ERBα expression. More members of the circadian clock components have been identified [13]–[16]. Because of the complexity of circadian timekeeping, mathematical modeling has emerged as an important tool to understand data and make novel predictions [17]–[19]. In particular, a recently published mathematical model reproduces much of the known data on circadian timekeeping (e.g., mutant phenotypes) and correctly predicts the pharmacological manipulation of circadian rhythms [20],[21].
Although it is widely believed that the major components of the mammalian circadian clock have been identified, the search for additional clock components continues. In our previous study [22] and in others [23],[24], the systematic screening by using ChIP (Chromatin immunoprecipitation)–Chip and ChIP-seq revealed several uncharacterized genome-wide BMAL1 targets. Among strong BMAL1 binding sites, only one novel gene, Gm129, was found besides known clock and clock-controlled genes such as Per1, Per2, Cry1, Cry2, Dbp, and Tef. Gm129 expression shows a robust circadian rhythm antiphasic to Bmal1. In light of its circadian expression, Gm129, now renamed Chrono (ChIP-derived Repressor of Network Oscillator), appeared to be a core clock gene.
Here, we study the functional role of Chrono in the circadian clock. CHRONO binds to the promoters of clock genes and functions as a negative regulatory component of the circadian clock. In vivo loss-of-function of Chrono including an Avp neuron-specific knockout (KO) mouse model displays a longer circadian period of locomotor activity. We demonstrate that Chrono is a core-clock component similar to Cry2, although its repression mechanism operates through an independent pathway. In silico study using a modified Kim–Forger model predicts that the recently identified residual rhythmicity in the Cry1, Cry2 double KO [25], is dependent on Chrono. Remarkably, Chrono is involved in glucocorticoid receptor (GR)–mediated metabolic physiology. We conclude that Chrono is part of the negative feedback loop of the mammalian circadian clock and a potential link between the clock and stress metabolism.
Our previous ChIP-based genome-wide analyses using a core clock transcription factor BMAL1 identified hundreds of target molecules [22]. Among these targets, Gm129, now called Chrono, was one of the groups with the strongest binding including core clock proteins PER and CRY. Another ChIP-seq experiment using in vivo brain samples also identified Chrono as a BMAL1 target. Chrono exists only in mammals, is well conserved among mammals (Figure S1), and consists of 375 amino acids with no functional domains. To examine whether Chrono encodes a polypeptide, we performed an in vitro translation experiment. Bands of approximately 45 kDa (CHRONO) and 46 kDa (CHRONO–FLAG) were observed as an in vitro translation product (Figure 1A).
Our previous study showed robust circadian oscillation of Chrono mRNA in the mouse SCN and liver [22]. We further examined its expression in five different mouse peripheral tissues (heart, lung, stomach, kidney, and testis) by quantitative RT-PCR. After entrainment of mice housed for 2 wk under a 12–12 h light–dark (LD) cycle, samples were collected every 4 h starting at circadian time (CT) 0 (n = 3 at each time point) in the third dark–dark (DD) cycle. The temporal expression of Chrono transcripts in all tested tissues except testis displayed robust circadian rhythms peaking at approximately CT 12 (Figure 1B), which were antiphasic to Bmal1. This result supports that Chrono encodes a component of the circadian clock loops [26]. The ChIP-seq experiment using brain samples revealed that BMAL1 strongly binds to CpG islands on the Chrono promoter in vivo (Figure 1C). Studying differently sized Chrono promoter constructs (−195, −138, −87, −52, and −16 bp from the transcriptional start site (TSS) of Chrono/PGL3B) showed that the closest E-box to the TSS is necessary to generate circadian oscillation of Chrono in NIH3T3 cells and that all of the three E-boxes contribute to robust circadian rhythms (Figure 1D and E). These results suggest that BMAL1 strongly binds to the E-boxes on the Chrono promoter and regulates circadian expression of Chrono, making Chrono a novel clock gene.
We next asked whether CHRONO is also expressed rhythmically at the protein level. We prepared liver samples at CT 2, 8, 14, and 20. We raised a specific antibody against the CHRONO protein. CHRONO showed circadian rhythm antiphasic to BMAL1 as in the mRNA expression (Figure S2A and C). This oscillation was observed in both the mouse CHRONO antibody we generated and the human C1orf51 antibody (ab106120, Abcam) (Figure S2B).
Because CHRONO showed a similar rhythmic expression profile to other core clock proteins, we asked if CHRONO binds directly with clock proteins. Various clock proteins with tags were expressed in COS7 cells and the expression was assessed by immunoprecipitation (IP) and blotting with anti-tag antibodies. CHRONO bound to BMAL1, PER2, CRY2, and DEC2 but not to PER1, CRY1, and DEC1 (Figure 2A and B and Figure S2D). Among these interactions, we asked if CHRONO–BMAL1 binding occurs endogenously in vivo. We observed a band of BMAL1 in the CHRONO antibody IP from mouse liver lysate that was absent in the IP from Chrono-deficient mouse liver (Figure 2C). These results suggest that CHRONO endogenously binds to BMAL1 in vivo.
Next, we asked how Chrono is involved in circadian transcription. The luciferase activity of the Per2 promoter (∼−2,817+110 bp from TSS/PGL3B) in NIH3T3 cells was repressed by co-expression with Cry2 and Dec2 (Figures 2D and S3A). The basic transcription activity of Per2 was increased by Bmal1 and Clock co-expression, and this activation was repressed by Chrono as well as Cry2. This repression was also seen in the Chrono promoter (−1,333 bp from TSS/PGL3B) (Figure S3B). Moreover, overexpression of Chrono reduced the transcriptional amplitude of expression on the Dbp promoter, just as Cry2 (Figure S3C and D). These results suggest that Chrono functions as a negative element of circadian transcription, similar to Cry2.
Histone modification by histone deacetylase (HDAC) is one of the mechanisms of transcriptional regulation [27]. HDAC is often recruited during the transcriptional repression process. We hypothesized that CHRONO is in a repressor complex that includes CRY2. To investigate the potential role of HDAC in Chrono-mediated transcriptional repression, we treated cells with trichostatin A (TSA), an HDAC inhibitor, in the reporter assay. Chrono, as well as Cry2, repressed the enhanced luciferase activity of the Per2 promoter. However, in the presence of TSA, Chrono did not repress activity but rather enhanced activity, whereas Cry2 did not change its repression (Figure 2E). To confirm the involvement of HDAC in Chrono-mediated transcriptional repression, we investigated the interaction of CHRONO with HDAC1. We expressed and showed co-IP of CHRONO–HA and HDAC1–FLAG in COS7 cells, indicating that CHRONO is bound with HDAC1 (Figure 2F and G).
We then asked if endogenous CHRONO participates in clock function as a core clock component. A ChIP experiment with the CHRONO antibody in NIH3T3 cells after induction with dexamethasone showed endogenous binding of CHRONO to the E-boxes on Per2 (Figure 3A) and Dbp (Figure 3B) promoters. The levels of chromatin occupancy around the E-boxes on Per2 and Dbp were significantly different between 32 h and 44 h after dexamethasone stimulation (Figure 3A and B, *p<0.05, **p<0.01). Moreover, endogenous CHRONO occupancy showed circadian oscillation antiphasic to BMAL1 (Figures 3C and S3G). These results strongly suggest that Chrono behaves as an auto-regulated clock component.
To evaluate the physiological and circadian clock function of Chrono in vivo, we generated Chrono KO mice by using a gene trap method (Figures 4A and S4). After entrainment in the 12–12 h LD condition, the locomotor activity rhythm under DD was recorded. In DD the average circadian period length of wild-type (WT) mice was 23.81±0.08 h (mean ± standard deviation, n = 13), whereas that of Chrono-deficient mice was significantly longer (23.96±0.11 h, n = 12) (*p<0.001; Student's t test) (Figure 4B–D). To logically confirm that the behavioral Chrono KO phenotype is an outcome of the observed biochemical and in vitro characteristics of Chrono (Figure 2), we adopt a mathematical modeling approach. We used a recently developed mathematical model of mammalian circadian clock (Kim–Forger model) because this model successfully reproduced and predicted the circadian period change in response to mutations of clock genes (e.g., Per1/2, Cry1/2, Bmal1, Clock, and etc.) and pharmacological inhibition of kinase (e.g., CK1δ/ε) [20],[21]. The model is extended to include biochemical mechanisms of Chrono, such as binding with other clock components (Figure 2A and B), transcriptional repression by Chrono (Figure 2D–G), and rhythms of Chrono (Figure 1B) (see details in Text S1). Although CHRONO acts similarly to CRY2, we also incorporate key differences, including their very different mRNA time profiles (Figure S5) and the fact that CHRONO does not bind PER1. When the transcription of Chrono is inhibited in the model, the model predicts that Chrono KO lengthens the period (Figure 4D, right), which indicates that the biochemical mechanisms we have identified for Chrono-mediated repression of BMAL1–CLOCK (Figure 2D) are expected to cause the Chrono KO phenotypes. There was no difference of basal locomotor activity between WT and Chrono KO mice during the light phase and dark phase after 1 wk in LD (p>0.5; Student's t test) (Figure 4E). We also examined how Chrono KO mice responded to shifts in the LD cycle. When the lighting cycle was advanced 6 h, both WT and Chrono KO mice re-entrained progressively over 10 d, and there were no difference between the genotypes (Figure S6A and B). Moreover, no induction of Chrono mRNA in the SCN was observed after light stimulation for 30 min (Figure S6C).
Because Chrono is a putative repressor in the circadian clock mechanism, we next asked whether the expression of clock genes is altered after the deletion of Chrono in mice. In mouse embryonic fibroblasts (MEFs) derived from Chrono KO mice (Figure S7A), Per2 expression was increased at 48 h and 52 h after dexamethasone stimulation (Figure S7B), compared with WT MEFs. In the liver derived from Chrono KO mice (Figure S7C), Per3, Cry1, Dbp, and Rev–erbs were increased at CT12 (*p<0.05, **p<0.01, Student's t test), consistent with the idea that Chrono is involved with negative feedback regulation of the core clock component via E-box. The similar trend was observed in the Chrono-deficient SCN (Figure S7D). Moreover, Acetyl-Histone H3 occupancy on Per2 and Dbp promoters was enhanced in Chrono KO MEFs compared to WT (Figure S7E and F; **p<0.01, Student's t test). This result suggests that Chrono changed the epigenetic modification of cells with the expression status.
Chrono's role in circadian timekeeping seems to mimic that of Cry2 both in vivo and in cells. Because CHRONO interacts with CRY2 and DEC2, we hypothesized that CHRONO represses BMAL1–CLOCK activity only when partnered with CRY2 or DEC2 to form a complex. To confirm this hypothesis, we performed a luciferase reporter assay using Cry2 KO MEFs (Figure 5A) or Cry2 knockdown (Cry2sh) NIH3T3 cells (Figures 5B and S8) and Dec2 KO MEFs (Figure S3E). The BMAL1–CLOCK complex induced Per2 transcription in all cells and Chrono repressed the BMAL1–CLOCK activity in all cells, including those without Cry2 and Dec2, indicating that CHRONO acts as a transcriptional repressor via an independent pathway from CRY2 and DEC2. Moreover, we established a stable NIH3T3 cell line with Bmal1 promoter luciferase and Cry2 knockdown. The rhythm of fibroblasts with overexpression of Flag as a control showed a robust circadian oscillation with a lengthened period (27.29±0.06 h; mean ± S.E.M.; n = 4) due to Cry2 knockdown. As expected, constitutive overexpression of Cry2 driven by the CMV promoter partially restored the period to a shorter value (26.91±0.07 h; n = 4; *p<0.01, Welch's t test) in the Cry2-sh/NIH3T3 cells. Consistent with the idea that Chrono mimics the Cry2 phenotype, overexpression of Chrono similarly shortened the period (26.66±0.15 h; n = 5; *p<0.01, Welch's t test) of oscillation in the Cry2-sh/NIH3T3 cells (Figure 5C and D). These relative modulations in periods are predicted in parallel by the extended Kim–Forger model (Figure 5E), under the simulated experimental conditions of Cry2 knockdown to 30% of its original value and/or constitutive overexpression of either Cry2 or Chrono. These results indicate that the circadian phenotypes of Chrono are similar to Cry2, although the repression mechanisms operate via different pathways.
Given the phenotypic similarities between Chrono and Cry2, we postulated that Chrono could overtake the role of Cry2 under KO conditions. To test this idea, we generated Chrono, Cry1 double KO mice by mating Chrono KO with Cry1 KO mice. The Cry1, Cry2 double KO mouse shows a circadian arrhythmicity in DD [28], and we expected a Chrono, Cry1 double KO would exhibit similar arrhythmicity. However, the circadian locomotor activity persisted under DD. In DD condition, Cry1 KO mice showed a shorter period than WT, similar to Cry1, Chrono double KO mice (Figure 5F). There was no significant difference of period between Cry1 KO and Cry1, Chrono double KO mice (Figure 5G). These results are predicted by our mathematical model, which shows short period rhythms under the double Chrono, Cry1 KO (22.95 h) and Cry1 single KO (22.89 h), respectively (Figure 5G, right). The model also predicted that the Chrono KO does not compromise the amplitude of Per2 mRNA rhythms (Figure 5H). Interestingly, the model predicts initially oscillating but damping Per2 rhythms in Cry1, Cry2 double KO (Figure 5H, blue). However, the triple, Chrono, Cry1, Cry2 KO completely removed any residual oscillations (Figure 5H, red). These results led us to conclude that, although CHRONO acts as a repressor of the BMAL1–CLOCK complex just as CRY2, the mechanism of action of CHRONO is independent of CRY2 and that CHRONO may be required for rhythmicity in certain mutant backgrounds.
The Cryptochromes (Cry1 and Cry2) have been reported to mediate rhythmic repression of the GR [29]. Because Chrono behaves as a transcriptional repressor with an effect similar to Cry2, we next verified the physiological role of the Chrono KO against stress responses. We first examined the expression of serum/glucocorticoid-regulated kinase 1 (Sgk1) after dexamethasone stimulation. We found that Sgk1 mRNA was up-regulated in Chrono KO MEFs when compared to WT MEFs after 1-h and 4-h exposures to dexamethasone (Figure 6A, *p<0.05, **p<0.01, Student's t test). The relative expression of Sgk1 was also activated by 4-h exposure to dexamethasone (Figure 6B; *p<0.05, Student's t test). Serum corticosterone levels in vivo were also robustly increased in Chrono KO when compared with WT mice after 0.5 h and 1 h of restraint stress (Figure 6C; *p<0.05; one-way ANOVA followed by Tukey–Kramer's multiple comparisons test, as follows, *p<0.05 versus 0.5 h time after restraint stress; **p<0.01 versus 1 h time after restraint stress). In a WT background, Chrono mRNA was up-regulated in MEFs (Figure 6D; **p<0.01, Student's t test) and hypothalamus (Figure 6E; *p<0.05, Student's t test) after 1 h of exposure to dexamethasone and restraint stress, respectively, whereas Cry2 mRNA was not significantly increased under restraint stress in both WT and Chrono KO hypothalamus as well as MEFs (Figure S9). Furthermore, the IP-Western experiment indicated CHRONO endogenously interacted with GR in vivo (Figure 6F). These results suggest that Chrono is up-regulated by stress-dependent behavior and contributes to repressive function in the glucocorticoid response.
To identify the role of Chrono in the SCN, we generated mice carrying a conditional Chrono KO allele using the Cre-loxP system (Figures 7A and S10). Avp-Cre mice express Cre specifically in arginine vasopressin (AVP) neurons (GENSAT project) [30]. AVP is the most abundant neuropeptide in the hypothalamus and specifically in the SCN [31],[32]; therefore, Avp-Cre Chronoflx/flx mice can be considered as SCN-targeted Chrono KO mice. The average locomotor activity rhythm of Avp-Cre Chronoflx/flx male mice (24.04±0.12 h, n = 6) was significantly longer than that of Chronoflx/flx littermates (23.84±0.06 h, n = 8) (*p<0.01, Student's t test) (Figure 7B–D). The basal locomotor activity was measured during the light phase and the dark phase after 1 wk in LD by the infrared beam breaking. The activity amount of the Avp-Cre Chronoflx/flx mouse was not altered compared with Chronoflx/flx mouse (Figure 7E). These results demonstrate that Chrono expression in the Avp neurons plays a central role in behavioral rhythms.
In this study, we characterized a gene called Chrono that serves as a component in the negative arm of the core mammalian circadian clock. Indeed, based on our genomic analysis of circadian promoters, CHRONO appears to be one of the last remaining components of the clock. The most interesting feature of CHRONO is its regulation by epigenetic control and behavioral stress. These findings place CHRONO in a central position to couple stress metabolism to clock regulation.
Our results suggest that CHRONO operates as a repressor of the core circadian feedback loop through the recruitment of HDAC. The repressive effects of CHRONO were also seen in Cry2 KO MEFs and Dec2 KO MEFs, which shows that CHRONO repression does not require cooperative interactions with CRY2 or DEC2 [33],[34]. It has been reported that the recruitment of histone-modifying enzymes is regulated in a circadian manner [35]–[39] and some complexes are observed in transcriptional repression as in PER and SIN3–HDAC [14]. HDAC1 rhythmically bound to the Per2 promoter (Figure S3F), and this result was consistent with a recent report [40]. Acetyl-Histone H3 occupancy around E-boxes was enhanced in Chrono KO MEFs compared to WT (Figure S7E and F), suggesting that Chrono also has the potential to form complexes with other histone-modifying enzymes. However, HAT (histone acetyltransferase) activity by CLOCK was not affected by CHRONO in vitro experiments (Figure S11). Further mechanistic study is required. CHRONO endogenously binds to the E-boxes of circadian genes with circadian rhythmicity, suggesting that Chrono is a core circadian repressor that operates as an auto-regulated component of the clock.
The Chrono KO mouse showed a lengthened circadian period similar to the Cry2 KO mouse [28]. Overexpression of Chrono restored the period of Bmal1-luc oscillations in Cry2 knockdown cells in a manner comparable to Cry2 overexpression. On the other hand, the Chrono, Cry1 double KO mouse did not show arrhythmic behavior as seen in the Cry1, Cry2 double KO mouse [28]. Together, these results suggest that Chrono plays a similar but independent role from Cry2. A conditional Chrono KO mouse driven by the Avp promoter (AvpCre Chronoflx/flx) also showed a lengthened circadian period. This points to Chrono's central role in the core clock and that its expression in Avp neurons is critical for the determination of behavioral circadian period. It also demonstrates that the AvpCre system (GENSAT line number QZ20_CRE) can potentially prove useful in dissecting the output pathway of the SCN that may perform coding by internally distributed periods [41].
The role of Chrono was also tested in silico by modeling mechanisms of Chrono-mediated repression of BMAL1–CLOCK (Figure 2D) within the most comprehensive and realistic mathematical model of the mammalian circadian clock [20]. The extended model successfully predicted that the reduced (or enhanced) Chrono expression results in a lengthened (or shortened) period (Figures 4D and 5E), matching the phenotypes of Chrono KO or overexpression (Figures 4D and 5D). This indicates that the proposed biochemical mechanisms of Chrono-mediated repression of BMAL1–CLOCK (Figure 2D) are consistent with the overall phenotypes observed when Chrono levels are changed.
Chrono is likely to be a physiological response-dependent regulator. In response to dexamethasone application, Sgk1 was up-regulated in Chrono KO MEFs when compared with WT MEFs. In vivo serum corticosterone levels were also increased in the Chrono KO mouse when compared with the WT mouse under restraint stress and CHRONO itself interacted with the GR. Along with the observation that Chrono mRNA expression was induced by the stress response in MEFs and hypothalamus, we conclude that Chrono is a potential repressor activated by behavioral stress and can couple the clock with the hypothalamic–pituitary–adrenal (HPA) axis. Our previous results showed acute physical stress also elevated Per1 mRNA through a glucocorticoid-responsive element [42]. Further experimental work is needed to reveal the detailed molecular interactions between the circadian clock and the HPA axis.
Our discovery of the role of Chrono opens up many possibilities for future work. Our mathematical modeling raises the interesting possibility that the Per2 oscillations from Cry1, Cry2 double KO neonatal SCN [25] is mediated by Chrono. In further studies, it would be interesting to see if the Cry1, Cry2, Chrono triple KO eliminates the ability to oscillate in neonatal SCNs as predicted by our model (Figure 5H). Further work should also examine the role of Chrono in linking the HPA to the circadian clock. Taken together, we conclude that Chrono is a novel circadian clock gene that acts as a repressor in the circadian system and modulates physiology.
All protocols of animal experiments followed in the present study were approved by the Animal Research Committee of Hiroshima University and Animal Care and Use Committees of the RIKEN Brain Science Institute.
NIH3T3, COS7 cells and MEFs were maintained in Dulbecco's Modified Eagle Medium (DMEM; Nacalai Tesque, Kyoto, Japan) supplemented with 10% fetal bovine serum (FBS) and penicillin–streptomycin antibiotics at 37°C and 5% CO2.
NIH3T3 and MEF cells were stimulated with 100 nM dexamethasone containing medium at each time point. Cells were fixed in 1× PBS containing 0.5% formaldehyde. Glycine was added to a final concentration of 0.125 M, and the incubation was continued for an additional 15 min. After washing the samples with ice-cold phosphate-buffered saline, the samples were homogenized in 1 mL of ice-cold homogenize buffer (5 mM PIPES [pH 8.0], 85 mM KCl, 0.5% NP-40, and protease inhibitors cocktail) and centrifuged (15,000× g, 4°C, 5 min). The pellets were suspended in nucleus lysis buffer (50 mM Tris-HCl [pH 8.0], 10 mM EDTA, 1% SDS, protease inhibitors) and sonicated 20 times for 30 s each time at intervals of 60 s with a Bioruptor (Diagenode, Inc.) or sonicated 10 times for 10 s each time at intervals of 50 s with a MICROSON (Misonix, Inc.) for brain samples. The samples were centrifuged at 15,000 rpm at 4°C for 5 min. Supernatants were diluted 10-fold in ChIP dilution buffer (50 mM Tris-HCl [pH 8.0], 167 mM NaCl, 1.1% Triton X-100, 0.11% sodium deoxycholate, protease inhibitor).
Whole brain samples from mice (C57BL/6J) were used for ChIP-seq. BMAL1-bound DNA was purified by SDS-PAGE to obtain 150–200 bp fragments and sequenced on an Illumina GA sequencer at the Research Center of Research Institute for Radiation Biology and Medicine (RiRBM), Hiroshima University. We generated 15,000–20,000 clusters per “tile,” and 26 cycles of the sequencing reactions were performed according to the manufacturer's instructions. The identification of each DNA fragments was performed using Genome Studio software (Illumina Inc.).
CHRONO, BMAL1, HDAC1 (ab7028, Abcam), Acetyl-Histone H3 (06-599, Millipore), and IgG-bound DNA were used for quantitative real-time reverse-transcription PCR (RT-PCR). The primers were designed for amplifying the E-box-like regions in Per2 and Dbp promoters.
NIH3T3 cultures at the concentration of 1×105 cells in Opti-MEM (Gibco) supplemented with 10% FBS in a 35 mm dish were transfected with the desired plasmids by using the Lipofectamine reagent (Invitrogen). The medium was exchanged 24 h after transfection with 100 nM dexamethasone containing medium, and 2 h later this medium was replaced with Opti-MEM supplemented with 1% FBS and 0.1 mM luciferin–10 mM HEPES (pH 7.2). Bioluminescence was measured by using the IV-ROMS (Hamamatsu Photonics) as described previously [26],[43].
NIH3T3 cells and MEFs were cultured and transfected with the desired plasmids by using Lipofectamine 2000 (Invitrogen) or Nucleofection (Lonza). Cells were harvested 24 h after transfection, and cell lysates were prepared and then used in the dual luciferase assay system (Promega). For exogenous expression, we transfected cells with pcDNA3 driven by ubiquitous cytomegalovirus (CMV) promoter.
Rabbit antibody against HA-tag and mouse antibodies against Flag-tag and Myc-tag were subjected to Western blot according to the manufacturer's protocol. For IP, COS7 cells transfected with the desired plasmids by using Lipofectamine 2000 (Invitrogen) were lysed in TNE buffer with protease inhibitor. Following the standard protocol, lysates were precleared with Dynabeads Protein G (Invitrogen) and then immunoprecipitated with rabbit anti-HA antibodies (Cell signaling), mouse anti-Flag antibodies (Sigma), or mouse anti–c-Myc antibodies (Sigma). After washing three times, the precipitates were resuspended in the 2× SDS-PAGE sample buffer, boiled for 5 min, and run on a 10% SDS-PAGE gel followed by Western blot analysis. Immunoreactive bands were detected by ODYSSEY Infrared Imaging System (LI-COR). These experiments were independently performed three times. The intensity of the band (BMAL1 and CHRONO rpar; was calculated by using ODYSSEY Infrared Imaging System.
Circadian rhythms of locomotor activity were analyzed as previous described [43]. Each mouse was individually housed for 2 wk in 12–12 h LD cycles, and then for 4 wk in constant darkness (DD). Locomotor activity was monitored with a locomotor activity recording apparatus (Biotex, Kyoto, Japan) that measures events of infrared beam breaking in 1-min bins. The data from the first week under DD were used to estimate the period of circadian locomotor activities of WT and Chrono KO mice. For RNA sampling, dissected tissues were immediately frozen in liquid nitrogen and stored at −80°C until processing.
The experiment was performed as described previously [44]. Male ICR mice (SLC, Japan) were exposed to an incandescent light stimulus (1,000 lux, 30 min) at CT16 in the second DD cycle. Animals were sacrificed 60 min after the initiation of the light exposure. Total RNA was prepared from six pooled pairs of SCN at each time point using PicoPure RNA Isolation kit (Applied Biosystems).
CHRONO protein was synthesized using the TNT T7 Coupled Reticulocyte Lysate System (Promega). We added 2 µg of template DNA (Chrono/pcDNA3 or Chrono–Flag/pcDNA3) to an aliquot of the TNT Quick Master Mix and incubated it in a 50 µl reaction volume for 60–90 min at 30°C. Synthesized proteins were detected by 10 mCi/ml (specific activity, 30 TBq/mmol) [35S]methionine, and resolved on SDS-PAGE (10%) gels using one-fifth of each translation reaction product mixed with an equal volume of sample buffer (15% glycerol, 5% β-mercaptoethanol, 4.5% SDS, 100 mM Tris-Cl, pH 6.8, 0.03% bromophenol blue). Gels were fixed, dried, and exposed to Hyper-film (Amersham) for 16 h to 2 d. The molecular masses (in kDa) of the translated proteins were derived using standard curves generated from protein size standards (Bio-Rad).
Each quantitative real-time RT-PCR was performed using the ABI Prism 7900HT sequence detection system as described previously [42],[45]. The PCR primers were designed with the Primer Express software (Applied Biosystems). The reaction was first incubated at 50°C for 2 min and then at 95°C for 10 min, followed by 40 cycles at 95°C for 15 s and 60°C for 1 min.
Blood samples were collected by tail bleeding at time points of 0, 0.5, 1, 2, and 4 h after restraint stress. Corticosterone in mouse serum was measured using YK240 corticosterone enzyme immunoassay (EIA) kit (Yanaihara Institute).
BLOCK-iT Lentiviral RNAi System (Invitrogen) was used for RNAi experiments. NIH3T3 cells at the density of 5×104 were infected with a lentiviral vector and cultured. We selected stably transfected cells with zeocin. The shRNA/NIH3T3 cells were infected with Bmal1 promoter-driven luciferase lentiviral vector and selected for stable expression with blasticidin.
Antibody against BMAL1 was generated as described previously [22]. Purified glutathione S-transferase (GST) –Gm129 N-terminal (amino acids 1 to 187) protein was produced as a recombinant protein in the competent cells BL21 (DE3) (Stratagene). After removing GST by using GSTrap FF and PreScision Protease (GE Healthcare), the produced antigen was used to immunize rabbits. The antiserum was subjected to affinity purification using Affi gel 10 (Bio-Rad) conjugated with the antigen. The anti-CHRONO antibody recognized its target protein in immunochemical analysis (Figure S4D).
C57BL/6 gene trap ES cell clone IST11761C7 was an embryonic stem cell provided by the gene trap method [46]. Long terminal repeat (LTR) –splice acceptor–βgeo–polyA–LTR sequence was inserted between exons 1 and 2 in one allele of the Chrono (Gm129) gene (Figure 4A) (TIGM, Texas A&M Institute for Genomic Medicine). In substitution for mRNA producing CHRONO protein, mRNA producing fusion protein of the neomycin-resistant gene product and β-galactosidase (B-geo) was transcribed from this variation allele. We generated chimeric mice from C57BL/6 gene trap ES cell clone IST11761C7, mated the chimeric mice with WT C57BL/6 (+/+), and subsequently obtained heterozygous KO mice (+/−). Ultimately, we obtained homozygous KO (−/−) mice by mating heterozygous KO mice. Absence of Chrono mRNA and CHRONO protein was confirmed by PCR, RT-PCR, and Western blotting (Figure S4B–D).
WT and Chrono KO MEFs were stimulated with 100 nM dexamethasone containing medium. After 32 h, mRNA were extracted by TRI reagent (Molecular Research Center, Inc.) every 4 h. Quantitative real-time RT-PCR was described above. Adult WT and Chrono KO mice were exposed to 2 wk of LD cycles and then kept in complete darkness as a continuation of the dark phase of the last LD cycle. mRNA expression was examined CT12 and next CT0 from the third DD cycle.
The Chrono conditional KO mice (Accession No. CDB0913K; http://www.cdb.riken.jp/arg/mutant%20mice%20list.html) were generated as described (http://www.cdb.riken.jp/arg/Methods.html). To generate a targeting vector, genomic fragments of the Chrono locus were obtained from the RP23-385O12 BAC clone (BACPAC Resources). A 950 bp region containing exons 2 and 3 of the Chrono gene was flanked by loxP sites (Figure 7A). Targeted ES clones were microinjected into ICR eight-cell stage embryos, and injected embryos were transferred into pseudopregnant ICR females. The resulting chimeras were bred with C57BL/6 mice, and heterozygous offspring were identified by PCR. Primers for 5′ loxP were used—forward E1F (5′-CAGACAGTGAAGAAGCTGCATA-3′) and reverse loxR2 (5′-CAGACTGCCTTGGGAAAAGC-3′)—yielding no product for WT allele and 603 bp products for the targeted allele, respectively. Primers for 3′ loxP were used—forward loxF1 (5′-GGCATGGGCTATTCTGTTTG-3′) and reverse loxR1 (5′-TTGAGGGAAACAGCAGAGGT-3′)—yielding 121 bp products for WT allele and 179 bp products for the targeted allele, respectively (Figure S10A and C).
We incorporated Chrono into a recently developed mathematical model of the mammalian circadian clock [20] based on our biochemical findings. We assumed that the mRNA degradation rate of Chrono is the same as Per2, as the Chrono mRNA time course is similar to Per2 mRNA [42]. We further assumed the time course of PER2–CHRONO nuclear entry is similar to the PER–CRY complexes. However, as found in our experiments, the CHRONO protein binds to the PER2 protein but not to PER1 protein (Figure 2A) and that the CHRONO protein interacts with BMAL1–CLOCK to inhibit its E-box activation on Per1, Per2, Cry1, Cry2, and Rev–erbs promoters. In the model, we did not consider binding between CHRONO and CRY2, as BMAL1–CLOCK repression by CHRONO did not depend on the presence of CRY2 (Figure 5). In total, 38 variables were newly added to the Kim–Forger model, which account for all the possible complexes involving CHRONO (details of computer simulation are described in Text S1, Tables S1, S2, S3, and Appendix S1). All the simulations were performed with Mathematica 8.0 (Wolfram Research).
|
10.1371/journal.pgen.1008325 | NKILA represses nasopharyngeal carcinoma carcinogenesis and metastasis by NF-κB pathway inhibition | The role of long non-coding RNA (lncRNA) in the progression of Nasopharyngeal carcinoma (NPC) has not been fully elucidated. The study was designed to explore the functional role of NKILA, a newly identified lncRNA, in the progression of NPC. We performed a lncRNA expression profile microarray using four NPC and paired para-cancerous tissues. NKILA was identified as a potential functional lncRNA by this lncRNA expression profile. We used 107 paraffin-embedded NPC tissues with different TNM stages to detect the expression of NKILA and analyzed the survival data by Log-rank test and Cox regression. The role of NKILA and its underlying mechanisms in the progression of NPC were evaluated by a series of experiments in vitro and vivo by silencing or expressing NKILA. Compared with control tissues, NKILA expression was identified to be decreased in NPC tissues. Low NKILA expression was correlated with unfavorable clinicopathological features and predicted poor survival outcome in NPC patients. After adjusting for potential confounders, low expression of NKILA was confirmed to be an independent prognostic factor correlated with poor survival outcomes. Furthermore, we found that NKILA overexpression in high-metastatic-potential NPC cells repressed motile behavior and impaired the metastatic capacity in vitro and in vivo. In contrast, RNAi-mediated NKILA depletion increased the invasive motility of cells with lower metastatic potential. Further experiments demonstrated that NKILA regulated the metastasis of NPC through the NF-κB pathway. Taken together, NKILA plays vital roles in the pathogenesis of NPC. The unique histological characteristics of NPC indicate that local inflammation plays a vital role in carcinogenesis of nasopharyngeal carcinoma.
| NF-κB is a pivotal link between NPC and inflammation. Importantly, NF-κB was found to be overexpressed in nearly all NPC tissues, and inflammatory cytokines have also been observed in NPC tissues. Inflammatory cytokines promote the susceptibility of NPC cells to metastasize via constant NF-κB activation. Here, we found that NKILA, a newly identified lncRNA, is upregulated by inflammatory cytokines and is significantly downregulated in NPC. By a series of in vitro and in vivo experiments, we show that NKILA exerts its effect as a tumor suppressor via inhibiting tumorigenesis and metastasis of NPC. Further studies indicate that NKILA regulates the metastasis of NPC through NF-κB pathway. Our research demonstrates that NKILA plays a critical role in the progression of NPC. These findings are particularly important as they provide new insights into the effects of inflammation on the biology of NPC. NKILA might be a candidate molecular marker and a novel therapy target for NPC patients.
| Nasopharyngeal carcinoma (NPC), a metastasis-prone cancer, which is particularly common in southeast Asia and southern China [1–4]. Due to the high radiosensitivity, radiotherapy has become the main treatment for locoregional NPC. Radiation oncology has improved the locoregional control(the tumor control of nasopharynx and neck lymph nodes), the development of distant metastasis becomes the major reason for treatment failure and occurs in 30–40% of patients with locoregional advanced NPC [5]. Thus, the assessment of the metastatic potential of NPC is vital for determining prognosis and treatment.
Long non-coding RNAs play pivotal regulatory roles in the physiological and pathological processes. Most lncRNAs regulate gene expression by RNA decay control, chromatin remodeling, and enhancer transcription in cis and epigenetic regulation [6–10]. Several lncRNAs are aberrantly expressed or play important roles in NPC, such as HOTAIR, ENST00000438550, and AFAP1-AS1 [11–15]. Inflammatory cytokines have been observed in NPC tissues and can promote the susceptibility to metastasis of NPC cells via constant NF-κB activation [16–18], therefore NF-κB is a pivotal link between NPC and inflammation. Interestingly, NF-κB is found to be overexpressed in nearly all NPC tissues [16, 19, 20]. NKILA is an NF-κB-interacting lncRNA [21], our previously study found it can be upregulated by inflammatory cytokines in breast cancer. By interacting with NF-κB/IκB, NKILA forms a stable complex, subsequently it masks the IκB phosphorylation motifs to repress the phosphorylation of IκB induced by IKK, then repress NF-κB pathway activation. But the role of NKILA in nasopharyngeal carcinoma remains unknown.
In our study, we examined NKILA expression in normal nasopharyngeal tissue, NPC tissue and cell lines. We proved that low NKILA expression predicts poor patient prognosis and that NKILA regulates the metastasis of NPC by the NF-κB pathway. In addition, we explored the role of NKILA in NPC carcinogenesis and metastasis.
To evaluate the role of lncRNA in the progression of NPC, we performed lncRNA expression profiles using four paired NPC and para carcinoma tissues by microarray. We observed that 2107 lncRNAs were upregulated while 2090 lncRNAs were downregulated by more than 2-fold, NKILA among these downregulated lncRNAs (Fig 1A, GSE95166). Quantitative RT-PCR verified the significant reduction in the expression of NKILA in NPC (Fig 1B). To confirm the results, NKILA expression levels were detected in fresh frozen tissues (26 NPC and 10 control tissues) by qRT-PCR. Compared with control tissues, NKILA was significantly downregulated in NPC tissues (Fig 1C). Furthermore, patients with developed distant metastasis have a lower NKILA expression than patients with non-metastatic NPC (P < 0.05, Fig 1C, Table 1). The results imply that low expression level of NKILA is correlated with the progression of NPC.
We further detected the expression of NKILA in 107 paraffin-embedded NPC tissues to evaluate the clinical significance of NKILA in patients with NPC. Scattered NKILA staining was observed in NPC cells cytoplasm, and the mean optical density (MOD) was used to quantify the NKILA staining. NKILA expression levels were compared in normal nasopharyngeal epithelia and samples from different stages of NPC. As shown in Fig 2A, NKILA was found abundantly expressed in normal nasopharyngeal epithelia and nasopharyngeal cells of metaplasia with atypical hyperplasia (Fig 2A) (P < 0.01), with a significantly higher MOD of NKILA staining compared with NPC tissues. Furthermore, NKILA staining decreased significantly with advanced disease staging in TNM stage I to III NPC, and staining was almost absence in stage IV tumors (Fig 2B).
NKILA expression was associated with the clinicopathological features of NPC patients (Table 1). Low expression of NKILA was correlated with metastasis (P < 0.05), larger tumor size (P <0.05), and late clinical stage of NPC patients (P < 0.005, Table 1). Other parameters (i.e. age, gender and lymph node status) were not found direct association with NKILA expression (P > 0.05).
This study suggested that NPC patients with high expression of NKILA had a significantly longer survival, the median follow-up time is 83 months (Fig 3A, P < 0.001). Furthermore, high expression of NKILA predicts a longer disease-free survival (DFS) (Fig 3B, P < 0.001), distant metastasis-free survival (DMFS) (Fig 3C, P <0.01), and local recurrence-free survival (LRFS) (Fig 3D, P <0.01). After adjusting for potential confounders, the multivariate analysis showed that high expression of NKILA was significantly correlated with improved OS (HR, 0.327; 95% CI, 0.171–0.623; P < 0.001), DFS (HR, 0.290; 95% CI, 0.153–0.549; P < 0.001), DMFS (HR, 0.353; 95% CI, 0.159–0.781; P = 0.010), and LRFS (HR, 0.227; 95% CI, 0.077–0.670; P < 0.01, Table 2).
NKILA suppressed the activation of NF-κB, which in turn inhibited tumorigenesis induced by aberrant NF-κB signaling by modulating apoptosis and invasion [21–23]. We observed that enforcing NKILA expression increased apoptosis in S18 cells (S18 vec vs S18 NKILA: 10.1% vs 19.3%, p<0.01). Conversely, silencing NKILA in S26 cells reduced apoptosis (Fig 4A, P < 0.05), suggesting that NKILA modulates apoptosis in NPC cells. We next verified the association of the NKILA expression level with apoptosis. We examined the apoptosis level in clinical NPC tissues using a TUNEL assay and the expression level of NKILA using in situ hybridization. The expression level of NKILA was positively associated with the tumor apoptosis level (Fig 4B, P < 0.001). Next, the regulation of NKILA in the tumorigenic activity during anchorage-independent growth in NPC cells was evaluated. As shown in Fig 4C and 4D, overexpression NKILA induced significant inhibition of anchorage-independent growth in S18 cells, as revealed by a decrease in the number and size of colonies (P < 0.001). Conversely, the depletion of endogenous NKILA in S26 cells induced a significant increase in the number and size of colonies (P < 0.001).
In addition, to explore the role of NKILA in tumorigenesis in vivo, xenograft tumor experiments were performed. S18 cells overexpressing NKILA or carrying a control vector were injected subcutaneously into the flank of nude mice, and then we measured the tumor size every 2 days to calculate the tumor volume. As shown in Fig 4E, NKILA overexpression in S18 cells greatly inhibited the tumor growth (P < 0.001), demonstrating that downregulation of NKILA is required for the malignant transformation of nasopharyngeal epithelial cells.
Next, we detected NKILA expression in NPC cell line to further explore the regulatory function of NKILA in NPC. Two paired NPC cell lines with high metastatic potential (S18 and 5-8F) and low metastatic potential (S26 and 6-10B) were used in the experiment. We found that NKILA expression levels increased by 2.6-fold in S26 (P < 0.001, versus S18) and by 4.1-fold in 6–10B (P < 0.001, versus 5-8F) (Fig 5A). It was shown that NPC cell lines with low metastatic potential cells had a higher NKILA expression level.
To explore the effect of NKILA on metastatic potential of NPC cells, we overexpressed NKILA in S18 NPC cells and examined the resulting metastatic potential. As shown in Fig 5B, S18 NPC cells overexpressing NKILA exhibited reduced migration and invasiveness (Fig 5B). Conversely, silencing NKILA significantly promoted the mobility of S26 cells, the metastatic potential of cells was enhanced.
Experimental metastasis assay was performed to evaluate the effect of NKILA on metastasis in vivo. We injected S18 cells with enforced overexpression of control vector or NKILA into the lateral tail vein of nude mice (4 weeks old), metastatic nodules in lungs were evaluated, numbers of metastatic nodules in lungs were markedly decreased in mice injected with S18 cells overexpressing NKILA, as shown in Fig 5C–5E. The number and volume of micrometastases in lungs of mice injected with S18 cells overexpressing NKILA were proven significantly reduced by H&E staining (Fig 5C). The results suggest that NKILA is extremely important for the metastatic development of S18 cells.
NKILA is up-regulated by inflammatory cytokines, which resulted in the inhibition of IKK -induced IκB phosphorylation and then repressed NF-κB pathway activation [21]. Nuclear translocation of P65 is emerging as a central feature of NF-κB pathway activation. Moreover, NKILA binding to the NF-κB: IκBα complex is essential for inhibition of NF-κB activation [21]. Thus, we evaluated the NF-κB activation by detecting NF-κB transcription activity and P65 translocation in NPC cells stimulated by inflammatory cytokines. We found that NKILA suppressed the enhancement of NF-κB transcriptional activity by TNFα. Conversely, silencing NKILA increased NF-κB transcriptional activity by more than 3-fold (Fig 6A). The enhancement of NF-κB transcriptional activity in S26 was suppressed by 40%, but further increased by 3.5-fold after NKILA was silenced. In addition, upregulation of the expression of NKILA resulted in retention of most of the P65 in the cytoplasm upon TNF-α stimulation, whereas nearly all of the P65 translocated to the nucleus in cells carrying empty vector (Fig 6B and S2 Fig). Conversely, silencing of endogenous NKILA expression led to prolonged P65 translocation to the nucleus upon TNF-α stimulation in S26 cells. Our results suggest that NKILA inhibits the activation of NF-κB pathway in S26 cells.
Subsequently, we evaluated the role of NKILA on IκBα and IKK phosphorylation and explored the mechanisms by which NKILA inhibits NF-κB activation in NPC. It revealed that both basal phosphorylation and TNF-α-induced phosphorylation of IκBα were repressed by enforced expression of NKILA in S18 cells; additionally, silencing NKILA increased the phosphorylation of IκBα with or without the stimulation of TNF-α in S26 cells (Fig 6C and S1 Fig). However, the phosphorylation of IKK was not influenced by NKILA expression (Fig 6C). Our study indicates that NKILA inhibits the activation of NF-κB primarily by inhibiting the phosphorylation of IκBα but not IKK in S18 and S26 cells. To confirm that NKILA works by inhibiting NF-κB, we used sc-3060 and JSH-23 to abrogate the nuclear translocation of NF-κB. As shown in Fig 6D and 6E, NKILA did not further increase the apoptosis of S18 cells, nor did it reduce the migration and invasion of S18 cells, suggesting that role of NKILA in NPC cell lines is dependent on IκBα.Collectively, our data show that NKILA represses the progression of NPC by inhibition of NF-κB pathway (Fig 7).
NPC, a malignant tumor with high tendency for metastasis, is very common in southern China. Patients of NPC presented with distant metastasis at diagnosis accounts for 5–8% of all cases; furthermore, after standard treatment, the proportion of distant metastasis in stage III–IV NPC is still as high as 30% [5, 24]. This tendency for metastasis emphasizes the urgency of elucidating the molecular mechanism underlying tumorigenesis and metastasis and possibly to develop new treatment for NPC. In our study, we first discovered that a long non-coding RNA named NKILA which was down-regulated in nasopharyngeal carcinoma, and we have demonstrated that overexpression of NKILA repressed the motile behavior and metastatic capacity of NPC cells. As previously reported, NKILA represses NF-κB activation by directly or indirectly inhibiting phosphorylation of IκBα in breast and hepatocellular carcinoma[21, 25]. We further demonstrated that NKILA could repress the metastasis of NPC by inhibiting NF-κB pathway. Furthermore, we identified that decreased NKILA expression was correlated with unfavorable clinicopathological features and poor survival outcomes in NPC patients.
The roles of lncRNAs (eg: HOTAIR, MALAT1, HOTTIP) have been confirmed by functional studies [26–30]. Dysregulated lncRNAs have been observed in NPC, and clusters of lncRNAs are dysregulated during the metastasis cascade [11, 12]. Nevertheless, the expression profiles of lncRNA in paired NPC and para carcinoma tissues have never been reported, and most of the dysregulated lncRNAs in NPC are largely unknown. Human lncRNA microarray was firstly performed in our study to detect the expression profiles in paired NPC and para carcinoma tissues and found that 4197 lncRNAs were dysregulated by more than 2-fold. Among these dysregulated RNAs, NKILA was found significantly downregulated in NPC, the decline of NKILA was more pronounced in patients with distant metastases, consistent with previously reported in breast cancer, Hepatocellular carcinoma and laryngeal cancer [21, 25, 31–33]. Second, we showed that NKILA expression decreased significantly with advanced disease staging in NPC clinical tissue samples, further research showed low expression of NKILA was correlated with metastasis (P< 0.05), larger tumor size (T stage, P <0.05), and late clinical stage (TNM stage, P < 0.01). No association was observed between NKILA expression and lymph node states (N stage) in NPC, but NKILA expression was significantly correlated with the lymph node in breast cancer (P < 0.001) [21]. As previously reported in other type of cancers, our study indicated that NPC patients with high NKILA expression survived significantly longer (OS, P < 0.001) or longer DFS (P < 0.001) [21, 31, 32]. We further provided evidence that patients with high expression of NKILA had longer DMFS and LRFS (P = 0.01, P <0.01 respectively), which was clinically significant for patients with NPC. Our study is the first to demonstrate that low NKILA expression predicts poor prognosis of NPC patients, with reinforcement data from multivariate analysis (Table 2). Additionally, aberrant activation of NF-κB may promote chronic inflammation even tumorigenesis in certain conditions which is correlated with NPC progression [18, 34–36]. Here, we demonstrate that overexpression of NKILA promotes apoptosis and represses the invasion of NPC cell lines as reported in breast cancer cells. NKILA can also enhance the effect of baicalein on cell apoptosis and metastasis in HCC [21, 25]. Furthermore, NKILA is firstly proved to be positively associated with apoptosis in human NPC tissues. The present study further confirms that NKILA represses tumorigenic and metastatic ability of NPC cells.
NKILA is firstly identified as upregulated by inflammatory cytokines through NF-κB pathway in breast cancer, in return NKILA regulate the metastasis of breast cancer via NF-κB pathway [21]. Whether NKILA regulates tumorigenesis and metastasis in NPC via NF-κB and its mechanism remains unclear. In the present study, we verified that NKILA suppressed the enhancement of NF-κB transcriptional activity by TNFα. In addition, upregulation of the expression of NKILA resulted in retention of most of the P65 in the cytoplasm upon TNF-α stimulation, In contrast, the depletion of NKILA expression significantly prolonged the sustained activation time of NF-κB pathway, our study firstly demonstrated that NKILA can regulate the NF-κB activation in NPC. The results are consistent with our previous study in breast cancer [21]. P65 was also found upregulated in laryngeal cancer tissues. P65 positively regulates the NKILA expression, however, NKILA inhibits the translocation of P65 to reduce the resistance of cells to X-ray radiation in laryngeal cancer [33]. Subsequently, we showed that NKILA mainly inhibits the phosphorylation of IκBα, rather than activating IKK to repress NF-κB activation in NPC cells. In addition, we used sc-3060 and JSH-23 to abrogate P65 nuclear translocation, and no further increased apoptosis or reduced migration and invasion was observed in NKILA overexpressing NPC cells. Our previous study shows that NKILA has a high affinity for P65 [21]. By binding to P65, NKILA forms a complex with the IKB/NF-κB complex, and then masks the IKK phosphorylation site to inhibit IκB phosphorylation. In the present study, it is confirmed that NKILA exerts its effect on NPC by inhibiting NF-κB activation. The important effect of NF-κB pathway in NPC progression is well known, but most studies focus on the activation of P50/P65 in NPC cells[37–40].It has been reported that FN1 regulates apoptosis of NPC cells though P65 in the NF-κB pathway [41]. Epstein-Barr virus (EBV) expresses high levels of BamHI-A rightward transcripts (BARTs) in NPC.LMP1 binds P50 to NF-κB sites in the promoter of BART and activates the BART promoters via NF-κB pathway, an autoregulatory loop is formed in NPC cells to maintain EBV latency [42]. The upstream component of NF‐κB or the targeted gene of NF-κB has also been reported to contribute to aberrant NF-κB pathway activation [37, 43, 44], MiR-125b has been shown to regulate NF-κB pathway activity by targeting A20 in NPC cells [45]. To our knowledge, no studies have demonstrated that NF-κB activation associated with NPC progression is primarily regulated by inhibition of IκBα phosphorylation in NPC, and more importantly, in the current study, we extended the understanding of the autoregulatory loop of inflammatory factors and NF-κB activation. Consistent with the important regulatory role of NKILA in breast cancer [21, 46], we demonstrate that NKILA regulates the progression of NPC cells by regulating NF-κB pathway activity.
Nearly all NPCs are EBV-associated tumor. EBV-encoded LMP1, a transmembrane protein, has been identified as a viral oncogene of NPC [47]. LMP1 induces constitutive activation of NF-κB by transforming effector sites 1 and 2, which is required for efficient B-lymphocyte transformation. NF-κB activation maintains the survival of EBV-transformed lymphoblastoid cells, and blocking NF-κB signal leads to the death of these malignant cells [48]. Interestingly, we demonstrate for the first time that NKILA, a lncRNA that is upregulated by inflammatory cytokines, inhibits NF-κB activation by repressing IκB phosphorylation induced by IKK in NPC. In addition, NKILA exerts its effect as a tumor suppressor via inhibiting tumorigenesis and metastasis of NPC, and overexpressing NKILA reverses tumorigenesis and metastasis of NPC. NKILA might be a vital gene for repressing the role of EBV and become one of the most important therapeutic targets for patients with nasopharyngeal carcinoma.
In summary, NKILA plays a critical role in NPC progression. The unique histological features of NPC indicate that local inflammation is essential in NPC tumorigenesis. The present study provides new insights into the effects of inflammation on NPC biology. NKILA might be a candidate molecular marker and a novel therapy target for NPC patients.
This study was performed in accordance with the Institutional Review Board of Sun Yat-sen University Cancer Center (GZR2016-210). Written informed consent was obtained from each patient, including signed consent for tissue analysis and consent to be recorded for potential medical research at the time of sample acquisition.
S18 and S26 are subclones of NPC cell lines CNE-2, 6-10B and 5-8F are subclones of NPC cell lines SUNE-1 that were reported previously [49]. S18 and 5-8F has high metastatic potential, whereas S26 and 6-10B has low metastatic potential. All cell lines were cultured in complete RPMI 1640 medium.
Tissue specimens were obtained from the Nasopharyngeal Department of Sun Yat-sen University Cancer Center. A total number of 107 paraffin-embedded NPC and 20 normal control tissues obtained between August 1999 and February 2001 were examined in the present study. Fresh frozen normal nasopharyngeal epithelial and NPC tissues were obtained by biopsy.
Total RNA extraction and qRT-PCR were performed as we have reported[13]. Primer sequences of NKILA were as follows: sense, 5′-AACCAAACCTACCCACAACG-3′; antisense: 5′-ACCACTAAGTCAATCCCAGGTG-3′.
It was performed by soft agar colony formation assay and foci formation assay. Soft agar colony formation assay was performed in Six-well plates with a layer of 0.66% agar. Preparation of cells in 2 -fold concentration of 1640 complete medium and 0.33% agar, after evenly mixed and seeded at 3 different dilutions:1×103, 2×103,3×103. After culturing for 12–14 days, clone(>50 cells) numbers were assessed at an original magnification of ×100, the scale bar is 100μm, which is approximately the diameter of 50 cell clusters. Spheres in ten random fields of view were counted each well. Foci formation assay was performed in six well plates, cells were seeded in triplicate at 3 different dilutions: 100, 200, 300. Cells were cultured for a period of 14 days. Clone (>50 cells) numbers were calculated. All experiments were repeated separately at least 3 times.
The expression of NKILA in paraffin-embedded samples was detected by in situ hybridization (ISH). ISH was performed and analyzed as we previously reported[21]. An SI score of 2 was used as the cut-off value, SI>2 was defined as high NKILA expression, SI≤2 was defined as low NKILA expression. The probes used in the ISH assays were as follows: NKILA: TCTCCAGACAGAATCAACTTCG; NKILA antisense: CGAAGTTGATTCTGTCTGGAGA.
S18 and S26 cells stably expressing lentiviral particles with NKILA or control vector were obtained from Gene Pharma (Shanghai, China). For cell transduction, NPC cells (30–50% confluency) were infected according to the instructions. After an incubation of 12–16 h, the medium was changed. Forty-eight hours later, the selection reagent puromycin (Sigma-Aldrich, St. Louis, MO) was used to select stably transfected clones. After 21days of continuous selection, cells infected with LV5-NKILA were designated S26 NKILA, cells infected with LV5-NC were designated S26 Vec.
Luciferase activity was determined by the Dual-Luciferase Reporter Assay System (E1910, Promega, Madison, WI)). Cells were transfected with the pGL3-basic or pNFκB-luc constructs together with pRL-TK at 50:1, and then then the indicated treatments were performed. Cells were harvested and assayed 24h later, and all of the experiments were performed in triplicate.
2 × 104 NPC cells in 100μl RPMI 1640 medium with 2% FBS were plated in the upper chamber of transwell (Corning, New York, NY, USA), the bottom chamber was filled with 600μl RPMI 1640 medium with 10% FBS. After 20h of incubation, fixed the cells on the lower membrane surface in 4% paraformaldehyde and stained with crystal violet, then counted. The cells were counted in ten random optical fields (×200 magnification), the average number was obtained from triplicate filters. Data are shown as the average ±SD. For the invasion assay, coated the upper chamber with Basement Membrane (R&D, Minneapolis, MN, USA) first. All experiments were performed at least three times.
Female BALB/c(nu/nu) nude mice (4–6 weeks of age) were used. For tumorigenesis experiments, cancer cells (105/mouse) were injected subcutaneously into the flank of the mouse, and measured the tumor size every two days. Calculated the tumor volumes as follows: tumor volume (mm3) = length×width2×0.5. Metastasis of NKILA-overexpressing cancer cells and control cells was evaluated by tail vein intravenous injection. 4 weeks after the first inoculation, the experiment was terminated. After euthanasia, the lungs of mouse were harvested and weighed separately, then prepared for H&E staining. Metastatic nodules in mouse was counted respectively.
1 × 103 S26 cells overexpressing NKILA or NKILA shRNA were cultured on coverslips overnight prior to the experiment. After fixing with 4% paraformaldehyde, IF was done and imaged as previously reported [50], primary antibodies against P65 was used, followed by FITC-conjugated secondary antibodies (Invitrogen, Carlsbad, CA). Sc-3060(Santa Cruz, CA) and JSH-23(Millipore, Billerica, MA) are drugs which inhibit NF-kB nuclear translocation. 30min before specified treatment,10uM Sc-3060 and 5uM JSH-23 were added into the culture.
Statistical analyses were done by SPSS 18.0. Correlation between NKILA expression and clinicopathological features was analyzed by chi-square test. The survival data were plotted and analyzed by Kaplan-Meier, log-rank test, and multivariate Cox regression analyses. All experiments for cell culture were performed in triplicate at least three times. The data were shown as means ± SD. P values were calculated by Student's t-test. A P value of no more than 0.05 was considered statistically significant.
|
10.1371/journal.ppat.0030064 | Identification of a Novel Polyomavirus from Patients with Acute Respiratory Tract Infections | We report the identification of a novel polyomavirus present in respiratory secretions from human patients with symptoms of acute respiratory tract infection. The virus was initially detected in a nasopharyngeal aspirate from a 3-year-old child from Australia diagnosed with pneumonia. A random library was generated from nucleic acids extracted from the nasopharyngeal aspirate and analyzed by high throughput DNA sequencing. Multiple DNA fragments were cloned that possessed limited homology to known polyomaviruses. We subsequently sequenced the entire virus genome of 5,229 bp, henceforth referred to as WU virus, and found it to have genomic features characteristic of the family Polyomaviridae. The genome was predicted to encode small T antigen, large T antigen, and three capsid proteins: VP1, VP2, and VP3. Phylogenetic analysis clearly revealed that the WU virus was divergent from all known polyomaviruses. Screening of 2,135 patients with acute respiratory tract infections in Brisbane, Queensland, Australia, and St. Louis, Missouri, United States, using WU virus–specific PCR primers resulted in the detection of 43 additional specimens that contained WU virus. The presence of multiple instances of the virus in two continents suggests that this virus is geographically widespread in the human population and raises the possibility that the WU virus may be a human pathogen.
| We have identified a novel virus, referred to as WU virus, in the family Polyomaviridae by screening of human respiratory secretions. Two human polyomaviruses, BK and JC, were identified in 1971 and infect the majority of humans around the world. These two viruses are closely related to each other and are both are pathogenic in immunocompromised individuals. Earlier this year, a third polyomavirus, KI, was described in human clinical specimens, although its pathogenicity and prevalence in humans has not yet been established. The discovery of WU virus brings the number of polyomaviruses detected in humans to four. WU differs from BK and JC significantly in its genome sequence and in its relative tissue tropism, suggesting that it is likely to have unique biological properties. This discovery raises many questions for further investigation, such as, Is WU virus a human pathogen? If so, what kind of disease does it cause? Where in the body does WU virus reside? At what age does infection typically occur? Perhaps most importantly, there are likely to be many more as of yet unidentified viruses infecting the human body.
| Viral infections of the respiratory tract are responsible for significant mortality and morbidity worldwide [1]. Despite extensive studies in the past decades that have identified a number of etiologic agents, including rhinoviruses, coronaviruses, influenzaviruses, parainfluenzaviruses, respiratory syncytial virus, and adenoviruses, approximately 30% of all cases cannot be attributed to these agents, suggesting that additional respiratory pathogens are likely to exist [2]. In fact, since 2001, six previously undescribed viruses have been identified by analysis of clinical specimens from the human respiratory tract: human metapneumovirus [3], SARS coronavirus [4], coronavirus NL63 [5], coronavirus HKU1 [6], human bocavirus [7], and the recently described KI virus [8]. In some instances, new molecular methods such as VIDISCA [5], pan-viral DNA microarrays [9], and high throughput sequencing [7,8] have played key roles in the identification of these agents. The advent of these new technologies has greatly stimulated efforts to identify novel viruses in the respiratory tract and in other human disease states.
Viruses in the family Polyomaviridae possess double-stranded DNA genomes and infect a variety of avian, rodent, and primate species. To date, two polyomaviruses, BK virus and JC virus, have been unambiguously described as human pathogens. BK and JC viruses are ubiquitous worldwide, and in adult populations, seroprevalence rates approaching 75% and 100%, respectively, have been reported [10]. Although human polyomaviruses have been suggested to utilize a respiratory route of transmission, detection of BK and JC polyomavirus nucleic acids in the respiratory tract has rarely been reported [11,12]. Infection with these two viruses is predominantly asymptomatic, although in the context of immunosuppression a number of syndromes have been clearly linked to these viruses. JC virus causes primary multifocal leukoencephalopathy, while BK virus has been associated with a variety of renal and urinary tract disorders, most importantly tubular nephritis, which can lead to allograft failure in renal transplant recipients and hemorrhagic cystitis in hematopoietic stem cell transplant recipients [13]. These viruses are believed to persist in a latent phase primarily in the kidney and can periodically undergo reactivation. Excretion of BK and JC viruses in urine has been reported in up to 20% of the general population [14,15]. Besides JC and BK virus, a very recent report has described a novel polyomavirus, KI, detected in human respiratory secretions and stool [8]. However, the pathogenicity and prevalence of this virus has not yet been established. In addition, in the late 1950s, ∼100 million people in the United States, and many more worldwide, may have been exposed to SV40, a polyomavirus that naturally infects rhesus monkeys via contaminated polio vaccines, leading to widespread debate about whether or not SV40 is capable of sustained infection and replication cycles in humans [16].
Much of the interest in polyomaviruses and SV40 in particular derives from the transforming properties carried by the early transcriptional region of the viral genome that encodes for the small T antigen (STAg) and and large T antigen (LTAg). T antigen is capable of binding both p53 and Rb proteins and interfering with their tumor suppressor functions. The early region alone is sufficient to transform established primary rodent cell lines [17] and in concert with telomerase and ras transforms primary human cells [18]. This has lead to controversy over whether any human tumors are associated with SV40 infection [19].
We describe the identification and characterization of a novel polyomavirus initially detected by high throughput sequencing of respiratory secretions from a patient suffering acute respiratory disease of unknown etiology. The virus was detected in the respiratory secretions from an additional 43 patients in two continents, and the complete genomes of multiple isolates were sequenced.
A nasopharyngeal aspirate (NPA) from a 3-year-old patient admitted to the pediatric ward of the Royal Children's Hospital in Brisbane with pneumonia was collected in October 2003. The patient had no other remarkable clinical traits other than the respiratory features of pneumonia. Testing of nucleic acid extracted from the NPA using a panel of 17 PCR assays for known respiratory viruses as described [20] yielded negative results. Total nucleic acid from the NPA was randomly amplified and cloned as described previously [9]. One 384-well plate of clones was sequenced using a universal M13 primer, and the resulting sequence reads were analyzed as described in Materials and Methods. Of the 384 reads, there were 37 poor quality sequences that were rejected from further analysis, 327 human sequences, six bacterial sequences, six viral sequences, and eight sequences of unknown origin that could not be classified. The bacterial sequences had greater than 97% nucleotide identity to known bacterial species, including Haemophilus influenzae (three reads), Streptococcus pneumoniae, Corynebacterium pseudodiphthericum, and Leifsonia xyli (unpublished data). Upon further examination, the six viral reads were collapsed into three unique regions, each of which possessed only limited homology to known polyomavirus proteins (sequences available in Figure S1). The highest scoring BLASTx hits for each of these three contigs possessed 35%, 50%, and 34% amino acid identity to JC virus STAg, BK virus LTAg, and SV40 VP1, respectively. At the time these experiments were performed, the KI virus genome had not yet been published. Subsequent analysis revealed amino acid identities of 66%, 65%, and 69% to KI virus for the three contigs. Furthermore, three of the eight previously unclassified sequence reads were determined to have between 58%–84% amino acid identity to KI virus VP1 and VP2 proteins by BLASTx analysis. Based on the limited sequence homology to known viruses, we tentatively assigned the name WU to the unknown polyomavirus.
The complete genome of WU was sequenced to 3× coverage using cloned fragments of the viral genome generated by a series of PCR primers. Analysis of the DNA sequence revealed genomic features characteristic of polyomaviruses. First, the WU genome size of 5,229 base pairs (bp) was quite comparable to those of the primate polyomaviruses BK (5,153 bp), JC (5,130 bp), and SV40 (5,243 bp). In addition, the overall GC content of the WU genome was 39%, which is quite similar to the GC content of BK (39%), JC (40%), and SV40 (40%). The genome organization included an early region coding on one strand for STAg and LTAg, and a late region coding on the opposite strand for the capsid proteins VP1, VP2, and VP3 (Figure 1). These two regions were separated by a regulatory region that contained typical polyomavirus features. The regulatory region contained an AT-rich region on the late side of the putative replication origin. Three repeats of the consensus pentanucleotide LTAg binding site GAGGC were present, as was one copy of the non-consensus LTAg binding site TAGGC. While most polyomaviruses contain four copies of the consensus, baboon polyomavirus (simian agent 12) is a primate polyomavirus that contains only three copies of the canonical binding sequence and one non-consensus binding site [21]. Unusual features in the WU regulatory region included the presence of two partially overlapping LTAg binding sites and slightly variant spacing between the LTAg binding sites as compared to SV40, BK, and JC (Figure S2).
In the early region, an unspliced open reading frame of 194 amino acids was detected that possibly encodes for the STAg. As the paradigm in other polyomaviruses is that STAg is expressed from a spliced message, analysis of potential splice sites revealed the presence of a putative splice donor sequence just one nucleotide 5′ of the initially predicted stop codon. Splicing to a downstream putative splice acceptor site would excise an intron of 70 nucleotides and generate a slightly larger STAg of 217 amino acids (Figure S3). While the precise carboxyl terminus of the WU STAg has not yet been experimentally verified, sequence analysis revealed the presence of a highly conserved cysteine-rich motif, CX5CX7–8CXCX2CX21–22CSCX2CX3WF, that was present in both of the predicted isoforms of WU STAg. This motif, which is present in all STAgs, was perfectly conserved in WU virus with the exception of the initial cysteine residue.
In all polyomaviruses, the initial ∼80 amino acids of the N-terminus of the STAg and LTAg are identical; the LTAg is generated by alternative splicing of the early mRNA transcript. In WU virus, a conserved splice donor site was identified immediately after amino acid 84 of the early open reading frame. The position of the splice site is similar to that found in SV40, BK, and JC virus, which occur after amino acids 82, 81, and 81, respectively. Splicing to a conserved splice acceptor site would generate a predicted protein of 648 amino acids (Table 1). The predicted WU virus LTAg contained conserved features common to T antigens, including a DnaJ domain in the N terminus with the highly conserved hexapeptide motif HPDKGG; the LxCxE motif necessary for binding Rb; a canonical DNA binding domain; a zinc finger region; and conserved motifs GPXXXGKT and GXXXVNLE in the ATPase-p53 binding domain [22].
Based on comparative sequence analysis of LTAgs, the polyomaviruses are classified into two subclasses: a primate-like group exemplified by SV40, and a mouse polyoma-like group exemplified by murine polyoma virus [22]. Using these criteria, the T antigen of WU appeared to more closely resemble the mouse polyoma-like class of virus than the primate class. First, the mouse polyoma-like viruses have insertions of varying length after amino acids 66 and 113 of SV40 as compared to the primate class. In the amino terminal domain of the WU virus LTAg, multiple sequence alignment revealed the presence of a two–amino acid and a ten–amino acid insertion at these two loci, respectively. Furthermore, the primate-like class typically contains an extension of the carboxyl terminus termed the host range domain that is absent in the mouse polyoma-like class. In contrast to SV40, BK, JC, and baboon polyomavirus, WU virus did not appear to encode a carboxyl terminal extension (Figure S4).
In addition to encoding LTAg and STAg, murine and hamster polyomaviruses utilize alternative splicing to generate an intermediate-sized protein referred to as middle T antigen. The WU virus early region was scanned for splicing motifs similar to known murine and hamster polyomavirus splice donor and acceptor sequences, but no obvious combination of splice sites was detected that would yield a middle T antigen sequence in the size range of known middle T antigens. In addition, SV40, JC, BK, and baboon polyomavirus all encode a fourth late protein termed the agnoprotein. There was no open reading frame present in WU with any detectable homology to the known agnoproteins. Thus, our sequence analysis suggests that neither middle T antigen nor agnoprotein are encoded by WU virus, although it is possible that the sequences have diverged beyond our ability to recognize the appropriate splice sites or protein products.
Multiple sequence alignments of the predicted STAg, LTAg, VP1, and VP2 open reading frames revealed that WU virus was clearly a novel virus that is most closely related to KI virus (Figure 2). Neighbor-joining analysis suggested that these two viruses appear to form a new subclass of polyomaviruses. In the early region and VP1 protein, the WU/KI branch was most closely related to the known primate polyomaviruses BK, SV40, JC, and baboon polyomavirus (Figure 2A–2C). Finally, the VP2 open reading frame was so divergent that its evolutionary relationship to other polyomaviruses aside from KI could not be reliably established (Figure 2D). Analysis of the VP3 amino acid sequence, which is completely contained within VP2, gave similar results as VP2 (unpublished data).
PCR primers were designed to specifically amplify WU. The initial screen used primers targeting the VP2 region, which possessed less than 20% amino acid homology to JC and BK virus to minimize the possibility of cross reactivity with the known human polyomaviruses. Empirical testing of the primers on samples known to contain BK and JC confirmed that the primers did not cross react with either of these genomes (unpublished data). Positives in the initial screen for WU virus were sequenced and then further confirmed by a second PCR reaction using primers targeting the 3′ end of the WU virus LTAg coding sequence. All 43 positive samples in the initial screen were confirmed using the second pair of PCR primers. A subset of samples that tested negative in the initial screen was also tested with the second PCR primer pair, and none of those samples were positive.
In order to assess the prevalence of WU polyomavirus, a cohort of 1,245 respiratory specimens collected in 2003 in Brisbane was examined. Thirty-seven out of the 1,245 (3.0%) samples tested were positive for the virus (Table 2). In this cohort, patients that tested positive ranged in age from 4 months to 53 years. The vast majority of the patients (33/37) were age 3 and under. In 12 patients with clear clinical evidence of respiratory tract infection, WU was the sole virus detected. Strikingly, in 25 of the 37 positive samples, one or more additional respiratory viruses were also detected. The most common co-infections were with rhinovirus (15 cases) and human bocavirus (ten cases). Furthermore, in one sample, a total of four viruses (WU, bocavirus, rhinovirus, and adenovirus) were detected, and in six other samples, a total of three viruses were detected (Table 2).
In addition, we examined two cohorts of patients from St. Louis, Missouri, United States. In one set of upper respiratory specimens collected in 2006, five out of 410 were positive for WU virus in the PCR assay. In addition, 480 bronchoalveolar lavage samples from patients (mostly adults) with severe acute respiratory illness were tested, yielding one positive. Of the positive samples, all six were co-infected with other viruses (Table 2). The age range of the positive cases varied from 4 months to 51 years.
To assess the sequence variation within different isolates, we analyzed the 250-bp region encompassed by the initial screening primers for all 43 cases (Figure 3). Several divergent strains were detected, including one sample that had five mutations (2%) within this region. In another case, a 12-bp deletion was observed. The fact that many isolates were identical in sequence was not surprising, given the relatively short length of the amplicon and the double-stranded DNA nature of the genome. In addition, we sequenced the complete genome of five additional isolates from five independent patients. Unfortunately, efforts to completely sequence the two most divergent isolates (based on the 250-bp sequence, B2 and B3) have been unsuccessful, presumably due to low viral titers in these samples. All six complete genomes were 5,229 bp in size, and overall, there was between 0.08% and 0.23% sequence variation from sample to sample, well above that expected from Taq PCR, ruling out the possibility that the additional positives were artifacts of PCR contamination. Moreover, the majority of the observed mutations were synonymous substitutions or in non-coding regions, lending further support to the argument that these were authentic strain variants. For JC virus, the reported intratype sequence variation is of a similar magnitude, ranging between 0.1% and 0.5% [23].
Because BK and JC virus are frequently excreted in urine, we examined urine samples from patient cohorts in both St. Louis and Brisbane for the presence of WU virus by PCR. In the St. Louis cohort, urine samples from 200 adult patients participating in a study of polyomavirus infections in kidney transplant recipients were tested [24]. For most patients, samples were tested at three time points: prior to transplant, 1 mo post transplant, and 4 mo post transplant, although for some patients the pre-transplant specimen was not available. Zero out of 501 samples tested were positive for the WU polyomavirus. As a control, using previously validated BK primers, we were able to amplify BK virus in a subset of these urine samples, confirming the integrity of the specimens themselves (unpublished data). Similarly, from the Brisbane cohort, none of the 226 urine samples tested were positive for WU virus.
We used a high throughput sequencing strategy to search for novel agents that were present in respiratory tract infections of unknown etiology. The focus of this study was on individual clinical specimens that still lacked a diagnosis after analysis with an extensive panel of diagnostic assays for known respiratory viruses. In one such patient sample, novel sequences with limited homology to known polyomaviruses were detected. Complete genome sequencing and phylogenetic analysis revealed that the new virus clearly had the genomic organization typical of polyomaviruses but was divergent from all previously described polyomaviruses. In keeping with the two-letter virus names for human polyomaviruses, we have named this novel polyomavirus WU virus [25,26]. Overall, the predicted amino acid sequences of WU virus proteins were most similar to the newly described KI virus (Table 1). Outside of KI, WU shared only ∼15%–49% identity to its closest relatives (Table 1).
Detailed analysis of the viral DNA sequence and genomic organization confirmed the novelty of WU virus. At all loci, WU virus was most similar to KI virus, but the degree of divergence between WU and KI was greater than the divergence between SV40 and BK, indicating that WU and KI are clearly distinct viruses (Figure 2). Based on the phylogenetic analysis, it appears that WU and KI define a novel branch within the Polyomaviridae family (Figure 2). Relative to the established polyomaviruses, some analyses suggested that the WU/KI branch might be more closely related to the primate polyomaviruses, while other features of the WU genome suggested that it might be more similar to murine polyomavirus. For example, neighbor-joining phylogenetic analysis suggested that the predicted STAg, LTAg, and VP1 open reading frames of both KI and WU were most closely related to SV40, JC, BK, and baboon polyomaviruses. Analysis of the VP2/VP3 region was more equivocal, as the proteins were too divergent to reliably assess. The apparent absence of the C-terminal “host range” domain in the LTAg and the agnoprotein open reading frame, both of which are present in the known primate polyomaviruses, suggested that WU virus was more similar to murine polyomavirus than the primate polyomaviruses by these criteria. While the evolutionary history of this virus is not clear at the moment, the totality of the analysis indicates that WU is clearly a unique virus.
We detected WU in 37 out of 1,245 (3.0%) patient specimens in Brisbane (excluding the original case) and in six out of 890 (0.7%) patient specimens tested in St. Louis. As the positive specimens were all collected from 2003 through 2006, it appears that WU is currently circulating, and its presence in both North America and Australia suggests that the virus is geographically widespread in the human population. The age range of patients that tested positive for WU virus spanned from 4 months to 53 years. The majority (86%) of the cases were found in children 3 years of age and under. Of the four positive specimens from adult patients (S1, S6, B1, and B3 in Table 2), three clearly had altered immune status. One patient was HIV-positive, one was immunosuppressed due to treatment for Wegener granulomatosis, and one was pregnant. The fourth adult patient (S1), while not obviously immunosuppressed, also suffered from liver cirrhosis, hypertension, type 2 diabetes, and co-infection with herpes simplex virus, and required mechanical ventilation. In addition, there were two other positive patients older than 3 years of age: a 6-year-old child who had previously been a bone marrow transplant recipient (Table 2, B27) and a 6-year-old child diagnosed with acute lymphoblastic leukemia (Table 2, B9). While preliminary, the age distribution of the positive cases in this study combined with the established paradigms for BK and JC virus suggest a model where acute infection with WU virus may occur relatively early in life and result in a latent infection. Immunosuppression or other insults such as viral infection could then lead to reactivation of WU virus in older individuals.
The patients who yielded positive specimens suffered from a wide range of respiratory syndromes, including bronchiolitis, croup, and pneumonia as well as other clinical maladies (Table 2). Detection of WU virus sequences in these patients is merely the first step in assessing the potential etiologic role of WU virus in acute respiratory tract disease. It is not yet known whether WU is infectious or whether it is capable of replication in the respiratory tract. One possibility is that WU is not involved at all in respiratory disease, but rather is simply transmitted by the respiratory route. The human polyomaviruses BK and JC are hypothesized to be transmitted by the respiratory route before taking up residency primarily in the kidneys. Latency in the kidneys of BK and JC is believed to be the reason that both viruses are excreted in the urine of up to 20% of asymptomatic individuals [14,15]. In this study, using the same PCR assays that were effective in respiratory secretions, we did not detect WU in any of the 727 urine samples we tested. The lack of detection of WU virus in the urine may reflect sensitivity issues, a bias in the cohorts tested, or simply that WU is unlike BK and JC viruses and is not secreted in the urine. A similar tissue profile to that of WU virus has been reported in initial studies of KI virus [8]. Future experiments will aim to determine the tissue tropism of WU and whether any tissue reservoirs for WU virus exist.
In the literature, there is one animal polyomavirus that has been found extensively in lung tissue. Infection of suckling mice with the mouse pneumotropic polyomavirus (MPPV) causes interstitial pneumonia and significant mortality. MPPV also differs from other polyomaviruses in that besides the kidneys, it can also be detected in the lungs, liver, spleen, and blood of suckling mice [27]. Thus, there is precedence for an animal polyomavirus causing respiratory disease, suggesting at least the possibility that WU virus could be similarly pathogenic in humans.
One striking observation from these studies is the relatively high frequency of co-infection detected in the respiratory secretions: 72% overall (100% in the St. Louis cohort and 68% in the Brisbane cohort). Although more extensive studies are necessary to confirm the generality of this observation, this raises several intriguing non-mutually exclusive possibilities to consider: 1) WU may be an opportunistic pathogen; 2) WU infection may predispose or facilitate secondary infection by other respiratory viruses; and 3) WU may be a part of the endogenous viral flora that is reactivated by inflammation or some other aspect of viral infection. Recent studies of the prevalence of the newly identified human bocavirus have also reported higher levels of co-infection than previously described for other viruses found in the respiratory tract, with co-infection rates as high as 50% reported [28,29]. In addition, five of six samples positive for KI virus were reported to be co-infected with other known respiratory viruses [8]. As detection methods improve in sensitivity and more comprehensive efforts are made to examine the diversity of viruses found in the respiratory tract, a greater appreciation for the rates of dual or multi-infection is gradually emerging. For example, the use of extensive panels of PCR assays in this study revealed that one of the positive specimens was quadruply infected; adenovirus, rhinovirus, and bocavirus and WU virus were all present. Further investigations that aim to systematically define the spectrum of viruses present in the respiratory tract are clearly warranted so that the possible roles that co-infections may play in disease pathogenesis can be explored.
Extremely high sequence divergence was observed in the capsid proteins VP1 and VP2 of WU virus and KI virus as compared to the other known polyomaviruses. This divergence may reflect a different “lifestyle” for the WU/KI branch as compared to known polyomaviruses. Our data demonstrating the presence of WU in respiratory secretions and its absence in urine samples suggest that the mode of transmission or the sites of persistence of WU may be distinct from the other human polyomaviruses. As such, the structure of the virion must be optimized to enable the virus to survive dramatically distinct physiological and environmental conditions. This may partially explain the observed sequence divergence in the capsid proteins.
Another question raised by this study relates to the potential antigenic cross reactivity of the WU capsid proteins. In terms of establishing the seroprevalence of WU itself and determining whether seroconversion accompanies acute infection with WU, it will be essential to conduct these studies with consideration for potential cross reactivity to KI, BK, JC, and SV40 antibodies. In addition, it is tantalizing to speculate whether serum antibodies to WU have the potential to cross react to SV40-derived antigens, and if so, whether they may at least partially account for some of the studies that report the presence of SV40 antibodies in the human population that is too young to have suffered exposure from contaminated polio vaccination [30–32].
In conclusion, we have identified and completely sequenced the genome of a novel polyomavirus. This virus appears to be geographically widespread in the human population as evidenced by the detection of 44 distinct cases in two continents. Based on preliminary analysis, WU and KI virus share some strikingly similar properties, including their complement of genes, phylogenetic relationship, and physical sites of detection in the human body. These data suggest that WU virus and KI virus define a novel branch within the Polyomaviridae family with unexplored biology and pathogenicity. Another implication of these results is that the diversity of viruses in this family may be far greater than currently realized. Further experimentation is now underway to determine the relative pathogenicity of WU virus in humans and to understand the molecular properties of the virus. Since the T antigen of WU is predicted to have transforming properties by analogy to other polyomavirus T antigens, one question currently under investigation is whether a subset of human tumors may be associated with WU.
Brisbane cohort. A total of 1,245 specimens (predominantly NPAs) were collected between January 1, 2003, and December 22, 2003, from patients presenting to the Royal Children's Hospital in Brisbane, Queensland, Australia, with symptoms consistent with acute lower respiratory tract infection.
St. Louis cohort #1. A total of 480 BAL specimens were tested. These included samples from a retrospective and a prospective collection. The retrospective specimens were from a sequential collection of BAL specimens submitted routinely to the Virology Laboratory at St. Louis Children's Hospital between December 2002 and August 2003 [33]. For the present study, an effort was made to select specimens from this collection from patients with acute respiratory illness, and to exclude specimens collected as routine post–lung transplant surveillance. The prospective specimens were from an ongoing study of the etiology of severe acute respiratory illness and were collected between October 2005 and October 2006. Both collections included specimens from patients of all ages, although the large majority were from adults.
St. Louis cohort #2. This collection was made up of respiratory specimens, mostly nasopharyngeal swabs, submitted for routine virologic testing to the Virology Laboratory at St. Louis Children's Hospital between September 2005 and June 2006. The majority of these specimens were from children. Of the 410 specimens in this collection, 200 were selected because they had been found to be positive by fluorescent antibody staining or culture for influenzavirus A or B, respiratory syncytial virus, parainfluenza virus, rhinovirus, or adenovirus.
Brisbane cohort. Urine specimens (226) that were submitted during 2003 to the diagnostic laboratory for routine investigation were collected. These represented a diverse mixture of donors, including those from (i) sexual health clinic (n = 50), (ii) pediatric clinic (n = 52), (iii) antenatal clinic (n = 33), (iv) indigenous health clinic (n = 36), and (v) bone marrow transplant patients (n = 55).
The St. Louis urine specimens were from a study of polyomaviruses in adult renal transplant recipients [24]. A total of 200 individuals were enrolled in the study between December 2000 and October 2002. From each patient, up to three specimens were tested, including a specimen obtained before the transplant and specimens obtained at 1 and 4 mo after transplantation.
Brisbane cohort. Nucleic acids were extracted from 0.2 ml of each specimen using the High Pure Viral Nucleic Acid kit (Roche Diagnostics Australia, http://www.rochediagnostics.com.au) according to the manufacturer's instructions. PCR assays for 17 known respiratory viruses were performed as described [20].
St. Louis cohort. All respiratory specimens were tested originally by fluorescent antibody staining using a panel of monclonal antibodies directed against influenza A and B, respiratory syncytial, parainfluenza 1–3, and adenoviruses (Simulfluor Respiratory Screen; Chemicon, http://www.chemicon.com). Specimens that were negative were also cultured using cell culture systems that could detect the same group of viruses plus rhinoviruses, cytomegalovirus, and herpes simplex virus. Total nucleic acid extracts were purified using a Qiagen M48 instrument (http://www.qiagen.com). Nucleic acid extracts were tested for a panel of respiratory viruses using the EraGen MultiCode-PLx respiratory virus panel (EraGen Biosciences, http://www.eragen.com), a multiplex PCR assay that detects the following viruses: influenza A and B, respiratory syncytial virus A and B, parainfluenza 1–4, human meatpneumovirus, adenovirus subgroups B, C, and E, rhinoviruses, and coronaviruses OC43, 229E, and NL63.
Samples were prepared in the following manner for high throughput sequencing analysis. A total of 200 ul of neat NPA sample was thawed and directly treated with DNase I (Fermentas, http://www.fermentas.com) for 60 min at 37 °C. Total nucleic acid was extracted using the Masterpure Complete DNA and RNA Purification Kit (Epicentre Biotechnologies, http://www.epibio.com). Then, 100 ng of total nucleic acid was randomly amplified using the RdAB protocol exactly as described [9]. RNA in the total nucleic acid preparation was converted to cDNA by reverse transcription with primer-A (5′ GTTTCCCAGTCACGATANNNNNNNNN). Two rounds of random priming with primer-A and extension with Sequenase (United States Biochemical, http://www.usbweb.com) enabled second strand cDNA synthesis as well as random priming of DNA originally present in the total nucleic acid sample. Amplicons were then generated via 40 cycles of PCR using primer-B (5′ GTTTCCCAGTCACGATA) with a cycling profile of: 94 °C 30 s; 40 °C 30 s; 50 °C 30 s; 72 °C 60 s. The primer-B–amplified material was TOPO cloned into pCR4.0 (Invitrogen, http://www.invitrogen.com) and transformed into bacteria, and white colonies were picked into 384-well plates. DNA was purified by magnetic bead isolation and sequenced using standard Big Dye terminator (v3.1) sequencing chemistry. Reaction products were ethanol precipitated, resuspended in 25 ul of water, and loaded onto the ABI 3730xl sequencer.
Sequences were assessed for quality using Phred [34], and reads that contained less than 50 contiguous bases with a score of phred 20 or greater were rejected. The remaining reads were analyzed in the following steps: 1) reads were aligned to the human genome using BLASTn with an e−10 cutoff; 2) remaining reads were aligned to a bacterial database using BLASTn with an e−10 cutoff; and 3) remaining reads were aligned to the viral RefSeq protein database using BLASTx with an e−2 cutoff [35].
The WU genome derived from the index case was sequenced to 3× coverage using six unique pairs of PCR primers for the amplification. Amplicons were cloned into pCR4.0 and sequenced using standard sequencing technology. All primers used for amplification and sequencing are listed in Table S1 and their positions depicted in Figure S5. Additional complete genomes were sequenced to at least 2× coverage using the same primers listed in Table S1. Completed genome sequences have been deposited into GenBank (see Supporting Information for accession numbers).
Protein sequences associated with the following reference virus genomes were obtained from GenBank: BK virus, JC virus, bovine polyomavirus, SV40, baboon polyomavirus (simian agent 12), finch polyomavirus, crow polyomavirus, goose hemorrhagic polyomavirus, African green monkey polyomavirus, budgerigar fledgling polyomavirus, murine pneumotropic virus, hamster polyomavirus, and murine polyomavirus (see Supporting Information for accession numbers). For WU virus, predicted open reading frames were used. For STAg, the predicted open reading frame of 194 amino acids was used for analysis. Multiple sequence alignment was performed using ClustalX (1.83). Neighbor-joining trees were generated using 1,000 bootstrap replicates.
For all PCR assays, standard precautions to avoid end product contamination were taken, including the use of PCR hoods and maintaining separate areas for PCR set up and analysis. For initial screening of WU virus, PCR primers AG0044 5′ tgttacaaatagctgcaggtcaa and AG0045 5′ gctgcataatggggagtacc were used with Accuprime hot start Taq (Invitrogen) to amplify 1 ul of template using the following program: 40 cycles of 94 °C 30 s; 56 °C 30 s; 72 °C 60 s. For every 88 samples tested, seven no-template negative controls were interspersed between the actual samples. Products were visualized following electrophoresis on 1% agarose gels. The resulting 250-bp amplicon was sequenced directly in both directions using primer AG0044 and AG0045. These sequences have been deposited in GenBank (see Supporting Information for accession numbers). Secondary confirmation was performed using primers AG0048 5′ TGTTTTTCAAGTATGTTGCATCC and AG0049 5′ CACCCAAAAGACACTTAAAAGAAA that generate a 244-bp amplicon in the 3′ end of the LTAg coding region. The same cycling profile of 40 cycles of 94 °C 30 s; 56 °C 30 s; 72 °C 60 s was used. For detection of both BK and JC viruses, primers AG0068 5′ AGTCTTTAGGGTCTTCTACC and AG0069 5′ GGTGCCAACCTATGGAACAG were used with a profile of 40 cycles of 94 °C 30 s; 56 °C 30 s; 72 °C 60 s.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) protein sequences used in this paper are as follows:
LTAg: African green monkey (NP_848008); baboon polyomavirus 1 (YP_406555); BK (YP_717940); bovine (NP_040788); budgerigar (NP_848014); crow (YP_529828); finch (YP_529834); goose (NP_849170); hamster (NP_056730); JC (NP_043512); KI Stockholm 60 (ABN09921); murine (NP_041264); murine pneumotropic (NP_041232); SV40 (NP_043127).
STAg: African green monkey (NP_848009); baboon polyomavirus 1 (YP_406556); BK (YP_717941); bovine (NP_040789); budgerigar (NP_848015); crow (YP_529829); finch (YP_529835); goose (NP_849171); hamster (NP_056732); JC (NP_043513); KI Stockholm 60 (ABN09920); murine (NP_041266); murine pneumotropic (NP_041233); SV40 (NP_043128).
VP1: African green monkey (NP_848007); baboon polyomavirus 1 (YP_406554); BK (YP_717939); bovine (NP_040787); budgerigar (NP_848013); crow (YP_529827); finch (YP_529833); goose (NP_849169); hamster (NP_056733); JC (NP_043511); KI Stockholm 60 (ABN09917); murine (NP_041267); murine pneumotropic (NP_041234); SV40 (NP_043126).
VP2: African green monkey (NP_848005); baboon polyomavirus 1 (YP_406552); BK (YP_717937); bovine (NP_040785); budgerigar (NP_848011); crow (YP_529825); finch (YP_529831); goose (NP_849167); hamster (NP_056734); JC (NP_043509); KI Stockholm 60 (ABN09918); murine (NP_041268); murine pneumotropic (NP_041235); SV40 (NP_043124).
WU complete genome sequences have been deposited under accession numbers EF444549–EF444554. VP2 partial sequences have been deposited under accession numbers EF444555–EF444593. |
10.1371/journal.pcbi.1000114 | Probing the Extent of Randomness in Protein Interaction Networks | Protein–protein interaction (PPI) networks are commonly explored for the identification of distinctive biological traits, such as pathways, modules, and functional motifs. In this respect, understanding the underlying network structure is vital to assess the significance of any discovered features. We recently demonstrated that PPI networks show degree-weighted behavior, whereby the probability of interaction between two proteins is generally proportional to the product of their numbers of interacting partners or degrees. It was surmised that degree-weighted behavior is a characteristic of randomness. We expand upon these findings by developing a random, degree-weighted, network model and show that eight PPI networks determined from single high-throughput (HT) experiments have global and local properties that are consistent with this model. The apparent random connectivity in HT PPI networks is counter-intuitive with respect to their observed degree distributions; however, we resolve this discrepancy by introducing a non-network-based model for the evolution of protein degrees or “binding affinities.” This mechanism is based on duplication and random mutation, for which the degree distribution converges to a steady state that is identical to one obtained by averaging over the eight HT PPI networks. The results imply that the degrees and connectivities incorporated in HT PPI networks are characteristic of unbiased interactions between proteins that have varying individual binding affinities. These findings corroborate the observation that curated and high-confidence PPI networks are distinct from HT PPI networks and not consistent with a random connectivity. These results provide an avenue to discern indiscriminate organizations in biological networks and suggest caution in the analysis of curated and high-confidence networks.
| A protein–protein interaction network represents the set of pair-wise associations that have been discerned between the constituent proteins of an organism. There are three main types of such networks: (i) those determined from a single high-throughput experiment; (ii) curated, where interactions are compiled from the literature; and (iii) high-confidence, which contain subsets of interactions from total sets that may comprise any from types (i) and (ii). The latter are deemed to better represent those interactions actually occurring in a cell. Through the use of graph-theoretic analyses and a random network connectivity model, we find that biological networks of type (i), determined from a single high-throughput experiment, contain random, indiscriminate, binding patterns. However, networks of type (ii) and type (iii) are not representative of the random model, suggesting that they contain biased influences upon the protein associations. These conclusions have been suspected for some time but are further clarified in this work. Our findings provide an avenue to detect unconstrained or completely random network structures and lend insights into the identification of preferentially connected networks resulting from the underlying biological processes or manual curation.
| Protein interaction networks are key to the understanding and modeling of many biological processes. At the highest level, networks enable the conceptualization of the different physiological, biological, and chemical functions that typically occur in a cell. At the core of a network description lie the connections, or relationships, between the components present in a system, such as interactions, reactions, and modifications. Using high-throughput (HT) experimental techniques, large sets of component connections (blueprints) are now becoming available. Ultimately, for a cellular system, we desire the complete set of interactions between the constituent proteins (interactome) [1],[2]. The architectures of protein interaction networks, or their modes of assembly, are a consequence of how biological functions and processes have evolved and adapted over time. As such, it is imperative to analyze experimentally discovered biological networks from a number of perspectives, including mathematical.
Efforts to elucidate entire protein-protein interaction (PPI) networks for species have emerged in the forms of experimental HT technologies [3]–[6], large-scale curation [7], and predictive, or inferring, methodologies [8],[9]. To date, extensive PPI networks have been experimentally determined for a number of organisms, including Saccharomyces cerevisiae [10],[11], Escherichia coli [12],[13], Helicobacter pylori [14], Drosophila melanogaster [15], Caenorhabditis elegans [16], Plasmodium falciparum [17], Campylobacter jejuni [18], and Homo sapiens [7]. A number of efforts to compile and, in some cases, curate the data have emerged [7], [19]–[23], and the topological properties of these networks have been widely explored using a range of theoretical techniques [24]–[27]. A common feature of almost all biological networks is that their degree distributions roughly resemble a power law: P(k)∼k−β, where P(k) is the probability of any component having k direct interactions (or degree k) and β is usually between one and three [28]–[30]. In fact, many real-world systems show power-law property distributions [31]. Whether or not PPI networks have a power-law degree distribution is under debate [32]; however, it is clear that in PPI networks proteins that have very low degrees (one or two) are prevalent, while there are very few proteins that have especially many interactions (tens to hundreds). A number of graph construction models are able to generate networks having power-law-type degree distributions, including those based on preferential attachment [33],[34], duplication [35]–[37], and hierarchical [38],[39] approaches. However, use of these models to reproduce a desired degree distribution, such as that observed for a particular experimentally determined PPI, is not straightforward. Therefore, it is difficult to ascertain precise levels of correlation between the models and the observed biological networks. In this respect, models that generate networks with given degree distributions are desirable. It is well known that Erdös-Rényi (ER) random graphs [40],[41] do not have power-law degree distributions, but variations of this model are able to generate random-type networks with desired degree distributions [42]–[45]. However, this type of graph has been reported to have topological properties that are generally different from PPI networks [46]–[48].
Many studies have aimed to discover biological insights from PPI networks. Avenues pursued to this end include the identification of salient protein clusters and functional modules [49]–[53]. Such biological entities usually occur as dense sub-graphs that are highly intraconnected but loosely connected to the remainder of the network. Consequently, procedures for identifying them have utilized graph-theoretical algorithms that analyze local and global topological network properties [50],[52],[53] and methods that include protein functional information [49]. Therefore, comprehension of the general organizational principles of PPI networks may serve to enhance the discernment and evaluation of biological modules.
In a previous study, we investigated the extent of preferential attachment, or degree-weighted (DW) behavior, in nine PPI networks [54]. It was demonstrated that, overwhelmingly, the probability of interaction of two proteins is proportional to the product of their degrees, i.e., Pij∝kikj, where ki and kj are the degrees of proteins i and j, respectively. It was also surmised that degree-weighted behavior is a characteristic of randomness. Here, we expand upon these findings by utilizing a random network construction model that generates a DW network, while attempting to duplicate a given degree distribution. We show that networks generated with this DW model have topological properties that are consistent with PPI networks determined from single HT experiments. The results suggest that these experimental PPI networks exhibit random connectivity. However, the model fails to reproduce properties of curated and high-confidence PPI networks, suggesting that these are composed of multiple single-experiment modules, or, if not, that they exhibit constraints in their organizations.
It should be stressed that the actual probability of two proteins physically interacting, or binding, is unlikely to be random. Such an event is dependent on many factors, including the types of residues, or domains, on each protein, their conformations, and the presence of perturbing proteins. Here, we are investigating PPI networks of experimentally identified protein interactions from which the degrees of the proteins are given properties. The frequency, or likelihood, of interaction between two proteins of particular degrees is then a secondary quantifiable property. It is the latter characteristic that we find to be indicative of randomness. However, the degree distributions of PPI, and many real-world, networks are known to resemble power-law scaling and not Poissonian, or random, distributions. Hence, the non-random degree distributions seem anomalous with respect to the random connectivities. We reconcile this discrepancy by describing a model for the evolution of protein degrees that consists of sequential duplication and random mutation steps. This evolution process converges to a steady state for which the degree distribution is identical to one that has been calculated by averaging over eight HT PPI networks. The results suggest that our interpretation of random connectivities in PPI networks is consistent with a randomly influenced evolution of their degree distributions.
Degree-weighted behavior, simply put, implies that the higher the degree of a node is the more likely it is to have an edge with any other node. Thus, the likelihood of an edge between two nodes is proportional to the product of their degrees, where the exact probability can be given by Pij = γ(kikj)θ. In order to conserve the degree distribution, θ must equal one and γ = E/Σi<j(kikj), where E is the total number of edges in the network. It has been shown that these probabilities and constraints are overwhelmingly incorporated in PPI networks [54]. The only nodes that seem to show any deviation from DW behavior are those with very many connections, also known as hubs. Although the DW nature is less pronounced for these nodes, hub-hub interaction probabilities are still high. However, it is important to note that the level of noise in the hub-hub region varies from network to network. In fact, for some PPI networks, such as P. falciparum [17], the DW behavior is exemplary throughout [54]. Figure 1 shows the DW nature of two PPI networks not included in the previous study, H. pylori [14] and C. jejuni [18]. Note that we are plotting the dependence of the probability of interaction P(k1,k2) between two nodes of degrees k1 and k2 upon the product of their degrees k1k2. These probabilities have been calculated by counting the total number of interactions occurring between all proteins of degree k1 and k2, and dividing this by the total number of all pairs of combinations that can be made.
The relation between DW behavior and the previously noted disassortive nature of biological networks [24],[55] is worth commenting on. Disassortiveness implies that high-degree nodes prefer to connect to low-degree nodes. Seemingly in contrast, DW behavior implies that if a node is given a choice of two potential interacting partners, it will more often connect to the one of higher degree. However, in a typical PPI network the number of high-degree nodes is magnitudes less than the number of low-degree nodes. Therefore, while a high-degree node may make many connections to low-degree nodes, and appear disassortive, the observation that it makes any connections with other high-degree nodes is significant. We have demonstrated that in PPI networks, high-degree nodes are almost always within one or two steps from each other [54]. This characteristic is exemplified in Video S1, which contains a three-dimensional animation of the HT E. coli PPI network determined by Arifuzzaman et al. [13].
A recently reported model, denoted “STICKY,” has been likened to PPI networks [47]. This model uses the probability of interaction between two nodes to be proportional to the product of their input weights, which are the experimentally observed degrees [47]. By allowing for self-interactions and normalizing for the total number of edges, E, the probability of an edge between nodes i and j is given by Pij = (kikj)/(4E), where the factor four arises due to double looping. The STICKY procedure enumerates through all pairs of nodes twice (once for i = j) and assigns an edge if a uniformly generated random number is larger than Pij. However, it was not reported that this procedure produces degree distributions that are different from the experimental input degree distribution. We observe that due to the nature of the edge-sampling procedure, the eventual degree of a chosen node can be modeled by a probability curve that is Poissonian about its expected, or input, degree. In other words, if a node has input degree k, then after many realizations of the STICKY procedure, i.e., multiple complete network constructions, the set of observed degrees for that node will follow(1)where P(λ) is the fraction of networks in which the node has a degree λ. For nodes of low degree (one or two), their Poissonian distributions (Equation 1) are substantially skewed towards λ = 0. This means, for example, that in a typical STICKY network construction, 36.8% of nodes of input degree one will remain degree one, while 36.8% will become degree zero. Our computational simulations of the STICKY procedure consistently generate observed degrees in line with those predicted from Equation 1. Generally, in PPI (and most real-world) networks, nodes of degree one are most prevalent. Therefore, a model that strictly preserves their degrees, rather than letting many become zero, is desired for a fair comparison between model and experiment. If two networks have varying degree distributions, it is likely that their underlying architectures are different, regardless of any similarities in other global topological properties.
A total of 12 PPI networks were included in the study and these were partitioned into two groups. The references immediately following the network species/labels represent the direct sources. These encompass original publications [11]–[18],[20], the Database of Interacting Proteins (DIP) [19], and the Human Protein Reference Database [7]. The first group contains eight PPI networks that have each been determined from an individual HT experiment using either yeast two-hybrid (Y2H) or tandem affinity purification (TAP) methodology: C. jejuni [18], E. coli (HT1) [12], E. coli (HT2) [13], C. elegans (Y2H) [16], S. cerevisiae (Y2H) [11], H. pylori [14], P. falciparum [17], and D. melanogaster [15]. Only the two PPI networks of E. coli were evaluated using TAP technology, all others were determined using Y2H methodology.
The second group contains four PPI networks that are either (i) merged experimental datasets: H. sapiens [7] and S. cerevisiae (DIP) [19]; (ii) inferred high-confidence from multiple datasets: S. cerevisiae (CORE) [19],[20]; or (iii) high-confidence from an individual experimental study: C. elegans (CORE) [16]. The number of proteins and interactions in each network is given in Table 1.
Given a set of nodes and their degrees, we consider a DW model that constructs a corresponding network while conserving the node degrees. Rather than considering every unique pair of nodes once (in any order) and (for each pair) generating a uniform random number to test whether an edge is assigned between them, as in the ER random [40],[41] and STICKY [47] construction procedures, we consider each unassigned edge once, for a given node, and use uniform random numbers to determine which other node it will connect to. This principle is not unlike that of previous preferential attachment models [33],[34] except that here it is used to generate networks for which each node has a specified degree, instead of growing, or evolving, them from seeds [56]. In the degree-conserving degree-weighted (DCDW) model, each node is considered once, in a random order, and a set number of edges are placed between itself and a DW random selection of the rest of the nodes. For each considered node, the remaining (potentially interacting) nodes are sampled for by using their input degrees as probability weights. However, none of the nodes are allowed to have more interactions than their given, or input, degrees. For an input degree sequence, which defines the desired degree ki of each node i, the DCDW model is defined by the following procedure:
This model generates a DW network, such that all nodes have the desired degree without including self interactions. In our computations, there are very rare instances when step (2) is unable to complete. In this case, we retain the edges that have been set and skip to the next node, which is determined at random. However, such an occurrence is extremely rare and has no real impact on the final degree distribution.
The DCDW model appears similar in style to the random network model of Newman, Strogatz, and Watts [44], which generates a random graph with a given degree distribution. In this latter model, each node is assigned a number of stubs equal to the desired degree of the node. These stubs represent incomplete edges that emerge from their respective nodes. The random network is then constructed by choosing pairs of stubs (on different nodes) at random and placing edges between them. Thus, it can be construed that the probability weight of a node, at any time, is proportional to the number of unconnected stubs. Therefore, the probability weight of each node will slowly diminish as its stubs are used up. In contrast, the DCDW model uses constant probability weights for the nodes (proportional to their input degree) throughout the network construction procedure. As a result, the DCDW method is more likely to generate a true DW graph in which the probability of an edge between two nodes is proportional to the numerical product of their eventual degrees. In a way, the DCDW model can be thought of as being a mode of implementation of the method proposed by Newman, Strogatz, and Watts [44], although strictly speaking, the DCDW method generates a random DW graph.
We demonstrate, using two examples, that the DCDW model effectively generates true DW graphs. Figure 2 illustrates the DW nature of the P. falciparum and the D. melanogaster PPI networks (black points) together with their equivalent (same input degree distributions) DCDW networks averaged over 100 constructions (red points). Two elements are evidenced from the plots: (i) The DCDW model, as expected, generates DW networks, and (ii) the PPI networks exhibit very similar DW behavior to their DCDW equivalents. We observe similar plots for all PPI networks studied here. The network of D. melanogaster shows slightly more noise than its DCDW counterpart in the hub-hub region; however, this is expected from previous observations [54]. It is also important to note that because PPI networks are generally construed from a single measurement of the interactions, they are prone to more noise.
We recently illustrated, through simulations, that Erdös-Rényi (ER) random graphs [40],[41] show near-perfect DW behavior [54]. It can be analytically shown that any random graph will show DW connectivity. Details are provided in Text S1. We conclusively demonstrate this by constructing an ER random graph equivalent of the P. falciparum network (where the probability of any edge is determined from the number of nodes and edges in the P. falciparum PPI network) 104 times, and for each construction we use the resultant degree distribution as input for the generation of a DCDW network. For each pair of networks, ER and DCDW, in each simulation, we calculate the number of assigned edges, average clustering coefficients (〈C〉), average shortest path lengths (〈L〉), and diameters (largest shortest path length) and then we average these properties over the 104 simulations. The clustering coefficient of a node i is defined as the fraction of possible edges between neighbors that are present, where a neighbor of node i is any other node that shares an edge with it [57]. The average clustering coefficient of a network, 〈C〉, is determined by averaging the clustering coefficients of all nodes, where nodes of degree one are defined here to have a clustering coefficient of zero. The shortest path length between two nodes is the minimum number of steps (or edges) that must be traversed in order to go from one to the other. The average shortest path length of a network, 〈L〉, is the average of all shortest path lengths that are not undefined. The results are given in Table 2. It is found that all of the aforementioned properties are essentially identical for the ER random and the DCDW networks. Thus, it appears that, given an input degree distribution indicative of an ER random network, the DCDW model will regenerate the ER random connectivity. The methods of construction for the two networks are very different; the ER random model uses a constant probability for the assignment of an edge between any two nodes, whereas the DCDW model scales the probability of an edge between two nodes with the product of their degrees.
We must conclude from the above findings that random networks are inherently DW and, conversely, that DW behavior implies randomness in the connectivities. A question is immediately realized: is it possible for a graph to not show uniform DW behavior? If our conclusion that DW behavior and randomness are synonymous is true, then removal of random and DW elements from a network construction process might yield networks that are not uniformly DW in their connectivities. We illustrate such an instance by modifying the DCDW procedure described above in two ways: firstly, in step (1), rather than enumerating all nodes once in random order, we enumerate all nodes i in order of decreasing degree; and secondly, in step (2), rather than weighting each of the possible interacting nodes (for node i) by their input degree, we weight them by the inverse of their input degree, i.e., P(i−j)∝1/kj. We use the degree distribution of the P. falciparum network as input and average probabilities of interaction over 1000 network constructions. Figure 3 illustrates the resulting dependence of the probability of interaction P(k1,k2) between two nodes of degrees k1 and k2 upon the product of their degrees k1k2. Probabilities that are exactly zero, e.g., P(1,1) = P(2,1) = 0, are not shown. It is clear that this modified network construction procedure generates networks for which the connectivities deviate significantly from uniform DW behavior. Therefore, we can surmise that if a network does not show uniform DW behavior, it likely has been generated with some limiting condition(s).
It has been previously established that PPI networks have a DW nature [54]. We have seen above that the DCDW model generates networks that are DW, while conserving a desired degree distribution. It was also demonstrated that random networks are intrinsically DW and that the DCDW model produces random networks for a given degree distribution. It remains to discover whether the DCDW model generates graphs that share topological characteristics with PPI networks. As a first step, we compute global properties of PPI networks and their equivalent DCDW networks (same input degree distributions). Table 1 provides the average clustering coefficients, 〈C〉, and the average shortest path lengths, 〈L〉, for the 12 PPI networks (undirected with no self interactions) and their DCDW equivalents, where values for each of the latter are averaged over 100 realizations. As described previously, the PPI networks in Table 1 are partitioned into two groups. The top eight are each taken from an individual HT experiment, whereas the bottom four are either (i) merged experimental datasets (H. sapiens and S. cerevisiae (DIP)), (ii) inferred high-confidence from multiple datasets (S. cerevisiae (CORE)), or (iii) high-confidence from an individual experimental study (C. elegans (CORE)). The PPI networks in each group have been arranged by decreasing average clustering coefficient.
We find that, for the first group, or the top eight networks in Table 1, average clustering coefficients of the experimental networks and the DCDW model are in excellent agreement. In fact, the DCDW model has values within 0.003 of the experimental for four systems (C. jejuni, C. elegans (Y2H), P. falciparum, and D. melanogaster) and within 0.007 for two (E. coli (HT1) and S. cerevisiae (Y2H)). The largest discrepancy of 0.021 is observed for the E. coli (HT2) network; however, the DCDW-determined value of 0.085 is still quite close to the experimental value of 0.064. The results clearly indicate that the DCDW model is accurately simulating the global clustering in PPI networks determined from individual HT experiments. In terms of the average shortest path lengths, the DCDW model predicts values within 0.14 for five systems in the first group (C. jejuni, E. coli (HT2), H. pylori, P. falciparum, and D. melanogaster). For the remaining three systems, the DCDW model predicts average path lengths that are somewhat smaller than the experimentally observed values. One reason for this is that in some PPI networks the DW behavior tends to level off in the hub-hub interaction region. As the DCDW model uses DW behavior throughout, it may generate slightly more hub-hub connections than are actually present. In such a case, and if many of the shortest path lengths utilize the hub proteins, one might expect the DCDW model to produce networks having slightly shorter path lengths than the actual PPI networks. However, this is observed for only three out of the eight PPI networks in the first group. Overall, the DCDW model predicts average clustering coefficients and shortest path lengths that are in good agreement with those of PPI networks determined from individual experiments. Furthermore, the orderings of the predicted and experimental values of each topological property are almost identical. These results lend further support to the presumption that DW behavior is intrinsic to these networks.
For the second group of PPI networks, which are either merged from multiple experimental datasets, high-confidence, or a combination of both, the DCDW-predicted clustering coefficients are far smaller than the actual values. In fact, for three systems (S. cerevisiae (CORE), H. sapiens, and S. cerevisiae (DIP)), the DCDW predictions are about a magnitude smaller. Average path lengths determined from the DCDW model are also consistently smaller than the true values, by over 0.50 in two instances (S. cerevisiae (CORE) and C. elegans (CORE)). The discrepancy is slightly less for the network of S. cerevisiae (DIP) (0.21). However, it must be concluded that the DCDW model fails to reproduce global properties of the PPI networks in this second group.
Given the success of the DCDW model with regard to the first group of PPI networks, the subsequent failure of this model when applied to the networks of the second group is initially unexpected. The networks in the first group are different from those of the second group in that each of the former are derived from a single experiment. Three of the networks in the second group (S. cerevisiae (CORE), H. sapiens, and S. cerevisiae (DIP)) have been assembled by the merging of multiple datasets. If the numbers of common proteins, or overlapping nodes, between pairs of datasets comprising a merged set are small, then, in effect, this merged set incorporates somewhat separated PPI sub-networks. For such a case, one would not expect the DCDW model to perform adequately because it does not incorporate constraints about which nodes are able to interact. Similar reasoning, in terms of artificially introducing selective connectivity, may be used to explain why the DCDW model cannot reproduce properties of the two high-confidence PPI networks S. cerevisiae (CORE) and C. elegans (CORE). Examination of the average shortest path lengths for the PPI networks in the second group indicate that they are much larger than for networks in the first group that have a similar clustering coefficient. This observation seems to corroborate the notion of multiple PPI sub-network contents and/or constrained connectivity.
It is clear that the DCDW model generates graphs that have similar global properties to PPI networks determined from a single HT experiment. Given this affinity, it is worth comparing their inner architectures further. We accomplish this by examining the behavior of node degree versus clustering coefficient and average shortest path length. No additional analyses are performed upon the networks of the second group given that their global properties, in particular the clustering coefficients, are substantially different from those of the DCDW model.
Clustering coefficient profiles are determined by evaluating the average clustering coefficient for nodes having the same degree. In this way, we elucidate the behavior of degree versus clustering coefficient. This type of analysis has been reported previously for S. cerevisiae PPI networks [46],[49] and metabolic networks [39]. Clustering profiles for the four largest PPI networks of the first group are shown as solid black lines in Figure 4. It is immediately apparent from the plots that the clustering coefficients do not vary smoothly with degree. Fluctuations start out small, relatively, in the low-degree regions but become wild as the degree is increased. However, there appears to be an overall trend in that clustering coefficients seem to decrease, somewhat, as the degree becomes large. This trend has been noted previously [39],[46],[49], and, although masked by large deviations, is most apparent here for the E. coli (HT2) network (Figure 4C) and least pronounced for the PPI network of D. melanogaster (Figure 4A). Clustering profiles are also shown for the corresponding DCDW model for the tenth (blue) and fiftieth (red) network realizations. We find that profiles for the two realizations are similar although it is clear that the DCDW model allows for some variation. The DCDW model, to some extent, reproduces the wild fluctuations of the experimental data. Correlation coefficients between profiles for the two DCDW realizations can be unexpectedly low, 0.05 and 0.15 for D. melanogaster (Figure 4A) and C. jejuni (Figure 4B), respectively, or considerable, 0.76 and 0.45 for E. coli (HT2) (Figure 4C) and E. coli (HT1) (Figure 4D), respectively. These correlations suggest that clustering coefficients are less constrained in the degree distributions of D. melanogaster and C. jejuni, while more limited for the distributions of both E. coli networks. These variabilities are reflected in the correlation coefficients between the experimental and the two DCDW profiles, which are lowest for D. melanogaster, −0.04 and 0.18, and highest for E. coli (HT2), 0.59 and 0.75. However, the large fluctuations in all profiles make adequate comparisons difficult. Nonetheless, no striking differences are observed and the overall DCDW profiles tend to follow those of the PPI networks, especially for the E. coli (HT2) network. Therefore, we can conclude that the DCDW model is reproducing features of the intrinsic clustering for these PPI networks. Analogous plots for the remaining four PPI networks of the first group are provided in Figure S1 and similar conclusions can be drawn from them.
The average path length for a node, also known as closeness, is evaluated as the average number of steps connecting it to all other nodes. Path length profiles are determined by averaging closeness over nodes having the same degree. The dependence of closeness upon the degree has been studied previously for three PPI networks [58]. Path length profiles for the four largest PPI networks of the first group are shown as solid black lines in Figure 5. It is clear that the average path length consistently, and smoothly, varies inversely with the degree, indicating that nodes of higher degree are more central in the networks. This observation has been noted previously [54],[58]. Path length profiles are also shown for the corresponding DCDW model for the tenth (blue) and fiftieth (red) network realizations. It is evident that the DCDW model is reproducing the path length features of the PPI networks. While values for the DCDW model are consistently less than those of the corresponding PPI networks, the lines run almost parallel. Near-perfect agreement is observed for the C. jejuni network (Figure 5B) and the greatest variation is seen for the E. coli (HT2) network (Figure 5C). Not only does the DCDW model have a very similar path length dependence upon the degree as the PPI networks, it also incorporates the characteristic increased fluctuations noted at higher degree. Similar conclusions are drawn from corresponding path length profiles of the other four PPI networks in the first group and these are shown in Figure S2.
The affinities in clustering and path length profiles between the DCDW model and the PPI networks of the first group corroborate the findings of the previous section, in which similar corresponding global properties were observed. The DCDW model consistently produces networks that have similar global clustering coefficients to the corresponding PPI networks and the model also reproduces important features of the clustering profiles. Although global average path lengths can be smaller for the DCDW model, the generated path length profiles almost parallel those of the PPI networks. Therefore, we must conclude that the DCDW model is a plausible representation of PPI networks determined from a single HT experiment. As was demonstrated earlier, the DCDW model is representative of randomness and so we must conclude that these PPI networks incorporate a substantial random element. However, it is clear that the PPI networks of the second group, which are merged, curated and/or high-confidence datasets, are not well described by the DCDW model. The DCDW model only incorporates the degrees of the proteins in that there are no other precepts used in the sampling, or determination, of the interactions. Therefore, we must conclude that there are other factors involved in the assemblies of the second group of networks. These factors may be artificial or biological. With regard to the former, it is known that there are very small interaction overlaps between HT experimentally determined networks for the same species [13],[59]. While each individual HT network may be representative of the DCDW model, a combined set will not be and, hence, will appear multi-modular. Alternatively, the manual curation of a PPI network may involve a search, or verification, of interaction partners for proteins already present in the intermediate network. Such a process may unintentionally introduce preferential attachments. In the event that a PPI network is not representative of the DCDW model, and any artificial influences can be discounted, then there must be biological actions leading to preferential, or selective, interactions.
If PPI networks from a single HT experiment incorporate a significant random element, as indicated above, then it is aberrant that they do not have degree distributions that are Poissonian in nature. Rather, HT experiments consistently generate PPI networks that have degree distributions that resemble power-law scaling. Therefore, they must contain some elements that distinguish them from ER random graphs. The findings described above suggest that the organizations of these PPI networks may be dependent only upon their degree distributions; i.e., it is the protein degrees that determine the observed interactions, rather than the converse. Such an interpretation would imply that the HT experiments are observing the ability to bind, rather than specific interactions that occur in the cell. If so, it should be possible to evolve degree distributions of PPI networks without the use of a network framework, i.e., we wish to model the evolution of the proteins' “binding affinities.”
There are well known network models that are able to generate graphs with power-law-type scaling degree distributions. These are based on a number of concepts, including preferential attachment [33],[34], duplication [35]–[37], and hierarchical [38],[39] approaches. However, these have typically not been shown to reproduce degree distributions of actual PPI networks. We use a degree distribution averaged over the eight individual HT datasets (listed in the top group of Table 1) as a template for PPI networks. This degree distribution, illustrated in black in Figure 6, is subsequently referred to as the normal PPI degree distribution (NPPI-DD). The NPPI-DD is not shown for degrees higher than 30 since this region includes more noise. It is clear that, overall, the NPPI-DD resembles power-law scaling; however, this scaling is somewhat more level in the low-degree region. This type of deviation from perfect power-law scaling has been noted previously for PPI networks [32].
Our model of protein degree evolution initializes by setting all protein degrees to be equal to one, i.e., the degree distribution at time t = 0 is represented by Pt = 0(k = 1) = 1, where Pt(k) represents the fraction of proteins having a degree k after time step t. During the first phase of the next time step, we postulate that all protein degrees are able to randomly mutate, and we model the total effect of the mutations into the degree distribution by use of the Poisson distribution:(2)Here, is the resultant degree distribution from the random mutation phase, and the term in braces is analogous to that seen in Equation (1), i.e., the probability that a protein of initial degree k will become degree λ. For t = 0, the summation reduces to one term, but for latter time steps the degree distribution is more diverse. We note that this procedure will result in some proteins having zero degree. During the second phase of this time step, after the mutation phase, we postulate that all proteins of degree one will duplicate. The reasons for duplicating only proteins of single degree are discussed below. This duplication phase is mathematically represented by:(3)where δλ1 is the Dirac delta function and is equal to one if λ = 1 and is zero otherwise. The final degree distribution at the end of the time step is obtained by discarding proteins with zero degree and renormalizing:(4)There are two reasons for eliminating proteins of zero degree: (i) all proteins of degree zero will remain degree zero in subsequent time steps, and (ii) most experiments do not report the proteins that have no observed interactions.
The two-step procedure described above, random mutation followed by duplication of degree-one proteins (RM+DD1), can be iterated over a number of time steps by cycling through Equations 2–4. Figure 6 shows RM+DD1 degree distributions for time steps t = 1 through t = 5 as blue lines. At t = 0 the distribution is simply P0(1) = 1. The curves clearly show that the degree distribution decays less abruptly with additional time steps and approaches some converged profile. Our computations indicate that the degree distribution is essentially converged after 100 time steps; therefore, the evolved distribution following a billion steps, shown in red in Figure 6, is representative of the steady state. This steady-state distribution, for the RM+DD1 model, is seen to almost exactly overlap the NPPI-DD.
The near-perfect agreement between the RM+DD1 steady-state and the NPPI-DD may be coincidental; however, an analysis of its foundations is warranted. Firstly, we are modeling only the evolution of protein degrees, i.e., the number of binding partners a protein has. There is no attempt to describe any network of interactions or any specific sub-network of protein interactions describing a particular biological function. The degree strictly represents the ability of proteins to bind, regardless of whether such an interaction is actually utilized in a biological system. Secondly, the justification for the random mutation phase is straightforward and is manifested in the long history of gene mutation research [60]. Here, the mutation concept is applied directly to the degree of a protein rather than to its sequence. We surmise that changes in sequence coincide with changes in behaviors, and the latter includes the degree. Lastly, we include a duplication, or growth, phase that is well substantiated from Ohno's hypothesis on genome growth by duplication [61] and more recent genomic studies [62]. There are two ways to justify the doubling of the degree-one phase: (i) proteins with degree one are purported to evolve faster [58],[63] and (ii) “new” proteins are likely to have a small number of interacting partners. However, there is no strict justification for duplicating proteins of only degree one. There is, obviously, a mathematically infinite number of ways to grow the number of proteins of each degree. Nonetheless, it is curious that exact duplication of only degree-one proteins yields a steady-state degree distribution nearly identical to the NPPI-DD. Whether or not there is biological justification for this element requires further investigation.
There are other non-network-based schemes [64]–[66] that generate property distributions, such as flicker noise or fitness of species, which converge to a critical point that resembles power-law scaling. However, these approaches rely upon the specifications of barrier thresholds that govern whether a spill over, or catastrophic event, occurs. Our model requires no such parameters. In fact, besides defining the mode of growth, the model is based on completely random events modeled by the Poisson distribution. This aspect complements the apparent random natures, discerned above, of the single-experiment determined PPI networks and supports the interpretation that these experiments may be witnessing the ability to bind.
Here we have expanded on a previous study that demonstrated that the interactions in PPI networks incorporate DW elements, i.e., that the probability of an interaction between two proteins is generally proportional to the product of their degrees [54]. This finding prompted the employment of a network model that constructs a DW network, while preserving an input degree distribution. This DCDW approach can be considered similar to a previously reported random-type graph model [44] in that a comparable construction procedure is used, however, a subtle difference is that the DCDW model maintains consistent nodal weights, equal to their input degrees, throughout the construction procedure. The DCDW model was shown to exactly reproduce properties of ER random graphs, when provided with degree distributions for the latter, and therefore we utilize it as a random network model. This DCDW model is shown to closely reproduce the topological properties of eight PPI networks, each assembled from an individual HT experiment. Furthermore, the PPI networks and the DCDW model were shown to contain similar clustering and path length profiles, which illuminated the relationships with degree. The results lend further support to the premise that DW behavior is intrinsic to these PPI networks and, therefore, indicative of a significant random element. Thus, it is reasonable to conclude that the connectivities in these PPI networks have substantial random characteristics for the observed degree distributions. We are not implying that the experiments are generating random interactions, rather we perceive that the interactions have evolved using a random-influenced mechanism over time and that the experiments may be observing the ability of proteins to bind. While Y2H data are known to be noisy, the two PPI networks of E. coli have been determined using a different methodology (TAP), yet each also has a close similarity to the DCDW model. Consequently, these findings may be relevant to any HT technology.
The apparent inclusion of randomness in the individual HT PPI networks precipitated the development of a model to describe the evolution of protein degrees, or binding affinities. We show that by initializing all nodes to have a single interacting partner and iteratively applying mutation and growth modulations (of degree-one nodes), a steady-state degree distribution that resembles a power law results. Moreover, this steady-state degree distribution is found to be almost identical to a degree distribution computed by averaging over eight HT experimental PPI networks. Therefore, we postulate that the resemblance of the observed PPI degree distributions to power-law scaling is simply a result of growth and random mutation over time. This type of evolution mechanism is not surprising, but the exceptional agreement suggests that we are capturing the essence of the process. The model is consistent with an evolutionary process driven by single gene duplications followed by slow continuous genetic drift of all proteins. This interpretation is compatible with the following observations on PPI networks: (i) essential proteins typically have high degrees [29]. From the evolution of the network, the fundamental genes that can sustain life appeared first. As they evolved via gene duplication and mutation, they acquired more degrees. Hence, they may now be found among the highest degree nodes. (ii) The evolutionary rate of a protein correlates inversely with its degree [63],[67]. Proteins with a greater number of interactions are more likely to have existed longer and, therefore, more likely to have incorporated additional mutations. As such, their rates of change may have slowed, being closer to a steady state. We anticipate that the mutation and growth model can be generalized and applied to other types of evolving real-world systems to provide qualitative and quantitative simulations.
In contrast to the HT PPI networks, the curated and high-confidence PPI networks have global properties that vary significantly from the DCDW model. These differences can be attributed to two main reasons. Firstly, if a curated network includes interactions from more than one HT data set, and the overlap between these sets is very small, then the curated network may be essentially multi-modular. While the individual HT data sets may be well represented by the DCDW model, the combined network may not be due to unintentionally introduced partiality in the interactions. An exception exists if the HT datasets are highly complementary and the merged set is representative of a single DW module. Secondly, curated and high-confidence PPI networks have been manually manipulated and, therefore, include biases, or preferential influences, upon the protein interactions which may or may not be representative of the underlying biology. For these networks, the DCDW model will not be an accurate representation as it includes no such constraints.
An important consideration is that HT methods may generate many false-positive interactions. If these false positives far outnumber the true, or real, interactions, then the total PPI network will appear systematically biased depending upon the mode of generation of the false positives. If so, the DCDW model is mimicking this bias rather than the true biology and, therefore, provides clues as to the origin of the false-positive interactions.
Many studies infer biological properties, or traits, by contrasting PPI networks against corresponding NULL networks, which are akin to DCDW networks. The PPI networks used are often downloaded from databases that have curated interactions from a number of sources, including HT experiments. While the individual HT datasets will have affinities to their corresponding NULL networks, as demonstrated in this work, the curated datasets will not. Therefore, any conclusions or inferences drawn from these studies should be treated with caution. The elucidation of guiding principles in biology is frequently contingent upon contrasts to randomness. However, care must be taken to ensure that the data are not artificially modulated, as in the case of many curated PPI networks.
The findings reported here indicate that HT PPI networks incorporate random interactions between proteins of varying binding affinities. The evolution of the proteins' affinities can be modeled by a mechanism based upon duplication and random mutation for which the steady degree distribution is almost identical to one averaged over eight HT experimental PPI networks. However, curated and high-confidence PPI networks are found to contain influences exogenous to the HT experiments, leading to preferential associations between protein pairs. These results provide a means to distinguish uninhibited network organization with respect to the observed degree distribution and may shed light for the identification of consistent influences leading to preferentially connected networks representing manual curation and/or the underlying biology. |
10.1371/journal.pntd.0006540 | Quantification of dengue virus specific T cell responses and correlation with viral load and clinical disease severity in acute dengue infection | In order to understand the role of dengue virus (DENV) specific T cell responses that associate with protection, we studied their frequency and phenotype in relation to clinical disease severity and resolution of viraemia in a large cohort of patients with varying severity of acute dengue infection.
Using ex vivo IFNγ ELISpot assays we determined the frequency of dengue viral peptide (DENV)-NS3, NS1 and NS5 responsive T cells in 74 adult patients with acute dengue infection and examined the association of responsive T cell frequency with the extent of viraemia and clinical disease severity. We found that total DENV-specific and DENV-NS3-specific T cell responses, were higher in patients with dengue fever (DF), when compared to those with dengue haemorrhagic fever (DHF). In addition, those with DF had significantly higher (p = 0.02) DENV-specific T cell responses on day 4 of infection compared to those who subsequently developed DHF. DENV peptide specific T cell responses inversely correlated with the degree of viraemia, which was most significant for DENV-NS3 specific T cell responses (Spearman’s r = -0.47, p = 0.0003). The frequency of T cell responses to NS1, NS5 and pooled DENV peptides, correlated with the degree of thrombocytopenia but had no association with levels of liver transaminases. In contrast, total DENV-IgG inversely correlated with the degree of thrombocytopenia and levels of liver transaminases.
Early appearance of DENV-specific T cell IFNγ responses before the onset of plasma leakage, appears to associate with milder clinical disease and resolution of viraemia, suggesting a protective role in acute dengue infection.
| In order to understand the role of dengue virus (DENV) specific T cell responses in protection against infection, we studied T cell cytokine production in relation to clinical disease severity and resolution of viraemia in a large cohort of patients with varying severity of acute dengue infection. We found that DENV-specific T cell responses were higher in patients with dengue fever, when compared to those with dengue haemorrhagic fever. In addition, early appearance of DENV-specific T cell responses was significantly associated with milder clinical disease (p = 0.02). DENV peptide specific T cell responses inversely correlated with the degree of viraemia, which was most significant for DENV-NS3 specific T cell responses (Spearman’s r = -0.47, p = 0.0003). The frequency of NS1, NS5 and pooled DENV peptides, correlated with the degree of thrombocytopenia but had no association with liver transaminases. Our data suggest that early appearance of DENV-specific T cell IFNγ responses appear to associate with milder clinical disease and resolution of viraemia, suggesting a protective role in acute dengue infection.
| Dengue virus is the cause of the most common mosquito-borne viral infection worldwide, indeed over half of the global population live in areas where there is intense dengue transmission putting them at risk of dengue infection [1]. Dengue virus causes 390 million infections annually, of which nearly a quarter are clinically apparent causing a spectrum of disease phenotypes ranging from mild dengue fever (DF) to dengue hemorrhagic fever (DHF). DHF is defined by a transient increase in vascular permeability resulting in plasma leakage, with high fever, bleeding, thrombocytopenia and haemoconcentration, which can lead to shock (dengue shock syndrome (DSS))[2]. It is however not fully understood why some people develop more severe forms of the disease, with patient history, immunity, age, viral serotype, sub-strain and epidemiological factors all postulated to play a role [3]. It was highlighted during a recent summit to identify correlates of protection for dengue, that dengue virus (DENV) specific T cell immunity should be studied in more detail, in order to develop safe and effective dengue vaccines [4].
Although a dengue vaccine (Denvaxia) is now licensed in several countries, the efficacy is low in dengue seronegative individuals and provides only partial protection against DENV2 [5]. Although it is now generally believed that DENV specific T cells are protective, it is important that dengue vaccines should not induce “harmful” T cell immunity [4, 6–8]. Indeed, a significant hurdle in developing an efficacious dengue vaccine has been our limited understanding of the protective immune response in acute dengue infection and the added complexity of the presence of four DENV serotypes that are highly homologous. Although the evidence for T cells playing a possible protective role in DENV infections is emerging, there is still conflicting evidence as to the role of antigen-specific T cells during dengue infection is reported in the literature.
T cell responses to DENV are predominantly directed towards the nonstructural proteins (NS), with the majority of the CD8+ T cell responses directed towards NS3 followed by NS5 and CD4+ T cell responses to envelope, PrM and NS1 proteins [9–11]. It was believed that highly cross-reactive T cells specific to DENV-NS3, and other proteins, associate with severe clinical disease (DHF), and it was thought that these cells contribute to DHF by inducing a ‘cytokine storm’[12–15]. It is hypothesized in the ‘original antigenic sin’ theory that T cell responses against the initial DENV serotype of primary infection persist and dominate during subsequent infections; and that these T cells are suboptimal in inducing robust antiviral responses upon re-challenge [13, 14, 16]. However, it has been shown that DENV-NS3 specific T cell responses were at very low frequency during acute disease, and only detected in the convalescent phase pointing away from a role in vascular leak [14, 16, 17]. Recently it was observed that DENV-specific T cells are found in large numbers in the skin during acute dengue infection, and it is speculated that highly cross-reactive, pathogenic skin T cells could be contributing to DHF, despite being absent or present at low frequencies in the peripheral blood [8, 18]. As the frequency of skin resident DENV-specific T cells was investigated in a small patient cohort, it is not yet clear whether the frequency of the skin T cells associated with clinical disease severity.
Conversely some studies in both humans and mouse models have shown that DENV-specific T cells in the blood are likely to be protective [19–23]. It was shown that in individuals who were naturally infected with DENV, polyfunctional CD8+ T cells responses of higher magnitude and breadth were seen for HLA alleles associated with protection [21]. Similar findings were seen with DENV-specific CD4+ T cell responses [23]. Our previous studies have also shown that the magnitude of IFNγ-producing DENV NS3-specific memory T cell responses was similar in those who had varying severity of recovered past dengue infection, suggesting that the magnitude of the memory T cell response does not correlate with clinical disease severity[22]. While many studies have been carried out to elucidate the functionality of T cell responses in dengue, these have been limited to studying T cells specific for particular HLA types by using tetramers/pentamers [16, 18], or to investigating T cell responses in individuals with unknown severity of dengue.
To aid the generation of an effective vaccine it will be important to understand the role, phenotype and frequency of dengue-specific T cell responses in relation to clinical disease severity and clearance of viraemia [6, 7]. Therefore, here we investigate T cell responses to immunodominant DENV NS proteins in patients with DHF and DF, and analyse the association of such responses with resolution of viraemia.
The study was approved by the Ethical Review Committee of The University of Sri Jayewardenepura. All patients were adults and recruited post written consent.
We recruited 74 adult patients with acute dengue infection from the National Infectious Diseases Institute, between day 4–8 of illness, following informed written consent. The patients were followed up throughout their hospital stay and all clinical features were recorded several times each day, from time of admission to discharge. Ultra sound scans were performed to determine the presence of fluid leakage in pleural and peritoneal cavities. Full blood counts, and liver transaminase measurements were performed serially through the illness. Clinical disease severity was classified according to the 2011 WHO dengue diagnostic criteria [24]. Accordingly, patients with ultrasound scan evidence of plasma leakage (those who had pleural effusions or ascites) were classified as having DHF. Those who did not develop any clinical or laboratory features of plasma leakage throughout their hospital stay, were diagnosed as having DF. Shock was defined as having cold clammy skin, along with a narrowing of pulse pressure of ≤ 20 mmHg. Based on this classification, 45 patients had DHF and 29 patients had dengue fever (DF) of the 74 patients recruited for the study.
Acute dengue infection was confirmed in serum samples using a PCR (see below) and dengue antibody detection. Dengue antibody assays were completed using a commercial capture-IgM and IgG ELISA (Panbio, Brisbane, Australia) [25, 26]. Based on the WHO criteria, those who had an IgM: IgG ratio of >1.2 were considered to have a primary dengue infection, while patients with IgM: IgG ratios <1.2 were categorized under secondary dengue infection [27]. The DENV-specific IgM and IgG ELISA was also used to semi-quantitatively determine the DENV-specific IgM and IgG titres, which were expressed in PanBio units.
DENV were serotyped and viral titres quantified as previously described [28]. RNA was extracted from the serum samples using QIAamp Viral RNA Mini Kit (Qiagen, USA) according to the manufacturer’s protocol. Multiplex quantitative real-time PCR was performed as previously described using the CDC real time PCR assay for detection of the dengue virus [29], and modified to quantify the DENV. Oligonucleotide primers and a dual labeled probe for DENV 1,2,3,4 serotypes were used (Life technologies, India) based on published sequences [29]. In order to quantify viruses, standard curves of DENV serotypes were generated as previously described in Fernando, S. et.al [28].
The peptide arrays spanning DENV NS1 (DENV-1 Singapore/S275/1990, NS1 protein NR-2751), NS3 (DENV-3, Philippines/H87/1956, NS3 protein, NR-2754) and NS5 proteins (DENV-2, New Guinea C (NGC), NS5 protein, NR-2746) were obtained from the NIH Biodefense and Emerging Infections Research Resource Repository, NIAID, NIH. The DENV NS3 peptide array consisted of 105, 14–17 mers peptides, NS1 and NS5 proteins were comprised of 60 and 156 peptides respectively. The peptides were reconstituted as previously described [30]. NS1, NS3 and NS5 peptides were pooled separately to represent the DENV- NS1, NS3 and NS5 proteins. In addition, total NS1, NS3 and NS5 peptides were combined to represent a ‘DENV-all’ pool of peptides.
Ex vivo IFNγ ELISpot assays were carried out as previously discussed using freshly isolated peripheral blood mononuclear cells (PBMC) obtained from 74 patients [22]. Fresh PBMCs were used due to concerns of possible reduction/alterations in IFNγ production by cryopreserved PBMCs [31]. DENV-NS3, NS1, NS5 and the combined DENV-ALL peptides were added at a final concentration of 10 μM and incubated overnight as previously described [16, 32]. All peptides were tested in duplicate. PHA was included as a positive control of cytokine stimulation and media alone was applied to the PBMCs as a negative control. The spots were enumerated using an automated ELISpot reader (AID Germany). Background (PBMCs plus media alone) was subtracted and data expressed as number of spot-forming units (SFU) per 106 PBMCs.
Quantitative ELISA for TNFα (Biolegend USA) and IL-2 (Mabtech, Sweden) were performed on ELISpot culture supernatants according to the manufacturer’s instructions.
PRISM version 6 was used for statistical analysis. As the data were not normally distributed, differences in means were compared using the Mann-Whitney U test (two tailed). Spearman rank order correlation coefficient was used to evaluate the correlation between variables including the association between DENV-specific T cell responses and platelet counts, degree of viraemia and liver transaminases.
To investigate the role of T cells in the progression of dengue infection we stratified patients based on disease severity. The clinical and laboratory features of the 74 patients recruited to the study are shown in Table 1. There was no statistically significant difference in the age of the individuals who had DF (median 29, IQR 29 to 42 years), when compared to those who had DHF (median 33, IQR 33 to 39 years). Of those who had DF, 18 (62.1%) were males and in those who had DHF 31 (68.9%) were males. Of these 74 patients, 45 had DHF and 29 had DF, and all 45 patients with DHF had ascites with 10 of them also experiencing pleural effusions. None of the patients developed shock and only one person progressed to bleeding manifestations (Table 1). The median duration of illness when recruited to the study was similar for patients with DF (median 5, IQR 5 to 6 days) and DHF (median 5, IQR 4 to 6 days).
To evaluate the role of T cell derived cytokine in the immunopathology or regulation of acute dengue infection, we stimulated PBMCs isolated from patients with either DF or DHF with peptides constituting DENV-derived non-structural protein (NS) and assessed cytokine production by ELISPOT. We stimulated the patient PBMCs with different pools of overlapping peptides making up either full length NS1, NS3 or NS5 protein, or a pool of total NS1, 3 and 5 peptides (DENV-all). NS3 and NS5 were selected for investigation as CD8+ T cell responses have been shown to be directed to these proteins and CD4+ T cells have been shown to target structural proteins and NS1 as the main non-structural protein[11, 21]. This combination of NS proteins from the particular DENV strains has previously been used to study DENV specific T cell responses [33]. We used this ex vivo ELISPOT method to model antigen presentation of dengue-derived peptides to antigen-specific T cells in vitro and assessed T cell activation by IFNγ production, as a representative cytokine produced by T cells during dengue infection. T cell responses to the pooled DENV peptides (DENV-ALL) (p = 0.02) were higher in PBMCs derived from patients with DF than DHF patients and the NS3-specific responses showed a trend to be higher in those with DF than DHF (Fig 1A). T cell responses to DENV-NS1 peptides were similar in patients with DF and DHF. 26/45 patients with DHF and 12/29 patients with DF, had zero responses to NS5. We did not detect TNFα in the ex vivo ELISpot culture supernatants, which is in contrast to studies performed by others on T cell clones that implied TNFα producing DENV-specific T cells contribute to disease pathogenesis [15]. We also did not detect significant quantities of IL-2. The T cell responses to the overlapping peptides were not different in males compared to females, and none of the peptide-specific responses correlated with the age of the individuals.
To assess if detection of DENV specific T cell responses before the onset of the critical phase (vascular leakage phase), was associated with a reduced likelihood of developing leakage, we isolated the data from DF and DHF patients recruited on day 4 post the onset of illness and analysed IFNγ production by peptide stimulated PBMCs. None of the patients had evidence of vascular leakage on day 4 of illness and those who developed leakage (patients with DHF), did so on day 5 or 6. DF patients (those who did not develop any plasma leakage throughout their illness) had a significantly higher IFNγ secretion response (p = 0.02) to the DENV-all peptide pool (median 42.5, IQR = 22.5 to 945 SFU/106 PBMCs), when compared to DHF patients (median 0, IQR = 0 to 12.5 SFU/106 PBMCs) (Fig 1B). As such, significantly higher DENV-specific T cell responses were seen in those who did not develop fluid leakage, and those who had lower DENV-specific T cell responses proceeded to develop fluid leakage (DHF). Responses to DENV-NS3, NS1 and NS5 also appeared higher in patients with DF at this time point, although this did not reach statistical significance (Fig 1B).
To further assess the time-course of the response we obtained a second blood sample from eight patients within our cohort two days after collection of the first sample. T cell responses to DENV-ALL and DENV-NS3 peptides increased from the first sample (day 4) to the second (day 6), but it was not statistically significant (p>0.05) (S1 Fig)
Thrombocytopenia is associated with clinical disease severity and higher degrees of thrombocytopenia are seen in those with DHF compared to those with DF[24]. We found that DENV peptide specific T cell responses correlated with the degree of thrombocytopenia. While this correlation with T cell responses and platelet counts was significant for DENV-NS1 (Spearmans r = 0.26, p = 0.01) (Fig 2A), NS5 (Spearmans r = 0.4, p = 0.0002) (Fig 2B) and DENV-All (Spearmans r = 0.31, p = 0.005) (Fig 2C), it was not significant for NS3 (Spearmans r = 0.18, p = 0.09) (Fig 2D). No association was seen with DENV-peptide specific T cell responses and aspartate transaminase (AST) and alanine transaminase (ALT) (S2 Fig), which are indicators of liver dysfunction [28, 34].
While some studies report that certain DENV serotypes associate with DHF [35, 36], others have shown that the risk of DHF is similar regardless of serotype [37]. Therefore, we proceeded to determine whether there were differences in the T cell responses to DENV-proteins based on the viral serotype that the patients were infected with. Within our cohort 30 (40.5%) patients were infected with DENV1, 19 (25.7%) with DENV2, 4 (5.4%) with DENV-3 and 2 (2.7%) with DENV-4 (Table 1). The serotype could not be determined in 19 (25.7%) patients, as they were not viraemic at the time of recruitment. DHF developed in 14/30 (46.7%) of the patients infected with DENV-1 and 15/19 (78.9%) of those infected with the DENV-2 and in 11/19 (57.9%) who were aviraemic at the time of recruitment (Fig 3A). Thus, it appeared that DENV-2 infection was more likely to lead to development of DHF (odds ratio 3.3, 95% CI 0.93 to 12.1), however the association was not statistically significant (p = 0.08) in this cohort. Aviraemic individuals displayed significantly higher IFNγ T cell responses to NS1 (p = 0.002), NS3 (p = 0.02), NS5 (p = 0.02) and DENV-ALL pooled peptides (p = 0.0004) when compared to those who were viraemic at the time of recruitment. In addition, those who were infected with the DENV-2 serotype, with a trend towards increased DHF susceptibility, had significantly lower responses to NS1 (p = 0.002), NS3 (p = 0.04), NS5 (p = 0.003) and DENV-All (p = 0.0003) peptides when compared to those who were infected with DENV-1.
As the PBMCs of patients infected with different DENV serotypes, were stimulated with the non-structural proteins from different virus serotypes (NS1-DENV1, NS3-DEN3 and NS5-DEN2), we sought to assess the sequence identity of these proteins between serotypes. Multiple alignment of the NS5 protein sequences of DENV2 (58 sequences) and DENV3 (28 sequences) was performed using virus variation resource [38] and analysed using Clustal omega and showed a sequence identity of > 72.1% between the NS5 proteins of these viral serotypes [39]. Multiple alignment of the NS3 protein of DENV2 (61 sequences) and DENV3 (28 sequences) showed a sequence identity of > 72.02%. Multiple alignment of the NS1 protein of DENV2 (62 sequences) and DENV3 (28 sequences) showed a sequence identity of >65.11%. The homology between DENV1 and DEN2 NS5 was >71.83% (comparison of 102 DENV1 sequences and 58 DENV2 sequences) [38], while the homology between DENV1 and DENV3 NS3 was >76.9% (comparison of 102 DENV1 sequences and 28 DENV3 sequences) [38]. Therefore, the T cell responses to different DENV serotypes is unlikely to be profoundly influenced by using non-structural protein sequences of different DENV serotypes. However, since there was some concern that stimulating using different serotypes would influence the results, we carried out further analysis. Since the most immunodominant protein was NS3, which was from DENV3, we also compared NS3 specific T cell responses in patients with DENV1 (n = 30) and DENV2 (n = 19). Although we did not find any significant difference in the T cell responses to NS3 based on infecting DENV serotype (p = 0.06), the responses to DENV3 NS3, appeared higher in patients with an acute DENV1 infection (median 85, IQR 5 to 435 SFU/million PBMBs) compared to those with an acute DENV2 infection (median 10, IQR 0 to 66.25 SFU/million PBMBs) (S3A Fig).
We then went on to compare responses to NS5 peptides, which were derived from DENV2, in patients with an acute DENV1 and DENV2 infection. Again, those with an acute DENV2 infection had significantly lower responses to the NS5 (p<0.0001) compared to patients with DENV1 (S3B Fig). We believe this difference is due to 15/19 (78.9%) patients with DENV2 having DHF (and thus poor T cell responses), whereas only 14/30 (46.7%) of those with DENV1 had DHF. We carried out a similar comparison within the group of patients with acute DENV1 for NS1 (which was from a DENV1 strain) and again did not see any difference within the DENV1 group in those with DF and DHF (S3C Fig), which was compatible with responses of the overall results.
DHF patients have been shown to have higher viral loads, exhibit prolonged viraemia [40, 41] and persistent DENV-NS1 antigenaemia [42, 43]. As such we attempted to elucidate a correlation between T cell cytokine responses and viremia. DENV specific T cell responses to NS1, NS3 and NS5 peptides in addition to the pooled peptides (DENV-ALL) inversely correlated with the degree of viraemia, which was most significant for DENV-NS3 specific T cell responses (Spearman’s r = -0.47, p = 0.0003) (Fig 3B and S4 Fig). The viral loads significantly inversely correlated with the platelet counts (Spearmans r = -0.34, p = 0.01), with the platelet counts being lowest in individuals with the highest viral loads.
It is thought that a second dengue virus infection with a different viral serotype is a risk factor for developing DHF[44]. To determine the effect of secondary infection on the resulting T cell response, we characterized patient infection history and assayed patient blood for the presence of dengue specific IgM and IgG. Primary infection was defined by DENV- specific IgM:IgG >1.2 [24]. Accordingly, 19 (25.7%) patients were classified as experiencing a primary dengue infection and 48 (64.9%) were defined as secondary dengue infection. The antibody results were inconclusive for 7 (9.4%) patients. Our results showed no significant difference in DENV specific T cell responses between primary and secondary dengue infection patient groups (p>0.05) for any of the DENV peptide pools (Fig 3C).
We semi-quantitatively determined the DENV-specific IgM and IgG antibody titres in all patients with DF and DHF, and we found that neither the DENV-IgM nor IgG antibody titres correlated with T cell responses to DENV-NS1, NS5 and NS3. However, the DENV-specific IgG antibody titres inversely correlated with viral loads in those with DHF (Spearman’s r = -0.37, p = 0.03) (Fig 4A), but not in those with DF (Spearmans r = -0.25, p = 0.16). In analysis of the IgG antibody titres of all patients (n = 74) they too inversely correlated with the degree of thrombocytopenia (Spearmans r = -0.29, p = 0.009) (Fig 4B). DENV-specific IgG also correlated with the highest aspartate transaminase (AST) (Spearmans r = 0.51, p = 0.004) (Fig 4C) and alanine transaminase (ALT) levels (Spearmans r = 0.4, p = 0.03) in all patients with acute dengue infection (Fig 4D).
In this study we set out to investigate the role of T cells in dengue immunity and found that DENV-specific T cells are present at low frequency during acute infection, consistent with previous reports published by us and by others [16, 17, 45]. IFNγ production was significantly higher in patients with DF as opposed to DHF, especially during early infection. Those who had lower DENV-NS3 specific T cell responses on day four since the onset of illness (before development of fluid leakage), were significantly more likely to subsequently develop vascular leakage of DHF. In addition, the frequency of pooled DENV-peptide specific, in particular DENV-NS3 specific, T cell responses was associated with resolution of viraemia. Aviraemic patients had significantly higher DENV- specific T cell responses when compared to those who were viraemic. However, overall the majority of patients with both DF and DHF, had a very low magnitude of DENV specific T cell responses. T cell IFNγ responses to DENV NS1, NS5 and pooled DENV (NS1, NS3 and NS5) peptides inversely correlated with the degree of thrombocytopenia, but we did not show any relationship with liver transaminases (AST and ALT levels). Both the degree of thrombocytopenia and a rise in both AST and ALT, have been shown to associate with dengue severity [24, 28]. Therefore, our data show that the early appearance (on day 4 of illness before any patient developed plasma leakage) of DENV-NS3 specific T cell responses is associated with milder disease, which is compatible with recent studies regarding the role of T cells in DENV infection [18, 21–23]. This suggests that DENV-peptide specific T cells are possibly have a protective role against developing severe forms of dengue infection.
Although Appana et al also evaluated ex vivo IFNγ to selected peptides of structural and non-structural DENV proteins by ELISpot assays, they did not find any differences in the frequency of DENV-specific T cell responses in patients with DF when compared to those with DHF [46]. However, only peptides that were predicted to bind to certain major HLA alleles were included in the authors’ peptide pools used in the ELISpot assays [46], whereas here we utilised peptides spanning the entire length of DENV NS1, NS3 and NS5, proteins. This difference in experimental approach may have affected the cytokine production profile of the responding T cells in the different disease states. In addition, as the viability and function of T cells have been shown to be affected in those with acute dengue infection [31], we used freshly isolated PBMCs in all our experiments to limit extraneous cellular stress in contrast to previous studies [15, 21, 46, 47].
In this study, some of the non-structural peptides used were of a different DENV serotype than that causing the acute infection as we used fresh PBMCs in our assays before the infecting serotype was known. As many previous studies on T cell responses have used peptides of unmatched DENV serotypes to assess T cell responses[33, 46] and also as it was shown that the breadth or the magnitude of DENV specific T cell responses, were similar even when peptides of unmatched serotypes were used [9], we too proceeded with the same approach. However, there is a possibility that using unmatched peptides could affect the magnitude of the T cell responses. Subsequent subanalysis of T cell responses to DENV1-NS1 and DENV3-NS3 of patients with an acute DENV1 and DENV2 infection, carried out by us also did not show any difference. However, analysis of T cell responses to DENV2 derived NS5 overlapping peptides, showed that patients with an acute DENV2 infection (matched peptides) had significantly lower responses than patients with DENV1 infection. We believe this difference is possibly due to 78.9% patients with DENV2 having DHF (and thus poor T cell responses), whereas only 46.7% of those with DENV1 had DHF.
As with most viral infections, the severity of the illness and T cell responses to the DENV has also been shown to depend on the HLA type of the individual [21, 23, 48]. However, we did not assess the T cell responses in relation to their HLA type as the primary aim of this study was to assess the overall DENV specific T cell responses based on clinical disease severity and viraemia, irrespective of their HLA types in a large cohort, with relevance to pathogenesis and vaccine design. As the HLA type is known to influence the magnitude of the T cell response, there is a possibility that the differences in the magnitude of T cell responses to different DENV overlapping peptides, could also be due to the differences in their HLA types, which were not assessed here.
In general, more severe forms of dengue infection are observed during a secondary heterologous dengue infection [4], which gave rise to the hypothesis that cross reactive T cells responding to the primary infecting DENV serotype are suboptimal in clearing the secondary virus, and lead to development of more severe disease [14, 16]. In these studies, it was shown that a tetramer of different viral specificity to the current infecting DENV serotype, sometimes had a higher affinity to the DENV specific T cells [16]. In our study, we did not observe any difference in IFNγ production in overall ex vivo ELISpot assays from PBMCs derived from patients with primary or secondary dengue infection; however, we did not examine variant peptide-specific responses. The broad differences we observed in DENV-specific T cell responses correlated only with clinical disease severity. Interestingly, the DENV-specific IgG levels, which were measured semi-quantitatively, inversely correlated with the degree of thrombocytopenia and also AST and ALT levels, which are known to associate with liver damage. DENV-specific IgG levels are known to be significantly higher in patients with secondary dengue, compared to primary dengue, indeed it is one of the criteria for definition of a secondary dengue infection. However, as we assessed the relationship of total IgG with these laboratory parameters and did not assess the presence of neutralizing antibodies, which are more likely to be protective, this needs to be further evaluated.
In summary, we found that DENV-specific T cell IFNγ responses, were associated with milder clinical disease severity and resolution of viraemia, suggesting a protective role for peptide specific T cells early in acute dengue infection.
|
10.1371/journal.ppat.1007504 | Immune-inducible non-coding RNA molecule lincRNA-IBIN connects immunity and metabolism in Drosophila melanogaster | Non-coding RNAs have important roles in regulating physiology, including immunity. Here, we performed transcriptome profiling of immune-responsive genes in Drosophila melanogaster during a Gram-positive bacterial infection, concentrating on long non-coding RNA (lncRNA) genes. The gene most highly induced by a Micrococcus luteus infection was CR44404, named Induced by Infection (lincRNA-IBIN). lincRNA-IBIN is induced by both Gram-positive and Gram-negative bacteria in Drosophila adults and parasitoid wasp Leptopilina boulardi in Drosophila larvae, as well as by the activation of the Toll or the Imd pathway in unchallenged flies. We show that upon infection, lincRNA-IBIN is expressed in the fat body, in hemocytes and in the gut, and its expression is regulated by NF-κB signaling and the chromatin modeling brahma complex. In the fat body, overexpression of lincRNA-IBIN affected the expression of Toll pathway -mediated genes. Notably, overexpression of lincRNA-IBIN in unchallenged flies elevated sugar levels in the hemolymph by enhancing the expression of genes important for glucose retrieval. These data show that lncRNA genes play a role in Drosophila immunity and indicate that lincRNA-IBIN acts as a link between innate immune responses and metabolism.
| Drosophila melanogaster is a powerful genetic model for studying the innate immune mechanisms conserved from flies to humans. With recent methodology, such as whole transcriptome analyses, novel non-protein coding genes in addition to protein coding genes are being increasingly identified. These long and short non-coding RNA genes are located between and within protein coding genes in the genome, and their functions are largely uncharacterized. In humans, such RNA genes have been shown to affect numerous physiological processes including immune responses. In Drosophila, very few non-coding RNA genes have so far been characterized in detail. In this study, we have identified and characterized an immune-inducible long non-coding RNA gene, lincRNA-IBIN. lincRNA-IBIN is induced by exposure to bacteria as well as the parasitoid wasp, Leptopilina boulardi, suggesting a general role in humoral and cellular innate immunity. Accordingly, forced expression of lincRNA-IBIN enhances the expression of genes involved in carbohydrate catabolism and elevates hemolymph glucose levels in Drosophila. These results indicate that lincRNA-IBIN acts as a link between immunity and metabolism in Drosophila. As research in Drosophila has often resulted in the identification of evolutionarily conserved mechanisms also in mammals, it remains to be studied whether long non-coding RNA genes regulate metabolism upon an infection also in humans.
| The fruit fly Drosophila melanogaster (D. melanogaster) is a widely used model system in immunological studies [1]. Drosophila has an elegant innate immune response that includes both the cellular and the humoral arms [2,3]. Activation of the cellular immune response involves mechanisms such as recognition, phagocytosis, encapsulation and the killing of parasites [4,5]. The humoral immune response is based on microbial recognition primarily by peptidoglycan recognition proteins leading to the production of antimicrobial peptides (AMPs)[6–9]. The humoral immune response is mainly mediated by two evolutionarily conserved NF-κB signaling pathways, the Toll and the Immune deficiency (Imd) pathway [10–12].
Recently, it has become evident that beside the protein coding genes that positively or negatively regulate the humoral and cellular innate immune responses, there is a multitude of short and long non-coding RNA genes that affect innate immune responses [13–16]. In between and within protein coding genes in the genome, there are thousands of uncharacterized non-coding RNA genes. Small non-coding RNAs (<200 nucleotides) are considered to have more of a “housekeeping RNA” role. However, the functions of long non-coding RNA (lncRNA, >200 nucleotides) genes are more diverse [17]. Although the number of lncRNAs is still a matter of debate, recent meta-analyses posit the human genome to give rise to >60,000 lncRNAs, albeit the majority is probably expressed at low levels [18,19]. In fruit flies, there are fewer lncRNAs in the genome and the ratio of lncRNAs to protein coding genes is lower than in humans [20]. The current lncRNA numbers can be found in the NONCODE Version v5.0 database (www.noncode.org).
The expression patterns of lncRNAs are highly specific to tissue, developmental stage and environmental conditions (reviewed in [14,15]) and they are thought to have tightly controlled biological roles. Recent studies have indicated that lncRNAs play an important functional role in innate immune responses, and specifically in innate immune cells. In mammals, lncRNA genes are expressed in monocytes, macrophages, dendritic cells, neutrophils, T-cells and B-cells [13]. A growing list of lncRNA genes, for example LincRNA-Cox2 [21], Lethe [22], PACER [23] and TNFα regulating hnRNPL interacting lncRNA (THRIL)[24], has been found to control gene expression in immune cells [13].
To study the role of lncRNA genes in the Drosophila immune response, we performed transcriptome analysis in D. melanogaster upon a bacterial infection with the Gram-positive Micrococcus luteus (M. luteus), giving particular emphasis to long non-coding RNA (lncRNA) genes. The most responsive of all transcripts was the lncRNA gene CR44404, which was upregulated 1300-fold upon a M. luteus infection. Here, we show that CR44404 is highly induced by both Gram-positive and Gram-negative bacteria in Drosophila adults and by a parasitoid wasp infection in Drosophila larvae. Because of the inducible nature of the CR44404 gene, we named it lincRNA-IBIN (Induced By INfection). Finally, we show that lincRNA-IBIN acts as a link between innate immune responses and metabolism by modulating the expression of genes regulating carbohydrate and peptide metabolism and affecting glucose levels in the hemolymph.
To investigate the importance of long non-coding RNAs in Drosophila immunity, we carried out a transcriptome analysis (RNAseq) of flies 24h after infection with the Gram-positive M.luteus in comparison to age and sex-matched uninfected controls. The RNA sequencing method used in this study recognizes polyadenylated non-coding RNAs, which are thought to represent the majority of long non-coding RNAs, although also ones without poly-A tails exist [25,26]. Prior to the transcriptome analysis, one of the Toll pathway target genes IM1 (Immune induced molecule 1), was measured from females and males upon M. luteus infection. IM1 was robustly induced in both male and female Drosophila (S1A Fig), and males were chosen for the transcriptome analysis. LYS-type peptidoglycan containing Gram-positive bacteria are known to induce the classical Toll pathway target genes including a number of antimicrobial peptides (AMPs) (e.g. [27]). As expected, AMPs were strongly upregulated in the transcriptome analysis upon a M. luteus infection, including Dro, Mtk, Drs, and multiple IMs (Fig 1A, S1 Table). Noteworthy, the highest upregulation in infected flies was seen in a previously unannotated long non-coding RNA gene, CR44404 (Fig 1A). The baseline expression of CR44404 is very low, and upon a M. luteus infection, it is induced by about 1300-fold. The induction of CR44404 expression was also shown to be comparable between males and females upon M. luteus infection (S1B Fig).
Besides CR44404, there were only 15 other lncRNAs that were more than 3-fold upregulated upon infection (Fig 1B, S2 Table). While findings from vertebrates indicate that lncRNAs have wide and important functions in immune responses [13–16], cancer and metabolism [28,29], the role of lncRNAs in Drosophila immunity has only begun to emerge. CR44404 was chosen for further analysis based on its intriguing expression pattern.
CR44404 is 228 nucleotides long (genomic loci 2R:17,671,068..17,671,295 [+]) and it is located between two protein coding genes; P32 and CG30109. Therefore, CR44404 is classified as a long non-coding intergenic RNA (lincRNA) molecule. Although CR44404 is very close to the protein-coding gene P32, the genes do not overlap. To confirm that CR44404 is an independent transcript, the expression levels of the adjacent genes were examined in the transcriptome analysis. Neither P32 nor CG30109 were affected by infection in the same way as CR44404, the expression of which was ~1300-fold upon a M. luteus infection. Instead, P32 (1.17-fold) and CG30109 (1.27-fold) were not significantly induced by M. luteus infection at 24h time point, indicating that CR44404 is expressed independently from them.
CR44404 is polyadenylated; it has a highly conserved cleavage signal sequence AAUAA towards the end of the full-length transcript. CR44404 does not contain open reading frames and based on the NCBI domain search tool [30], it does not contain any predicted protein domains. According to RNA secondary structure predictions, CR44404 is multibranched (contains 3–4 GC-rich branches) and contains a variable amount of smaller (hairpin) loops connected to a bigger loop (S2 Fig). Based on the high expression of CR44404 upon infection and its genomic location, we named the gene lincRNA-Induced By INfection (lincRNA-IBIN).
As lincRNA-IBIN was shown to be strongly induced by Gram-positive bacteria 24h p.i, we next infected male flies with either the Gram-positive M. luteus or the Gram-negative Enterobacter cloacae (E. cloacae) to measure the gene expression kinetics of lincRNA-IBIN during multiple time points ranging from 0-24h after infection (Fig 1C). This revealed that lincRNA-IBIN is also induced by Gram-negative bacteria, and in both infections, the induction occurred within the first hours of infection and gradually increased towards the 24h time point (Fig 1C). To further study the role the Toll pathway and the Imd pathway [10,11], in the expression of lincRNA-IBIN, we knocked down MyD88 (an adaptor protein functioning downstream of the Toll receptor in the Toll pathway), cactus (a negative regulator of the Toll pathway) and Relish (an Imd pathway NF-κB factor). Thereafter, we infected the flies with M. luteus or E. cloacae and measured the lincRNA-IBIN RNA levels. Upon a M. luteus infection, the expression of lincRNA-IBIN was shown to be dependent on the expression of MyD88, i.e the functional Toll pathway (Fig 1D). Knocking down MyD88 upon a E. cloacae infection had no effect on the expression of lincRNA-IBIN (Fig 1E), whereas knocking down Relish or using the RelishE20 null mutant inhibited the expression of lincRNA-IBIN, showing that it requires a functional Imd pathway in this context (Fig 1E). Knocking down Relish or using the RelishE20 null mutant upon a M. luteus infection did not inhibit the expression of lincRNA-IBIN (Fig 1D). The role of the Toll pathway activation to the expression of lincRNA-IBIN was further confirmed in uninfected flies by knocking down the inhibitor of the κB factor cactus, which strongly induced the expression of lincRNA-IBIN (Fig 1F). Also, during the larval stage, lincRNA-IBIN was induced by the ectopic expression of the constitutively active form of the Toll receptor, Toll10b (Fig 1G) and by overexpression of the Imd molecule with the ubiquitous da-GAL4 driver (Fig 1H).
Next, we tested if the Osa-containing Brahma (BAP) complex is needed for the expression of the lincRNA-IBIN. The BAP complex is a group of protein-coding genes working together in remodeling chromatin [31], and the complex has previously been reported to affect the Toll pathway-induced Drs-luc reporter in vitro in Drosophila [32,33]. Interestingly, when osa expression was knocked down, lincRNA-IBIN expression was strongly inhibited upon both a M. luteus (Fig 1I) and a E. cloacae (Fig 1J) infection. The knockdown of another BAP complex component brahma (brm) also reduced lincRNA-IBIN expression upon both infections (Fig 1I and 1J). Moreover, because lincRNA-IBIN is strongly induced by a bacterial infection in Drosophila adults, indicating a role in the humoral immune response, we next studied whether lincRNA-IBIN is also induced during the cellular immune response by infecting Drosophila larvae with Leptopilina boulardi (L. boulardi) parasitoid wasps. Also in this context, the expression of lincRNA-IBIN was strongly induced (Fig 2).
In conclusion, lincRNA-IBIN seems to have a rather broad role in the immune response, being induced by a bacterial infection in flies and by parasitoid wasps in larvae. The M. luteus -mediated induction of lincRNA-IBIN expression was shown to be dependent on the Toll pathway, whereas the E. cloacae -mediated induction requires the Relish/Imd pathway. In each studied case, lincRNA-IBIN expression was dependent on a functional BAP complex. This type of unspecific induction via both NF-κB pathways is rather uncommon in Drosophila and argues for a general immunity related function for lincRNA-IBIN.
To understand the role of lincRNA-IBIN in Drosophila immunity, we investigated where lincRNA-IBIN was expressed and whether its effects were tissue-specific. Since lincRNA-IBIN expression is strongly infection-inducible, we reasoned that it is most likely expressed in immune-responsive tissues (the fat body and hemocytes, the Drosophila blood cells). Wasp infection of Drosophila larvae led to the induction of lincRNA-IBIN expression in the fat body and hemocytes (Fig 2A and 2B). Like in flies, osa RNAi in larval hemocytes (HH> osaIR) and fat bodies (C564> osaIR) kept the expression of lincRNA-IBIN close to the basal level (Fig 2A and 2B). Because lincRNA-IBIN is a short gene and very strongly induced upon infection like AMPs, we next investigated whether lincRNA-IBIN is secreted into the plasma in similar manner as AMPs (Fig 2C). First, we confirmed that hemocytes and plasma were separated by centrifugation (S3 Fig). We also checked the expression of a hemocyte specific gene Hemolectin (Hml) and a fat body-specific gene Larval serum protein 1 alpha (Lsp1α) in each tissue sample (Fig 2C, i and ii). A Hml signal was detected in the hemocyte fraction, whereas Lsp1α levels were high in fat body samples but not in hemocytes or in the plasma (Fig 2C, i and ii). We did not detect lincRNA-IBIN in the plasma fraction in large quantities (Fig 2C, iii).
Pin-pointing the cellular localization of a lncRNA reveals typically more about its function than does the structure of the RNA. A RNA FISH (RNA Fluorescent In Situ Hybridization) protocol was performed with 3rd instar larval hemocytes. Uninfected w1118 larval hemocytes were used as a control for imaging the basal expression level and localization of lincRNA-IBIN (Fig 2D, ii). Hemocytes from larvae overexpressing lincRNA-IBIN1 (HH>lincRNA-IBIN1) and infected with L. boulardi were used to induce the expression of lincRNA-IBIN (Fig 2D, iii), and they showed that lincRNA-IBIN (pink labelling) was primarily expressed in the nuclear compartment (blue labelling) of the cell. Therefore, we conclude that lincRNA-IBIN is expressed in immune responsive tissues, is not secreted into the plasma in large amounts and its cellular localization is mainly nuclear. This suggests that the function of lincRNA-IBIN may be in the regulation of gene expression, which is typical for lncRNAs [34–36].
To study the function of lincRNA-IBIN in uninfected and infected flies, we generated UAS-lincRNA-IBIN overexpression fly lines. Two of the generated lines, lincRNA-IBIN1 and lincRNA-IBIN7, were selected for the following experiments. lincRNA-IBIN overexpression in the lincRNA-IBIN1 and lincRNA-IBIN7 lines was induced using the C564-GAL4 driver, which is expressed strongly in the fat body [37,38](Fig 3A). lincRNA-IBIN expression in uninfected flies was significantly increased in both overexpression lines (Fig 3A, white bars), with higher expression levels in the lincRNA-IBIN7 line. 24 h after a M. luteus infection, the effect of the overexpression on the expression of lincRNA-IBIN was masked by the overwhelming endogenous expression of lincRNA-IBIN (Fig 3A, black bars). To study the effect of the long-term exposure of flies to elevated levels of lincRNA-IBIN, we monitored the lifespan of flies overexpressing lincRNA-IBIN with the C564-GAL4 driver and controls. To ensure maximal lincRNA-IBIN expression, flies were cultured at +29°C for the duration of the experiment. Neither one of the lincRNA-IBIN overexpression lines (lincRNA-IBIN1 and lincRNA-IBIN7) showed a statistically significant difference in the lifespan between flies overexpressing lincRNA-IBIN and controls (S4 Fig).
As lincRNA-IBIN was the most strongly induced gene upon a M. luteus infection, we first investigated whether overexpressing lincRNA-IBIN affected the survival of the flies against a septic infection with Gram-positive bacteria. For the survival experiment, we chose lincRNA-IBIN7 flies as these produced the highest overexpression without an infection. We first infected lincRNA-IBIN7 flies with M. luteus to prime the Toll pathway. 24h later, the flies were infected with the more pathogenic bacteria, Enterococcus faecalis (E. faecalis) [39]. Overexpression of lincRNA-IBIN7 (C564-GAL4>lincRNA-IBIN7) improved the survival of the flies from the infection compared to the flies not overexpressing lincRNA-IBIN (Fig 3B). MyD88 (a positive regulator of the Toll pathway) and cactus (a negative regulator of the Toll pathway) knock-down flies were used as controls. This indicates that lincRNA-IBIN positively affects immunity against the pathogenic Gram-positive bacteria E. faecalis.
Next, we investigated whether lincRNA-IBIN regulates the central feature of the fly immune defense against Gram-positive bacteria, namely the production of AMPs via the Toll pathway. The expression of the Toll pathway mediated genes was monitored in flies overexpressing lincRNA-IBIN and controls after exposure to M. luteus for 24 hours (Fig 3C and 3D). lincRNA-IBIN overexpression using both the lincRNA-IBIN1 and lincRNA-IBIN7 lines with the C564-GAL4 driver resulted in significantly elevated levels of IM1 upon infection (Fig 3C). The expression of Drosomycin was elevated in C564>lincRNA-IBIN7 flies, whereas in the lincRNA-IBIN1 line the trend was similar, yet not significant (Fig 3D). As expected, MyD88 knockdown decreased the expression of IM1 and Drosomycin upon infection, whereas cactus knockdown caused a strong induction of IM1 and Drosomycin expression also in the uninfected flies (Fig 3C and 3D).
To address the importance of lincRNA-IBIN in a situation where the expression of endogenous lincRNA-IBIN is prevented, we utilized the following experimental approach. Upon an E. cloacae infection, lincRNA-IBIN expression is fully dependent on Relish (Fig 1E). Relish RNAi flies do not produce endogenous lincRNA-IBIN upon an E. cloacae infection, but in the Relish RNAi flies combined with the lincRNA-IBIN7 construct, lincRNA-IBIN is overexpressed (Fig 3E). Next, we monitored the survival of Relish RNAi flies and Relish RNAi flies with the lincRNA-IBIN7 construct from an E. cloacae infection (Fig 3F). Fig 3F demonstrates that lincRNA-IBIN overexpression provides protection against an E. cloacae infection. lincRNA-IBIN overexpression does not itself induce antimicrobial peptides (Fig 3G and 3H), indicating that the protection is independent of the AMPs. Taken together, lincRNA-IBIN overexpression enhances the expression of target genes of the Toll pathway. Upon infection, lincRNA-IBIN overexpression gave flies a survival advantage. However, this is not due to the induction of AMPs itself, but results from a mechanism that prompts further investigation.
As shown in Fig 2, lincRNA-IBIN is expressed in immunogenic tissues in the fly, such as the fat body and hemocytes. Next, we examined the role of lincRNA-IBIN overexpression in the cellular response, i.e. the hemocytes. Phagocytic plasmatocytes are the main hemocyte type in uninfected larvae. Lamellocytes, which are formed upon a parasitoid wasp infection, function in the encapsulation of the wasp eggs and larvae [5,40,41]. To further investigate if lincRNA-IBIN has a role in two major components of the cellular immune response, namely the increase in hemocyte numbers and differentiation of lamellocytes, we utilized the hemocyte reporters (msnCherry,eaterGFP) to detect hemocytes with flow cytometer. The combination of the reporters with the hemocyte (MeHH> for short, see materials and methods) and fat body (MeC564>) drivers enabled us to detect the hemocytes and overexpress lincRNA-IBIN in these tissues. Driving lincRNA-IBIN expression in hemocytes (MeHH>lincRNA-IBIN) resulted in an increase in total hemocyte numbers in uninfected larvae (S5A Fig), but did not induce ectopic lamellocyte formation (S5A’ Fig). lincRNA-IBIN overexpression in the fat body (MeC564>lincRNA-IBIN) did not have an effect on hemocytes (S5B–S5B’ Fig).
lincRNA-IBIN could enhance the proliferation of hemocytes or their release from a reservoir located in segmental bands under the larval cuticle, called the sessile compartment [42,43]. To that end, we imaged whole larvae and checked for the existence of sessile bands. We did not observe any noticeable loss of sessile bands that could explain the increased hemocyte numbers (S5C Fig). In L. boulardi-infected larvae, there was a slight decrease in the numbers of hemocytes in the lincRNA-IBIN7 line (S5A Fig), but lamellocytes were not affected (S5A’ Fig). Taken together, the overexpression of lincRNA-IBIN in hemocytes increases the hemocyte numbers, but does not affect hemocyte differentiation, in unchallenged Drosophila larvae.
To identify the downstream pathways and targets of lincRNA-IBIN, we performed transcriptome profiling of flies overexpressing (OE) the lincRNA-IBIN7 construct with the C564-GAL4 driver. lincRNA-IBIN OE and control flies were either infected with M. luteus or E. cloacae, or they were left uninfected. In uninfected lincRNA-IBIN OE flies, lincRNA-IBIN expression was induced 166-fold (Fig 4A (472±33 normalized number of reads for C564>IBIN vs. 2.8±0.62 normalized number of reads for w, IBIN). Also in this transcriptome analysis, target genes of the Toll pathway were induced in M. luteus-infected flies, and lincRNA-IBIN OE further elevated their levels (S3 Table). Candidate target genes under lincRNA-IBIN regulation were searched for in the uninfected lincRNA-IBIN7 transcriptome data using a cut-off value of 2 for fold change. Genes with induced expression level from medium to high (>10 reads) in the treatment of interest were included in the analysis. Based on this criteria, 45 genes (including lincRNA-IBIN) were upregulated (S4 Table) and 21 genes were downregulated (S5 Table) in unchallenged lincRNA-IBIN OE flies. The top upregulated genes included one of the copies of Major heat shock 70 kDa protein, Hsp70Bb (32.9-fold), Niemann-Pick type C-2 (Npc2e, 29.6-fold) and Amyrel (6.8-fold; S4 Table). Among the most downregulated genes were εTrypsin (5.3-fold) and βTrypsin (3.2-fold) both involved in proteolysis (S5 Table). A cluster analysis of the up- and downregulated genes revealed two major gene clusters, both of which are involved in metabolism (Fig 4B). Eight out of forty-five genes upregulated in lincRNA-IBIN overexpressing flies belong to a carbohydrate metabolism/glycoside hydrolase gene cluster, and six out of twenty-one downregulated genes belong to a proteolysis / peptidase S1 gene cluster (Fig 4B).
To investigate the identified metabolism related gene clusters in more detail, we plotted normalized read values of selected genes including all the treatments (uninfected, lincRNA-IBIN7 OE, E. cloacae and M.luteus infections with and without lincRNA-IBIN7 OE; Fig 4C & 4D). As shown in Fig 4C, overexpression of lincRNA-IBIN7 in uninfected flies causes the upregulation of six maltase genes (Mal-A1, Mal-A6, Mal-A7, Mal-A8, Mal-A2, Mal-B1), whose function is to catalyze the hydrolysis of maltose (disaccharide) to glucose units (monosaccharide). In addition, lincRNA-IBIN7 overexpression increased the expression of amylases (Amy-p, Amy-d, Amyrel), which hydrolyze dietary starch into disaccharides. tobi (target of brain insulin) and Gba1a (Glucocerebrosidase 1a) are also involved in sugar metabolism, and were elevated by lincRNA-IBIN7 overexpression (Fig 4C). Of note, the expression of some of these genes involved in carbohydrate metabolism was also elevated upon M. luteus and E. cloacae infections (Fig 4C, blue and orange bars). This demonstrates that enhanced sugar metabolism is needed to fight infections as shown before [44–46], and indicates that lincRNA-IBIN is involved in providing this metabolic switch. In contrast, many genes involved in peptide and protein catabolism were downregulated upon lincRNA-IBIN7 overexpression and infection (Fig 4D). The enhanced (S6B and S6C Fig) or downregulated (S6D Fig) expression of selected metabolic genes upon lincRNA-IBIN overexpression (S6A Fig) was confirmed by qPCR also with the other lincRNA-IBIN overexpressing fly line, lincRNA-IBIN1.
According to the Flybase high throughput expression data (www.flybase.org), all the identified metabolism genes are almost exclusively expressed in the midgut. Therefore we next analyzed whether lincRNA-IBIN is expressed in the adult midgut in response to septic injury. To ensure maximal induction of lincRNA-IBIN, we infected adult flies by pricking them with E. cloacae, and 24h later dissected the midgut tissues of the infected flies and controls. An E. cloacae infection induced the expression of lincRNA-IBIN also in the adult midgut (Fig 4E). Furthermore, strong expression of the C564-GAL4 driver in the adult midgut was demonstrated (S7A Fig), indicating that besides the fat body, lincRNA-IBIN is also overexpressed in the midgut when the driver C564-GAL4 is used (S7B Fig). This is in line with the observed transcriptional changes when the overexpression of lincRNA-IBIN was driven using the C564-GAL4 driver.
To further study changes in glucose metabolism during an infection, we measured glucose and trehalose levels from the hemolymph 24h after a septic injury with E. cloacae. Hemolymph from uninfected and infected flies (50 flies in each sample) was collected for the analysis (Fig 4F). An infection induced an increase in the glucose level in circulating hemolymph (Fig 4G), whereas the level of circulating trehalose was unaltered (S7C Fig). This suggests that increasing the hemolymph glucose level is important for the immune response. Next, to investigate the role of lincRNA-IBIN on the hemolymph glucose level, hemolymph from flies overexpressing lincRNA-IBIN7 with the C564-GAL4 driver and controls (w; lincRNA-IBIN7) was collected as described (Fig 4F). In addition to an infection (Fig 4G), also lincRNA-IBIN7 overexpression caused a statistically significant increase in the glucose level in circulating hemolymph (Fig 4H). In conclusion, these data imply that a septic infection with Gram-negative bacteria enhances the transcription of genes involved in carbohydrate metabolism, which leads to elevated sugar levels in the hemolymph. lincRNA-IBIN regulates in part this metabolic shift to ensure sufficient energy resources for the needs of the immune cells and tissues.
Drosophila melanogaster has been one of the most fruitful models for studying the immune response [1]. For example, identification of the regulators of the signaling cascades of innate immunity in Drosophila has greatly advanced our understanding of the control of mammalian immune responses [3,47]. Today, key pathways and proteins controlling the immune reactions in Drosophila are well documented [10,11,48,49]. However, while findings in vertebrates indicate that lncRNAs have wide and important functions in immune responses [13–16], cancer and metabolism [26, 27], the role of lncRNAs in Drosophila immunity has only begun to emerge.
Based on our Drosophila transcriptome analysis, only few lncRNA genes were up- or downregulated in response to an infection with the Gram-positive LYS-type peptidoglycan containing bacteria M. luteus. However, expression of the gene CR44404, named as lincRNA-IBIN (Induced By INfection), was highly upregulated during a M. luteus infection in flies. lincRNA-IBIN expression was also induced by the Gram-negative DAP-type peptidoglycan containing bacteria E. cloacae, indicating that activation of either the Toll or the Imd pathway induces its expression. As an infection with the parasitoid wasp L. boulardi also induced the expression of lincRNA-IBIN, this indicates that lincRNA-IBIN might have a general role in immunity.
The expression patterns of lncRNAs have been found to be highly specific for tissue, developmental stage and context (reviewed in [14,15]). When studying the expression levels of lincRNA-IBIN in larvae, the highest expression levels were found in the fat body. lincRNA-IBIN was also expressed in hemocytes, but was not secreted in considerable amounts into the plasma. The basal level of lincRNA-IBIN expression in these tissues was very low. However, the lincRNA-IBIN response to an infection was fast, as already at two hours after a bacterial infection the expression of lincRNA-IBIN was highly induced. This shows similar expression kinetics to AMPs and further argues for an important function for lincRNA-IBIN in immunity.
The functional importance of lincRNA-IBIN was studied by overexpressing it in the fat body and in hemocytes with the UAS-GAL4-system. Overexpressing lincRNA-IBIN locally in the fat body resulted in elevated levels of the antimicrobial peptide Drosomycin and another Toll pathway target, IM1, upon a M. luteus infection. lincRNA-IBIN overexpression also led to enhanced resistance against an infection with Gram-positive bacteria. To study the function of lincRNA-IBIN further, we performed a full transcriptome analysis of uninfected and infected flies overexpressing lincRNA-IBIN. The most upregulated gene in lincRNA-IBIN overexpressing flies was Hsp70Bb. The main function of the Hsp70 proteins, like other heat shock proteins, is to maintain the proper trafficking and folding of proteins [50]. In addition, Hsp70 expression has been shown to be induced by the Gram-negative Erwinia carotovora carotovora [51] and by medium from γ-irradiated Escherichia coli bacteria [52]. As a stress-responsive gene, the upregulation of Hsp70Bb could indicate that lincRNA-IBIN overexpressing flies are experiencing stress. However, no other heat shock genes were induced in lincRNA-IBIN OE flies, indicating that a general stress response is unlikely. Another highly upregulated gene in lincRNA-IBIN overexpressing flies was Npc2e. Npc2e binds microbial components, and it has been indicated to play a role in the Imd pathway [53]. In our data, Npc2e was also upregulated after E. cloacae and M. luteus infections (S4 Table), but these changes were not significant after an FDR correction.
Importantly, the lincRNA-IBIN OE transcriptome analysis revealed two main gene clusters, which were affected by lincRNA-IBIN overexpression, namely glycoside hydrolases (upregulation) and peptidases and proteases (downregulation). Inflammatory responses from infection to cancer involve changes in metabolic pathways [54,55]. The activated immune cells switch from oxidative phosphorylation toward aerobic glycolysis; this is a well-known phenomenon in cancer cells called the Warburg effect, and it is also recognized as important for immune cells upon an infection [44,56,57]. The metabolic switch toward aerobic glycolysis leads to increased glucose consumption [58]. In Drosophila, the polysaccharide starch can be hydrolyzed into the disaccharide maltose and further on into the monosaccharide glucose by maltases. lincRNA-IBIN overexpression elevated the expression of multiple maltase genes, indicating that lincRNA-IBIN has a role in enhancing the catabolism of starch. Here we showed that besides the fat body cells and hemocytes, lincRNA-IBIN was also expressed in the Drosophila gut upon infection, which is the main site of starch metabolism. lincRNA-IBIN overexpression with the C564-GAL4 driver, which is also strongly induced in the midgut, led to elevated levels of free glucose in adult hemolymph providing an energy source for immune cells. lincRNA-IBIN did not seem to affect the insulin signalling pathway, which in humans, and in flies, controls peripheral as well as central nervous system -related aspects of metabolism [59,60].
Glucose and trehalose are the major circulating sugars in Drosophila, and the majority of the total sugar in the hemolymph is trehalose [61]. lincRNA-IBIN overexpression enhanced the expression of glycoside hydrolases and led to slightly elevated levels of hemolymph glucose, but not trehalose. Therefore, lincRNA-IBIN may improve the retrieval of glucose from dietary sugars from the gut, for example by enhancing the expression of maltases. More elevated glucose levels were seen by a septic infection. It is likely that upon infection, also other factors influence the elevation of glucose levels compared to lincRNA-IBIN overexpression alone; for example, the enhanced release of glucose from glycogen storages upon infection has been shown [44,46].
The expression of genes related to proteolysis was downregulated upon lincRNA-IBIN overexpression, indicating that the uptake of amino acids from food is reduced. Immune responses are tightly regulated as prolonged inflammation is costly to the host [62,63]. Dionne and coworkers showed that flies infected with Mycobacterium marinum undergo a process resembling wasting, where the flies progressively lose metabolic stores in the form of fat and glycogen and become hyperglycemic [45]. Hence, there must be delicately controlled mechanisms for the direct control of energy allocation to the immune response upon infection or inflammation. Here, we have characterized lincRNA-IBIN as one of the potential regulators of this switch.
It is important to notice, however, that we have used UAS/GAL4-based overexpression of lincRNA-IBIN to study its function. We trust that this artificial overexpression system provides valid information about the effect of lincRNA-IBIN on its target genes, but one cannot exclude that some of the observed effects may be caused by non-physiological lincRNA-IBIN expression. For example, overexpressing a short RNA molecule, such as lincRNA-IBIN, may trigger antiviral immune responses. However, the normal life span and the lack of elevated expression of known virus infection-responsive genes such as Vago, AGO2 or Sting [64–66] in lincRNA-IBIN OE flies argues against a non-specific viral response. However, the more definite conclusion about the role of lincRNA-IBIN in Drosophila immunity will require the generation of a loss-of-function mutant. lincRNA-IBIN mutant flies would also be essential to address whether lincRNA-IBIN is required for the observed metabolic changes during an infection and to address if lincRNA-IBIN is required for normal innate immunity in Drosophila.
lincRNA-IBIN, like most lncRNAs, demonstrates low evolutionary sequence conservation, and the lack of homologous sequences prevented the use of similar sequences for the identification of a function for lincRNA-IBIN. lncRNAs are composed of domains that permit either protein binding and/or base-pairing with RNA or DNA sequences [34,67–70]. Based on a secondary structure prediction, it is difficult to estimate whether lincRNA-IBIN interacts with protein, DNA or RNA molecules. The primary functions of lncRNAs can be traced according to their cellular localization. Based on our RNA fluorescent in situ hybridization analysis, lincRNA-IBIN was mostly localized within the nucleus. The nuclear localization suggests that lincRNA-IBIN localizes to its genomic target site(s) through RNA-DNA or RNA-protein interactions in the chromatin, where it could modulate chromatin regions or the expression of target genes by functioning as a guide, decoy or a scaffold for interacting molecules.
An overview of lincRNA-IBIN functions is presented in Fig 5. Taken together, lincRNA-IBIN has a role in both humoral and cellular innate immune pathways in larvae and adults. The basal expression level of lincRNA-IBIN is very low and it responds rapidly to an infection, but further studies are required to fully understand its role in different infection contexts. lincRNA-IBIN is expressed in immune responsive tissues and its expression is regulated by NF-κB signaling and the chromatin modeling BAP complex. In the gut, lincRNA-IBIN has a role in the activation of glycoside hydrolases. Finally, expression of lincRNA-IBIN elevates the levels of free glucose in the hemolymph. Based on our findings we postulate that lincRNA-IBIN has an important role in the metabolic switch required to provide additional glucose for immune cells during a systemic infection in Drosophila.
The UAS-GAL4 -based system in Drosophila was utilized in most experiments to achieve the silencing or overexpression of genes in the F1 progeny [71]. The brm (transformant ID #37720), osa (#7810), MyD88 (#25399) and Relish (#108469) UAS-RNAi lines, and the isogenized w1118 flies (w1118iso) that were used as controls, were obtained from the Vienna Drosophila Resource Center (www.vdrc.at). cactus UAS-RNAi flies (5848R-3) were obtained from the National Institute of Genetics Fly Stock Center in Japan. RelishE20 mutant flies were a kind gift from prof. Dan Hultmark. The Imd overexpression line that overexpresses the Imd protein under the control of UAS, was originally a kind gift from prof. Jules Hoffmann. The constitutively active Toll10b mutant [72] was originally obtained from the Bloomington Drosophila Stock Center at Indiana University. C564-GAL4 flies, driving the expression of a UAS-construct in the fat body [37,38,73] and some other tissues, were a kind gift from Prof. Bruno Lemaitre (Global Health Institute, EPFL, Switzerland). The Daughterless-GAL4 (Da-GAL4) flies drive the expression of a UAS-construct ubiquitously [74]. The combination of the HmlΔ-GAL4 (w1118; P{Hml-GAL4.Δ}2P{wUAS-2xEGFP}AH2) driver [75] and the He-GAL4 (P{He-GAL4.Z}) driver [76](HH-GAL4) was used to drive the expression of the UAS-constructs in hemocytes [77]. The GAL4 drivers were backcrossed into the w1118 background that was used as a control in crosses without a GAL4-driver. The hemocyte reporter lines eaterGFP (for plasmatocytes) [78] and MSNF9mo-mCherry (for lamellocytes, hereafter called msnCherry) [79] were obtained from Robert Schulz’s laboratory. The lines were recombined to create the msnCherry,eaterGFP reporter line. The msnCherry,eaterGFP reporter was further crossed with C564-GAL4 to obtain msnCherry,eaterGFP;C564-GAL4 (MeC564> for short). msnCherry, eaterGFP; HmlΔ-GAL4; the He-GAL4 (MeHH>) line was a kind gift from I. Anderl.
To create lincRNA-IBIN overexpressing fly lines, the full-length gene for lincRNA-IBIN was cloned into the EcoRI and BcuI (SpeI) restriction sites in the pUAST vector using the following primers (with restriction sites underlined): CR44404_F: TAAGCAGAATTCCACAATCTAAAGTTAACTTGCC and CR44404_R: CACACAACTAGTGTTTATTTTCTTTCTATGGTTG. The produced plasmids were injected into the w1118 background in Best Gene Inc., USA (thebestgene.com). Ten lines producing red-eyed transformants were generated, and two lines (lincRNA-IBIN1 and lincRNA-IBIN7) with good overexpression of lincRNA-IBIN were selected for experiments.
For the experiments, 10–15 virgin females were crossed with 5–7 males per vial containing mashed-potato, syrup and yeast-based fly food medium. Crosses were kept at +25°C and flies transferred daily into fresh vials. The vials with eggs were transferred to +29°C after one day of egg laying and kept there until the experiments at the larval or adult stage unless otherwise stated. Test groups and controls were kept at the same conditions at all times. After testing that target genes of the Toll pathway were induced in a similar manner in the progeny male and female flies, male flies were used for the transcriptome analysis with and without a M. luteus infection. The expression of CR44404 was tested in both female and male flies and found to be equivalent, after which male flies were used in all of the subsequent experiments.
For the lifespan experiment, lincRNA-IBIN overexpressing flies (lincRNA-IBIN1 and lincRNA-IBIN7 lines) were crossed with C564> driver flies at +25°C. After one day, the eggs were transferred to develop at +29°C for a maximum lincRNA-IBIN overexpression for the entire lifespan of the flies. Twice a week, the number of the flies was recorded and the flies transferred to fresh food.
Micrococcus luteus (M. luteus) was cultured on Luria-Bertani (LB) agar plates under Streptomycin selection (final concentration 100 μg/ml) and left to grow at 29°C for 2–3 days. Enterobacter cloacae (E. cloacae) was cultured on LB agar plates under Nalidixic acid selection (final concentration 15 μg/ml) and the plates were incubated overnight at 37°C. The concentrated bacterial culture used for pricking the flies was prepared by collecting the colonies from the plate into 100μl of 50% glycerol in phosphate buffered saline (PBS; 137 mmol/l NaCl, 2.7 mmol/l KCl, 10 mmol/l Na2HPO4, 1.8 mmol/l KH2PO4).
Enterococcus faecalis (E. faecalis) was cultured in Brain-Heart-Infusion (BHI) medium and incubated at 37°C with shaking (225 rpm) overnight. The absorbance of the E. faecalis bacterial culture grown overnight was measured with a spectrophotometer at 600nm after which it was diluted 1:25 in 5 ml of BHI medium and left to grow for 2–3 hours at 37°C with shaking (225 rpm) until the absorbance at 600nm was 0.75. Then 2 ml of the bacterial culture was centrifuged at 700 x g for 5 min and the supernatant was discarded. The pellet was resuspended into 100 μl of 50% glycerol in PBS and the bacterial concentrate was used for pricking the flies.
For bacterial infections, 0–2 day old male flies were collected and placed at +29°C for 48h, after which a septic injury to the flies was caused by pricking them in the thorax with a thin sharp tungsten wire dipped into a concentrated bacterial culture. For the activation of the Toll pathway, flies were infected with M. luteus (Gram-positive bacteria) and incubated for 24h at 25°C. For measuring AMP expression levels, infected flies and non-infected controls were incubated at 25°C for the duration of the infection, harvested, and their RNAs were extracted as described below. For survival experiments, M. luteus infected flies (24h, 25°C) were subsequently infected with E. faecalis and incubated at RT. The survival of the flies was monitored for 48h, as described earlier [39]. To activate the Imd pathway, the flies were infected with the Gram-negative bacterium E. cloacae, and the flies were incubated at 25°C for the duration of the infection.
2nd instar larvae were infected with strain G486 of L. boulardi parasitoid wasps by placing 20 female wasps in vials with larvae. After two hours at room temperature, the wasps were removed and the larvae were transferred back to +29°C. 48 hours later the larvae were dissected to collect the hemocytes, plasma and fat bodies. The infection status of the larvae was checked by visually confirming the presence of L. boulardi eggs or larvae.
For the first transcriptome analysis (Fig 1A and 1B), total RNAs from uninfected or M. luteus -infected (24h p.i.) w, osaIR male flies were extracted with the TRI reagent. For the second transcriptome analysis (Fig 4), total RNAs were extracted from uninfected male flies with lincRNA-IBIN overexpression (C564>lincRNA-IBIN7), uninfected controls (w1118,lincRNA-IBIN7), M. luteus -infected lincRNA-IBIN OE and control flies (24 h p.i.) and E. cloacae -infected lincRNA-IBIN OE and control flies (6 h p.i.). All the sample groups were crossed at the same time and kept in the same conditions until collection. The resulting RNA samples were DNase treated with the RapidOut DNA removal kit (Thermo Scientific). The quality of the total RNA samples was ensured with the Advanced Analytical Fragment Analyzer and found to be good. Total RNA samples were pure, intact and all samples were of similar quality.
The preparation of the RNA libraries and Illumina HiSeq 2500 sequencing were carried out in the Finnish Microarray and Sequencing Centre (Turku, Finland). The RNA libraries were prepared according to the Illumina TruSeq Stranded mRNA Sample Preparation Guide (part # 15031047): Firstly, the poly-A containing RNA molecules were purified using a poly-T oligo attached to magnetic beads. Following purification, the RNA was fragmented into small pieces using divalent cations under an elevated temperature. The cleaved RNA fragments were copied into first strand cDNA using reverse transcriptase and random primers. Strand specificity was achieved by replacing dTTP with dUTP in the Second Strand, followed by second strand cDNA synthesis using DNA Polymerase I and RNase H. The incorporation of dUTP in second strand synthesis quenches the second strand during amplification, because the polymerase used in the assay is not incorporated past this nucleotide. The addition of Actinomycin D to First Stand Synthesis Act D mix (FSA) prevents spurious DNA-dependent synthesis, while allowing RNA-dependent synthesis, improving strand specificity. These cDNA fragments then have the addition of a single 'A' base and subsequent ligation of the adapter: the Unique Illumina TruSeq indexing adapter was ligated to each sample during the adapter ligation step for later pooling of several samples in one flow cell lane. The products were then purified and enriched with PCR to create the final cDNA library. Typically, the RNAseq library fragments are in the range of 200–700 bp and the average size of the fragments is 250–350 bp. The samples were normalized, pooled for the automated cluster preparation and sequenced with an Illumina HiSeq 2500 instrument using TruSeq v3 sequencing chemistry. Paired-end sequencing with a 1 x 50 bp read length was used, followed by a 6 bp index run. The technical quality of the HiSeq 2500 run was good and the cluster amount was as expected.
In both transcriptome analyses, the reads obtained were aligned against the Drosophila melanogaster reference genome (BDGP6 assembly, downloaded from the Illumina iGenomes website and originally derived from Ensembl). The reads were associated with known genes based on RefSeq annotations derived from UCSC database and the number of reads associated with each gene was counted using the featureCount method. The counts were normalized using the TMM normalisation method of the edgeR R/Bioconductor package. The number of reads is represented as RPKM values (Reads Per Kilobase of exon per Million reads mapped). RPKM = total gene reads / [mapped reads (millions) x total length of gene exons (kb)]. Genes with expression values (read number) of less than 0.125 across the treatments were considered to be expressed at low levels and excluded from the analysis.
For extracting RNA from whole flies or larvae, 3 x 5 individuals per phenotype were collected and snap-frozen on dry ice or in liquid nitrogen. For RNA extraction from the fat body, fat bodies from 3rd instar larvae were dissected with forceps under a stereomicroscope and washed by dipping them three times into a 20 μl drop of 1 x PBS. In total, three biological replicates were prepared and pools of whole fat bodies from ten larvae per each biological replicate were used. Samples were snap-frozen in liquid nitrogen and stored at -80°C until RNA extraction. For RNA extraction from hemocytes and plasma, 55–60 larvae per replicate were washed, placed in a drop of 1 x PBS on a multiwell glass slide and dissected with forceps to release hemolymph. To separate hemocytes and plasma from hemolymph, suspensions were centrifuged at 2500 x g for 10 min, after which the plasma was carefully pipetted into a clean tube. Hemocytes and plasma samples were snap-frozen in liquid nitrogen and stored at -80°C until RNA extraction. For RNA extraction from adult guts, flies were dipped in 70% ethanol and dissected on a glass slide in 15 μl of 1 x PBS. The midgut region of the gut was separated and washed in a second drop of 1 x PBS. Guts from 10 flies per sample were pooled and centrifuged at 2000 x g for 2 min, after which PBS was removed and guts snap-frozen in liquid nitrogen and stored at -80°C until extraction.
To start the RNA extraction, a sufficient amount of the TRI reagent (MRC, Fisher Scientific) was added to the frozen whole flies, larvae or tissues. Whole flies, larvae, fat body and gut tissues were quickly thawed and homogenized in the TRI reagent using a micropestle (Fisher Scientific). Hemocytes were homogenized in the TRI reagent by pipetting up and down for a minimum of ten times. Plasma samples were quickly thawed and suspended in the TRIzol LS reagent (Thermo Fisher Scientific) by pipetting up and down ten times. Thereafter, total RNAs were extracted according to the manufacturer’s (TRI reagent or TRIzol LS) instructions. RNA pellets were dissolved in nuclease-free water, and the RNA concentrations and the purity were determined by a Nano-Drop 2000 (Thermo Scientific) measurement.
Quantitative real-time PCR (qRT-PCR) was carried out with the iTaq Universal SYBR Green One-step kit (Bio-Rad, Hercules, CA, USA) using total RNAs (approximately 40 ng/sample) as templates. RpL32 or ND-39 was used as a housekeeping gene to normalize differences in RNA amounts between samples. In the experiments presented in Fig 3C, the amounts were standardized to 40 ng of total RNA/sample. This is because no products/mRNA from genes that are considered to have a housekeeping role are normally found in the plasma. Expression levels of genes in the test groups and controls were measured within the same qRT-PCR experiment. If the samples within an experiment did not fit in one 96-well plate, a reference sample was measured in all plates to make internal normalization between plates possible. In the qRT-PCR experimental figures, one uninfected control sample (indicated in the figure legend) was set to 1, to calculate fold-induction values. The primers used are listed in Table 1.
Individual 3rd instar wasp-infected and uninfected msnCherry,eaterGFP;C564>lincRNA-IBIN (MeC564>lincRNA-IBIN), msnCherry, eaterGFP; HmlΔ>; He > lincRNA-IBIN (MeHH>lincRNA-IBIN) and control larvae were placed in a 20 μl drop of cold 8% BSA in 1 x PBS and dissected carefully with forceps. Carcasses were removed and the bled hemolymph was pipetted into a vial with 80 μl of 8% BSA in 1 x PBS. Ten larvae were dissected per cross and each cross was replicated three times. The samples were run with a BD Accuri C6 flow cytometer (BD, Franklin Lakes, NJ, USA), using a gating strategy established in [80]. In short, GFP-positive cells were detected in the FL1 (510/15 BP filter) and mCherry-positive cells in the FL3 (610/20 BP filter). GFP-only, mCherry-only and non-labelled hemocytes were used to establish the gates. Some of the GFP fluorescent signal was detected in the non-primary FL3 detector, and this was corrected for by subtracting 9% from the signal.
To check how well centrifuging separated the hemocyte and plasma fractions, five late 3rd instar HH>GFP larvae were bled in 100 μl of 1 x PBS. The vials were centrifuged for 10 minutes at 2500 g at +4°C. The supernatant containing the plasma was pipetted into another vial (~90 μl) and the hemocyte pellet was re-suspended in 90 μl of 1 x PBS. Plasma and hemocyte samples were run with a flow cytometer and the numbers of GFP-positive hemocytes in both fractions were determined.
Late 3rd instar larvae were gently washed in a drop of water with a brush, dried on a piece of tissue paper and placed on a glass slide dorsal side facing up in a drop of 70% ice-cold glycerol. A coverslip was placed on the larvae and they were stored at +4°C overnight. The next day, the immobilized larvae were imaged with a Zeiss AxioImager M2 with Apotome 2, with an EC Plan Neofluar 5x/0.16 objective. A Colibri LED light source was used to excite GFP (LED 470 nm) and mCherry (LED 555 nm) and images were captured with an AxioCam HRm CCD camera. Images were processed with ImageJ (Version: 2.0.0-rc-59/1.51j) and Adobe Photoshop CS4. Ten larvae per cross were imaged.
For detecting the cellular localization of lincRNA-IBIN, the RNA fluorescence in situ hybridization (RNA FISH) method with Cy3-tagged probes labeling the lincRNA-IBIN molecules was used. Late 3rd instar male larvae were washed in a drop of water with a brush and the hemolymph of two larvae per sample type (four biological replicates) was carefully bled out from the larvae in 20 μl of ice-cold 1 x PBS on a multiwell glass slide well, avoiding contamination from other tissues. Hemocytes were left to adhere for one hour in a humidified chamber at RT. Samples were fixed with cold 3.7% paraformaldehyde in 1 x PBS for 5–10 min and washed with 1 x PBS for 3 x 5 min. Samples were permeabilized with 1 x PBS + 0.1% Triton X-100 for 5 minutes and washed with 1 x PBS until there was no foam, and the mask around the wells was dried carefully with a tissue paper. The samples were blocked with 3% BSA in 1 x PBS at +4°C o/n. The RNA FISH protocol was performed by using the QuantiGene ViewRNA Assay (Affymetrix) and the probes for lincRNA-IBIN and RpL32 for Drosophila are now available in their catalog. For the hybridization of lincRNA-IBIN and RpL32 probes (control) and a negative “no probe” control, pre-warmed diluents and humidified chambers were used, and the incubator temperature (+40°C) was monitored. The Working Probe Set Solution was prepared by diluting each probe set 1:100 in Probe Set Diluent QF: 20 μl drops were prepared for each sample by combining 0.2 μl of Probe Set and 19.8 μl of Probe Set diluent QF. The previous solution was aspirated from the wells and replaced with 20 μl of the Working Probe Set Solution and the samples were incubated in humidified chambers for three hours at +40°C. Working Probe Set Solution was aspirated and the wells were washed three times with Wash Buffer (this was used in all the washes). 20 μl of PreAmplifier Mix solution per sample was prepared by diluting PreAmplifier Mix 1:25 in Amplifier Diluent QF and added to samples and incubated at +40°C for 30 min. After washing three times, Amplifier Mix solution was prepared by diluting Amplifier mix 1:25 in pre-warmed Amplifier Diluent QF, added to the samples and incubated at +40°C for 30 min. After three washes, the Label Probe Mix Solution was prepared by diluting Label Probe Mix 1:25 in Label Probe diluent QF, added to the samples and incubated at +40°C for 30 min. Samples were washed three times and were left for 10 min in the wash buffer for the final wash. The samples were mounted with 20 μl of ProLong Gold Antifade Mountant with DAPI (Thermo Fisher Scientific). Cover glasses were pressed on and the slides were left to harden overnight in the dark, transferred to +4°C for a day and imaged.
The samples were imaged with a Zeiss LSM 780 confocal microscope with a Plan Apochromat 63 x/1.4 oil immersion objective. A pulsed diode laser was used to excite DAPI (405 nm) and a diode laser (561 nm) was used to excite Cy3 for imaging lincRNA-IBIN and RpL32. Images were captured using a Quasar spectral GaAsP PMT array detector and camera allowing fast spectral imaging. Images were processed with ImageJ (Version: 2.0.0-rc-59/1.51j) and Adobe Photoshop CS4.
lincRNA-IBIN overexpressing (C564>lincRNA-IBIN7) and control flies (w1118,lincRNA-IBIN7) were allowed to eclose for 2 days, collected in fresh vials and kept at 29°C for two days prior to collecting the hemolymph. For experiments with infected and uninfected flies, w1118 flies were collected as above. w1118 flies were kept in fresh vials at 25°C for one day, after which half of them were infected by septic injury with a E. cloacae -contaminated needle. Flies were kept at 25°C for another 24 h prior to collecting the hemolymph. The hemolymph was collected by pricking the flies in the thorax with a thin sharp tungsten wire sterilized in 70% ethanol. Pools of 50 pricked flies were collected on ice in 0.5 ul microtubes with small holes punctured in them and placed in 1.5 μl microtubes. The flies were centrifuged at 5000 x g for 5 min, after which 0.8 μl of hemolymph was collected from the bottom of the 1.5 ml tube and diluted 1:100 in Trehalase Buffer (TB; 5 mM Tris pH 5.5, 137 mM NaCl, 2.7 mM KCl). The samples were snap-frozen in liquid nitrogen and stored at -80°C. Glucose and trehalose were analyzed using a colorimetric assay (Sigma Glucose (GO) assay kit, GACO20) based on the glucose oxidase (GO) enzyme following the protocol described in [81]. First, a trehalase stock was prepared by diluting 3 μl of porcine trehalase (Sigma-Aldrich; T8778-1UN) with 1 ml of TB. Samples were heat-inactivated for 5 min at 70°C, and divided into two 40 μl aliquots; one treated with an equal amount of trehalase stock to break down trehalose into free glucose, and the other left untreated by adding an equal amount of TB only. The samples were then incubated at 37°C overnight. Glucose standards were prepared by diluting 16 μl of a 1 mg/ml glucose stock solution with 84 μl of TB. 2-fold standard dilution curves were generated. Next morning, a 30 μl aliquot of each sample, the dilution series and a blank were loaded onto a 96-well plate and 100 μl of the GO reagent (GAGO20 Glucose assay kit, Sigma-Aldrich) was added. The plate was sealed with parafilm and incubated at 37°C for one hour. To stop the reaction, 100 μl of 12 N sulfuric acid (H2SO4) was added on the samples, after which the absorbance at 540 nm was measured using the Wallac Envision 2104 Multilabel Reader (PerkinElmer). The amount of glucose and trehalose (glucose + trehalose—glucose) in the samples were determined according to the glucose standard curve.
To verify that the C564-GAL4 driver is expressed in the guts of adult flies, C564>GFP males and females were dissected in a drop of 1 x PBS and their guts were removed. The guts were checked for GFP expression using a stereomicroscope fluorescence adapter (NIGHTSEA, MA, USA) with a Royal blue LED (440–460 nm) for excitation and a 500 nm long-pass filter. Images were captured with Nikon DS-Fi2 camera.
The first transcriptome analysis (Fig 1A and 1B) data was analyzed using the R package Limma. The package uses a modified t-test to generate an FDR (false discovery rate) corrected p-value (adjusted p-value) for each comparison. In the second transcriptome data analysis, the comparison between lincRNA-IBIN overexpression and controls (Fig 4A, S4 Table and S5 Table) was done using a two-tailed t-test (unequal variances assumed) with a 5% false discovery rate (FDR) correction using the Benjamini-Hochberg method [82]. In Fig 4B, genes that had a normalized read number >10 in the treatment of interest and an expression fold change >2 were included in the cluster analysis performed with the DAVID Bioinformatics resources 6.8 (https://david.ncifcrf.gov) [83,84] online tool. For Fig 4C and 4D, pairwise comparisons between uninfected control sample and different treatments were carried out using a two-tailed t-test assuming unequal variances.
Statistical analyses of gene expression by qRT-PCR results were carried out using a two-tailed t-test for two samples assuming equal variances. Statistical analyses of fly survival experiments were carried out using the log-rank (Mantel-Cox) test with Prism 6 (GraphPad). Data on hemocyte quantifications and types were plotted and analyzed with R version 3.3.2 (2016-10-31), Copyright 2015 The R Foundation for Statistical Computing. Data were analyzed using analysis of variance (ANOVA) followed by Tukey’s HSD post hoc test when requirements for normality and homoscedasticity were met, and in other cases a non-parametric Kruskal-Wallis rank sum test followed by Dunn’s post hoc test were applied. The level of statistical significance was established as p < 0.05.
|
10.1371/journal.ppat.1006040 | The Two-Component System ArlRS and Alterations in Metabolism Enable Staphylococcus aureus to Resist Calprotectin-Induced Manganese Starvation | During infection the host imposes manganese and zinc starvation on invading pathogens. Despite this, Staphylococcus aureus and other successful pathogens remain capable of causing devastating disease. However, how these invaders adapt to host-imposed metal starvation and overcome nutritional immunity remains unknown. We report that ArlRS, a global staphylococcal virulence regulator, enhances the ability of S. aureus to grow in the presence of the manganese-and zinc-binding innate immune effector calprotectin. Utilization of calprotectin variants with altered metal binding properties revealed that strains lacking ArlRS are specifically more sensitive to manganese starvation. Loss of ArlRS did not alter the expression of manganese importers or prevent S. aureus from acquiring metals. It did, however, alter staphylococcal metabolism and impair the ability of S. aureus to grow on amino acids. Further studies suggested that relative to consuming glucose, the preferred carbon source of S. aureus, utilizing amino acids reduced the cellular demand for manganese. When forced to use glucose as the sole carbon source S. aureus became more sensitive to calprotectin compared to when amino acids are provided. Infection experiments utilizing wild type and calprotectin-deficient mice, which have defects in manganese sequestration, revealed that ArlRS is important for disease when manganese availability is restricted but not when this essential nutrient is freely available. In total, these results indicate that altering cellular metabolism contributes to the ability of pathogens to resist manganese starvation and that ArlRS enables S. aureus to overcome nutritional immunity by facilitating this adaptation.
| The ubiquitous pathogen Staphylococcus aureus is a serious threat to human health due to the continued spread of antibiotic resistance. This spread has made it challenging to treat staphylococcal infections and led to the call for new approaches to treat this devastating pathogen. One approach is to disrupt the ability of S. aureus to adapt to nutrient availability during infection. During infection, the host imposes manganese and zinc starvation on invading pathogens. However, the mechanisms utilized by Staphylococcus aureus to overcome this host defense are unknown. We report that ArlRS, a global staphylococcal virulence regulator, is important for resisting manganese starvation during infection. Loss of ArlRS does not prevent S. aureus from acquiring metals but instead renders the bacterium incapable of adapting to limited manganese availability. ArlRS mutants also have metabolic defects and a reduced ability to grow on amino acids. When using glucose as a carbon source S. aureus is more sensitive to manganese starvation and increases the expression of manganese transporters relative to when amino acids are provided suggesting a higher demand for manganese. These observations indicate that ArlRS contributes to resisting nutritional immunity by altering metabolism to reduce the staphylococcal demand for manganese.
| Staphylococcus aureus is a ubiquitous pathogen that colonizes 30% of the population at any given time and can infect virtually every human tissue [1]. These facts and the continued spread of antibiotic resistance have led both the Centers for Disease Control and the World Health Organization to state that S. aureus poses a serious threat to human health [2, 3]. Both organizations have highlighted the need to identify new strategies for treating S. aureus and bacterial infections in general. Elucidating how pathogens overcome nutritional immunity, a critical component of the immune response in which the host restricts essential nutrients from the invading pathogen, has the potential to address this need.
Transition metals such as iron (Fe), manganese (Mn) and zinc (Zn) are essential for virtually all forms of life. Their importance is emphasized by the prediction that 30% of all enzymes interact with a metal cofactor [4, 5]. During infection, invading microorganisms must acquire all of their nutrients from the host. Vertebrates take advantage of this fact and combat invading pathogens by restricting the availability of essential metals [6, 7]. While the most well characterized aspect of nutritional immunity is the Fe-withholding response, it has recently become apparent that the host also restricts access to Mn and Zn during infection [7–12]. The prototypic example of Mn and Zn restriction is the staphylococcal abscess, which is rendered devoid of these two essential metals [8, 13]. This depletion starves S. aureus for these metals resulting in the inactivation of metal-dependent enzymes, such as the staphylococcal superoxide dismutases [8, 9]. A critical component of this withholding response is the Mn- and Zn-binding protein calprotectin (CP). CP comprises ~50% of the cytosolic protein in neutrophils and at sites of infection it can reach concentrations in excess of 1 mg/ml [14, 15]. Mice lacking CP have defects in metal sequestration and are more susceptible to a range of bacterial and fungal pathogens, including S. aureus, Acinetobacter baumannii, Klebsiella pneumoniae, and Candida albicans [8, 9, 16–19].
CP, a member of the S100 family of proteins, is a heterodimer comprised of S100A8 and S100A9 (also known as calgranulin A/calgranulin B and MRP8/MRP14), and has two transition metal-binding sites [9, 20]. The first site, known as the Mn/Zn site, is capable of binding either Mn or Zn with nanomolar and picomolar affinities (Kd), respectively [9, 20–22]. The second site, known as the Zn site, binds Zn with picomolar affinity [9, 21–23]. CP exerts antimicrobial activity against a variety of bacterial and fungal pathogens in vitro, including S. aureus, by starving them for metals [8, 9, 16–19]. While the sequestration of both Mn and Zn contributes to the antimicrobial activity of CP, the Mn/Zn site is necessary for maximal antimicrobial activity by CP against a wide range of Gram-positive and Gram-negative pathogens including S. aureus [21].
The ability of S. aureus and other successful pathogens to cause disease indicates that they possess adaptations that allow them to minimize and overcome host-imposed Mn and Zn starvation. One mechanism employed by pathogens to cope with nutrient limitation is the expression of dedicated high-affinity acquisition systems. Mn and Zn import systems that contribute to pathogenesis are found in numerous pathogens including: S. aureus, Brucella abortus, Campylobacter jejuni, Salmonella enterica, Yersinia pestis, Streptococcus pneumoniae, Streptococcus pyogenes, Streptococcus suis and A. baumannii [6, 24–34]. S. aureus expresses two dedicated Mn import systems: MntH, an NRAMP family member, and MntABC, an ABC-type transporter [13, 31, 35]. In S. aureus, MntH is constitutively expressed, while MntABC is induced by Mn limitation [13, 31]. High-affinity Mn acquisition systems play a critical role in resisting Mn starvation during infection, and staphylococcal mutants lacking these systems are attenuated in several models of infection [13, 31, 36]. The virulence defect of a staphylococcal mutant lacking both Mn importers is reversed in CP-deficient mice, indicating that these systems specifically contribute to resisting host-imposed Mn starvation [13]. The ability of a mutant lacking dedicated Mn importers to cause comparable disease to wild type bacteria when Mn is not limited also highlights the critical importance of Mn sequestration to host defense.
While high-affinity metal acquisition systems contribute to infection, they do not prevent the host from imposing Mn starvation. This is evident by the increased bacterial burdens observed in CP-deficient mice and by inhibition of staphylococcal SOD activity during infection [8, 9]. S. aureus and other pathogens are able to successfully cause infection despite experiencing Mn and Zn starvation, thus they must possess additional adaptations that allow them to resist nutritional immunity. In this study, we identified the S. aureus two-component signal transduction system ArlRS as an important factor in resisting CP-imposed Mn starvation. Infection studies using wild type and CP-deficient mice revealed that ArlRS is necessary for establishing invasive S. aureus infection and resisting Mn starvation in vivo. Additionally, we discovered that S. aureus is more sensitive to Mn starvation when using glucose as a carbon source as compared to when amino acids are provided. Furthermore, ArlRS appears to play a critical role in facilitating the use of amino acids as a carbon source. These results indicate that altering core metabolic pathways is critical to overcoming host-imposed metal starvation.
S. aureus experiences Mn and Zn starvation during infection, yet it is still able to successfully cause infection. This fact indicates that S. aureus possesses adaptations that allow it to overcome this host defense. To identify the factors that allow S. aureus to resist Mn and Zn starvation during infection, a transposon library was screened for mutants with enhanced sensitivity to CP. This screen identified a mutant that has an insertion in arlRS (arlRS:erm), which is more sensitive to CP than wild type S. aureus (Fig 1A). ArlRS is a two-component system and global virulence regulator that influences many staphylococcal processes, including autolysis, toxin expression, surface protein expression and biofilm formation [37–40]. As with most two-component systems, the signal sensed by ArlS is currently unknown. Subsequent assays using an ArlRS deletion mutant (ΔarlRS) and an arlR insertion mutant (arlR:erm) created in S. aureus Newman produced similar results to those obtained with the transposon mutant (Figs 1B and S1A). Similar to what was observed by Walker and colleagues, loss of ArlRS did not impact hla expression or hemolysis on blood agar plates [40] (S1B Fig). Expressing ArlRS from a plasmid reversed the increased sensitivity of ΔarlRS to CP (Fig 1C). Increased sensitivity to CP was also observed in arlR:erm derivatives of the community-acquired MRSA strain USA300 (JE2) as well as the methicillin-sensitive strain SH1000 (Fig 1D and 1E). In total, these results indicate that ArlRS promotes resistance to host-imposed metal starvation in both methicillin-sensitive and methicillin-resistant strains of S. aureus.
CP sequesters both Mn and Zn preventing the individual impact of withholding either metal from being evaluated with wild type protein. To circumvent this issue, the sensitivity of ΔarlRS to a series of engineered CP variants with altered metal-binding properties was assessed [9, 21]. Specifically, CP variants lacking the Mn/Zn site (ΔMn/Zn), the Zn site (ΔZn), or both sites (ΔMn/ZnΔZn) were utilized. When incubated with the ΔMn/ZnΔZn double site mutant, which cannot bind Mn or Zn, ΔarlRS grew as well as wild type (Figs 1F and S1C). This result confirms that the increased sensitivity of ΔarlRS to CP is due to an inability to cope with either Mn or Zn starvation. Similar to WT CP, ΔarlRS was more sensitive than wild type bacteria to the ΔZn site mutant, which can bind either Mn or Zn. However, the increased sensitivity of ΔarlRS was almost completely abrogated when grown in the presence of the ΔMn/Zn mutant, which can only bind Zn (Figs 1F and S1C). These results indicate that loss of ArlRS impairs the ability of S. aureus to cope with host-imposed Mn starvation.
ArlRS has been shown to repress expression of the staphylococcal autolysins LytM, LytN and Atl. As a result, loss of ArlRS can result in increased autolysis of methicillin-sensitive strains of S. aureus [38, 39, 41]. As such, cell lysis could potentially explain the enhanced sensitivity of ΔarlRS to CP. Previous studies revealed that the increased autolysis of strains lacking ArlRS can be reversed by individually deleting Atl or LytM in the ΔarlRS mutant background [41]. To determine if the diminished ability of ΔarlRS to resist Mn limitation is the result of increased autolysis, ΔarlRS lytM:erm, ΔarlRS atl:erm and ΔarlRS lytN:erm double mutants were assessed for CP sensitivity. Loss of LytM, LytN or Atl did not diminish the sensitivity of ΔarlRS to CP (Fig 2A). Control experiments revealed that loss of Atl, LytM, or LytN alone did not alter the sensitivity of S. aureus to CP (Fig 2B). These results indicate that increased sensitivity to CP of strains lacking ArlRS is not a result of increased autolysis. This idea is further supported by the increased sensitivity of the USA300 (JE2) arlR:erm mutant to CP (Fig 1D), as loss of ArlRS does not result in increased autolysis of methicillin-resistant isolates [41]. Cumulatively, these results indicate that increased autolysis does not contribute to the diminished ability of strains lacking ArlRS to resist Mn starvation.
Previous work demonstrated that loss of MntABC and MntH renders S. aureus twice as sensitive to CP as wild type bacteria [13]. Initially, to determine if the increased sensitivity of strains lacking ArlRS is due to decreased expression of Mn importers, the transcript levels of mntA and mntH were assessed by qRT-PCR. Following growth in metal-replete medium, wild type and ΔarlRS expressed comparable levels of both mntA and mntH (Fig 3A). Consistent with previous studies, CP treatment significantly increased mntA transcript levels in wild type bacteria (Fig 3A) [13]. CP also increased mntA expression in the strain lacking ArlRS, suggesting that the increased sensitivity of ΔarlRS is not due to reduced expression of Mn importers. We also assessed the impact that loss of ArlR had on a strain lacking the Mn importers (ΔmntCΔmntH arlR:erm) to grow in the presence of CP. As before, the arlR:erm mutant was more sensitive to CP treatment than wild type bacteria (Figs 3B and S2). However, loss of ArlR in the ΔmntCΔmntH mutant background further increased sensitivity of the transporter double mutant to CP, suggesting that ArlRS and the Mn transporters function independently to promote resistance to Mn starvation. To further evaluate if loss of ArlRS impacts the ability of S. aureus to acquire Mn or Zn, intracellular metal levels in wild type and ΔarlRS were directly assessed using inductively coupled plasma optical emission spectrometry (ICP-OES). This analysis revealed that loss of ArlRS does not impair Mn or Zn acquisition in the absence of CP (Fig 3C), as intracellular metal levels were the same in wild type bacteria and in ΔarlRS. In the presence of CP both WT and ΔarlRS had reduced levels of intracellular Mn (Fig 3C). No reduction in intracellular Zn or Fe were observed in the presence of CP. This observation is consistent with prior studies, which indicated that Mn binding is necessary for maximal antimicrobial activity [21]. Combined, these results indicate that a defect in metal acquisition is not responsible for the increased sensitivity of ΔarlRS to host-imposed metal starvation, suggesting that loss of ArlRS prevents S. aureus from adapting to limited Mn availability.
ArlRS contributes to hematogenous pyelonephritis and endocarditis in mouse and rabbit models of infection [39, 40, 42]. To determine whether ArlRS also contributes to systemic infection, wild type (C57BL/6) mice were retro-orbitally infected with wild type S. aureus Newman or ΔarlRS. During the course of the infection mice infected with ΔarlRS lost significantly less weight than mice infected with wild type S. aureus (Figs 4A and S3A). Consistent with the weight loss, the ΔarlRS mutant had significantly diminished bacterial burdens in the liver, heart, and kidneys when compared to wild type bacteria (Fig 4B and 4C) indicating that ArlRS plays an important role in establishing systemic disease. To evaluate the contribution of ArlRS to resisting Mn starvation during infection, CP-deficient (C57BL/6 S100A9-/-) mice, which do not remove Mn from liver abscesses [8, 13], were infected with wild type bacteria and ΔarlRS. Compared to C57BL/6 mice, the CP-deficient mice infected with ΔarlRS lost significantly more weight (Figs 4A, S3B and S3C). The CP-deficient mice infected with ΔarlRS also had increased bacterial burdens in the liver when compared to wild type C57BL/6 mice. Notably, despite the substantial virulence defect of the mutant in wild type mice, there was only a minimal difference between wild type S. aureus and ΔarlRS in the livers of CP-deficient mice (less than half a log difference vs. a 4 log difference). These results indicate that ArlRS contributes to systemic disease and that this two-component system is critical for resisting host-imposed Mn starvation during infection.
While the results so far demonstrate that ArlRS contributes to resisting host-imposed Mn starvation both in culture and during infection, the underlying mechanism is not apparent. ArlRS is a global regulator that is involved in many cellular processes including virulence factor gene regulation [37–39]. It is unlikely that the regulation of toxins and other virulence factors whose targets are absent in media would have an effect on resisting metal limitation in culture. In addition to controlling virulence factor expression, ArlRS negatively regulates the expression of genes encoding for several phosphotransferase systems (PTS) and positively regulates the expression of enzymes potentially involved in amino acid utilization [39]. This includes a locus that encodes for a putative alanine dehydrogenase, threonine/serine deaminase, and amino acid importer. This locus is induced upon exposure to CP and this induction is dependent on ArlRS (Fig 5). Glucose and other sugars are the preferred carbon source utilized by S. aureus and many other bacteria to generate energy [43, 44]; however, energy can also be generated using amino acids. While the metal dependency of glycolytic enzymes in S. aureus is unknown, Mn is a critical cofactor involved in sugar utilization by both Bradyrhizobium and S. pneumoniae [45, 46]. In contrast to sugars, amino acids can bypass the potentially Mn-dependent steps of glycolysis by being directly converted to pyruvate or TCA cycle intermediates [44, 47, 48]. Cumulatively, these observations lead to the hypothesis that Mn and Zn starvation may impair glycolysis. Furthermore, they suggest that the carbon source utilized could impact staphylococcal CP sensitivity and that ArlRS contributes to resisting Mn limitation by shifting S. aureus away from using sugars as an energy source to amino acids. If this hypothesis is correct, S. aureus would be expected to be more resistant to CP-induced Mn sequestration when using amino acids as opposed to glucose as a carbon source. Furthermore, loss of ArlRS would be expected to reduce the ability of S. aureus to grow when amino acids are provided as the sole carbon source and alter staphylococcal metabolism when both nutrient types are present.
To test this hypothesis, a defined medium compatible with CP growth assays, which allows the carbon source to be altered, was developed (Fig 6A). This medium was then used to assess the sensitivity of S. aureus to CP when glucose or casamino acids were provided as the sole energy source. These assays revealed that S. aureus Newman is more sensitive to CP when glucose was provided as the sole carbon source compared to bacteria using casamino acids (Figs 6B and S4A). Similar results were also observed when a defined medium containing purified amino acids as an energy source was used (S4B Fig). USA300 (JE2) was also more sensitive to CP when only glucose was available as a carbon source. (Figs 6C and S4C). To determine whether the increased sensitivity of S. aureus to CP when glucose is used as a carbon source is dependent on Mn or Zn sequestration, the experiment was also performed with the ΔMn/Zn and ΔZn site mutants. While decreased growth in glucose-containing medium was observed when bacteria were growing in the presence of the ΔZn mutant (binds both Mn and Zn), growth in the presence of the ΔMn/Zn mutant (binds only Zn) was comparable to that of growth in medium containing amino acids, meaning that no growth defect was observed (Fig 6D). Additionally, the addition of excess Mn to glucose-containing medium reversed the phenotype (Fig 6E). Combined, these results suggest that Mn sequestration is responsible for the reduced ability of S. aureus to grow in glucose-containing medium in the presence of CP. If using glucose as a carbon source requires more Mn than amino acids, mntA expression would be expected to be higher in medium containing only glucose as an energy source relative to medium containing amino acids when Mn availability is restricted. Consistent with our hypothesis, mntA levels increased in the presence of intermediate concentrations of CP, but not in Mn-replete medium, when bacteria were grown in the presence of glucose but not in amino acids (Fig 6F). Cumulatively, these results indicate that utilizing glucose as carbon source increases the cellular demand for Mn when compared to when amino acids are used. Vitko et al. have previously shown that lactate is produced when bacteria are grown in glucose-containing medium but not when they are grown in amino acid-containing medium [49]. When Newman and USA300 (JE2) were grown in medium containing both glucose and amino acids in the presence of CP, lactate production decreased with increasing CP concentrations (Fig 6G and 6H). This observation is consistent with the idea that S. aureus shifts away from utilizing glucose as a carbon source when Mn is limiting and that utilizing amino acids as a carbon source minimizes the cellular demand for this metal.
Next, the ability of ArlRS mutants to utilize amino acids as an energy source was assessed. Analysis of the ΔarlRS derivative of Newman revealed that this strain was severely delayed in growth when utilizing amino acids as a sole carbon source (Figs 6A and S4D), suggesting a role for ArlRS in amino acid utilization. More modest but still significant reductions were also observed with arlR:erm derivatives of S. aureus USA300 (JE2) and SH1000 when only amino acids were provided as an energy source (S4E and S4F Fig). To evaluate if loss of ArlRS alters staphylococcal metabolism, the production of lactate was assessed in the ΔarlRS and arlR:erm derivatives of Newman and USA300 (JE2) following growth in the presence and absence of CP (S4G–S4J Fig). Both of the mutants had decreased lactate production relative to the parent strain regardless of CP treatment. Notably, differing from wild type, the strains lacking ArlRS did not reduce their production of lactate at inhibitory concentrations of CP (Fig 6G and 6H). In conjunction with the growth phenotypes, these results suggest that loss of ArlRS disrupts staphylococcal metabolism and results in reduced growth in the presence of amino acids as a carbon source. Combined, these results support the idea that switching from utilizing sugars to amino acids as an energy source reduces the staphylococcal demand for Mn enhancing the ability of S. aureus to resist host-imposed metal starvation. They also suggest that ArlRS critically contributes to this process.
Transition metals such as Fe, Mn and Zn are important for virtually all forms of life, as they are involved in numerous biological processes ranging from metabolism to regulation of virulence factor expression [4–6, 50]. To combat invaders, the host takes advantage of this essentiality by starving invaders for these metals. Recent work has revealed that in addition to restricting Fe availability, the essential transition metals Mn and Zn are also withheld by the host. Despite expressing high-affinity Mn acquisition systems, S. aureus, and presumably other successful pathogens, experience metal starvation during infection [8, 9]. However, the adaptations that allow pathogens to overcome host-imposed Mn and Zn starvation are unknown. The investigations in this study revealed that to successfully cope with host-imposed Mn starvation, S. aureus must alter core metabolic pathways. Previous studies have emphasized the importance of sugar consumption and fermentation to staphylococcal virulence [49, 51–54]. These obsevations include the finding that the catabolite control protein (CcpA), which promotes the consumption of sugars, and an expanded repertoire of glucose importers enhances the ability of S. aureus to cause disease [53, 54]. At the same time other studies sugest that uptake of amino acids facilitate the development of staphylococal disease [55]. In this study, we found that when bacteria encounter Mn starvation the presence of amino acids enhances the growth of the bacterium. These prior observations in conjunction with the current results emphasize the dynamic nature of sites of infection and further highlight the important contribution of metabolic plasticity to staphylococcal virulence and bacterial pathogenesis in general [49, 51, 52, 56, 57]. This work also revealed that the two-component system ArlRS enhances the ability of S. aureus to grow when amino acids are available as a sole carbon source and contributes to the ability of S. aureus to resist host-imposed Mn starvation during infection. This observation significantly expands the contribution of this two-component system to staphylococcal disease, which is canonically associated with regulation of toxin production and biofilm formation [37–40].
Recently, CP has been reported to bind Fe (II) with high affinity leading to the suggestion that the antimicrobial activity of the protein is derived from the ability to bind Fe, not Mn [58]. However, consistent with prior studies of A. baumannii [19], analysis of metal levels in S. aureus revealed that CP does not reduce intracellular Fe levels (Fig 3C). In contrast, in both S. aureus and A. baumannii CP reduced the accumulation of Mn [19]. These results suggest, at least for these two pathogens, that Fe sequestration is not a major contributor to the antimicrobial activity of CP. Additionally, the virulence defects in wild type mice of staphylococcal ΔmntCΔmntH and ΔarlRS mutants, which are more sensitive to Mn starvation in culture, are reversed in CP-deficient mice [13]. These results, in conjunction with the inhibition of Mn-dependent enzymes during infection [9], further support the body of work indicating that Mn sequestration by CP contributes to host defense.
Canonically, glycolysis is thought to be a magnesium-dependent process. However, many bacteria contain a Mn-dependent isoform of phosphoglycerate mutase and other glycolytic enzymes such as enolase and pyruvate kinase that are either dependent on Mn or are highly activated by small amounts of Mn [24]. The increased sensitivity of S. aureus to Mn starvation when only glucose is available as a carbon source suggests that at least one essential step in glycolysis is dependent on Mn. This observation adds S. aureus to the growing list of organisms, including S. pneumoniae and Bradyrhizobium japoincum, which are dependent on Mn in order to consume glucose [45, 46]. The presence of Mn-utilizing glycolytic enzymes in a variety of microbes and the dependency of glycolysis in some pathogens on this metal suggests that host-imposed Mn starvation may also impede the ability of other pathogens to utilize sugars as an energy source.
The Fe-sparing response, the repression of Fe-rich enzymes when the availability of this metal is limited, enhances the ability of bacteria to grow in Fe-poor environments. This response allows bacteria to prioritize the usage of Fe by reducing the production of non-essential Fe-dependent proteins thereby preserving the limited quantity of available Fe for essential functions [59, 60]. The preferred carbon source of S. aureus and many other bacteria is glucose [43, 44]; however, energy can also be generated by utilizing amino acids. As such, glucose utilization is not strictly essential in S. aureus. In contrast to glycolysis, which can require Mn, the degradation of amino acids (e.g., alanine, serine and threonine) utilizes enzymes that do not employ this metal as a cofactor [44, 47, 48]. This observation suggests that relative to utilizing sugars, amino acid consumption should decrease the cellular demand for Mn, increasing the availability of this metal for essential Mn-dependent enzymes (Fig 7). Both the reduced expression of mntABC in Mn-limited medium when S. aureus is utilizing amino acids vs. glucose as a carbon source and the observation that growth on amino acids diminished staphylococcal sensitivity to Mn starvation support this idea. The latter observation suggests that consumption of amino acids instead of sugars is functionally a Mn-sparing response analogous to that of Fe. At higher concentrations of CP, S. aureus is equally sensitive to metal limitation regardless of whether the bacteria were grown in glucose- or amino acid-containing medium. Inhibition of Mn-dependent processes, which cannot be circumvented by switching carbon source utilization, could explain this observation. In response to Mn limitation, S. pneumoniae also downregulates glycolytic enzymes and increases the expression of amino acid utilization genes [46]. This observation and the current studies suggest that switching from utilizing sugars to amino acids is likely a conserved response to host-imposed Mn starvation.
While the utilization of amino acids as an energy source reduces the cellular demand for Mn, in most bacteria, including S. aureus, catabolite repression prevents them from utilizing non-preferred carbon sources, such as amino acids, when a preferred carbon source is present [61, 62]. ArlRS represses the expression of alternative sugar uptake systems and stimulates the expression of genes encoding for enzymes involved in amino acid utilization [39]. As such, ArlRS provides a mechanism by which S. aureus can override the normal carbon source preference of the cell. In response to CP, ArlRS positively regulates the expression of proteins predicted to be involved in the catabolism of alanine and serine. Differing from Liang et al, loss of ArlRS did not impact the expression of these proteins in the absence of CP. This difference can likely be explained by differences in growth conditions. These two amino acids can be converted directly to pyruvate, bypassing any metal-dependent enzymes in glycolysis [39, 44, 47, 48]. Notably, a global screen for staphylococcal factors that contribute to abscess formation identified alanine and serine importers as contributing to the ability of S. aureus to cause disease [55]. As with most two-component systems, the signal sensed by ArlS is currently unknown. The necessity of ArlRS to resist Mn starvation suggests that Mn availability alters the activity of this system. ArlRS may respond directly to Mn availability or indirectly by sensing a disruption of glycolysis or other Mn-dependent processes induced by Mn limitation. However, additional experimentation is necessary to evaluate this possibility.
ArlRS contributes to the ability of S. aureus to cause disease in several models of infection [42]. In addition to regulating metabolic processes, ArlRS increase the production of surface proteins and influences the expression of numerous virulence factors, biofilm formation, as well as autolysis and cell growth. It also directly and/or indirectly interacts with other regulators [37, 39]. Thus, a reduced ability to grow on amino acids may not be the only factor that contributes to the diminished ability of strains lacking ArlRS to resist CP. As ArlRS regulates the expression of other staphylococcal regulators, including Agr, LytSR, MgrA and Rot, the factors that are directly vs. indirectly regulated by this system are unknown [37–39]. While the direct targets are unknown, it does link Mn availability to virulence factor expression. Even though the benefit to S. aureus of co-regulating presumably Mn-independent virulence factors is not immediately apparent, this does appear to be a common theme amongst bacterial pathogens. In Borrelia burgdoferi, Mn influences BosR expression, which in turn regulates expression of the alternative sigma factor, RpoS. This alternative sigma factor facilitates the adaptation of B. burgdoferi to the mammalian host [63]. S. pneumoniae uses the Mn-responsive regulator PsaR to regulate the expression of adhesins [64–67].
Due to the continued emergence of antibiotic resistance, bacterial pathogens remain a serious threat to human health. The current study provides new insight into the mechanisms utilized by pathogens to overcome nutritional immunity. It suggests that alterations in carbon source utilization and reducing the cellular demand for Mn is important for resisting host-imposed Mn starvation. These results significantly broaden our understanding of how bacteria overcome nutritional immunity. The continued study of this bacterial response and the associated metabolic changes has the potential to identify new opportunities for therapeutic intervention.
All experiments involving animals were reviewed and approved by the Institutional Animal Care and Use Committee of the University of Illinois Urbana-Champaign (IACUC license number 15059) and performed according to NIH guidelines, the Animal Welfare Act, and US Federal law.
For routine overnight cultures, bacteria were grown in 5 ml of tryptic soy broth (TSB) or Chelex-treated RPMI plus 1% Casamino acids (NRPMI) supplemented with 1 mM MgCl2, 100 μM CaCl2 and 1 μM FeCl2 [13] in 15 ml conical tubes at 37°C on a roller drum. As needed, 10 μg/ml of chloramaphenicol was added for plasmid maintenance. S. aureus strain Newman and its derivatives were used for all of the experiments, unless otherwise indicated. For experiments using USA300 (JE2) and derivatives (USA300 (JE2) arlR:erm, USA300 (JE2) lytM:erm, USA300 (JE2) lytN:erm and USA300 (JE2) atl:erm), strains were obtained from the Nebraska library [68]. Newman ΔarlRS was generated by amplifying the 5’ and 3’ flanking regions (~1 Kb up- and downstream) of arlRS using the indicated primers (Table 1). 5’ and 3’ fragments were cloned into the pKOR1 knockout vector via site-specific recombination. The deletions were created using allelic replacement, as described previously [69]. All constructs were verified by sequencing and all mutant strains were confirmed to be hemolytic by growth on TSA blood agar plates. To generate constructs for complementation studies, the arlRS coding sequence was amplified with the indicated primers (Table 1) and cloned into the pOS1 vector under the control of the Plgt promoter. The lytM, atl and lytN mutants in Newman and Newman ΔarlRS and arlR mutants in SH1000, Newman and Newman ΔmntCΔmntH were constructed by transducing the lytM:erm, atl:erm, lytN:erm and arlR:erm alleles via Φ85 phage from USA300 (JE2) lytM:erm, USA300 (JE2) atl:erm, USA300 (JE2) lytN:erm and USA300 (JE2) arlR:erm.
CP growth assays were performed, as described previously [9, 21]. Briefly, overnight cultures were back-diluted 1:50 into fresh TSB (5 ml in a 15 ml conical tube) and grown for 1 h at 37°C on a roller drum. The cultures were then diluted 1:100 into 96-well round-bottom plates containing 100 μl of growth medium (38% TSB and 62% calprotectin buffer (20 mM Tris pH 7.5, 100 mM NaCl, 3 mM CaCl2, 10 mM β-mercaptoethanol)) in presence of varying concentrations of CP. The growth medium was supplemented with 1 μM MnCl2 and 1 μM ZnSO4 except for assays with the Newman arlRS:erm transposon mutant. For all assays, the bacteria were incubated with orbital shaking (180 RPM) at 37°C and growth was measured by assessing optical density (OD600) every 2 hours. Prior to measuring optical density, the 96-well plates were vortexed. For CP growth assays using defined medium, a medium based on the one previously reported by Richardson et al. [51] was used. For these assays, the preculture was the same as a growth assay using TSB in the growth medium. The growth medium for assays using defined medium consisted of 38% medium and 62% CP buffer (20 mM Tris pH 7.5, 100 mM NaCl, 1 mM CaCl2, 10 mM β-mercaptoethanol). The defined medium (2.6X) consisted of 0.5 g/L NaCl, 1.0 g/L NH4Cl, 2.0 g/L KH2PO4, 7.0 g/L Na2HPO4, 0.228 g/L biotin, 0.228 mg/L nicotinic acid, 0.228 mg/L pyridoxine-HCl, 0.228 mg/L thiamine-HCl, 0.114 mg/L riboflavin, 0.684 mg/L calcium pantothenate, 0.104 g/L phenylalanine. 0.078 g/L lysine, 0.182 g/L methionine, 0.078 g/L histidine, 0.026 g/L tryptophan, 0.234 g/L leucine, 0.234 g/L aspartic acid, 0.182 g/L arginine, 0.078 g/L serine, 0.15 g/L alanine, 0.078 g/L threonine, 0.130 g/L glycine, 0.208 g/L valine and 0.026 g/L proline. The defined medium was then supplemented with 6 mM MgSO4, 1 μM FeCl2, 1 μM MnCl2 and 1 μM ZnSO4. Casamino acids (6.5%), glucose (1.3%) or glucose (1.3%) and 18 amino acids (1 mM alanine, serine, threonine, lysine, asparagine, glutamic acid, isoleucine, arginine, histidine, methionine, valine, proline, cystidine, glycine, phenylalanine, tyrosine, leucine and tryptophan) were provided as carbon sources as indicated. In the figures, “DM” refers to defined medium without a carbon source, “glc” refers to defined medium with glucose as a carbon source, “AA” refers to defined medium with casamino acids as a carbon source and “glc+18AA” refers to glucose and 18 amino acids as a carbon source. For complementation experiments, overnight cultures were back-diluted 1:50 into fresh TSB and grown for 2 h at 37°C [9, 21, 51]. When a metal starvation step was included the bacteria were grown overnight in NRPMI supplemented with 1 mM MgCl2, 100 μM CaCl2 and 1 μM FeCl2 and directly inoculated 1:100 in to the assay medium. Calprotectin was purified, as previously described [9, 21]. The initial ArlRS transposon mutant was identified during optimization experiments for screening a Tn917 mutant library. For these assays bacteria arrayed in 96-well plates were grown overnight in NRPMI supplemented with 1 mM MgCl2, 100 μM CaCl2 and 1 μM FeCl2. These cultures were then assayed for CP sensitivity, as described above.
To assess the expression of mntA, mntH and NWMN_1348, S. aureus was grown as for CP inhibition assays in complex medium in the presence and absence of 240 μg/ml of CP or in defined medium in the presence and absence of 120 μg/ml of CP. Bacteria were harvested during log phase growth (OD600 = ~0.1), samples were collected, an equal volume of ice-cold 1:1 acetone-ethanol was then added to the cultures, and they were frozen at -80°C until RNA extraction. RNA was extracted and cDNA was generated, as previously described [70–72]. Gene expression was assessed by quantitative reverse transcription-PCR (qRT-PCR) using the indicated primers (Table 1, [13]) and 16S was used as a normalizing control.
To assess intracellular metal levels in wild type and ΔarlRS, S. aureus strains were grown as for CP inhibition assays using complex medium in the presence and absence of 240 μg/ml of CP. Bacteria were harvested during log phase growth (OD600 = ~0.1), washed twice with 0.1 mM EDTA, washed twice with water, and digested with nitric acid. Prior to nitric acid digestion an aliquot was used to determine the total number of bacteria in the sample. After digestion, ICP-OES was performed by the Microanalysis facility at the University of Illinois Urbana-Champaign.
Mouse infections were performed, as described previously [8, 9], with the exception that mice were anesthetized with isoflurane. Briefly, 9-week old mice were infected retro-orbitally with approximately 1 x 107 CFU in 100 μl of sterile phosphate-buffered saline. Following injection, the infection was allowed to proceed for 96 h before the mice were sacrificed. Livers, hearts and kidneys were removed, the organs were homogenized, and bacterial burden was determined by plating serial dilutions.
L-lactate production was assayed, as described previously [49]. Briefly, bacteria were grown as for CP inhibition assays described above using NRPMI overnight cultures and back-diluted 1:100 into defined medium in the presence and absence of 60, 120, 240 and 480 μg/ml of CP. Samples were harvested every hour during log phase, heat inactivated (70°C for 5 min), and supernatants were collected. Samples were stored at -20°C. L-Lactic acid concentrations were measured using a Roche Yellow Line kit (R-Biopharm).
|
10.1371/journal.ppat.1007094 | Listeria monocytogenes InlP interacts with afadin and facilitates basement membrane crossing | During pregnancy, the placenta protects the fetus against the maternal immune response, as well as bacterial and viral pathogens. Bacterial pathogens that have evolved specific mechanisms of breaching this barrier, such as Listeria monocytogenes, present a unique opportunity for learning how the placenta carries out its protective function. We previously identified the L. monocytogenes protein Internalin P (InlP) as a secreted virulence factor critical for placental infection. Here, we show that InlP, but not the highly similar L. monocytogenes internalin Lmo2027, binds to human afadin (encoded by AF-6), a protein associated with cell-cell junctions. A crystal structure of InlP reveals several unique features, including an extended leucine-rich repeat (LRR) domain with a distinctive Ca2+-binding site. Despite afadin’s involvement in the formation of cell-cell junctions, MDCK epithelial cells expressing InlP displayed a decrease in the magnitude of the traction stresses they could exert on deformable substrates, similar to the decrease in traction exhibited by AF-6 knock-out MDCK cells. L. monocytogenes ΔinlP mutants were deficient in their ability to form actin-rich protrusions from the basal face of polarized epithelial monolayers, a necessary step in the crossing of such monolayers (transcytosis). A similar phenotype was observed for bacteria expressing an internal in-frame deletion in inlP (inlP ΔLRR5) that specifically disrupts its interaction with afadin. However, afadin deletion in the host cells did not rescue the transcytosis defect. We conclude that secreted InlP targets cytosolic afadin to specifically promote L. monocytogenes transcytosis across the basal face of epithelial monolayers, which may contribute to the crossing of the basement membrane during placental infection.
| Infections during pregnancy can lead to infections of the placenta, spread to the fetus, and cause fetal damage and death. Improving maternal-child heath is a global heath priority. Yet, progress to prevent and treat pregnancy-related diseases has lagged behind other medical fields. Using pregnant guinea pigs, which have a placental structure that closely resembles humans, we identified a protein (InlP) secreted by the bacterial pathogen Listeria monocytogenes that strongly promotes placental infection. In human placental organ cultures bacteria deficient in InlP were impaired in their ability to spread from infected placental cytotrophoblasts into the underlying fetal stroma. Here, we solved the crystal structure of InlP, and identified Afadin, a cytoplasmic protein that localizes to adherens junctions as a binding partner of InlP. We demonstrate that InlP decreases the magnitude of traction stresses epithelial cells exert on an underlying extracellular matrix, and furthermore, that InlP facilitates bacterial spread from infected epithelial monolayers into an underlying compartment. Our study provides new insights into the mechanisms of bacterial spread across the placental barrier.
| During pregnancy, the consequences of placental infection can be severe, ranging from maternal sepsis to miscarriage, and can lead to pre-term birth and lifelong disability [1]. Fortunately, such infections are relatively rare–which stands as a testament to the strength of the feto-maternal barrier. Despite serving such an important function, the molecular, cellular and histological components of feto-maternal barrier have only just begun to be elucidated. Because the barrier is so effective at preventing infection, pathogens that do manage to cross it must have evolved strategies of countering host defenses and thus provide a unique opportunity for addressing the mechanistic features that make the feto-maternal barrier so formidable [2].
Listeria monocytogenes is a well-characterized food-borne pathogen capable of placental crossing and is thus ideal for probing this barrier [3]. In the healthy, non-pregnant adult, it causes gastrointestinal illness, but in immunocompromised individuals meningitis can result, and in pregnant women, sepsis, spontaneous abortion, preterm labor, infant brain damage and death are possible outcomes [4]. After ingestion, L. monocytogenes infects epithelial cells of the intestine, mainly via interactions between the bacterial surface protein Internalin (encoded by the gene inlA) and the host cell receptor E-cadherin, which is exposed at the tips of intestinal villi [5–7]. After uptake by an intestinal epithelial cell, the bacterium escapes the resulting phagosome and replicates in the host cytoplasm, where it induces actin polymerization, forming actin comet tails that drive its rapid motility [8]. Upon reaching the host plasma membrane, L. monocytogenes can form a membrane-bound protrusion that pushes its way into a neighboring cell. Engulfment of the protrusion by the neighboring cell leaves the bacterium in a double-membrane vacuole from which it escapes, resetting its life cycle [8, 9]. From the initial site of infection in the intestinal epithelium, L. monocytogenes can spread via actin-based motility into immune cells, which facilitate spread throughout the host while protecting the pathogen from humoral immune defenses [10, 11]. Actin-based cell-to-cell spread is thought to contribute to the ability of L. monocytogenes to evade the host immune system and to penetrate a variety of protective barriers in the host [12], including the blood-brain barrier and the feto-maternal barrier in the placenta [13–15].
A clue to the protective nature of the feto-maternal barrier lies in the fact that almost all pathogens that cross it are known to have an intracellular aspect to their life cycle—they move through cells from mother to fetus [2]. Two interfaces are available; each is unlike any other part of the mammalian body. The first, where fetal extravillous trophoblasts (EVTs) anchor the placenta to the uterine decidua, has a unique immune environment [16]. The trophoblasts possess innate immune defense properties shown to prevent bacterial infection [17] and restrict growth of intracellular L. monocytogenes [18]. The second, much more extensive interface is composed of fetal syncytiotrophoblasts (STB), which form a vast, thin multinucleate layer without cell-cell junctions, bathed in maternal blood. This is the site of gas and nutrient/waste exchange. Underlying the STB is a second, single-celled layer of individual, self-renewing cytotrophoblasts (CTBs) that periodically fuse with the STB to allow its growth. Our previous work has shown that the STB’s lack of cell-cell junctions [19] and the syncytial stiffness generated by dense networks of actin filaments [20] act as significant deterrents to STB infection by pathogens. Rather, infection of human placental organ cultures suggests that the first interface, where EVTs contact uterine decidua, is the preferred route of placental infection by L. monocytogenes, which then spreads via actin-dependent cell-to-cell spread to the CTB monolayer underlying the STB [19]. But even once L. monocytogenes has reached the subsyncytial CTBs, most bacterial movement occurs laterally from cell to cell within the monolayer, and it only rarely transcytoses in a direction perpendicular to the monolayer to colonize the fetal stroma beneath it [19].
Recently, we used an unbiased genetic screen in a pregnant guinea pig model of listeriosis to identify L. monocytogenes genes that contribute specifically to infection of the placenta. [21]. One gene identified in this screen encodes InlP, a member of the internalin family of proteins. Pathogenic L. monocytogenes strains contain genes for 25 known members of this protein family, which share a common overall structure that includes a secretory signal sequence at the N-terminus and a large leucine-rich repeat (LRR) domain, a motif frequently implicated in protein-protein interactions [22]. The two best-characterized internalins, InlA and InlB, contribute to L. monocytogenes invasion of epithelial cells and hepatocytes by binding to their cognate host cell surface receptors, E-cadherin and c-Met respectively [5, 23]. While most internalins are predicted to be anchored to the bacterial cell surface via attachment to the peptidoglycan cell wall or to lipoteichoic acids, four of them, including InlP, lack obvious anchoring domains and are predicted to be secreted by the bacterium [24]. In order to address the mechanism by which InlP assists L. monocytogenes in crossing the feto-maternal barrier, we set out to find host protein binding partners. In this work, we identify the host cell cytoplasmic protein afadin as a major binding partner for InlP, and demonstrate that the InlP-afadin interaction specifically enhances L. monocytogenes transcytosis—that is, the ability of L. monocytogenes to exit from the basal face of an infected polarized epithelial monolayer, consistent with its potential contribution to placental infection.
Human subjects: This study was conducted according to the principles expressed in the Declaration of Helsinki. The study was approved by the Institutional Review Board at the University of California, San Francisco, where all experiments were performed (CHR# 11–05530). All patients were adults and provided written informed consent for the collection of samples and subsequent analysis.
All chemicals were purchased from Sigma-Aldrich unless otherwise stated. Center for the Structural Genomics of Infectious Diseases (CSGID) standard protocols were used for cloning, over-expression, and purification of InlP [25, 26]. Briefly, InlP (from residue 31–388) and Lmo2027 (from residue 31–367) were cloned into the pMCSG7 expression vector (http://bioinformatics.anl.gov/mcsg/technologies/vectors.html).
Following transformation into the BL21 (DE3) Magic E. coli strain, cells were grown in M9-selenomethionine medium or Terrific Broth (TB) at 37°C up to an OD600 of 1. At that point, the temperature was reduced to 25°C and protein over-expression was induced by the addition of isopropyl-1-thio-D-galactopyranoside (IPTG) to a final concentration of 1 mM. After 16 h, cells were harvested by centrifugation, suspended in a buffer containing 10 mM Tris-HCl pH 8.3, 500 mM NaCl, 10% glycerol, and 1 mM Tris (2-carboxyethyl) phosphine hydrochloride (TCEP) and lysed by sonication. InlP was purified by Ni-NTA affinity chromatography and eluted with a buffer containing 10 mM Tris-HCl pH 8.3, 500 mM NaCl, 1 mM TCEP. The His-tag was cleaved overnight at 4°C incubating the enzymes with His-tagged TEV protease. The protein samples were then reloaded onto the nickel column and the flow-through was collected. At this point, the protein was concentrated using Vivaspin centrifugal concentrators (GE Healthcare Life Sciences) and both its size and purity were checked by SDS-PAGE.
Sitting drop crystallization plates were set up at room temperature. InlP crystals were obtained with a 1:1 mixture of 8.5 mg/mL InlP in 10 mM Tris pH 8.3, 1 mM TCEP and condition D11 from the Qiagen PACT crystallization screen (0.2 M Calcium chloride, 0.1 M Tris pH 8, 20% (w/v) PEG 6000). Lmo2027 crystals were obtained with a 1:1 mixture of 12.6 mg/mL Lmo2027 and 10 mM Tris pH 8.3, 0.5 M NaCl, and 1 mM TCEP and condition D4 from the Qiagen Classics II crystallization screen (0.1 M Citric acid (pH 3.5) and 20% (w/v) PEG 3350). Harvested crystals were transferred to the mother liquor before being frozen in liquid nitrogen. Diffraction data were collected at 100o K at the Life Sciences Collaborative Access Team at the Advance Photon Source, Argonne, Illinois (APS BEAMLINE 21-ID-G). All structural work was performed by the CSGID [25]. Data were processed using HKL-2000 for indexing, integration, and scaling [27]. A selenomethionine derivative of InlP was used to phase the structure by single-wavelength anomalous diffraction. Lmo2027 was phased by molecular replacement, using the InlP structure as a starting model. Structure was refined with Refmac [28]. Models were displayed in Coot and manually corrected based on electron density maps [29]. All structure figures were prepared using PyMOL Molecular Graphics System, Version 1.3 (Schrödinger, LLC).
For ITC experiments, InlP was prepared by overnight dialysis in ITC buffer (50 mM HEPES pH 7.5 and 150 mM NaCl). Experiments were performed on the MicroCal ITC200 instrument (GE Healthcare). InlP was loaded into the cell at 0.12 mM and 25° C. Calcium chloride titrant (3.13 mM) dissolved in leftover dialysis buffer was loaded into the syringe. Syringe rotation was set at 1000 RPM, with titrant injections spaced at 2 min intervals. The initial 0.2 μL injection was excluded from the data and the nineteen subsequent 2.0 μL injections were used for data analysis. Binding parameters were obtained by fitting isotherms using the Origin 7 (OriginLab, Northampton, MA) software package.
Yeast two-hybrid analysis was performed by Hybrigenics (France). Briefly, InlP (aa 31–388) was used to screen a human placenta cDNA library. Clones expressing proteins with positive interactions with InlP were isolated and sequenced.
For the GST-pull down experiments bait proteins were prepared as follows: Cloning: inlP gene (from residue 31–388, oligo Fw2470 EcoRI-pGex CCGAATTCC GCTTCTGATTTATATCCACTACCT, Rv2470 XhoI-pGex aGCTCGAG TCATTAATAGTTACATTCCAATCATAAGAG) or Lmo2027 (from residue 31–367, oligo Fw2027 EcoRI-pGex CCCCGAATTCCGCATCCGATTTATATCCACT ACC, Rv2027 XhoI-pGex GAAGCTCGAGCTACTAATTAACCGTAAGTACCC) were cloned in pGEX-4T vector (GE Healthcare) using restriction enzymes following transformation in BL21 DE3 E. coli strain. InlPΔLRR5, InlPΔLRR7 and InlPΔLRR8, lacking the LRR5 (Δamino acids 174–195), the LRR7 motifs (Δamino acids 218–239) and the LRR8 (Δamino acids 240–261) respectively were generated using Site Directed Mutagenesis (SDM) on the inlP-pGEX-4T vector following In-Fusion HD Cloning Kit protocol (Takara Bio USA, Inc.) using the following primers: ΔLRR5-fw:GATTTCACTGGAATGCCTATTCCTTTACTTATTACGTTGGATCTAAG, ΔLRR5-rv: CGTAATAAGTAAAGGAATAGG CATTCCAGTGAAATCAGGGATAGA; ΔLRR7-fw CCTGATTTTCAAAATTTACCT AAATTAACTGATTTAAATTTAAGAC, ΔLRR7-rv: AGTTAATTTAGGTAAATTTTGAAAATCAGGAATAGTTGTCA, ΔLRR8-fw: CCTGATTTTCAAA ATAACTTACCTAGTTTAGAATCCTTAAACT ΔLRR8-rv: TAAACTAGGTAAGTTATTTTGAAAATCAGGTGTATTGGTTAA. Baits purification: BL21 DE3 E. coli carrying pGEX-4T, pGEX-4T-2027, pGEX-4T-inlP, pGEX-4T-inlPΔLRR5, pGEX-4T-inlPΔLRR7 or pGEX-4T-inlPΔLRR8, were grown until OD 0.6 in 250 mL of LB supplemented with 100 μg/mL ampicillin at 37°C and 180 rpm. GST or GST-InlP/2027/ΔLRR5/ΔLLR7/ΔLRR8 expression were then induced with 0.2 mM IPTG for 4 h at 30°C 180 rpm. Pellet from 250 mL of culture was lysed with 35 mL of cell lytic express (Sigma) and the soluble fraction was incubated overnight with 5 mL of Glutathione Sepharose 4B (GE Healthcare) under constant agitation. GST-Glutathione Sepharose 4B or GST-InlP/2027/ΔLRR5/ΔLLR7/ΔLRR8/—Glutathione Sepharose 4B were then collected and washed with 100 mL of PBS at 500 x g and stored at 4°C in Tris-HCl 50 mM, NaCl 50 mM pH 8.
For preparation of protein extracts for pull-down experiments with human placenta: 2.3 grams of placental villi (~21 weeks gestational age) were lysed in 10 mL of M-PER (Thermo Fisher Scientific) supplemented with one tablet of complete Protease Inhibitor Cocktail (Roche) and PhosSTOP phosphatase inhibitor (Roche). The sample was homogenized for 30 sec, sonicated for 4 min and incubated for 1 h on ice. The soluble fraction was collected after centrifugation 14000 x g for 20 min at 4°C, split in two and incubated overnight under constant agitation with 1 mL of GST-Glutathione Sepharose 4B, GST-InlP- Glutathione Sepharose 4B or GST-2027- Glutathione Sepharose 4B respectively (prepared previously). The next day, washes were performed with 50 mL PBS and elution fractions were collected with 3 mL Tris-HCl 50 mM, NaCl 50 mM free glutathione 10 mM, pH 8. Elution fractions were run on 10% gel (NuPAGE Novex 10% Bis-Tris Protein Gels) and analyzed by Western blot, Coomassie blue staining (SimplyBlue SafeStain, Life Technologies) or sent to Taplin Mass Spectrometry Facility at Harvard Medical school for protein identification from SDS-PAGE gel (S1 Table).
In preparation of protein extracts for pull-down experiments with MDCK cells: one 75 mm flask of MDCK or MDCK AF-6-/- cells polarized for three days was lysed in 180 μL of cell lytic M supplemented with complete Protease Inhibitor Cocktail (Roche). The soluble fraction was split in two and incubated overnight under constant agitation with 20 μL of GST-Glutathione Sepharose 4B or GST-InlP/2027/ΔLRR5/ΔLRR7/- Glutathione Sepharose 4B respectively (prepared previously). The day after beads were washed with 2 mL of Tris-HCl 50 mM, NaCl 50 mM, free glutathione 10 mM, pH 8.0. The bound proteins were eluted by boiling the beads in SDS sample buffer (NuPAGE sample reducing agent 10X and NuPAGE LDS sample buffer 4X) for 30 min and analyzed by Western Blot or Coomassie blue staining (SimplyBlue SafeStain, Life Technologies) or sent to Taplin Mass Spectrometry Facility at Harvard Medical school for protein identification (S1 Table). Raw mass spectrometry data were then further analyzed to determine differences in binding of proteins to InlP-GST versus InlPΔLRR5-GST. First proteins identified by only 1 peptide in 1 or 2 samples were excluded from further analysis. Then proteins found to bind to GST alone were also excluded from further analysis. Finally, proteins that showed 0 intensity for either InlP-GST or InlPΔLRR5-GST were also excluded from the analysis.
To detect afadin by Western Blot analysis, elution fractions were analyzed on 4–12% gels (NuPAGE Novex 4–12% Bis-Tris Protein Gels) run on ice in MOPS buffer (NuPAGE MOPS SDS Running Buffer) for 3 h at 200 V and transfer was performed at 100V for 2 h. Primary staining using mouse anti-afadin antibody (BD Transduction Laboratories, 1:2000) and mouse anti- β-actin antibody (SIGMA, 1:1000) were incubated overnight at 4°C. Secondary antibody goat anti-mouse IgG-HRP (1:10000, Santa Cruz Biotechnology's) was incubated for 1 h at room temperature. Protein detection was performed using ECL Western Blotting Substrate (Pierce).
For human placental organ cultures, placentas from elective terminations of pregnancy (gestational age 4 to 8 weeks) were collected and prepared as previously described [30]. Briefly, fragments from the surface of the placenta were dissected into 3–6 mm tree-like villi, placed on Matrigel (BD Biosciences)-coated Transwell filters (Millipore, 30-mm diameter, 0.4 μm pore size) and cultured in Dulbecco's modified Eagle-F12 medium (DMEM-F12; 1:1, vol/vol) supplemented with 20% FBS, 1% L-glutamine and 1% penicillin/streptomycin (Invitrogen). Organ cultures were fixed for 15–30 min in 4% paraformaldehyde at room temperature, flash frozen in Tissue-Tek O.C.T. Compound (Sakura Finetek, Torrance, CA). Sections were cut at 7 μM, permeabilized for 5 min in ice-cold acetone, dried, rehydrated in PBS, and blocked with Background Buster (Innovex Biosciences, Richmond, CA). Primary staining with mouse anti-afadin antibody (BD Transduction Laboratories, 1:2000) was incubated for 2 h at room temperature in PBS with 1% BSA and 5% normal goat serum (Jackson ImmunoResearch Laboratories). Secondary staining with Alexa-Fluor 594 goat anti-mouse (1:500, ThermoFisher Scientific) was incubated for 1 h at room temperature, nuclei stained with DAPI (Affymetrix), and sections mounted in Vectashield Mounting Medium (Vector Laboratories). Stitched high-power images were acquired on a Nikon Ti-E epifluorescent microscope with DS-Ri2 camera and NIS-Elements 4.30 software (Nikon Corporation).
MDCK (Madin Darby Canine kidney cell line, ATCC), MDCK-AF6-/- (a generous gift from Denise Marciano, UT Southwestern) cells were grown in Eagle's MEM (Minimal Essential Medium) with Earle's Balanced Salt Solution (BSS) supplemented with 10% FBS [69]. Invasion assays and growth curves were performed as previously described [31]. MDCK cells expressing a humanized version of InlP were generated as follows: the IRES-GFP sequence was amplified from the pIRES-EGFP-puro plasmid (Addgene) using oligo fw-IRESGFP-EcoRI: GGAATTCGAATTCTAAGCCCCTCTCCCTCCCCCC and rv-IRESGFP-BamHI GGAATTCGGATCCTTACTTGTACAGCTC GTCCATGCCG and cloned into the plasmid PCR2.1. The inlP gene was codon optimized for expression in mammalian cells (hinlP) and the NheI-hinlP-Flag-EcoRI construct was synthesized by GeneWiz. Using the restriction enzymes NheI and EcoRI the NheI-hinlP-Flag-EcoRI construct was then cloned into PCR2.1-IRES-GFP. The construct NheI-hinlP-Flag-EcoRI-IRES-GFP-BamHI was then cut from PCR2.1 using the restriction enzymes NheI-BamHI and cloned in pCW-Cas9-Blast (Addgene) plasmid by replacing the Cas9 gene and generating the plasmid pCW- hinlP-Flag-IRES-GFP. UCSF ViraCore was used to generate a Lentivirus strain carrying pCW-hinlP-Flag-IRES-GFP, which was used to transduce MDCK or MDCK AF-6-/- cells. Viral transduction was performed in 12-well plates as follows: freshly plated MDCK cells were allowed to adhere for 1 h at 37°C (4x104 cells/well), next nearly all media was removed and 350 μl of viral supernatant was added to each well, followed by overnight incubation at 37°C. The next morning 1 mL fresh media was added to each well without removing the viral supernatant and cells were grown for 48–72 h, followed by puromycin selection (5 μg/mL) for another 4 days. Protein expression was analyzed by Western Blot +/- addition of doxycycline 1 μg/mL. Primary staining: mouse anti-flag antibody (1:5000), and mouse anti-β-actin antibody (SIGMA, 1:1000), secondary antibody: goat anti-mouse IgG-HRP (1:10000, Santa Cruz Biotechnology). Protein detection was performed using ECL Western Blotting Substrate (Pierce).
mRNA was extracted from bacteria grown in BHI at O.D. 0.5. Bacterial pellet was resuspended in 1ml of TRIzol and lysed using lysing matrix B 2 ml tubes and FastPrep 2 cycle 6.5 m/s with 60 sec on ice between cycles. RNA was then purified using kit Directzol RNA MiniPrep (Zymo Research), treated with Dnase I (Roche) for 2.5 hrs at 37°C and further purified using RNA cleaning using KIT (Quiagen). cDNA was then prepared using the ImPromII reverse transcriptase (Promega). qPCR was performed using SsoAdvanced Universal SYBR Green Supermix (Biorad). Genes of interest were amplified using primers: AACCAATTGACAACTATTCCTGA, TGGTGTAGTTAACCATCGTACCAG for inlP (lmo2470) and inlP-ΔLRR5; CTTCCGCAATGGACGAAAGT, ACGATCCGAAAACCTTCTTCATAC for detecting ribosomal prokaryotic RNA 16S; on a StepOnePlus Real-Time PCR System. 16S was used as reference (housekeeping) gene.
Single cell TFM assays were performed as previously described [32, 33]. Two layered polyacrylamide hydrogels, the upper of which contained 0.04% carboxylate-modified red latex beads of 0.1 μm in diameter (FluoSpheres; Molecular Probes) were prepared and activated via SulfoSanpah crosslinker (Thermo Fischer Scientific, 22589) as previously described [32, 33]. Hydrogels were attached to 24-well glass bottom plates (MatTek) to enable monitoring of multiple conditions simultaneously. The stiffness of the hydrogels was 5 kPa, achieved with a 5% final concentration of acrylamide (Sigma, A4058) and 0.15% final concentration of bis-acrylamide (Fisher, BP1404-250). Activated hydrogels were coated with collagen I (Sigma-Aldrich, C3867) overnight at 4°C. Hydrogels were washed with PBS the next day and equilibrated with cell media for 30 min at 37°C prior to addition of cells.
For single cell TFM experiments, 2 mL containing 10x105 cells were added on six-well TC polystyrene plates for 24 h. Upon seeding 1 μg/mL doxycycline was added to the wells where InlP expression was desired. The next day, cells were lifted via 0.25% trypsin-EDTA, washed once in PBS and then cell pellets were re-suspended in media and 0.5x104 cells in 1 mL were added per hydrogel to achieve single cell attachment. Cells were allowed to adhere for 3 h at 37°C before imaging. To measure cell-ECM traction forces in confluent monolayers cells were added to a concentration of 4x105 cells per well directly onto the hydrogels 24 h prior to imaging. Upon seeding 1 μg/mL doxycycline was added to the wells where InlP expression was desired.
Multi-channel time-lapse sequences of fluorescence (to image the beads within the upper portion of the hydrogels) and phase contrast images (to image the cells) were acquired using an inverted Nikon Diaphot 200 with a CCD camera (Andor Technologies) using a 40X Plan Fluor NA 0.60 objective and the MicroManager software package (Open Imaging). In addition, at the beginning and end of the recordings an image of the cells’ nuclei stained with 1 μg/mL Hoechst and an image of InlP fluorescence were acquired to ensure similar cell densities and similar InlP expression across conditions, respectively. The microscope was surrounded by a cage incubator (Haison) maintained at 37°C and 5% CO2. Images were acquired every 5 min for approximately 4 h. Subsequently, at each time interval we measured the 2D deformation of the substrate at each point using an image correlation technique similar to particle image velocimetry [34]. We calculated the local deformation vector by performing image correlation between each image and an undeformed reference image which we acquired by adding 10% SDS at the end of each recording to detach the cells from the hydrogels. We used interrogation windows of 32 x 16 pixels (window size x window overlap) for cells forming a monolayer or 32 x 8 pixels for single cells. For single cell experiments, we used a custom algorithm using MATLAB (MathWorks) to identify the contour of the cells from the phase contrast images [35]. We calculated the two-dimensional traction stresses that single cells exert to the hydrogel as described elsewhere [35]. We calculated the strain energy (Us) as the mechanical work imparted by the cell to deform its hydrogel:
Us=12∫sτ→(z=h)∙u→(z=h)ds, where u→ is the measured displacement vector field on the free surface of the hydrogel and ∫s()ds represents a surface integral. For cell monolayer experiments, traction stresses were measured as described above and elsewhere [33, 36].
To quantify L. monocytogenes transcytosis, MDCK cells were seeded on Transwell inserts with a 3-μm pore size loaded onto tissue culture-treated 12-well plates (Corning, 3402) at a density of 5x105 cells per insert three days prior to infection. The day before infection, L. monocytogenes cultures were inoculated in 3 mL of Brain Heart Infusion (BHI) directly from glycerol stocks and grown at 30°C for 15 to 16 h without agitation. The day of the experiment, the optical density of the overnight cultures was measured to ensure that the MDCK cells were infected with a comparable multiplicity of infection (MOI) across bacterial strains. For each strain, 1.2 mL was centrifuged at 2000 x g for 5 min, and washed with 1 mL of DPBS. Bacterial pellets were resuspended in 16 mL of MEM. Host cells were washed with MEM once, infected with 0.5 mL of bacterial mix per Transwell, and incubated at 37°C for 10 min. The unused bacterial mix was serially diluted and plated onto BHI agar plates containing 200 μg/mL streptomycin for bacterial enumeration. Transwells were washed three times with MEM alone, then basal and apical media were both replaced with MEM + 10% FBS, and plates were incubated at 37°C for 1 h to allow bacterial invasion, followed by replacement with MEM + 10% FBS + 50 μg/mL gentamicin to kill extracellular bacteria, and incubated at 37°C for 15 min more. To assess subsequent transcytosis, Transwells were washed three times with MEM + 10% FBS, and were then transferred to a new 12-well plate containing 1 mL of fresh MEM + 10% FBS per well. This plate was incubated at 37°C for 1 h, at which point basal media was collected, vortex vigorously and 100 μL of each sample plated onto BHI agar plates containing 200 μg/mL streptomycin. This step served as a negative control to exclude for damaged monolayers. Meanwhile, the Transwells were transferred to another new 12-well plate containing fresh MEM + 10% FBS per well. After an additional 3 h at 37°C, basal media was plated on BHI agar plates containing 200 μg/mL streptomycin as before, this time to assess true transcytosis.
To quantify L. monocytogenes invasion, MDCK cells were seeded onto tissue culture-treated 12-well plates at a density of 5x105 cells per well three days prior to infection. MDCK cells were then infected with an ΔactA L. monocytogenes strain expressing mTagRFP under the actA promoter [37], which becomes transcriptionally active upon entry into the host cell cytoplasm. Invasion was assayed with flow cytometry as previously described [38].
We attempted to find host binding partners for InlP in placental cells using a yeast two-hybrid system to screen a human placenta library [39] (S2 Table). In parallel, we used InlP fused to glutathione-S-transferase (InlP-GST) and a mass spectrometric approach to pull-down and identify InlP binding partners in human placental tissue extracts (S1 Fig and S1 Table). Both methods should in principle be able to detect both host cell surface and cytoplasmic binding partners for InlP. While each of these approaches identified multiple candidates, the protein afadin, encoded by the MLLT4/AF-6 gene, was the only one identified by both. Afadin is an F-actin-binding protein that also binds to the cytoplasmic domain of the cell-cell adhesion molecule nectin, as well as to a wide variety of signaling proteins including Src and multiple members of the Ras superfamily [40]. Afadin has been shown to contribute to the formation of cell-cell junctions during mouse embryonic development [41, 42] and to apical constriction of adherens junctions during Drosophila morphogenesis [43]. In addition, afadin has been proposed to play multiple roles in cell polarization, migration, differentiation, and oncogenesis [40].
We confirmed specific binding of InlP to afadin by pull-down experiments followed by Western blot analyses (Fig 1A and 1B). Elution fractions obtained after pull-down using an InlP-GST fusion protein incubated with cell lysates from MDCK, MDCK AF-6-/- or human placental villi were analyzed. Using anti-afadin antibodies, we detected afadin signal in the input fraction and in the elution fractions after incubation of InlP-GST with protein extract from MDCK cell lines (Fig 1A) or from human placental villi (Fig 1B). No afadin was detected in the elution fractions when the GST control was used as bait (Fig 1A and 1B). Confirming the specificity of our antibodies, Western blot analysis on pull-down experiments performed using protein extract from the afadin knock-out MDCK cell line did not show a detectable signal (Fig 1A).
To determine the amount and spatial distribution of afadin present in human placental villi, we performed immunofluorescence microscopy staining for afadin and observed that this protein is abundant in the placenta (Fig 1C–1E). Localization was primarily observed at the CTB layer just beneath the syncytium, with small amounts in the stroma (see asterisk and arrow in Fig 1C–1E). The abundance of this potential InlP binding partner in the placenta, combined with our previous report that InlP plays an important role in L. monocytogenes infection at the maternal-fetal interface [21], are consistent with the hypothesis that InlP-afadin interactions are important for the pathogenesis of placental infections.
The internalins that are predicted to be secreted by L. monocytogenes include InlC (Lmo1786), InlP (Lmo2470), Lmo2027 and Lmo2445 [24]. Of these, only InlC has been extensively studied, but its primary sequence is not particularly similar to that of InlP (24% sequence identity). The two other internalins predicted to be secreted, Lmo2027 and Lmo2445, have not yet been characterized. Pairwise analyses of amino acid sequences indicated that InlP and Lmo2445 are only 25% identical and 40% similar; however, InlP and Lmo2027 are much less divergent, with 65% identity and 77% similarity (Fig 2A). Moreover, InlP and Lmo2027 share a similar C-terminal domain that is not found in any of the other internalin family members [24].
Because of its high degree of sequence similarity to InlP, we decided to test whether Lmo2027 also binds afadin. As before, we performed pull-down assays with MDCK cell lysates, followed by Western blots on elution fractions using Lmo2027-GST or InlP-GST fusion proteins as bait (Fig 2B and 2C). Unlike InlP-GST, Lmo2027-GST did not enrich the afadin signal in the elution fractions significantly more than the GST control. Even when diluted 10-fold, InlP-GST still pulled down considerable afadin in comparison to the GST control and Lmo2027-GST. These results show that, despite the relatively high sequence similarity of the two proteins, InlP, but not Lmo2027, binds afadin.
To identify structural distinctions that might account for differences in afadin binding affinity, we obtained 1.4 Å and 2.3 Å crystal structures of InlP and Lmo2027, respectively (S3 Table). Analysis of the structures reveals that InlP and Lmo2027 generally resemble previously characterized internalins (Fig 2D and 2E). Both proteins possess an N-terminal domain that adopts the helical structure characteristic of internalins. This is followed by a concave leucine rich-repeat (LRR) domain made up of 9 LRRs in InlP and 8 LRRs in Lmo2027 that can be further subdivided into a “cryptic” portion consisting of 3 LRRs, which were not anticipated to adopt an LRR-fold based upon bioinformatics analyses, and a “canonical” portion consisting of 6 LRRs in InlP and 5 LRRs in Lmo2027. Finally, the C-terminal domain in both proteins assumes an immunoglobulin-like fold.
In particular, two features distinguish InlP and Lmo2027 from each other and from previously characterized internalins. First, the additional LRR in InlP extends the LRR domain in this protein as compared to Lmo2027 and increases the distance between the N-terminal and Ig-like domains by ~4 Å. Such a fundamental distinction in the dimensions of the internalin could affect whether a prospective binding partner could dock to and bind the LRR. Second, the concave surface of previously crystallized internalin LRRs present an unbroken β-sheet [24], but in InlP and Lmo2027 the third cryptic LRR (3rd LRR loop) protrudes, adopting a twisted conformation that disrupts the continuity of this surface. Since this region in other LRR proteins is generally involved in engaging the protein binding partner [22, 44], the distinct 3rd LRR loop seems well situated to play a role in conferring binding specificity.
Unexpectedly, a calcium ion (from the crystallization solution) bound to the center of the 3rd LRR loop structure in InlP, but not Lmo2027 (Fig 2D and 2E; S2 Fig), suggesting that a calcium-stabilized loop conformation might be relevant for InlP protein-protein interactions. Two of the residues specifically involved in coordination of the calcium ion in the InlP crystal structure, D132 and G135, are not conserved in Lmo2027 (changed to T and E, respectively). To examine the interaction between InlP and calcium, we performed isothermal calorimetry experiments, and found that calcium binds with a relatively low affinity of 35 μM (S2 Fig). It is thus possible that the binding of calcium (or possibly another cation) influences InlP interactions in some physiological context.
Based on the alignment shown in Fig 2A, it is evident that the main differences between InlP and Lmo2027 are amino acid differences in that cluster to the C-terminal domain, as well as the absence of 22 amino acid residues of Lmo2027 corresponding to one of the LRR motifs of InlP. Given the high sequence similarity of the 6 “canonical” LRR motifs, it is difficult to definitively identify which one is missing in Lmo2027, even if the pairwise alignment between InlP and Lmo2027 suggests that the missing LRR motif might be LRR7. Therefore, we decided to systematically delete individual LRR motifs in the InlP protein and test the effect of these deletions on the protein’s ability to bind afadin. We successfully generated three InlP deletion mutants—InlPΔLRR5-GST, InlPΔLRR7-GST and InlPΔLRR8-GST—lacking the LRR5 (Δamino acids 174–195), the LRR7 (Δamino acids 218–239) and the LRR8 (Δamino acids 240–261) motifs, respectively. Each of these was then used as bait in pull-down experiments with MDCK cell lysates, followed by Western blot analysis of the elution fractions to detect afadin.
Resins prepared using E. coli lysate expressing InlPΔLRR7-GST (Fig 3A) pulled down afadin from MDCK cell lysates with an efficiency roughly comparable to that observed for InlP-GST (Fig 3B). In contrast, we detected lower amounts of afadin in the pull-down fraction when lysate from E. coli expressing the InlPΔLRR5-GST fusion protein was used to prepare the resin (Fig 3B). As the expression level of the InlPΔLRR5-GST fusion protein in E. coli appeared to be somewhat lower than that of either InlP-GST or GST alone (Fig 3A), we also performed a Western blot comparing the elution fractions after five-fold dilution for both InlP-GST and GST with no dilution for the InlPΔLRR5-GST elution fraction. Even compared to diluted InlP-GST, substantially less afadin was detected in the sample for InlPΔLRR5-GST (Fig 3C). This suggests that deletion of the LRR5 motif decreases afadin binding by at least ten-fold. The InlPΔLRR8-GST was expressed in E. coli lysates at ~30% of InlP-GST levels prepared under similar conditions (S3 Fig). While we did observe reduced afadin binding by resins prepared using E. coli lysates containing this fusion protein as compared to native InlP-GST (S3 Fig), this decrease was approximately proportional to the reduction in expression level. Overall this analysis suggested that InlPΔLRR5 represents a useful mutant for interrogating the InlP-Afadin interaction.
To examine the effects of the InlPΔLRR5 mutation on binding to host proteins, we used a GST control, InlP fused to GST (InlP-GST) or InlPΔLRR5 fused to GST (InlPΔLRR5-GST) as baits and identified proteins pulled down from MDCK epithelial cell lysates by mass spectrometry (Fig 3D and S1 Table). Of the MDCK cell proteins found specifically to bind InlP-GST, only afadin (MLLT4/AF-6) had also been identified in the screens described above using yeast two-hybrid screening on a human placental library (S2 Table) and mass spectrometry of proteins pulled down from human placental tissue extracts (S1 Table). Consistent with the Western blot analysis, mass spectrometry analysis confirmed that afadin associated much less strongly with InlPΔLRR5-GST than with InlP-GST (Fig 3D and S4 Fig). These findings are consistent with the hypothesis that the InlPΔLRR5 mutant is specifically deficient in interacting with afadin in MDCK cells, while otherwise retaining much of its normal structure and protein-protein interaction potential.
We sought to investigate whether the interaction between InlP and afadin could alter the structural integrity of epithelial cell monolayers in ways that might facilitate L. monocytogenes invasion or cell-to-cell spread. To test whether InlP affects the integrity of cell-cell junctions in a monolayer, we constructed a line of MDCK cells that could express a humanized version of InlP after induction by doxycycline (MDCK-InlP+). We cultured wild-type MDCK, MDCK AF-6-/-, and MDCK-InlP cells on Transwell filters for three days and measured the permeability of the cell monolayers using inulin-FITC. As expected after deletion of a protein necessary for proper formation of adherens junctions and tight junctions [41], MDCK AF-6-/- monolayers showed a substantial increase in permeability to inulin relative to wild-type MDCK monolayers. However, MDCK-InlP monolayers showed low permeability, comparable to the baseline level for wild-type MDCK monolayers, both with and without induction of the InlP transgene using doxycycline (Fig 4A). This result suggests that the interaction between InlP and afadin does not inhibit the ability of afadin to perform its normal function in establishment and maintenance of impermeable cell-cell junctions.
Next, we used these cell lines to examine the modulation of cell-substrate interactions by InlP binding to afadin. To this end, we also generated a line of MDCK AF-6-/- cells with the doxycycline-inducible InlP transgene. We used traction force microscopy (TFM) to measure the ability of MDCK cells to generate traction forces on their substrates with and without expression of Afadin and/or InlP. Uninduced MDCK-InlP cell monolayers without doxycycline (the positive control) grown on deformable substrates as confluent, polarized monolayers were able to generate significant strain energy on the order of 100–200 nN•μm (Fig 4B and 4C) and to exert traction stresses with peak values around 100 Pa, comparable to previous reports [45, 46]. Both the induction of the InlP transgene and the deletion of afadin caused a substantial reduction in strain energy. Induction of InlP expression in the background of the MDCK AF-6-/- cell line did not lead to any further reduction, consistent with the hypothesis that the primary effect of InlP expression on abrogation of traction force generation in MDCK cells is due to its interaction with afadin.
We have previously demonstrated that a decrease in monolayer traction force can be caused either by a direct inhibition of cell-substrate interactions or by an indirect inhibition of lateral tension at cell-cell junctions in the monolayer [33]. To distinguish between these two possibilities, we also performed TFM for isolated cells under the same four conditions. Just as for the monolayers, either the expression of InlP or the deletion of AF-6 caused a significant decrease in the ability of isolated individual cells to generate traction force (Fig 4D and 4E). Taken together, these results suggest that the InlP-afadin binding interaction specifically disrupts afadin’s normal activity in enhancing traction force generation at the cell-substrate interface, but has no effect on afadin’s normal function at cell-cell junctions in the monolayer.
Finally, we examined the influence of the InlP-afadin interaction on infection and spread of L. monocytogenes in MDCK monolayers. We infected wild-type and MDCK AF-6-/- cells with wild type or ΔinlP L. monocytogenes. Gentamicin was added at 30 min to kill extracellular bacteria, and remaining intracellular bacteria were enumerated at 2, 5, 8 and 24 h post-infection. At 2 h, the MDCK AF-6-/- cells harbored ~10 times the number of bacteria as compared to wild-type MDCK cells, although the rate of intracellular growth of the bacteria was unchanged. Deletion of inlP had no significant effect on invasion or intracellular replication of L. monocytogenes in either host cell background (Fig 5A).
The observation that MDCK AF-6-/- monolayers had substantially more intracellular bacteria at early time points in the course of infection as compared to wild-type MDCK cells raised the possibility that the deletion of AF-6 caused an increase in the efficiency of L. monocytogenes invasion. Because the cell receptor for L. monocytogenes invasion into epithelial cells, E-cadherin, is normally sequestered from the apical surface of cells in well-polarized, confluent monolayers, invasion under such circumstances is rare and typically takes place at sites of local, transient disruption of the epithelial cell-cell junctions, for example at sites of extrusion of apoptotic cells [7]. Consequently, any disruption of normal cell-cell junctions in epithelial monolayers typically increases the efficiency of L. monocytogenes invasion [38]. To determine if this was the reason for the increased bacterial load in MDCK AF-6-/- monolayers at early time points, we infected both wild-type and AF-6-knockout MDCK cells with ΔactA L. monocytogenes expressing mTagRFP under the actA promoter and counted the number of RFP-positive host cells by flow cytometry 5 h post-inoculation, which represents the number of bacteria-containing cells. Because these bacteria are not capable of actin-based cell-to-cell spread, this assay allows determination of the number of directly invaded host cells, independent of any later variations in bacterial spread [47]. Indeed, we found approximately 6-fold more invasion of MDCK AF-6-/- cells with L. monocytogenes as compared to wild-type MDCKs (Fig 5B).
Our previous work demonstrated that wild-type and ΔinlP L. monocytogenes form foci of similar size in MDCK monolayers [21], suggesting that horizontal cell-to-cell spread in a polarized epithelium is not affected by InlP. This is corroborated by our findings here that the InlP-afadin interaction does not disrupt the function of afadin at cell-cell junctions. However, we also showed that ΔinlP strains are impaired in spreading from the CTB epithelial layer into the placental stroma [21], and here we have found that the InlP-afadin interaction specifically disrupts the mechanics of MDCK cell traction force generation at the cell-substrate interface. We therefore wondered whether this protein-protein interaction might specifically affect the ability of intracellular L. monocytogenes to use actin-based motility to spread through the basal face of the CTB epithelial monolayer underlying the placental STB, rather than horizontally through the lateral faces of neighboring CTB epithelial cells. We designed an assay to mimic this particular type of spread by growing MDCK cells on Transwell filters with large (3 μm) pores through which bacterial protrusions could extend into the lower well. Under these culture conditions, MDCK cells form well-polarized monolayers and secrete a thick and well-organized basement membrane (basal lamina) [48], so bacterial protrusions reaching the lower well must have been able to form at the basal side of the cell and push through the basement membrane. Approximately 5 hours after inoculation from the apical side, bacteria that had successfully performed transcytosis through the monolayer were collected from the lower well. In this assay, ΔactA bacteria exhibited a substantial (>10-fold) reduction in transcytosis, demonstrating that this process is dependent on bacterial actin-based motility (Fig 5C). When we infected WT MDCK monolayers with the ΔinlP strain we found a statistically significant reduction in transcytosis (about 4-fold or 75%) as compared to wild-type L. monocytogenes (Fig 5C), consistent with the hypothesis that the InlP-afadin interaction specifically alters cell mechanics at the basal cell-substrate interface and thereby facilitates L. monocytogenes spread in a direction perpendicular to the monolayer.
If the nature of the InlP-afadin interaction were simply to sequester or inhibit afadin’s normal activity at the basal surface, we would expect that deletion of AF-6 in the host cells should be able to rescue the transcytosis defect of ΔinlP L. monocytogenes. Therefore, unlike wild-type MDCK monolayers showing a 75% decrease in transcytosis when infected with ΔinlP L. monocytogenes as compared to wild-type bacteria, we would expect this decrease to be minimal when infecting MDCK AF-6-/- monolayers. Consistent with this hypothesis, when we infected MDCK AF-6-/- monolayers with ΔinlP L. monocytogenes we observed just a 20% decrease in transcytosis as compared to wild-type L. monocytogenes (Fig 5D). This result confirms that the InlP-afadin interaction is important for efficient transcytosis, while the slight residual decrease in transcytosis suggests that there might either be additional binding partners of InlP that facilitate crossing of basement membranes, or that this decrease is due to some pleiotropic effect of AF-6 deletion in the MDCK cells. Consistent with the latter hypothesis, we did typically observe less transcytosis of wild-type L. monocytogenes for MDCK AF-6-/- monolayers as compared to wild-type MDCK monolayers for experiments performed in parallel, while the absolute amount of transcytosis for ΔinlP L. monocytogenes was fairly similar in both host cell types (S5 Fig). An intriguing alternative hypothesis is that the InlP-afadin protein-protein interaction might exert a specific effect to enhance transcytosis of L. monocytogenes that is not equivalent to the simple removal or sequestration of afadin from this site.
As an additional line of evidence that the InlP-afadin interaction is largely responsible for the decrease in transcytosis for MDCK cells infected with ΔinlP L. monocytogenes, we took advantage of our earlier determination that the InlPΔLRR5 mutation specifically disrupts afadin binding. Infection of wild-type MDCK cells with L. monocytogenes expressing the mutant InlPΔLRR5 resulted in a transcytosis defect comparable to the defect of the complete ΔinlP deletion in parallel experiments (Fig 5E). We also confirmed that inlPΔlrr5 is expressed at similar levels to WT inlP by RT-qPCR for bacteria grown in BHI media, unlike the ΔinlP mutants that show no expression of inlP (S5 Fig). Without an antibody to the InlP protein, we cannot directly confirm that the InlPΔLRR5 protein is produced or secreted at normal levels in these infection experiments. However, we note that the purified InlPΔLRR5-GST fusion protein binds normally to most protein binding partners from MDCK cell lysates, suggesting that it is stable and able to fold properly. Therefore, we conclude overall that the direct binding interaction between InlP and afadin enhances transcytosis of L. monocytogenes through the basal face of polarized epithelial cells.
En route to infection of the fetus, L. monocytogenes experiences multiple bottlenecks, from trafficking to the placenta [2] to surviving the placental innate immune defenses [49] to crossing the trophoblast monolayer and its associated basement membrane into the fetal stroma and fetal circulation [19]. InlP initially attracted our interest due to its identification in a screen for L. monocytogenes mutants that were defective in their ability to infect the placenta [21]. Two unbiased screening methods to identify potential placental binding partners, a yeast two-hybrid screen using a placental cDNA library and mass spectrometric identification of human proteins pulled down by InlP from placental extracts, converged on afadin as a significant candidate binding partner. Afadin is not found on the cell surface, but instead is primarily associated with the cytoplasmic face of cell-cell junctions containing nectin [41, 50]. Therefore, it seems likely that InlP is most helpful in breaching the trophoblast monolayer and accessing the fetal stroma.
Of the 25 known members of the L. monocytogenes internalin family, the only other one known to have host cell cytoplasmic binding partners involved in its virulence functions is InlC. InlC binds to one of the subunits of the IκB complex thereby modulating the host cell innate immune response [51]. InlC also binds the adapter protein Tuba, which cooperates with the actin polymerization regulator N-WASP to modulate tension at cell-cell junctions and thereby promote L. monocytogenes cell-to-cell spread in epithelial monolayers [52]. Intriguingly, InlC and InlP are two of only four internalin family members that are thought to be secreted in a soluble form rather than anchored to the bacteria cell surface [24]. As more internalin family members are characterized, it will be interesting to see whether the trend persists of surface-anchored internalins binding to host cell surface proteins and secreted internalins binding to host cell cytoplasmic proteins.
While we have not determined the structure of the InlP-afadin complex, our work does give some clues into the likely nature of this interaction. The internal in-frame deletion mutant InlP-ΔLRR5, which lacks one of the leucine-rich repeats, binds afadin less well than full length InlP, but this is probably not simply due to the shorter overall size of the concave face of the LRR domain because InlP-ΔLRR7 and InlP-ΔLRR8, which should have a similar overall length of the LRR domain concave face, seem to have little or no deficiency in afadin binding. It is thus likely that there are specific residues in LRR5 that stabilize the interaction. An interesting unanswered question concerns the role of the unusual protruding loop from LRR3 in determining binding partner specificity. Although we found that this loop between amino acids 132–136 in the InlP structure bound Ca2+ with an affinity of 35 μM, this binding affinity is probably too weak to be relevant in the host cell cytoplasm where the Ca2+ concentration rarely rises above ~2 μM [53]. This is not the case for other subcellular compartments like the acidifying phagosomes, where the Ca2+ concentration has been shown to be >70 μM [53]. Thus, it is possible that calcium binding by InlP plays a role during L. monocytogenes escape from the primary (single-membrane) or secondary (double-membrane) vacuole. It should be noted that Shaughnessy et al. [54] observed that listeriolysin O pores generated by the bacterium in the process of primary vacuolar escape allow Ca2+ in the vacuole to leak into the cytoplasm, albeit slowly. It is possible that calcium binding promotes changes in InlP behavior that are specific to different calcium microdomains in the polarized epithelium [55, 56], including microdomains potentially specific to basal protrusions and/or placental trophoblasts [57].
Other bacterial pathogens are also known to secrete virulence factors into the host cell cytoplasm that modulate host cell biology in ways that benefit bacterial survival or spread [33, 58–61]. One particularly relevant example is the virulence factor Sca4, secreted by Rickettsia parkeri. Analogous to InlC, Sca4 binds the host protein vinculin at cell-cell junctions and alters cortical tension, promoting actin-based cell-to-cell spread of the bacterium horizontally through host cell monolayers [33]. Given the precedent that both InlC and Sca4 bind to proteins that perform scaffolding functions at cell-cell junctions in polarized epithelia (Tuba and vinculin respectively) and modulate cortical tension to promote bacterial spread, our initial expectation on discovering that InlP binds the cell-cell junction-associated protein afadin was that its mechanism of action would be similar. However, we have found no evidence that InlP has any effect on cell-cell junctions in the host monolayer, or indeed that it makes any significant contribution to lateral actin-based spread of L. monocytogenes within the horizontal plane of the monolayer.
Instead, our results indicate that the relevant site of action of InlP in the host cell is the basal face, not the lateral face, and that the InlP-afadin interaction specifically modulates the nature of the host cell organization at the basal face or interaction with the underlying basement membrane in a way that promotes L. monocytogenes spread that is perpendicular, rather than parallel to the monolayer. Afadin has not been described as being localized to the basal face of polarized epithelial cells, or to cell-ECM junctions; instead its functional characterization has focused primarily on nectin binding at cell-cell junctions [40, 41]. Conditional knockouts in the mouse intestine lead to increases in intestinal permeability [62], consistent with the phenotype we have described here of afadin loss causing permeabilization of cell-cell junctions in MDCK cells. Whether afadin also plays a role in cell-ECM connections is unclear, but in certain cell types, removal of afadin can enhance cell migration [63] and invasive capacity [64], or suppress neuronal axon branching [65]. While these cellular behaviors might involve specific changes in cell-ECM interactions, it is also likely in these cases that the effects of afadin disruption are mediated indirectly through its interactions with critical signaling molecules, Rap1, Src and R-Ras respectively [63–65].
How does InlP binding modulate afadin activity at the basal face of epithelial cells? For both InlC and Sca4, deletion of the host cell binding partner is sufficient to (partially) rescue the bacterial deletion phenotype, consistent with the simple idea that binding of the bacterial virulence factors to their host cell partners effectively inhibits or sequesters those host cell partners [33, 52]. In contrast, our results indicate that the consequences of the interaction between InlP and afadin are more complex than simple sequestration. InlP expression in uninfected MDCK cell monolayers has no effect on integrity of cell-cell junctions, while deletion of afadin results in a significant increase in monolayer permeability, so the presence of InlP does not appear to disrupt afadin’s normal activities at the cell-cell junctions. In contrast, expression of InlP phenocopies afadin deletion with respect to the ability of epithelial cells to generate traction forces on their substrates. And most intriguingly we have found that AF-6 knockout MDCK cells are less able to support transcytosis of wild-type L. monocytogenes across the basement membrane than are wild-type host cells; if afadin normally simply strengthens the cell-substrate interface then the AF-6 knockout host cells should be more permissive for bacterial transcytosis, rather than less. This raises the interesting possibility that the InlP-afadin binding interaction results in a novel activity that specifically promotes bacterial formation of actin-based protrusions at the cell basal face.
In the context of placental infection, the spatial specificity for the activity of the InlP-afadin interaction at the basal face of polarized epithelial cells is relevant, as our previous work has suggested that vertical spread from the CTB monolayer into the underlying fetal stroma is a major barrier for infection of the fetus [19, 21, 66]. However, we previously showed that the ΔinlP mutant is not impaired in reaching the liver or spleen in orally-infected guinea pigs [21], raising the question of why the phenotype is apparently specific for placental infection when enhanced transcytosis through the basement membrane should also be relevant for L. monocytogenes crossing the intestinal epithelium or crossing the endothelial cells in blood vessels to invade the liver. In this context, we note that the intestinal barrier is organized differently from that of the placenta. The intestinal epithelium is rich in intercellular junctions that link neighboring cells together, while the placental syncytiotrophoblast that contacts maternal blood is a single, continuous plasma membrane extending over tens of thousands of square millimeters in later stages of human gestation. The invasive cytotrophoblasts in the uterine decidua do have exposed lateral surfaces and cell-cell junctions, but they lack the lymphatic and immune components of the intestine in that there are no Peyer’s Patches and M cells with phagocytic and migratory capabilities. While some white blood cells do traffic across this barrier in both directions [67], traffic is likely more limited than in the mucosal-associated lymphoid tissues of the gut. The lack of a role for InlP in crossing the intestinal barrier is therefore consistent with earlier work showing that L. monocytogenes traffics from the gut to the bloodstream by being carried by immune cells [68]. In this context, the limited immune cell traffic observed across the placental barrier might explain how ΔinlP L. monocytogenes are impaired in crossing the subsyncytial CTB monolayer—undergirded by a basement membrane—into the fetal stroma [21]. Alternatively, it is possible that the basement membrane underlying the CTB layer simply presents a more formidable physical barrier than that found underneath the intestinal epithelium. An interesting issue, to be addressed in future work, is whether there is impairment in the invasion of ΔinlP L. monocytogenes across the blood-brain endothelial barrier.
For decades, L. monocytogenes has been a useful tool for the study of mammalian cells, barriers, and immune functions. Increasingly, its capacity to reveal basic knowledge about higher levels of biological organization—mammalian tissues and organs—is becoming evident. Through investigations of how bacteria cross barriers and of the interactions between bacterial factors and host pathways, we learn not only more about what regulates pathogenesis, but about the structures of these tissues themselves, in how they are generated, their composition, and how they can be disrupted.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.