Variation in substitution rates across bat rabies lineages

We employed maximum likelihood (ML) and Bayesian phylogenetic analyses to define 21 subspecies, species or genus specific lineages of rabies virus for comparative analyses of evolutionary rates (see Materials and Methods for analytical details and operational definitions of lineages). A relaxed molecular clock analysis indicated that viral lineages were relatively young, ranging in age from 83–305 years, with a most recent common ancestor of all bat lineages dating back to 1585 (95% Highest Posterior Density, HPD: 1493–1663; Table S1). To describe the evolutionary rate variation among viral lineages, we focused on substitution rates in the third codon position (CP₃), as these predominately synonymous substitutions can indicate more clearly how viruses respond to processes affecting their replication rate and generation time between infections [22]. Average substitution rates estimated for each lineage independently (Independent Lineage Models, ILM) and by a hierarchical phylogenetic model (HPM) each spanned approximately one order of magnitude among viral lineages compartmentalized to different host species (ILM range: 8.31×10⁻⁵–2.08×10⁻³; HPM range: 2.16×10⁻⁴–1.07×10⁻³ substitutions/site/year). This indicates that select lineages exhibit approximately 5–22 fold acceleration of evolution relative to the slowest evolving viruses. The HPM substantially improved the precision of parameter estimates relative to the ILMs, with only negligible differences in point estimates for most lineages, as previously described in other host-virus systems [8], [23] (Figure 1). Notably, the more extreme values estimated by the ILMs, typically from viral lineages with less informative datasets, were drawn closer to the population mean, suggesting less susceptibility of the HPM to stochastic noise introduced by sampling error and potentially more accurate estimates.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Heterogeneity in evolutionary rates of host-associated bat rabies viruses.

Median substitutions per site per year in the 3^rd codon position of the nucleoprotein gene for estimates generated by the HPM (filled circles) and ILMs (open circles). Colors and dashed gray lines distinguish bat genera as indicated below the x-axis. Credible intervals denote the 95% highest posterior density interval on evolutionary rate.

https://doi.org/10.1371/journal.ppat.1002720.g001

Virus evolutionary rates are labile following host shifts

If evolutionary rate is a relatively static trait of viruses, it should be conserved in novel environments and would be expected to reflect the ancestral history of the virus, with closely related viral lineages having similar rates, regardless of their contemporary host environment. We tested the degree of phylogenetic signal in the evolutionary rates of bat rabies virus lineages by quantifying values of Blomberg's K (a common statistic for diagnosing phylogenetic non-independence in comparative analysis) [24], [25]. In the context of viral host shifts, significant values of K would indicate that the evolutionary rate tended to remain similar after establishment in the recipient species, whereas weaker values of K would indicate greater shifts in evolutionary rates than expected between the donor and recipient host species. We detected very low and non-significant values of K for both rates estimated under the ILMs (K = 0.36, P = 0.58) and the HPM (K = 0.39, P = 0.40) and these estimates of K did not differ significantly from expected values given our phylogenetic tree and the observed rates under a null model with randomly distributed rates (Figure 2B). Indeed, lineages that shared a most recent common ancestor sometimes had disparate rates of evolution (e.g., LcV and LxV and PhV and MyV2 in Figure 2A), although other closely related virus pairs showed minimal differences in evolutionary rate (e.g., LnV and PsV). Notably, viral lineages maintained by bat species in the temperate zone frequently had slower rates of evolution than lineages from tropical or subtropical bats (Figure 2A). The plasticity of evolutionary rate was corroborated by our Bayesian phylogenetic analysis, which, accounting for uncertainty in the evolutionary history of bat rabies lineages, found no correlation in the evolutionary rates along consecutive branches in the bat rabies virus phylogeny (covariance = 0.005, 95% HPD: −0.051–0.058). These analyses demonstrated that rates of viral evolution may be altered or conserved following establishment in new host species and point to host biology rather than the ancestral history of the virus as the most likely candidate for controlling rabies virus evolution. Using a phylogeny of bat hosts from mitochondrial sequence data, we found that viral evolutionary rates were similarly unconstrained by host evolutionary relatedness (ILMs: K = 0.07, P = 0.18; HPM: K = 0.21, P = 0.08), such that viruses associated with closely related bat species or sub-species often had dissimilar evolutionary rates (Figure 1, Figure S1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Evolutionary lability of viral substitution rates in the rabies virus phylogeny.

(A) Bayesian phylogenetic tree of bat rabies viruses with viral lineage names and climatic regions denoted in black (TE: temperate; ST: subtropical; TR: tropical). Host-associated lineages are condensed to triangles connecting the most recent common ancestor to the sampled branches. Lineages are colored along a blue (slowest) to red (fastest) continuum according to evolutionary rate in CP₃ using estimates from the Independent Lineage Models (ILMs). Bayesian posterior support values (>0.70) are given above branches to the lineage level only. All colored lineages received Bayesian posterior support values of ≥0.91. (B) Frequency histogram of expected values of Blomberg's K from 5000 random assignments of substitution rates estimated from the ILM to lineages. Dashed lines indicate 95% bounds of the null distribution and diamonds denote the median values of K for the randomized rate model, RRM (gray), the ILMs (blue), and the HPM (red).

https://doi.org/10.1371/journal.ppat.1002720.g002

Local host environment determines virus evolutionary rates

Because the evolutionary rates of rabies virus lineages could not be explained by the reservoir host or virus phylogeny alone, we tested whether physiological, ecological or environmental traits of hosts (Table S2) could instead determine viral evolutionary rates using a generalized linear model (GLM) comparison approach. The factors that we tested included both evolutionary conserved traits of bats (i.e., basal metabolic rate, coloniality, long-distance migration) and descriptors of sampling effort and climatic region of viral lineages that were largely independent of the bat phylogeny (Table S3). The GLM identified the climatic region of bat taxa as the single strongly supported predictor of viral evolution (Akaike importance weight = 1.0), alone explaining 66% of the variance in viral evolutionary rates (F_1,19 = 37.2, R² = 0.66, P<0.0001; Table S4). A phylogenetic generalized least squares regression approach, designed to control for any residual effects of the virus phylogeny on rates of viral evolution, yielded similar results (Table S5).

A weakness of the statistical methods described above was that they could not account for the often-substantial uncertainty in point estimates of evolutionary rate from the ILMs (Figure 1). We therefore incorporated the same categorical and continuous terms directly into a Bayesian HPM that allowed us to simultaneously quantify the posterior distribution of the rate of evolution for each viral lineage from the molecular sequence data, estimate regression model parameters and compare candidate models using Bayes factors (BF) while accounting for phylogenetic uncertainty. The Bayesian model echoed the strong support for accelerated viral evolution in the tropics and subtropics relative to viruses restricted to the temperate zone (log effect size, β = 1.24 [95% highest posterior density = 0.72–1.76]; BF = 466.54) with negligible support for all other predictors (BF≤1; Figure 3A). On average, rabies viruses found in tropical or subtropical bat species accumulated 9.44×10⁻⁴ (ILM: 1.15×10⁻³) substitutions per site per year (subs/site/year), while viruses in temperate bats accumulated only 2.53×10⁻⁴ (ILM: 2.92×10⁻⁴) subs/site/year (Figure 3B) – a nearly fourfold deceleration of viral evolution in temperate bats.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Predictors of viral evolutionary rate from the Bayesian hierarchical phylogenetic model.

(A) Effect sizes (β) on a log scale for each predictor were conditioned on the inclusion of that term in the model (i.e., β | δ = 1). Climatic region, coloniality, seasonal inactivity and long distance migration were categorical variables. Horizontal lines are the 95% highest posterior density intervals on conditional effect sizes and squares (median effect sizes) are proportional to effect sizes. (B) Violin plot showing the effect of climatic region on viral evolutionary rate. White points, black boxes and whiskers indicate the median, inter-quartile range and the total range of values for that group, respectively. The gray shading shows the probability density of evolutionary rate at different values.

https://doi.org/10.1371/journal.ppat.1002720.g003

RT-PCR amplification and sequencing of rabies viruses

Viral sequences were generated from tissue samples from naturally infected bats that were collected by state public health laboratories following human or domesticated animal exposures. Total RNA was extracted directly from bat brains without passage using Trizol (Invitrogen, Carlsbad, CA) according to the manufacturer's instructions. A 903 bp fragment comprising the last 687 bp of the N gene, a non-coding region following the 3′ end of N and a small fragment of the phosphoprotein gene was amplified by reverse transcription-polymerase chain reaction and sequenced using oligonucleotide primers 550F and 304R, as described previously [16]. To enable comparison with existing sequences in GenBank, only the coding region of N was used in subsequent analyses. Sequences generated herein have been deposited into GenBank under accession numbers JN594500–JN594503 and DQ445318–DQ445330, DQ445352 (updated sequences). An additional 650 complete or partial rabies virus nucleoprotein gene sequences that were associated with bats from North and South America and contained information on sampling date to year were downloaded from GenBank.

Biological and ecological data for bat host species

We collected information on the overwintering activity patterns, migratory behavior, roosting behavior and metabolic rates (basal and during seasonal torpor) of the bat species that served as reservoir hosts for the rabies viruses included here from the primary literature and existing databases (Table S2). Long distance migration was defined as seasonal movement of individual bats of at least 1000 km [34]. For 4 species for which basal metabolic rate (BMR) data were unavailable, we borrowed values from species within the same genus or family that had similar body mass. Notably, body mass explains >92% of variation in BMR in bats and phylogeny explains much of the residual variation [29]. When torpid metabolic rate (TMR) estimates spanned a range of temperatures or spatial locations, rates were selected to match the conditions that bats are likely to experience in northern latitudes of their range where hibernation/torpor is most important. To calculate TMR for species that lacked values in the literature, we estimated the relationship between BMR and TMR for the 9 species in our dataset for which both values were available. This relationship was remarkably consistent across species (TMR = 2.2–3.2% of BMR, mean = 2.8%), with the exception of Tadarida brasiliensis, for which TMR was 12.6% of BMR. Because the reported estimate of TMR for that species (and for the other subtropical and tropical species in our study) likely represented a daily torpor rather than longer-duration, seasonal torpor, it was excluded from the calculation of the average mentioned above [35]. Overwintering activity was challenging to classify because the frequency and duration of bat activity during winter are poorly understood for many temperate bat species and can vary substantially throughout their geographic range [30], [36]. Therefore, we classified species as inactive during winter if extended bouts of seasonal torpor or hibernation were reported in any part of their geographic range, recognizing that this classification may have been overly conservative. As a potentially more geographically sensitive proxy of year-round activity, the climatic region (tropical, subtropical, temperate) of the center of the geographic range of each viral lineage was also recorded. North American lineages circulating between 35° and 23.5° latitude and South American lineages circulating south of −23.5° latitude that support mean winter temperatures of ≥10°C were considered subtropical, and lineages found towards and away from the equator relative to these latitudes were classified as tropical and temperate, respectively [37].

Selection of lineages for hypothesis testing

To define phylogenetic lineages to be included in subsequent analyses, ML and Bayesian phylogenetic analyses were performed using Garli v.0.96b and BEAST v.1.6.1, respectively [38], [39]. The ML analysis used the General Time Reversible (GTR) model of nucleotide substitution with invariant sites (I) and Γ distributed rate variation among sites as suggested by Akaike's information criterion corrected for small samples size (AICc) in jModeltest [40]. The ML tree was estimated by 5 independent searches with random starting trees, followed by 5 additional searches using the best tree from the previous set of searches as the starting tree. For the Bayesian analysis, we linked substitution rates for the first and second codon positions (CP₁₂) and allowed independent rates in CP₃. Separate substitution models were selected for CP₁₂ and CP₃ in jModeltest using AICc after partitioning aligned sequences by codon position. The BEAST analysis therefore applied the TIM1ef+I+Γ substitution model to CP₁₂ and the TVM+Γ substitution model to CP₃. We used the Bayesian skyride model as a flexible demographic prior for viral effective population size and an uncorrelated lognormal relaxed molecular clock to accommodate rate variation among lineages. Five independent Markov Chain Monte Carlo (MCMC) analyses were run for 50 million generations each, with samples from the posterior drawn every 50,000 generations following variable burn-in periods based on convergence of likelihood values and model parameters. The results from the five runs were combined to generate a maximum clade credibility tree and divergence time summaries. Lineages for subsequent analyses of substitution rates included those that (i) contained at least 8 sequences (mean = 30.9 sequences), (ii) were supported by Bayesian posterior probabilities of >0.9 and (iii) were sampled over a minimum time span of 4 years (mean = 19.4 years). These conditions aimed to achieve a compromise between precision in estimates and hypothesis testing ability. To ensure that sparsely sampled viral lineages did not bias our central findings, statistical analyses were conducted with covariates designed to identify effects of sampling heterogeneity or using hierarchical phylogenetic modeling to incorporate uncertainty in rate estimates. Of the 28 viral lineages identified in the initial phylogenetic analyses, 21 fit our criteria for inclusion in subsequent analyses, amounting to a final dataset of 648 sequences collected between 1972 and 2009 from 21 bat species or sub-species.

Independent estimation of nucleotide substitution rates in bat rabies lineages

Sequence alignments were constructed for each viral lineage and nucleotide substitution models were selected for CP₁₂ and CP₃ as described above. For each lineage, the substitution rate was estimated in BEAST assuming an uncorrelated lognormal relaxed molecular clock to accommodate rate variation along branches and the Bayesian skyline model as a flexible demographic prior that could be applied to all viral lineages. The evolutionary rate in CP₃ was calculated by multiplying the mean substitution rate by the relative rate parameter for that partition. Each simulation was run for at least 100 million generations, with parameters sampled every 5,000–10,000 generations. The first 10% of each run was discarded prior to the construction of the posterior probability distributions of parameters. Each analysis was run sufficiently long that effective sample sizes for parameters were >200 and results of several independent runs were combined. Analyses of evolutionary rates focused on substitution rates in CP₃ since these largely synonymous substitutions reflect differences in evolution associated with generation time [22]. However, rates in CP₁₂ were closely correlated with CP₃ rates (r = 0.87, P<0.0001). Rate estimates for all codon partitions are shown in Table S6 and Table S7.

Bayesian hierarchical inference of nucleotide substitution rates and hypothesis testing

The separate estimation of substitution rates for each rabies virus lineage assumes complete independence of parameters across viral lineages, but this is unlikely the case given the close evolutionary relationships among lineages and biological similarities of the processes of infection and replication among lineages. Because the quantity of data varied among lineages (the number and temporal range of sequences), independent estimation in sparsely sampled lineages may lack power, causing imprecise estimates. HPMs have been proposed to improve the precision of parameter estimates for partially independent datasets such as these (e.g., populations of HIV within different patients) by assuming that individual lineage parameters vary around a shared unknown, but estimable population mean [23]. More recently, tools have been developed within BEAST to incorporate fixed effects into HPMs and to select among candidate models via Bayes factors [8]. These models take the general form of:(1)where θ is the evolutionary response variable of interest (here, the rate of molecular evolution in CP₃), β₀ is an unknown grand mean, δ is a binary indicator that tracks the posterior probability of the inclusion of predictor, P, in the model and β is the estimated effect size of predictor P. The use of binary indicator variables (δ) within the MCMC search allows for a Bayesian stochastic search variable selection approach that simultaneously estimates the posterior probabilities of parameters for all possible combinations of predictors and allows for calculation of the Bayes factor support for individual predictors as the ratio of the posterior odds to the prior odds of each predictor in the model.

We constructed a HPM for the 21 bat rabies virus lineages that assigned separate strict molecular clocks to CP₁₂ and CP₃ of each viral lineage and included fixed effect predictors of the evolutionary rate in CP₃. For each CP, the evolutionary rate parameter of the molecular clock, the parameters of the GTR substitution model and the shape parameter of the discrete Γ distribution were modeled hierarchically across lineages, with all other parameters varying independently across data partitions. Results were robust to simpler substitution models lacking Γ heterogeneity within each codon partition and to statistical analyses using the rate of evolution averaged across all three codon positions. The fixed effects in the full model included climatic region (temperate vs. tropics/subtropics), mass-independent BMR, mass-independent TMR, coloniality (solitary vs. colonial), seasonal activity (non seasonal vs. hibernation/periodic seasonal torpor) and long-distance migration (migrants vs. non-migrants) (Table S2). Climatic region was condensed to two categories in the HPM based on exploratory analyses that demonstrated no difference in evolutionary rate between tropical and subtropical viral lineages. All continuous variables were log transformed. The HPM was implemented in BEAST using four independent MCMC searches of 150 million generations each, with the posterior sampled every 5,000 generations. Results from the four runs were combined after discarding the first 10% of each. Effect sizes of predictors reported from the HPM were calculated conditionally on the portion of the posterior distribution for which the respective effects were included in the model (i.e., β_i|δ_{Effect i} = 1). Similarly, evolutionary rate estimates from the HPM were calculated conditionally on samples of the posterior in which the statistically supported predictor was included in the model (107,773/108,004 samples). Source files for the BEAST analysis and R scripts for conditional effect size and parameter estimation are available from the corresponding author upon request.

In addition to the Bayesian hierarchical hypothesis testing framework described above, we conducted a more traditional GLM analysis of host predictors of mean viral evolutionary rates and a phylogenetic least squares (PGLS) regression analysis. These analyses contained the factors included in the HPM above, but because these methods do not account for uncertainty in the estimation of evolutionary rates (Figure 1), we also included several factors to identify the influence of estimation error: the number of years spanned and the number of sequences that comprised each dataset. For the GLM analysis, an initial model containing all terms was simplified using an exhaustive search of possible models using AICc in the glmulti package of R [41], [42]. Models with Akaike weights within 10% of the highest were retained in the confidence set shown in Table S4. The PGLS regression was conducted in the caper package of R, using the Pagel's λ statistic to account for phylogenetic non-independence of viral evolutionary rates [43]. Because λ for climatic zone was low in the virus phylogeny (λ = 4.5e-5), we had little power to reconstruct ancestral climatic states. This precluded testing whether viral jumps between climatic zones were correlated with directional changes in viral evolution.

Phylogenetic signal in viral substitution rates and host traits

Blomberg's K measures the degree of phylogenetic non-independence of species traits, with values ranging from 0 to infinity [24], [25]. Values of K<1 indicate less phylogenetic signal (more trait lability) than expected under a Brownian motion model of evolution and K>1 indicate more correlation with phylogeny than expected. The K statistic was calculated for the ILM and HPM sets of rate estimates using the topology of the rabies virus phylogeny in Figure 2A in the picante package of R [42], [44]. The statistical significance of estimates of K was tested by comparing the expected distribution of values on our phylogenetic tree to 5,000 randomizations of observed rates along the tips of the tree. A similar analysis using the ML phylogeny of bats estimated from published mitochondrial cytochrome oxidase I sequences assessed whether the ecological and physiological similarity of closely related host species promotes evolution towards similar rates of viral evolution (Table S3, Figure S1).