Widespread Signals of Recent Selection

We calculated iHS for all SNPs with minor allele frequency >5%. Figure 3 shows a summary of the extreme values on Chromosome 2, with similar plots across all autosomes and the X chromosome provided in Figures S3–S25. The plotted points are those with |iHS| > 2.5, and correspond to the most extreme 1% of iHS values.

Open in a separate window

Figure 3

Plots of Chromosome 2 SNPs with Extreme iHS Values Indicate Discrete Clusters of Signals

SNPs with |iHS| >2.5 (top 1%) are plotted. The bottom plot combines signals for all three populations, plotting only SNPs with derived frequency >0.5 and iHS <−2.5. Such SNPs correspond to high-frequency-derived SNPs in the range for which our test is most powerful. The short vertical bars below each plot indicate 100-kb windows whose signals are in the top 1% of windows genome-wide.

There is clear clustering of extreme values into distinct regions where many SNPs show evidence for selection. One such region, at 135–138 Mb in Europeans, contains the lactase gene (LCT), previously noted as a target of strong selection [6,33]. In principle, clustering of unusual iHS scores might occur even under a completely neutral model. However, several lines of evidence indicate that selection is indeed producing widespread signals in the data.

First, simulations show that we observe more extreme values of the unstandardized iHS scores than expected under a range of neutral models (Figure 4). For each population, we performed neutral simulations that matched the observed SNP density and allele frequency spectrum, which included extensive recombination rate variation, including hotspots, and utilized a range of demographic models consistent with previous studies of demographic parameters (see Materials and Methods). The demographic models included a variety of bottleneck models for East Asians and Europeans, and models of constant size with recent growth for Yoruba [34,35]. Since previous genetic studies indicate that the Yoruba are likely to have the least complex demographic history [34–36], we focus mainly on simulation results for that population. We find that extreme values of the unstandardized iHS are more frequent in the real Yoruba data than in any of the simulated models, as expected if the largest iHS values are frequently due to selection. There is also an excess of extreme values in the Europeans, but in the East Asians some demographic models show as much variance as the real data (unpublished data).

Open in a separate window

Figure 4

Central 99% Range of Unstandardized iHS for SNPs in the Yoruba Data and for SNPs in Matched Neutral Simulations

The upper and lower lines mark the boundaries of the central 99% distribution of the unstandardized iHS ratio, as a function of derived allele frequency. The gray lines plot results for a range of plausible demographic models. The fatter tails in the real data are consistent with the action of selection.

Second, in all three populations, there is greater clustering of extreme (standardized) iHS values in the real data than in neutral simulations with heterogeneous recombination rates. For example, in simulations matching aspects of the Yoruba data, only 0.1% of windows of 50 consecutive SNPs had more than 16 SNPs with |iHS| > 20. In the actual YRI data, we observed a 14-fold enrichment of such windows. Since this calculation is based on empirically standardized scores, the signal of extra clustering is distinct from the previous signal of overdispersion of the unstandardized scores. Simulations designed to match the East Asian and European datasets also indicate that there is excess clustering of extreme iHS values in the real data, though for these populations the relative enrichment is quantitatively smaller due to the extra variance seen in neutral bottleneck models (the data show 2.3-fold and 2.7-fold enrichment of the top 0.1% of windows, respectively). In summary, the visual sense of clustering of high |iHS| scores in Figure 3 does indeed exceed the level of clustering expected under neutrality, supporting a model in which distinct selective events produce large |iHS| scores across discrete regions.

Third, in our data, extreme iHS scores frequently occur in regions where the frequency spectrum also indicates the action of selection (Figure 5). As shown in Figure 2, a version of Fay and Wu's H test [37] for ascertained SNPs (H_asc) provides a useful method for detecting sweeps where the selected site has reached high frequency. In simulations we find that high frequency selected sites tend to have both strongly negative iHS scores and strongly negative values of Fay and Wu's H_asc. However in neutral simulations, iHS and H_asc are essentially uncorrelated. In the Yoruba data, as many as one half of high frequency–derived SNPs with large iHS scores fall into the most extreme 1% of windows for H_asc, genome-wide. This correlation argues strongly that many of our extreme iHS scores are in fact the result of positive selection. In our data, an excess of significant H_asc scores is also seen for low-frequency alleles with positive iHS scores. This probably results from ancestral alleles that have hitchhiked to high frequency. Similar, but weaker, correlations are present in both the European and East Asian data (unpublished data).

Open in a separate window

Figure 5

Strong Correlation between iHS and H_asc for the Yoruba Data

The left-hand plot shows the probability that a 51-SNP window centered on a given SNP is in the lowest 1% of the empirical distribution for Fay and Wu's H in neutral simulated data, and in the Yoruba data overall. Notice that in neutral simulations, there is essentially no correlation between iHS and H. Right-hand plot: In contrast, in simulations with selection (cyan line, s = 100) there is a big increase in the rate of significant H values for high-frequency selected alleles with strongly negative iHS (<−2.5). The same pattern is seen for the real data (red line). In the real data, sites with strongly positive iHS (>2.5) show an increase in the rate of positive H scores at low derived allele frequencies (magenta line). The latter probably reflects instances of an ancestral allele hitchhiking to high frequency with a selected sweep.

Fourth, there is a highly significant enrichment for extreme iHS values in genic regions (p < 10⁻²⁰ in all populations; Table S1). This is to be expected if selection occurs most often in (or near) genes, though one might not expect a dramatic difference in rates, since simulated selective events tend to produce signals over quite wide regions which would include both genic and non-genic SNPs. The proportion of SNPs with |iHS| > 2 is 1.23-fold higher in genic SNPs than non-genic SNPs in Yoruba, 1.16-fold higher in Europeans, and 1.13-fold higher in East Asians. A further enrichment in extreme iHS values is found in SNPs in overlapping genes compared with SNPs in non-overlapping genes. To check that these results are not an artifact of the higher HapMap SNP density in genes, we thinned genic SNPs at random so that the proportion of genic SNPs matched the proportion of the genome containing genes. After reanalysis, the difference between genic and nongenic iHS values remained about the same as before (unpublished data).

Last, various regions identified previously as likely targets of sweeps-in-progress are detected by our survey, including signals in Europeans in the lactase region [6,33], in the 17q21 inversion [20], and at CYP3A5 [7]; in the ADH cluster on Chromosome 4 in East Asians [38]; and in olfactory receptor clusters on Chromosomes 11p15 and 11q11 in Yoruba [39]. However, we do not detect all previously identified selection candidates, for example failing to find signals at G6PD in Yoruba [3,5], and in two genes involved in brain development that were recently reported to be under recent positive selection [8,9].

Overview of Selection Regions

To identify the strongest signals of selection, we divided the genome into non-overlapping windows of 100 kb. In this analysis windows were defined by physical location to facilitate the comparison of signals across populations. For each population, we identified the windows in the highest 1% of the empirical distribution for proportion of SNPs with |iHS| > 2. The positions of these windows on Chromosome 2 are indicated by vertical bars below each panel of Figure 3. Henceforth, we consider these windows to be candidates for containing selective sweeps. We find that 8 of 14 genomic regions listed as selection candidates by the HapMap Project are among the top 5% of our signals (Table 10 in [18]). A summary of some of our strongest signals is shown in Table 1, and a complete list is provided online in Protocols S1–S3). As an illustration, Figure 6 depicts the haplotype patterns, decay of EHH with distance, and plot of iHS scores for three strong candidate regions identified by our genome-wide scan.

Table 1

Summary of Some of the Strongest iHS Signals Genome-Wide

Open in a separate window

Open in a separate window

Figure 6

Signals of Selection for Three Candidate Selection Regions Discussed in the Text

The columns show (left) scatter plots of negative iHS scores, (center) haplotype plots, and (right) decay of haplotype homozygosity. In each case the core SNP for the center and right-hand plots was chosen as a SNP with high negative iHS score (starred in the scatter plots); the allele marked in red is derived. For each signal, values are listed for the derived allele frequency (p_d) and the local deCode recombination rate estimate.

Analysis of the haplotype structure in candidate selection regions indicates that the events detected are typically very recent. We calculated the lengths of the haplotypes around the SNP with the largest negative iHS value in each region (Table S2). The average total distance between the first point to the left, and to the right of the core SNP at which EHH on the selected haplotypes drops below 0.25 is 0.52 cM in both East Asians and Europeans, and 0.32 cM in Yoruba. Hence, candidate sweep regions tend to be narrower in Yoruba than in the non-African populations, indicating that typical sweep events may be substantially younger in the non-African populations. (The size of the area affected by a strong sweep depends only weakly on the effective population size [40] and so the larger effective population size in Yoruba is not an explanation of the smaller average sweep size.)

A fully rigorous estimation of the ages of the candidate sweeps is difficult with the current data. However, making the simplistic assumption of a star-shaped genealogy for the favored haplotypes and assuming a generation time of 25 y, suggests average ages of 6,600 years and 10,800 years in the non-African, and African populations, respectively (Materials and Methods). Simulations using SelSim (Materials and Methods) suggest that these haplotype spans are consistent with selection coefficients of 0.01–0.04, assuming a population size of 10,000. These estimated ages should not be taken to imply a burst of selection at a particular time; instead, these ages and selection coefficients might represent areas in the parameter space in which we have good power. The longest haplotypes around derived alleles at >50% frequency extend 1.39 cM in East Asians (near the Gaucher disease gene, GBA), 1.25 cM in Europeans (near NKX2–2, which is involved in insulin regulation), and 0.97 cM in Yoruba (in a gene desert on Chromosome 5p15). These long haplotypes indicate extremely strong selection on recent mutations, though it is difficult to be confident about the actual genetic target of the selective events. In summary, the selection events that we detect are generally very recent, substantially postdating the separation times of these populations, and falling mainly within the agricultural phase of human evolution.

Figure 7 shows the extent of overlap of sweep regions across the three populations. Most of the sweep regions are found in only one of the three populations, consistent with the estimates indicating that these events generally postdate population separation. Nonetheless, there is a clear excess of sweeps that are shared between pairs of populations, or among all three populations. In principle, sharing of signals between populations might also be due to haplotypes that are inherited from the ancestral populations. However, this is probably a small effect since such unusually long haplotypes would be unlikely to survive the effects of recombination for >1,000 generations, separately in each population. Instead, the data suggest that most of the selective events that we detect are local to a single population, but that a significant fraction of the selective events are experienced by more than one population.

Open in a separate window

Figure 7

Sharing of iHS Signals between Populations

The numbers listed inside circles represent the numbers of 100-kb windows that are in the top 1% of the empirical distributions in at least one population. The numbers in the intersection regions are in the top 1% for one population, and the top 5% for one or both of the other populations. The counts that would be expected if signals were independent across populations are shown in parentheses. The number of windows not in any circle is reported in the upper-left corner.

This view is further supported by the relationship between iHS and allele frequency divergence measured using F_ST (Figures S26–S28). SNPs with high |iHS| in one population, but low |iHS| in another are likely to have high F_ST, indicating that the SNP has changed frequency rapidly since population separation. Among the modest number of SNPs that have extreme positive iHS or extreme negative iHS in both populations, there is not an excess of high F_ST, perhaps due to recent gene flow of alleles (and haplotypes) that are favored in multiple populations.

SNP annotation information

Annotation information was obtained from dbSNP build 123 for all SNPs, including physical positions (for genome build #34) and strand orientation. We also used the dbSNP classification of SNP functional states: non-synonymous, synonymous, within an intron or mRNA UTR, or within 2 kb of a gene (locus region). SNPs in any of these functional classes were considered genic. All other SNPs were considered nongenic.

Obtaining ancestral states from dbSNP and chimp sequence alignment

For ~30% of the SNPs, estimates of the ancestral state were already available from dbSNP using ftp://ftp.ncbi.nih.gov/snp/database/b124/mssql/data/SNPAncestralAllele.bcp.gz and an allele reference table at ftp://ftp.ncbi.nih.gov/snp/database/b124/mssql/data/Allele.bcp.gz. To obtain ancestral states for all SNPs, we obtained FASTA sequences for each SNP as specified for that rs entry in the dbSNP database, and aligned those to the draft genome build of the chimp sequence, obtained from the UCSC database [12] using blat, version 27 [48]. For each SNP, we selected the overall best alignment, preferring alignments with the highest overall blat score, and preferring alignments mapping to unique chimp chromosome over random or unmapped sequences when possible. We then inferred the ancestral state as the chimp allele at the appropriate position in the sequence, provided that the sequence quality score was ≥20 at that site, and that it matched one of the human alleles. The agreement between our ancestral state estimates and those in dbSNP was 94.5%. For each SNP, we accepted an ancestral state if it could be obtained either from dbSNP or from our own analysis.

Simulations overview

We simulated data under a variety of neutral and selection models using the coalescent programs ms (for neutral models) and SelSim (for selection) [49,50]. Overall, we aimed to match the simulated analysis to the real analysis as closely as possible. We performed separate simulations for each population sample, matching the numbers of sampled chromosomes, the site frequency spectrum, and the average spacing of markers in units of ρ = 4Nr to the data and considering a range of appropriate demographic models for each population based on previous literature. As detailed below, our approach to dealing with ascertainment bias was to simulate a moderately large number of candidate SNPs under an assumed demographic model, and then use rejection sampling to thin the SNPs to match the observed frequency spectrum and SNP density. The rejection sampling function serves implicitly as a model of the ascertainment process.

For every simulated dataset, we estimated genetic distances from the data (as described below) in order to calculate iHS values. For the East Asian sample (where haplotyping is most difficult due to the lack of genotype data from relatives), we also incorporated haplotype estimation into the simulations. For each model and combination of parameters, we simulated 1,000 independent regions. The lengths of simulated regions averaged 500 units of ρ (a little more than 1 Mb at the genome-average recombination rate).

Neutral demographic models

For each population we explored a range of demographic scenarios supported by the literature as plausible models for each population [16,35,36,51,52]. The models considered here spanned the range of models in the confidence sets identified by [35] for closely related populations. We treat the Hausa results from that study as being appropriate for Yoruba because Hausa and Yoruba are genetically very closely related. For the non-African populations, we simulated a family of bottleneck models where a population of constant size N_A instantaneously shrinks in size to b · N_A at time t_start generations before the present. The population remains at that size for t_dur generations and then instantaneously recovers to either its original size or to a larger (or smaller) size. For the African population, we considered models in which an ancestral population at equilibrium size N_A grows exponentially beginning t_onset generations in the past at rate α, such that the present population size is N_Ae ^αt_onset [34]. The ms command lines corresponding to each demographic model simulated are provided in Protocol S4.

Modeling variation in recombination rate

Since recombination rate variation plays an important role in determining patterns of LD [30–32], our simulations allowed for recombination hotspots, and modeled variation both in the background rate and in the spacing and intensity of hotspots. Hotspots were assumed to occur as a Poisson process, at an average rate of 1 hotspot per 100 kb [30,31]. The intensity of hotspots was modeled as a gamma distribution, with the mean intensity 28-fold (and the median 20-fold) the background rate, roughly matching that observed [30]. Hotspot length was gamma distributed with mean 2 kb. The background recombination rate was modeled with a gamma distribution, such that the total rate of recombination events in the background and in hotspots matched the average estimated rate from the data, for each population.

Modeling SNP ascertainment

Due to SNP ascertainment, the allele frequency spectrum in the HapMap differs greatly from the frequency spectrum in DNA sequence data. In all three samples there is a deficit of rare variants, and an excess of high frequency variants. Moreover, the actual frequency distributions are substantially different among the three samples (Figure S29). Since SNP ascertainment can have a major impact on analysis, it is important to model this in the simulations. One potential approach would be to model the ascertainment procedure explicitly, and to attempt to recreate the observed frequency spectrum [52,53]. However, this is very difficult to do in a fully convincing way for the present data, since the full history of ascertainment of SNPs into dbSNP and from there into HapMap is complex and not well-documented.

Thus, as a practical alternative, we simulated data with complete ascertainment and then used rejection sampling to obtain the observed frequency spectrum (separately for each population sample). In effect, the rejection sampling function serves as an empirical model of the unknown ascertainment process. Let f(a) be the probability that a SNP in our dataset has a copies of the derived allele (i.e., f(a) is the observed frequency spectrum), where 0 < a < n and n is the total haploid sample size. Let π(a) be the corresponding frequency spectrum in a neutral model with complete SNP ascertainment, assuming a specified demographic model. To achieve the observed frequency spectrum, we simulated data from the chosen demographic model with complete ascertainment, and then retained sites with a copies of the derived allele with probability f(a)/(cπ(a)), where c = max[f(a)/π(a)]. The mutation rate was chosen so that after ascertainment the expected SNP density would match the observed density.

Model for selective sweeps

We used the program SelSim to simulate data with selection and recombination [50]. The model assumes that a new mutation at a specified position in the sequence experiences a constant additive selection pressure σ = 2NS where N is the population size and s is the additive selective advantage per copy per generation. The data are sampled when the mutation reaches a specified frequency, or reaches fixation. The program only accommodates models of constant population size and hence is most appropriate for modeling the Yoruba (for whom a constant-size model falls within the confidence set of [35]). For each σ, we fixed the frequency of the selected site to p = 0.1, 0.2, … , 0.9, 1.0 and generated samples of size n = 120. We also simulated a model of selection where the sample was taken 200 generations after the fixation of the beneficial allele. 1,000 repetitions were generated for each scenario. The recombination hotspot model and mutation/ascertainment scheme were as described above. The rejection weights for the SNP ascertainment were determined under the neutral model, so that when σ = 0 the frequency spectrum matches the data (on average).

Haplotype phase uncertainty

For the simulations of the African and European data, we ignored uncertainty in haplotype phase since for those populations the availability of trio data makes the reconstructed haplotypes very accurate. For the East Asian data, the phasing error rates are higher since all the individuals are unrelated. To assess the impact of this, all of our simulation data for the East Asians were treated first as diploid genotypes, and then haplotype phase was estimated using a new, rapid-phasing algorithm, fastphase, kindly provided by Paul Scheet and Matthew Stephens. This algorithm uses a similar model and achieves nearly the same accuracy as Phase 2, which was used to phase the HapMap data [54].

Estimation of a fine-scale genetic map

To generate a high-resolution, genome-wide genetic map, we implemented a version of the Li and Stephens algorithm [24]. That algorithm was designed for estimating recombination rates over fine scales based on haplotype data. As input data, we used the physical position of each SNP, along with the phased haplotypes. This was performed separately for each of the three population samples as well as for each repetition of simulated data. The output is a high-resolution estimate of a genetic map for each population, where each SNP is assigned a position in units of ρ = 4Nr where N is the effective population size and r is the recombination rate per generation.

Our implementation follows the original algorithm closely, except in three regards. First, we maximized the likelihood using the E.M. algorithm, finding that this converges quickly and reliably, even on very large datasets. Second, we performed local smoothing of the rates as follows: the recombination rate between SNPs n and n + 1 is estimated as the maximum likelihood rate in an interval that extends from 20 kb to the left of SNP n to 20 kb to the right of SNP n + 1. This window size was chosen as a suitable compromise between wanting to detect recombination hotspots, and not wanting to be overly misled by random noise in the data. The smoothing has the effect of spreading out hotspots, to some extent, but over a scale that is much less than the physical length of “interesting” haplotypes. Third, we did not implement the bias correction suggested by [24] because this depends in part on the spacing of markers, which varies widely in the data, and because the bias cancels out in our test ratio anyway. Simulations (unpublished data) indicate that our implementation performs well, having only slight bias for the parameters relevant to HapMap, and correctly identifying regions of rate variation. Over large scales it is also highly correlated with the deCODE genetic map [55]. There is a modest downward bias in the estimated rates in partial sweep regions, but in simulations this bias has little discernible effect on power. This type of bias will tend to be conservative, in the sense of reducing signal.

Calculation of iHS

The iHS was computed as described earlier in this paper for every SNP with ancestral state information and with minor allele frequency >5%. To compute the iHH, what we do is equivalent to the following. We plot the decay of EHH to the left and right of the core SNP, for each allele, using all available SNPs. The x-axis is in units of genetic distance, 4Nr, estimated as described above. The EHH values at successive SNPs are joined by straight lines, and then we compute the total area under each curve, between the nearest points to the left and right of the core SNP where the EHH drops below 0.05. Since the population genetics of the X chromosome differs from that of the autosomes (e.g., due to smaller effective population size), we standardized the X chromosome separately from the autosomes, and excluded it from our overall genome-wide analyses.

The SNP data include occasional large gaps between successive SNPs. If the region spanned by EHH > 0.05 reached a chromosome end or the start of a gap >200 kb, then no iHS value was reported for the core SNP. If there was a smaller gap (20–200 kb), then the genetic distance spanned by the gap was reduced by a multiplicative factor 20/g, where g is the gap size in kb. This is an ad hoc fix designed to eliminate spurious signals produced by occasional large gaps, while minimizing the loss of data. After applying the various filters, we were left with 676,718, 701,633, and 621,915 SNPs with iHS scores in the European, Yorubans, and East Asian populations respectively.

Summaries of the frequency spectrum

Both Tajima's D and Fay and Wu's H tests use summaries of the site frequency spectrum to identify signatures of selection [37,56]. However, these tests were designed for full-sequence data, and not for pre-ascertained SNPs. Nonetheless, we have tested the value of using these statistics to identify regions whose frequency spectra are strongly different from the bulk of the genome, suggesting the influence of selection. To compute these statistics on ascertained data, we treated the ascertained genotypes as if they were the complete sequence data, and then computed D and H in the usual way. The ascertainment-biased summary statistics are denoted D_asc and H_asc to indicate that they are not expected to follow the usual distributions. Since the ascertained data are enriched for high-frequency variants, the average D_asc is highly positive, whereas the average H_asc is highly negative. We obtained empirical distributions of both D_asc and H_asc in all 50-marker windows across the entire autosomal genome for each population. The empirical p-value for a particular window represents the proportional rank of the statistic in that window compared with the overall genome distribution (i.e., low p-values indicate a more negative D or H than the bulk of the empirical distribution).

Identifying candidate signals of selection

In order to identify genes or genomic regions with signals of selection, we focused on windows of consecutive SNPs. Where possible, we used a standard window size of 50 SNPs. For analyses of genes with ≤50 SNPs, we created windows of 50 SNPs centered on each gene. For the ~5% of genes with >50 SNPs, we treated the window as containing all SNPs in the gene. To study the overlap of signals across populations, we modified our analysis to consider fixed, nonoverlapping, 100-kb windows so that the window positions would correspond exactly across populations.

For both kinds of window definition, the size of signal in each region was quantified by the proportion of SNPs with |iHS| > 2. Simulations indicated that this criterion provides a powerful signal of selection. Separately for each population and analysis, we determined empirical cutoffs for the top 1%, 5%, etc., of signals genome-wide, and considered these strongest signals to indicate candidate selection regions. Since the numbers of SNPs varied considerably across the fixed 100-kb windows, we grouped windows into bins with similar numbers of SNPs, and determined empirical cutoffs separately for each bin. Regions with <10 SNPs were dropped, as were bins with <20 regions. The X-chromosome data was analysed separately. The thresholds for empirical cutoffs for the X chromosome were based on the autosomal cutoffs. 100-kb windows and genes in the X chromosome that were above these thresholds were considered to be significant.

Estimating haplotype ages

To obtain a crude estimate of the ages of sweeps, we assumed (1) a star phylogeny for the selected haplotypes, (2) that the two chromosomes which descended from a common ancestor and carry a selected allele are identical between the core SNP and the point of the nearest recombination event on either lineage, (3) that the two chromosomes start to differ immediately beyond the nearest recombination event. Hence, Pr[Homoz] = e ^−2rg where Pr[Homoz] is the probability that two chromosomes are homozygous at a recombination distance r from the selected site, given a common ancestor g generations before the present. Taking the generation time to be 25 y, the ancestor time in years becomes t = 25g. Then, for example in Africans, when Pr[Homoz] = 0.25, we observed that 2r = 0.32% (the distances in the text are the sum of r in both directions). Then t = −(25)(100)ln(.25/.32 10,000 years. A similar calculation was performed for the other populations.

Gene ontology analysis

The PANTHER Gene Ontology database was downloaded from https://panther.appliedbiosystems.com in March 2005. For each PANTHER “biological-process,” category we determined both the total number of genes, and the number of genes in the top 10% of empirical signals (defined as above). One-tailed p-values for enrichment of particular categories were obtained either using the standard Chi-square approximation if the expected number of counts exceeded five, or otherwise by exact enumeration of the binomial probabilities. We used a Bonferroni correction to assess overall experiment-wide significance, though this is likely to be particularly conservative in this case since the 222 PANTHER categories are partially nested and overlapping.

We thank Graham Coop and other members of the Pritchard Lab, Anna Di Rienzo, Molly Przeworski, and the anonymous reviewers for very helpful discussions and comments; Paul Scheet and Matthew Stephens for prepublication use of their fastphase algorithm; Minghong Ward and others at NCBI for help obtaining some ancestral state file information from the dbSNP site; and the International HapMap Consortium for its work in creating this dataset.

Funding. Our project was supported by RO1 HG002772–1. BFV also received partial support from RO1 DK55889 to Nancy Cox.

Competing interests. The authors have declared that no competing interests exist.