Genic enrichment in the distribution of derived allele frequency differences

To investigate the influence of recent genetic adaptation in chimpanzee subspecies we compared population differentiation at putatively functional sites (genic sites, defined as +- 2kb from protein-coding genes) to differentiation at non-functional sites (here non-genic). Natural selection can only act on functional sites (although it can affect neutral sites tightly linked to functional sites), so differences between functional and non-functional genomic sites can be ascribed to natural selection. After binning every SNP by its signed difference in derived allele frequency between a pair of subspecies (δ), for each bin of δ we calculated the genic enrichment, defined as the ratio of genic SNPs vs. all SNPs for each bin of δ, normalized by the global genic SNP ratio [10, 11, 29]. This strategy has been deployed in the study of human local adaptation [10, 11, 29], and by not relying on the patterns of linked variation it is not strongly restricted to particular modes of selection. In all subspecies comparisons (Fig 2, orange line) the genic enrichment is greatest for SNPs with the largest δ, and the tail bins of δ exhibit significantly greater genic enrichments than any other bin. While not every genic SNP is in this bin due to positive selection, we expect these SNPs, which show the largest frequency differences between subspecies in the genome, to be strongly enriched in targets of positive selection that rose quickly in frequency in one of the two subspecies [10, 11, 29].

Download:

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

Fig 2. Genic enrichment in bins of signed difference in derived allele frequency (δ).

A, X-axis: δ is computed as the difference in derived allele frequency, for each pair of chimpanzee subspecies. Tail bins (the last bin in either end of δ) contain those SNPs with the largest allele frequency differences. Numbers are of the genic SNPs in each tail bin. Y-axis: genic enrichment in each δ bin (Methods). B, Genic enrichment eastern and central chimpanzee δ, plotted separately due to a different Y-axis limit. NC = Nigeria-Cameroon. The asterisk shows significance of the asymmetry in the genic enrichment (* = 0.01). Shading represents the 95% CI (i.e. alpha = 0.05 for a two-tailed test) estimated by 200kb weighted block jackknife. Grey dashed lines represent simulations under increasing levels of background selection that best match different aspects of the data: lightest to darkest shades: B = 0.93 (excluding δ tail bins), 0.92 (all δ bins), and 0.88 (unmodified genic B values form McVicker et. al. 2009 [34]).

https://doi.org/10.1371/journal.pgen.1008485.g002

The number of tail SNPs and the magnitude of genic enrichment in the tails of δ across subspecies pairs varies in accordance with their N_e and divergence times (Fig 2 and S1 Table). Calculated against western chimpanzees, the subspecies with the lowest long-term N_e [1, 20], the δ tail genic enrichment is the least, ranging from 1.05 to 1.10 (Fig 2A). This indicates that this strategy to identify the targets of local adaptation has reduced power in this subspecies. A greater tail genic enrichment, 1.21 to 1.29, is seen for δ calculated using Nigeria-Cameroon, the species with the second lowest long-term N_e (Fig 2A). This is comparable to the magnitude of the genic enrichment in the tails of δ between human populations (S1 Appendix, S1 Fig; see [10, 11, 29]; the genic enrichment across each bin of δ also resembles those observed in human populations (S1 Appendix, S1 Fig; see [10, 11, 29]. In these comparisons the tail genic enrichment is symmetric (Fig 2A) with symmetry defined as overlapping δ tail bin genic enrichment 95% CIs).

In marked contrast to these symmetric enrichments, we find a distinctive asymmetry between the tail bin genic enrichments of central and eastern chimpanzees (Fig 2B). The central δ tail exhibits a typical genic enrichment (1.19, 95 per cent confidence interval 1.01–1.36) but surprisingly, the eastern δ tail has a much stronger genic enrichment (1.58, 95 per cent confidence interval 1.47–1.68) that is significantly greater than the central tail (P < 0.005; weighted 200kb block jackknife, see Methods) and any other δ tail (all P < 0.0001; weighted 200kb block jackknife).

The large confidence interval of the central chimpanzee δ tail genic enrichment is due to the low number of SNPs in this tail. But we also identified a highly unusual 200kb genomic block on chromosome 3 (chr3:46508626–46708625) that contains 67 highly differentiated alleles between central and eastern chimpanzees, similarly distributed among the two tails (35 genic SNPs in the central tail and 32 genic SNPs in the eastern tail). Concerned that this block could bias our results, we repeated the enrichment analysis after excluding all SNPs contained within it. Removing this block reduces the genic enrichment slightly in the eastern tail (1.56) but substantially in the central tail (1.03) resulting in an even stronger asymmetry between the tails.

To directly quantify the asymmetry of the eastern and central chimpanzee δ tail genic enrichments, we tested if the log₂ ratio of each pair of δ tail bin genic enrichments departs from zero, with the expectation that log₂ ratios would be randomly distributed about zero if the δ tail genic enrichments are symmetric. Not surprisingly, we find a trend for the genic enrichment to be greater for the subspecies with the higher long-term N_e, although the log₂ ratios are similar and small (ranging from 0.03 and 0.07). The only exception is for eastern vs. central, where it is 0.41 (95% CI, 0.17–0.65, 200kb weighted block jackknife). This is six times larger than the highest ratio between other subspecies pairs (Fig 3; S2 Table, Bonferroni corrected p-value δ western vs. central = 0.16 all other Bonferroni corrected p-values < = 0.004, z-test). The eastern vs. central asymmetry in genic enrichment is thus a clear outlier (p-value < 2.2e-16, two-sided Kolmogorov-Smirnov test).

Download:

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

Fig 3. Direct quantification of δ tail bin genic enrichment asymmetry.

The asymmetry of the genic enrichments in the δ tails is measured by taking their log₂ ratio, thus 0 indicates a symmetric enrichment (equal enrichment in both δ tails). NC = Nigeria-Cameroon. Dot = observed asymmetry. Horizontal lines represent confidence intervals estimated by 200kb weighted block jackknife (light = 95%, i.e. alpha = 0.05 for a two-tailed test). Grey vertical marks represent the δ tail asymmetry in simulations, under increasing levels of background selection that best match different aspects of the data: lightest to darkest shades: B = 0.93 (excluding δ tail bins), 0.92 (all δ bins), and 0.88 (unmodified genic B values form McVicker et. al. 2009 [34]).

https://doi.org/10.1371/journal.pgen.1008485.g003

Recombination rate variation does not explain the δ tail asymmetry

Lower average recombination rates at genic versus non-genic sites may influence our analysis [30, 31], and we find a lower average inferred recombination rate at genic than non-genic sites (1.36 cM/Mb versus 1.61 cM/Mb) based on the existing, population-based chimpanzee recombination map [32]. While this could lead to uniform increases in genic differentiation in genic sites, it is difficult to imagine that it would lead to qualitatively differences between subspecies, especially in the closely related eastern and central chimpanzees: while the recombination rate of individual sites may change through time, the global recombination rate at genic and non-genic sites will not differ among them. Indeed, low recombination rate does not drive our signal: recombination rates for both the eastern and central δ tails are not lower than the genic average recombination rate or that recombination rate in the other δ tails (S2 Fig), indicating that outliers in the recombination rate distribution do not contribute disproportionately to the δ tails in this pairwise comparison, in contrast to other pairwise δ tails, (S2 Fig). To further test the effect of varying recombination rates in the results, we repeated the δ tests restricting the analysis to sites with recombination rate < 0.5 cM/Mb. By restricting to this rate range, we equalise the average genic and non-genic recombination rates (0.16 cM/Mb versus 0.15 cM/Mb), and remove the association between δ and recombination rates. Still, all δ tails are significantly enriched for genic sites (S2 Fig). Most importantly, the eastern δ tail genic enrichment is significantly greater than that of central chimpanzees (S2 Fig), showing that differences in recombination rate do not drive the asymmetric genic enrichment observed in this comparison.

Background selection does not explain the δ tail asymmetry

A certain level of δ tail bin genic enrichment (Fig 2) is, in principle, compatible with both recent positive selection and background selection (BGS) [10], the latter because linkage to sites under purifying selection reduces N_e locally in genic regions and increases the effects of random genetic drift [33]. BGS in the range estimated for humans [34] can, for example, explain the δ tail bin genic enrichment in human populations, suggesting that this pattern is not evidence for pervasive recent human adaptation [10, 11, 29]. In this light, we next explored if BGS can explain our observed δ tail bin genic enrichments, and in particular the observed asymmetry between eastern and central chimpanzee δ tail bin genic enrichments.

The strength of BGS can be quantified as a B value, the ratio of diversity at a neutral site linked to sites under purifying selection compared to the expected neutral diversity in the absence of purifying selection. Equivalently, this can be expressed as the ratio of the respective N_e [35, 36]. Previous attempts to simulate the effects of BGS on δ tail bin genic enrichments have simulated a single genome average B value [10, 11, 29], but the genome is heterogenous with regard to both the local density of sites under selection and the relative strength of selection, and thus B is expected to vary across the genome. We thus run coalescent simulations (Methods) with B values sampled from the empirical distribution estimated for the human genome by McVicker et. al. 2009 [34].

The B value that best explains the genic enrichments across all bins of δ (the B value that minimizes the summed square differences between observed and simulated enrichments across all pairwise δ bins, S2 Appendix) is 0.92 –i.e. a reduction of diversity of ~ eight per cent with respect to non-genic regions–decreasing to 0.93 (weaker BGS) when excluding the δ tail bins. We note that the vast majority of genic sites do not fall within the δ tail bins, thus a B of 0.93 explains most of the δ genic enrichment spectrum (Fig 2, lightest dashed line).

When fitting solely the twelve δ tail bins (Methods, S3 Table), 0.92 is still the best fitting B, followed closely by 0.90 with a 3.5% worse fit. This suggests that BGS that reduces diversity up to 10 per cent is enough to explain the observed δ tail bin genic enrichments. We note that simulations of BGS using a single genome average B value for all sites shows congruent results, but it requires much stronger BGS (B = 0.86) to match the 12 twelve δ tail bin genic enrichments (S2 Appendix, S2 Fig).

Previously, it was shown that BGS alone does not produce δ tail bin genic enrichment asymmetries in comparisons of human populations [11]. We also find that BGS does not result in significant eastern vs. central δ tail bin genic enrichment asymmetry. Simulations show a slight asymmetry in the tail genic enrichment (Figs 2B and 3) due to differences in their demographic histories (S2 Appendix, S4 Table). Nevertheless, no simulated value of B in the range (0.93–0.88) results in a tail genic enrichment log₂ ratio that falls within the 95% CI of the observed ratio (Fig 3). Further, we could not find a B value that results in a genic enrichment that lies within the 95% CI for both eastern and central chimpanzees. In contrast, the small (though statistically significant) asymmetries in all other pairwise δ tail bin genic enrichments are observed in simulations and thus fully explicable by demography and BGS (Fig 3). We note that the unmodified McVicker B values (which have an average B of 0.88) could explain the observed eastern δ tail bin genic enrichment, but they provide an extremely poor fit to all other δ bins (Fig 2)–including the central δ tail. We hypothesise that this is due to the B inference ignoring the effects of positive selection, which likely have a pronounced effect on the variance of B.

Previous work has shown that background selection varies little among the great apes [18]. Theory suggests that the diversity-reducing effect of BGS is independent of N_e, being determined by the distribution of fitness effects (s), except for the narrow range of N_e * s = 1 (Nam et al., 2017), while previous work suggests that more than 80% of deleterious mutations in chimpanzees have N_es » 1 [37] Thus, the expectation is that the diversity reducing effect of BGS should be the same across all four chimpanzee subspecies. Indeed, we find comparable effects of background selection across subspecies: the relative reduction in neutral variation linked to genes is comparable amongst chimpanzee subspecies (S4A Fig), and neutral diversity has similar dependency on recombination rate and density of functional features across subspecies (with the exception of western chimpanzees, S4B Fig). Further, using a population genetic statistical model [38] we estimate the same reduction in neutral diversity due to background selection in each chimpanzee subspecies, at 11%, in the highest likelihood model (S3 Appendix, S5 Table). Thus, despite their differing demographic histories (Fig 1), the effects of BGS are very similar across each chimpanzee subspecies. This justifies using the same average strength of BGS across subspecies above. Nevertheless, to explore if our conclusions are robust to this assumption, we also modelled a greater strength of BGS in eastern chimpanzees (B = 0.825, the value which best matches the eastern δ tail bin genic enrichment) than in the other subspecies (B range 0.900–0.850), using a single genome wide average B value for each scenario. Stronger BGS in eastern chimpanzees does not produce an eastern central δ tail bin asymmetry as large as that observed in the genomes (log2 ratio range 0.120–0.146), further illustrating that BGS cannot explain the greater tail genic enrichment in eastern chimpanzees (S5 Fig). Rather, this is most likely a signal of recent adaptation.

Population-specific branch lengths with PBSnj

Pairwise comparisons cannot determine which subspecies has changed. Direction, and therefore biological meaning, to allele frequency difference can only be garnered by assuming that derived alleles most often provide the basis for new adaptations. This approach is also limited by the collapsing of the shared history of lineages. For example, in the Nigeria-Cameroon vs. eastern comparison, 22% of the SNPs in the eastern δ tail are also in the central δ tail (for Nigeria-Cameroon vs. central comparison), whereas only 3.5% are highly differentiated to both Nigeria and central chimpanzees. Thus, δ summarises the allele frequency change across several parts of the phylogeny, limiting the biological interpretation of its tails.

To overcome this limitation, we developed a statistic that extends the widely used Population Branch Statistic (PBS) [8]. Briefly, large PBS values identify targets of positive selection as SNPs with population-specific allele frequency differentiation, as these sites result in unusually long branch lengths in pairwise F_ST-distance trees between three taxa. Small PBS values are due to very short branches, for example due to purifying, shared balancing selection or rare mutations. We extend this test to more than three taxa in the novel PBSnj statistic by applying the Neighbor-Joining (NJ) algorithm on the matrix of the per-SNP pairwise F_ST distances of the four subspecies (Methods, S4 Appendix). This way, PBSnj allows us to jointly compare the four subspecies and identify SNPs with very long branches (allele frequency differentiation) in one subspecies only. Additional advantages of PBSnj are that it does not rely on the specification of ancestral or derived states, and that the NJ algorithm does not require specification of a phylogenetic tree describing the relationship amongst taxa (S4 Appendix).

PBSnj allows us to determine within which lineage, eastern or central chimpanzees, allele frequencies have changed to result in the asymmetric δ genic enrichment. Analogous to the δ tail bins, we binned PBSnj scores and calculated the genic enrichment for each species PBSnj tail (Fig 4A). The PBSnj eastern tail has significantly stronger genic enrichment than the central tail (eastern: 1.35, central: 1.13, log₂ ratio = 0.27, p < 0.0005 estimated from weighted 200kb block jackknife, Fig 4B). This shows that the central vs. eastern asymmetry in the δ tail bin genic enrichments (Figs 2B and 3) is due to the drastic allele frequency rise of genic SNPs in eastern chimpanzees since their divergence with central chimpanzees.

Download:

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

Fig 4.

Genic enrichment in bins of PBSnj in eastern and central chimpanzees. A X-axes: PBS scaled to take values in the range 0–1. Y-axes: Genic enrichment computed as described in Fig 2. Shading represents the 95% CI (i.e. alpha = 0.05 for a two-tailed test) estimated by 200kb weighted block jackknife. B: log₂ ratio of the eastern and central PBSnj tail (PBS > = 0.8) genic enrichment. A,B Grey dashed (A) or vertical (B) lines represent the PBSnj genic enrichment in simulations, under increasing levels of background selection that best match different aspects of δ, as described in Figs 2 and 3: lightest to darkest shades: B = 0.93 (excluding δ tail bins), 0.92 (all δ bins), and 0.88 (unmodified genic B values form McVicker).

https://doi.org/10.1371/journal.pgen.1008485.g004

Again, recombination rate variation does not explain the observed eastern PBSnj enrichment, as it persists even after comparing genic and non-genic sites with similar recombination rates as described above (S6 Fig). Importantly, across the range of B values (0.88–0.93), simulations show that eastern and central chimpanzee PBSnj tail genic enrichments are expected to be equal or tend to be higher for central than eastern chimpanzees (Fig 4B). In fact, BGS would need to be much stronger in eastern chimpanzees than in central chimpanzees to produce the observed levels of PBSnj tail genic enrichments. BGS with B < 0.93 would be required to produce the genic enrichment exhibited in the eastern PBSnj tail, but B > = 0.93 produces PBSnj tail genic enrichments of equal or greater magnitude as those seen for central chimpanzees, and it provides a much better fit to the data in the rest of the PBSnj distribution for both species (Fig 4A). This result is replicated in simulations with a single average B value for all sites, (S5 Appendix and S6 Appendix, and S6 Table). Thus, we observe strong evidence of positive selection for eastern chimpanzees: they exhibit the greatest genic enrichment for highly differentiated SNPs, an enrichment that (unlike in other subspecies) we cannot explain by demography and background selection alone. By using the genomic blocks used to estimate the PBSnj tail Confidence Intervals in Fig 4A, we estimate that an additional 12–23 population specific sweeps are sufficient to explain this signature (Methods, S7 Fig). Although this is a conservative estimate, it shows that we do not require an unrealistically large number of selective sweeps to explain the distinct pattern of eastern chimpanzees.

PBSnj eastern tail SNPs have regulatory functions

While less than 1% of PBSnj eastern tail SNPs result in an amino acid change, compared to the genic background, PBSnj tail genic SNPs are significantly enriched in non-synonymous variants (PBSnj tail proportion amino acid change = 0.84%; genic background proportion = 0.06%, permutation test p = 0.001, S7 Table lists PBSnj eastern tail non-synonymous SNPs). This is not an indication that non-synonymous sites are especially important for local adaptation, but rather reflects an enrichment of PBSnj SNPs in exonic (PBSnj tail proportion = 1.84%; genic background proportion = 1.46%, permutation test p = 0.046) and CDS (PBSnj tail proportion = 3.97%; genic background proportion = 2.87%, permutation test p = 0.001) SNPs, and the observed proportion of exonic SNPs that are non-synonymous matches that of the genic background (PBSnj tail proportion 0.45, genic background proportion = 0.44, permutation test p = 0.53).

This puts the focus on regulatory changes. We used regulomeDB [39] to predict putative regulatory consequences of chimpanzee SNPs from the sequence context and biochemical signatures of homologous human sites. The PBSnj eastern tail genic SNPs are more likely to have strong evidence of regulatory function (3.7% vs. 3.0%, permutation test p = 0.012) and less likely to have no ascribed regulatory function (52.3% vs. 56.0%, permutation test p = 0.0001) than randomly sampled genic SNPs, S8 Table. In contrast, PBSnj central tail SNPs show no difference to the genic background for either category (S8 Table; Nigeria-Cameroon and western also exhibit weaker but significant enrichments). Interestingly, PBSnj eastern tail SNPs do not differ in functional constraint (as measured by phastCons scores [40], see Methods) from random genic SNPs (S9 Table). This suggests that while likely enriched in regulatory functions, these sites are not under particularly strong long-term constraint, perhaps because they do not affect functions that have been tightly conserved over long evolutionary times.

Potential biological functions of the PBSnj eastern tail SNPs

To understand the biological mechanisms and putative selective factors driving the recent adaptations in eastern chimpanzees, we investigated the genes containing the genic SNPs in the PBSnj eastern tail (hereafter PBSnj eastern genes). As there are very few studies of the functional effects of chimpanzee genetic variation, we are reliant on homology and conservation of function to other primate species. We therefore tested three different gene sets that are potentially biologically informative: Human Gene Ontology (GO) annotations [41, 42]; human Viral Interacting Proteins (VIPs) [4]; and genes with specific gene expression changes in the vervet monkey in response to SIV infection [43, 44].

Seven GO categories are significantly enriched (all p < = 0.00026, False Discovery Rate (FDR) < 0.1; GOWINDA; S10–S13 Tables), with a striking preponderance of immunity-related GO categories and genes involved in anti-viral activity. The top category is “cytoplasmic mRNA processing body assembly”, and three of the five PBSnj eastern genes in this category (DDX6 [45], ATXN2 [46] and DYNC1H1 [47]) are either key components of processing bodies (P-bodies) or regulate the assembly or growth of P-bodies in response to stress. Selection on the immune system is suggested also by the second category, “antigen processing and presentation of peptide antigen via MHC class I”. The signal in this category is due to six genes, of which only HLA-A is an MHC gene, with the other genes being B2M, ERAP1, PDI3, SEC13, and SEC24B. There are three more significant categories related with immunity: “T cell co-stimulation”, “negative regulation of complement-dependent cytotoxicity”, and “type I interferon signalling pathway”. Even the “cytoplasmic mRNA processing body assembly” category is potentially linked to virus infection as P-bodies are cytoplasmic RNA granules manipulated by viruses to promote viral survival and achieve infection [48, 49]. The enrichment in immune categories and virus-related genes is in perfect agreement with the PBSnj eastern genes being also enriched in three sets of VIPs [4]–genes with no annotated immune functions but that interact with viruses. Specifically, the enriched VIPs are for Dengue virus, Bovine leukaemia virus and human T‑lymphotropic virus (p < 0.01, FDR < 0.1, see Table 1 and S14–S17 Tables), which are genes with no annotated immune functions but that interact with viruses.

Download:

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

Table 1. VIP gene enrichment in the PBSnj eastern tail.

https://doi.org/10.1371/journal.pgen.1008485.t001

The genes from these VIP sets are not members of the five significantly enriched immunity related GO categories, and VIP sets generally do not contain immunity genes. This provides an independent signal for the relevance of viruses to PBSnj eastern genes. Together, these results suggest that adaptation to pathogens, and viruses in particular, may have had an important role in the recent adaptation in eastern chimpanzees.

In light of the suggestive evidence for virus-related adaptation, our attention was drawn to the simian immunodeficiency virus. Amongst chimpanzee viruses, SIV is intensively studied as it is the progenitor of the human immunodeficiency virus (HIV) that created the global acquired immune deficiency syndrome (AIDS) pandemic. It is also of interest here because it appears to only infect natural populations of eastern and central chimpanzees [50–53], and because it has mediated fast, recent adaptations in other natural hosts [54]. Specifically, Svardal et. al. (2017) investigated a set of genes that change expression in response to SIV infection in SIV natural hosts (vervet monkeys) but not in non-natural hosts that develop immunodeficiency (macaques) [43, 44], hereafter referred to as “natural host SIV responsive genes”. Natural host SIV responsive genes are likely involved in the specific immune response of natural hosts to SIV infection, which limits the effects of the virus and prevents subsequent immunodeficiency. These genes also show signatures of positive selection in vervet monkeys, suggesting that ongoing adaptation to the virus occurs even in natural hosts [54]. These observations also reveal natural host SIV responsive genes as a highly relevant candidate gene set for genes under positive selection in SIV infected species. Strikingly, the PBSnj eastern tail SNPs, but none of the other subspecies’ PBSnj tails, are significantly enriched in these same natural host SIV responsive genes [43, 44] (observed 118 genes, expected 100, p-value = 0.0195, GOWINDA, FDR = 0.064 see Methods, Table 2, S18 Table).

Download:

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

Table 2. SIV responsive gene enrichment in subspecies PBSnj tails.

https://doi.org/10.1371/journal.pgen.1008485.t002

In fact, the set of natural host SIV responsive genes can fully explain the unique eastern signature: the asymmetry in the PBSnj tail is abolished when this set of genes is removed from the analysis–the genic enrichment in the eastern PBSnj tail decreases from 1.35 to 1.26, and the 95% confidence interval of this point estimate now overlaps those of Nigeria-Cameroon and central chimpanzees (Methods). A reduction in the genic enrichment in the PBSnj tail is expected, as it is enriched in natural host SIV responsive genes; but this exercise allows us to show that in the absence of selection in natural host SIV responsive genes, the signature of recent positive selection in eastern chimpanzees would not be exceptional.

The natural host response in vervet monkeys is associated with changes in the expression of these natural host SIV responsive genes. In agreement with potential adaptations in gene expression, the set of PBSnjE SNPs in the natural host SIV responsive genes are further enriched in sites with putative gene regulatory function (p = 0.0485 when compared with other PBSnj eastern tail genic SNPs, p = 0.0089 with all genic SNPs) and strongly depleted of sites with no predicted regulatory function (p = 0.0001 when compared with other PBSnj eastern tail genic SNPs, p = 0.0001 with all genic SNPs, S19 Table).

While these genes were not identified in chimpanzees where SIV experiments are naturally severely limited, this suggests a potentially similar mechanism of adaptation to SIV (or to an unknown virus with a similar effect in gene expression) in vervet monkeys and chimpanzees.

Biological functions of the PBSnj central tail SNPs

Despite having a larger long-term N_e than eastern chimpanzees, central chimpanzees do not show a clear genomic signature of recent adaptation. Despite being naturally infected by SIV and being the source of pandemic HIV, they show no clear indication of selection in SIV responsive genes: the PBSnj central tail has a greater number of SNPs in SIV responsive genes than expected (36 vs. 29), but the enrichment is non-significant (p = 0.0763; resampling test, Table 1). Power to identify a significant enrichment is hampered by the low number of SNPs. However, highly differentiated SNPs in the PBSnj long branches of central chimpanzees are significantly enriched in one GO category, “chemokine receptor activity”, due to SNPs in CCR3, CCR9 and CXCR6 (p = 0.00001, FDR < 0.02, GOWINDA). Each of these genes is located within the large cluster of cytokine receptor genes on chromosome 3, but they appear to be associated with different sweep events (S8 Fig). Chemokine receptors facilitate the response to chemokine signalling, and their functions are important for both inflammatory and immunity responses [55]. For instance, CCR3 is involved in the recruitment of eosinophils in the airway [56] and CCR9 is implicated in T-cell development [57].

These genes are of also interest because the human paralogs CCR5 and CXCR4 are the two most common co-receptors for HIV-1 cell entry [58, 59]. Both CCR3 and CXCR6 can be used to enter the cell by some SIV, HIV-1 and HIV-2 subtypes [60–63], and the SIV of both sooty mangabey [64] and vervet monkey [65] use CXCR6. Although some results suggest that chimpanzee SIV cannot use chimpanzee CXCR6 [66], the full breadth of co-receptors used by SIV in chimpanzees is still being investigated. For instance, it is unknown how important coding sequence variation is to co-receptor function [66]. We note that one of the PBSnj tail SNPs in CCR3 results in an amino acid substitution (246 S/A) in transmembrane domain 6, and the paralogous region has been implicated in the modulation of CCR5 activity [67]. Thus, changes in these co-receptors may have the potential to affect the entry of SIV in chimpanzee cells.

Genotypes, haplotypes and genic regions

We analysed the 58 chimpanzee genomes described in de Manuel et. al. (2016), with sample sizes of: eastern 19, central 18, Nigeria-Cameroon 10, western 11 after excluding the hybrid Donald. For most tests based on allele frequencies, we used the chimpanzee VCF file from de Manuel et al., (2016) after removing every SNP with at least one missing genotype across all chimpanzees. For haplotype phasing, we also included the 10 bonobo genomes from [20]. To statistically phase haplotypes we used Beagle [79] v 4.1 (downloaded from https://faculty.washington.edu/browning/beagle/b4_1.html, May 2016). We used default parameters without imputation, except that after the initial 10 burn in iterations we performed 15 phasing iterations (default is five) using the following command line: java -Xmx12000m -jar beagle.03May16.862.jar gt = vcf out = vcf.phased impute = false nthreads = 1 niterations = 15

For the analysis of δ we chose to use the inferred ancestral chimpanzee allele from EPO alignments from ensembl. For comparison, and to show that our result was robust to AA inference method we also used the homologous human genome reference allele as the ancestral state for chimpanzee SNPs. We used the human genome from the 1000 Genomes project phase III human_g1k_v37.fasta, available from: http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/technical/reference/human_g1k_v37.fasta.gz

We used the UCSC liftover utility to convert chimpanzee SNPs’ coordinates from pantro 2.1.4 to human genome version 37 (hg19) coordinates, then used samtools faidx to retrieve the human allele for that position. Both of these inference methods recovered the same signal of a significantly greater δ tail bin genic enrichment for eastern vs. central chimpanzees, see S22 Table. Again, we also note that our new statistic PBSnj does not require inference of the ancestral allele. In total we analysed 29,778,147 autosomal SNPs. Of these there was EPO high confidence inferred ancestral states for 27,081,963 SNPs and human inferred ancestral states for 28,776,489 SNPs.

We considered protein-coding genes on the autosomes (17,530 genes) and define ‘genic sites’ by extending the transcription start and end coordinates from ENSEMBL biobank for pantro2.1.4 by 2kb on each side.

Genetic map

For statistics that required a genetic map, we used the pan diversity genetic map [32] inferred from 10 western chimpanzees. We downloaded the chimp_Dec8_Haplotypes_Mar1_chr-cleaned.txt files from birch.well.ox.ac.uk/panMap/haplotypes/genetic_map. These files consist of SNPs and their inferred local recombination rate. These map data were inferred from sequences aligned to the pantro2.1.2 genome, so we used two successive liftover steps to convert the coordinates of sites used to infer the genetic map to pantro2.1.4 coordinates: pantro2.1.2 to pantro2.1.3, then pantro2.1.3 to pantro2.1.4. Two steps are required as there are no liftover chains relating pantro2.1.2 to pantro2.1.4. Of the 5,323,278 autosomal markers, 33,263 were not lifted from pantro2.1.2 to pantro2.1.3. The remaining 5,290,015 were also successfully converted to pantro2.1.4 coordinates. After liftover we filtered sites that after the two steps were mapped to unassigned scaffolds or the X chromosome, which left 5,289,844 SNPs. Next, we sorted loci by position to correct cases where their relative order was scrambled. This left a final number of 5,289,460 autosomal SNPs. Recombination rates were then recalculated by linear interpolation between consecutive markers (marker x, marker y) using the average of their estimated recombination rates (rate x, rate y).

Signed difference in derived allele frequency (δ)

Using the derived allele frequency of each SNP for each subspecies we calculated, for each pair of chimpanzee subspecies, the signed difference in derived allele frequency (DAF) between them: δ = DAF_pop1 –DAF_pop2; DAF_pop1 > DAF_pop2: δ > 0; DAF_pop1 < DAF_pop2: δ < 0; -1 < = δ < = 1. We bin δ into 10 bins of 0.2. The choice of subspecies assigned to pop1 or pop2 is arbitrary and has no bearing on the results. To ensure that both tail bins are identically wide, we define them as Bin 1: -1 < = δ < = 0.8 and Bin 10 as 0.79 < δ < = 1. As a consequence, the Bin 5 (0.00 < δ < 0.2) is marginally narrower than the other bins (by 0.01), but it contains a large number of sites and the slight size difference has negligible impact on the analyses.

We estimate confidence intervals and infer p-values for δ genic enrichment using a weighted block jackknife [80] utilising the method of Busing et. al. [81]. This has been used for analogous tests, as it accounts for linkage disequilibrium, which means that SNPs in δ bins are not full independent of each other. We divide the genome into non-overlapping 200kb windows to capture the blocking effect of LD. We then recalculate, for each bin, the genic enrichment using a delete-1 window jackknife. We also weight the windows by the total number of SNPs in them, to downweigh, within each bin, blocks with large numbers of linked SNPs. We determine that two tails are differentially enriched if their 95% CIs of enrichment do not overlap. For directly testing asymmetry (or in the case of PBSnj, equality) using the log₂ ratio, we use the same weighted block jackknife, and use the 95% CI as a two-tailed test with alpha = 0.05. Other enrichment and resampling tests are described in Methods subsection “Statistics”.

Population Branch Statistic neighbour-joining

The Population Branch Statistic (PBS), [8] is a test of population specific natural selection. In the framework of a three-taxon distance tree, SNPs under selection specific to one population are detected as those that result in longer than expected branch lengths (large allele frequency differentiation). To generate the tree, for each site, the full distance matrix of pairwise F_ST is computed. A three taxa tree is unrooted and has only one possible topology, so simple algebra allows the calculation of each branch length in the tree. Extreme outliers in the distribution of PBS are considered candidates of positive selection.

We introduce Population Branch Statistic neighbour-joining (PBSnj) as a simple method to calculate population specific branch lengths when more than three taxa are being analysed. We note that related Methods have recently appeared in the literature [82, 83]. Full details are in S4 Appendix, but in brief, using the full matrix of pairwise F_ST, F_ST values are transformed to units of drift time as ln (1-F_ST) [8]. For fixed differences this transformation is mathematically undefined i.e. ln (0), and F_ST = 1 is replaced with the next largest observed F_ST value for a given population pair. Then the Neighbor-Joining algorithm [84] is used to infer the tree topology and calculate branch lengths. This overcomes errors in the inferred length of external branches due to misspecification of a fixed tree topology. To enable a binning scheme of PBSnj values that is comparable between subspecies, these scores are further normalised to be on the 0–1 scale.

F_ST for PBSnj was calculated using the estimator described in [85] because there are unequal sample sizes for the subspecies, and the classical Weir and Cockerham estimator can be biased with unequal sample sizes [85]. To calculate genic enrichments along the PBSnj distribution we bin SNPs in PBSnj bins 0.2 units wide. As for δ analyses, we use the 200 kb weighted block jack-knife to estimate confidence and significance levels. We provide a source code file, written in R, to calculate PBSnj (“PBSnj_function.R”), see Data availability.

Model of Chimpanzee demographic history

The most detailed exploration of chimpanzee demography comes from the work of de Manuel et. al. (2016). This paper describes the 58 chimpanzee full genome sequences we use here, and estimation of their inferred demographic model. As this paper took a primary interest in investigating chimpanzee-bonobo post speciation gene flow, and to reduce the number of parameters to be estimated, models were inferred using either Nigeria-Cameroon or western chimpanzees, but not both. Thus, de Manuel et. al. (2016) provides “bonobo, eastern, central, Nigeria-Cameroon” and “bonobo, eastern, central, western” models. These are referred to, respectively, as ‘becn’ and ‘becw’ models below.

For this investigation we use a merged demographic history. To begin the construction of this model, we recognised that there is little gene flow involving western chimpanzees in the ‘becw’ model, but that gene flow events are a key determinant of patterns of chimpanzee genetic diversity and differentiation in the ‘becn’ model. We therefore used the ‘becn’ model as a scaffold to which parameters relating to western chimpanzees (bottlenecks, expansions and N_e estimates) from the ‘becw’ model are “grafted” in, to create a merged ‘becnw’ model. To make sure that the N_e of western chimpanzees was appropriately scaled, all N_es 1000 ya pastwards for western chimpanzees specified in the ‘becw’ model were normalized by multiplying by the ratio of the inferred N_es of central chimpanzees specified from 1000 ya pastwards in the ‘becn’ and ‘becw’ models: scaled western N_e = western N_e * 3.66914400056 / 4.3158739382. Present western N_e was normalised by the ratio of the present central N_e: scaled western N_e = western N_e * 0.3092 / 0.30865.

Initially, we used the split time of the western and Nigeria-Cameroon lineages of 250ky reported by de Manuel et. al. which was estimated from sequence divergence data, but this gave a bad fit to F_ST values, being substantially lower than observed (S23 Table). We addressed this by increasing the western/Nigeria-Cameroon divergence time in proportion to the ratio of model:observed western/Nigeria-Cameroon F_ST. i.e. F_STObserved / F_STModel = timeX / 250kya = > timeX = F_STObserved / F_STModel x 250kya. We adjust the observed F_ST by -0.008 –to capture the average difference between model versus observed F_ST values for central/eastern/Nigeria-Cameroon chimpanzees. This simple calculation results in an adjusted time of 267kya for the western/Nigeria-Cameroon split. F_ST values for this new model show a much better fit to observed values (S23 Table), and it is this model that we use for all subsequent modelling of genic enrichments and the effects of background selection.

To determine model fit above, we calculated all pairwise average F_ST values for the simulated data and compared them to the empirical F_ST estimates. For each scenario, we simulated 1,000,000 2kb fragments (2 Gb of sequence).

All simulations of neutral diversity and background selection were performed with msms [86], and following de Manuel et. al. assuming a mutation rate of 1.2e^-8 and recombination rate 0.96e^-8, with the following command line:

msms 116 1 -t 0.96048 -r 0.768384 2001 -I 5 0 38 36 20 22 0 -n 1 0.0742 -n 2 0.3181 -n 3 0.3092 -n 4 0.0386 -n 5 0.08114434 -m 1 2 0 -m 1 3 0 -m 1 4 0 -m 2 1 0 -m 2 3 1.8181960943074 -m 2 4 0 -m 3 1 0 -m 3 2 2.02290154800773 -m 3 4 0 -m 4 1 0 -m 4 2 0 -m 4 3 0 -m 5 1 0 -m 1 5 0 -m 5 2 0 -m 2 5 0 -m 5 3 0 -m 3 5 0 -m 4 5 0 -m 5 4 0 -en 0.001 1 1.83290809268 -en 0.001 2 1.161030985567 -en 0.001 3 3.66914400056 -en 0.001 4 1.23640124358 -en 0.001 5 0.9132505 -em 0.020875 1 2 0 -em 0.020875 1 3 0 -em 0.020875 1 4 0 -em 0.020875 2 1 0 -em 0.020875 2 3 1.8181960943074 -em 0.020875 2 4 1.12888460726286 -em 0.020875 3 1 0 -em 0.020875 3 2 2.02290154800773 -em 0.020875 3 4 0.514005225416364 -em 0.020875 4 1 0 -em 0.020875 4 2 0.61034918826118 -em 0.020875 4 3 2.77081002950074 -em 0.042025 1 2 0 -em 0.042025 1 3 0.0447270935214584 -em 0.042025 1 4 0.00204350937063846 -em 0.042025 2 1 0 -em 0.042025 2 3 1.8181960943074 -em 0.042025 2 4 1.12888460726286 -em 0.042025 3 1 0.0340892941439601 -em 0.042025 3 2 2.02290154800773 -em 0.042025 3 4 0.514005225416364 -em 0.042025 4 1 0.00878072013784504 -em 0.042025 4 2 0.61034918826118 -em 0.042025 4 3 2.77081002950074 -en 0.104325 2 0.0402577179646081 -en 0.104325 3 0.192594746352967 -en 0.106325 3 8.73162876459514 -ej 0.106325 2 3 -em 0.106325 1 2 0 -em 0.106325 1 3 0.0177338314347154 -em 0.106325 1 4 0.00204350937063846 -em 0.106325 2 1 0 -em 0.106325 2 3 0 -em 0.106325 2 4 0 -em 0.106325 3 1 0.00723425109237692 -em 0.106325 3 2 0 -em 0.106325 3 4 0.193855714034029 -em 0.106325 4 1 0.00878072013784504 -em 0.106325 4 2 0 -em 0.106325 4 3 0.00771007640703268 -en 0.21195 5 0.1223036 -en 0.214175 5 0.194964 -en 0.267475 4 1.23640124358 -en 0.267475 5 0.194964 -ej 0.2675 5 4 -en 0.41955 1 0.158405393915496 -en 0.42155 1 0.299481445247702 -en 0.473075 4 0.0306317427630759 -en 0.475075 4 2.79429564470655 -en 0.480625 4 0.0872103733618782 -em 0.480625 1 2 0 -em 0.480625 1 3 0.0177338314347154 -em 0.480625 1 4 0.00204350937063846 -em 0.480625 2 1 0 -em 0.480625 2 3 0 -em 0.480625 2 4 0 -em 0.480625 3 1 0.00723425109237692 -em 0.480625 3 2 0 -em 0.480625 3 4 0.193855714034029 -em 0.480625 4 1 0.00878072013784504 -em 0.480625 4 2 0 -em 0.480625 4 3 0.00771007640703268 -en 0.482625 3 1.66920782430592 -ej 0.482625 4 3 -em 0.482625 1 2 0 -em 0.482625 1 3 0.241282075772286 -em 0.482625 1 4 0 -em 0.482625 2 1 0 -em 0.482625 2 3 0 -em 0.482625 2 4 0 -em 0.482625 3 1 0.0101771164248256 -em 0.482625 3 2 0 -em 0.482625 3 4 0 -em 0.482625 4 1 0 -em 0.482625 4 2 0 -em 0.482625 4 3 0 -en 1.5988 3 0.00336130452736601 -en 1.6008 3 1.47105091660349 -ej 1.6008 1 3 -em 1.6008 1 2 0 -em 1.6008 1 3 0 -em 1.6008 1 4 0 -em 1.6008 2 1 0 -em 1.6008 2 3 0 -em 1.6008 2 4 0 -em 1.6008 3 1 0 -em 1.6008 3 2 0 -em 1.6008 3 4 0 -em 1.6008 4 1 0 -em 1.6008 4 2 0 -em 1.6008 4 3 0

As a further assessment of the fit of the model, we plotted the observed and simulated site frequency spectrum (SFS), S9 Fig. In general, the model fit is good, being poorest for singletons (too high) and high frequency derived sites (too low). This is likely due to effects of selection on the genome, which is not incorporated into the neutral demographic model. We note too, that this model was computed using only the allele counts from regions of the genome under weak/no selection as inferred from GERP scores, further explaining the reduced fit at these two site classes.

Simulations of chimpanzee genetic data under neutrality and background selection.

We used msms to perform coalescent simulations of chimpanzee demography. To simulate the effects of background selection (BGS) we modified the estimates of effective population size (N_e) from the demographic model by multiplying them by a scaling factor, which represents the B score or effective reduction in N_e due to BGS. 0.8, for example, reduces the N_e and hence expected neutral diversity to 80% the level seen for neutral sites unlinked to regions under purifying selection [11]. To capture the possible variance in B across the genome, we sampled from the estimates of B for the human genome inferred by McVicker et. al. [34] downloaded from: http://www.phrap.org/software_dir/mcvicker_dir/bkgd.tar.gz. The coordinates for this file are for human reference genome assembly version hg18, and we used the coordinates of all annotated autosomal protein coding genes for hg18, +- 2kb, downloaded from ensembl.

McVicker et. al. estimated that autosomal diversity levels were reduced by ~ 20 per cent [34]. The mean B for genic regions is 0.75 and for non-genic regions B is 0.85, resulting in an average ratio (or effective genic B) of ~ 0.88, implying that diversity in genic regions should be reduced 12 per cent compared to non-genic regions. To shift this average B and explore the fit to the genomic data, we added a fixed constant to all genic B values, so that we could simulate BGS with different genome average B in the range 0.88–0.93, in 0.01 increments, in order to ascertain which average B best fit the observed δ genic enrichments. B scores were constrained to a maximum value of one, see S8 Fig for the distribution of non-genic and genic B scores. For non-genic regions and for each average B, we simulated 10 million 2.0 kb loci. After processing and calculating allele frequencies, we performed δ and PBSnj genic enrichments as described previously. To estimate a BGS strength that best matched the observed δ genic enrichments, we performed a simple sum of squared differences, summed for each δ genic enrichment bin for each pairwise comparison.

We also report results from simulations utilising a single genome average B value for all genic sites. We simulated non-genic regions with B = 1, and genic regions with different chosen strengths of BGS. We used B in the range 1–0.8, incremented by 0.025, with additional 0.0125 increments between 0.9–0.85. For neutral regions and for each B we simulated 25 million 2.0 kb loci. These serve a useful comparison to the simulations generated using McVicker derived B values.

Estimating the number of extra eastern chimpanzee adaptive events

We use the structure of the block jack-knife to estimate the number of adaptive events that are needed to result in the PBSnj eastern tail genic enrichment being greater than that of central chimpanzees or generated by BGS. Recall that to estimate the error variance on the genic enrichment in each bin of PBSnj, we divided the genome into non-overlapping 200 kb blocks. For each block we have the number of genic and non-genic SNPs per bin of the PBSnj distribution. For eastern chimps, there are 3528 genic SNPs contained within 832 blocks (i.e. 166 MB) in the PBSnj rhs tail i.e. with a PBSnj scaled length > = 0.8. NB: the majority of SNPs in this bin are not fixed, so we are not assuming that these are completed sweeps.

Of these, there are 448 blocks containing only 1 SNP i.e. 54% of blocks, 81 blocks with 10 or more outlier genic SNPs. i.e. 10% of blocks, with a block maximum count of 117 genic SNPs (S7A Fig).

We rank blocks by the number of genic SNPs that are outliers. Iterating over this sorted list we remove blocks and recalculate the enrichment for genic SNPs. We define matching as the number of iterations required to reduce the tail bin genic enrichment to below a magnitude less than the upper 95% confidence limit or the point estimate of the central PBSnj tail genic enrichment. We chose to order by the number of eastern tail genic SNPs as this results in a monotonically decreasing genic enrichment with each block being removed.

Measures of conservation and effects on gene regulation

We used phastCons [40] to infer highly conserved sites. We used the 20 mammalian multiz alignment to the human genome hg38, downloaded from UCSC (http://hgdownload.cse.ucsc.edu/goldenPath/hg38/multiz20way/maf/). To reduce the chance that polymorphism in chimpanzees affects inference of conservation, we removed both the chimp and bonobo reference genomes form these alignments. We estimated the phylogenetic models from fourfold degenerate (non-conserved model) and codon first position sites (conserved model). We then predicted base conservation scores and conserved fragments using the following options:—target-coverage 0.25—expected-length 30. Resultant conserved elements covered 69.24% of the human exome, or an enrichment of 17.27. We note that although we attempted to remove the Pan branch from our estimates, it is impossible to completely avoid the use of these genomes, for example, when converting predicted conserved elements from hg38 to pantro2.1.4. These results have been deposited on Dryad (see Data availability).

We used regulomeDB [39] to identify putatively regulatory role of genomic sites. Due to the close phylogenetical relationship between chimpanzees and humans, we argue that in lieu of any functional data for chimpanzees, inferred function from homologous positions in the human genome is a useful proxy for function in the chimpanzee genome. To obtain regulomeDB information for variable chimpanzee positions we used liftover to map SNP coordinates from pantro2.1.4 to hg19, keeping positions that reciprocally mapped to homologous chromosomes. Alan Boyle then kindly provided regulomeDB annotations for these positions. In regulomeDB, lower scores reflect higher confidence in regulatory function. We modified scores on the basis that scores 1a-f are given for positions that are human eQTLs, which we do not use as they refer to the specific allele change in humans rather than to the function of the site. Without eQTLS, scores 1a-c and 2a-c reflect the same biochemical signatures and location within transcription factor motifs. Thus, we combine these scores in to a new “high confident” regulatory function category. Our “non-regulatory” category includes positions with regulomeDB scores of 6 or 7, which have no evidence of being regulatory. We did not use sites with intermediate scores.

Gene set enrichment analyses

We used GOWINDA [87] to test for enrichments in Gene Ontology (GO) categories, which corrects for clustering and gene length biases. We used either GO categories or custom gene lists as candidate gene sets. GO categories for humans were obtained from the GO consortium [41, 42], while gene sets were manually created from published sets of Viral Interaction Proteins [4] and a set of genes that are differentially expressed in CD4 cells after SIV infection in the natural SIV host vervet monkey but not in that non-natural host macaque [43, 44, 54].

GOWINDA has an input file format which enables flexible usage of nonstandard gene sets. Genes are defined in a gtf file. We created a gtf from the ENSMBLE gene definitions, but restricted these to genes with clear 1–1 orthologs with humans. Our gtf file contained 16,198 of 17,530 protein coding genes. This gene set has been deposited on Dryad (see Data availability). Additional inputs are the PBSnj tail SNP set, and the background SNP set (of which the candidates are a subset). For all gene set enrichments, the background SNPs set was the full genome-wide set of genic variants for which PBSnj could be calculated.

GOWINDA was designed to reduce false positives that result from gene length bias (the probability of randomly containing an outlier SNP increases with gene length) and clustering of genes (such as paralogs) that share function. It achieves this by using resampling of background SNPs, which is the genome wide set of SNPs considered in a test. We use the—mode gene switch. In this case, background SNPs are randomly sampled until the number of overlapping genes matches the total number of genes overlapping the PBSnj tail SNP set. Empirical p-values are estimated for each GO category, as the proportion of resamples which contain the same or greater number of genes than the PBSnj tail SNP set, per GO category (for each random background sample a pseudo p-value per GO category is also likewise calculated). FDR at each p-value, p, is then estimated as the number of observed p-values less than or equal to p, R_obs, divided by the total number of resamples with a p-value less than p R_exp i.e. FDR = R_obs / R_exp.

It is important to note that only genic background SNPs that are within the candidate set of genes (e.g. genes with GO definitions) are used in the random sampling. For the GO enrichment, after filtering for gene sets with at least 3 genes, the GO definition file contains definitions for 15649 genes, and 95% of genic background SNPs are used for resampling. This is important, as therefore GOWINDA cannot be used to directly test for enrichment in a single or small set of candidate gene sets. Providing one category, for example, would reduce the background SNP set to only those background SNPs in the genes in that category. Resampling can only ever return the same number of genes in this case. Thus, for VIPs and for the SIV gene set, we included an additional category, which is the full set of genes in the gtf file (“all gene set”). This has no effect on empirical p-value estimation. Its effect on FDR correction is limited as R_obs is unchanged. For a candidate p-value, the all gene set will not be lower or equal to it unless the candidate p-value is itself 1. Thus R_obs is unchanged. The effect on R_exp is hard to determine, but for small empirical p-values should be proportionately small.

There are 98 VIP gene sets in [4], reduced to 53 when filtered for those containing at least 3 genes. For these and for the GO categories we used an FDR < 0.1 as a cut-off when discussing significant categories. There is only one SIV response genes set, so we only report the empirical p-value and treat p-value < 0.05 as significant. Note that this procedure does not allow the calculation of an FDR for the SIV set, nor over the family of tests (SIV gene set enrichment in all four subspecies) but we tested a strong a priori expectation that given the eastern PBSnj tail genes are enriched for viral immunity genes, this would be due to vervet SIV response genes. However, to estimate such an FDR, we used a resampling scheme: For each gene in the genome, we assign a weight, which is the proportion of SNPs in that gene compared to the genome as a whole. This is to correct for gene length bias. We make the intersect of all the SIV genes in each PBSnj tail. We then do weighted resampling from all genes in the genome to create sets of genes as large as the intersect set, and calculate an empirical p-value for each subspecies, as defined above. These empirical p-values are highly similar to those provided by GOWINDA, suggesting that our weighting scheme effectively controls for gene length bias. We then calculate the FDR for each empirical p-value, with R_exp summed over all four subspecies.

Natural Host SIV responsive genes underpin the eastern PBSnj tail genic enrichment

We wanted to test if selection on natural host SIV responsive genes could be the reason that eastern chimpanzees exhibit the strongest signal of genetic adaptation. Our simple test is to hypothetically propose that if selection had not acted on the natural host SIV responsive genes then those genes would not contribute a SNP to the PBSnj eastern genic tail. Thus, we removed the genic tail SNPs from the 118 genes that are natural host SIV responsive and have SNPs in the outlier bin of the eastern PBS scores. However, we don’t remove the genic SNPs within these genes that are in any of the other subspecies. This means we will affect the eastern genic enrichment, but not the enrichment of other subspecies. We argue that this answers the question “what would the eastern genic enrichment be if selection had not acted on these genes in eastern chimpanzees”.

Statistics

To test enrichment in phastCons scores and regulomeDB scores we use random resampling tests. For a candidate set of SNPs sized n, we randomly draw the same number of genic SNPs. For phastCons and regulomeDB we calculate the proportion of SNPs in a category.

For all resampling tests, p-values are estimated as 1 + n resamples > = observed (or < = observed as appropriate) / 1 + n resamples. Adding 1 to both the numerator and denominator ensures that resampling p-values do not equal 0, which is a downward biased estimate given finite resampling.

To test if the proportion of VIP category genes that were also vervet SIV responsive genes differed from null expectations, we used a binomial test of the proportion of VIP genes that were also SIV responsive compared to the null expectation, which is the proportion of all genes that are SIV responsive.