Abstract
Detailed modeling of a species’ history is of prime importance for understanding how natural selection operates over time. Most methods designed to detect positive selection along sequenced genomes, however, use simplified representations of past histories as null models of genetic drift. Here, we present the first method that can detect signatures of strong local adaptation across the genome using arbitrarily complex admixture graphs, which are typically used to describe the history of past divergence and admixture events among any number of populations. The method—called Graph-aware Retrieval of Selective Sweeps (GRoSS)—has good power to detect loci in the genome with strong evidence for past selective sweeps and can also identify which branch of the graph was most affected by the sweep. As evidence of its utility, we apply the method to bovine, codfish and human population genomic data containing multiple population panels related in complex ways. We find new candidate genes for important adaptive functions, including immunity and metabolism in under-studied human populations, as well as muscle mass, milk production and tameness in specific bovine breeds. We are also able to pinpoint the emergence of large regions of differentiation due to inversions in the history of Atlantic codfish.
Introduction
One of the main goals of population genomics is to understand how adaptation affects patterns of variation across the genome and to find ways to analyze these patterns. In order to identify loci that have been affected by positive selection in the past, geneticists have developed methods that can scan a set of genomes for signals that are characteristic of this process. These signals may be based on patterns of haplotype homozygosity [1, 2], the site frequency spectrum [3, 4] or allelic differentiation between populations [5, 6].
Population differentiation-based methods have proven particularly successful in recent years, as they make few assumptions about the underlying demographic process that may have generated a selection signal, and are generally more robust and scalable to large population-wide datasets. The oldest of these are based on computing pairwise FST [7, 8] or similar measures of population differentiation between two population panels across SNPs or windows of the genome [9, 10]. More recent methods have allowed researchers to efficiently detect which populations are affected by a sweep, by computing branch-specific differentiation on 3-population trees [6, 11, 12], 4-population trees [13], or arbitrarily large population trees [14–16], or by looking for strong locus-specific differentiation or environmental correlations, using the genome-wide population-covariance matrix as a null model of genetic drift [17–21].
Although some of these methods for detecting selection implicitly handle past episodes of admixture, none of them uses “admixture graphs” that explicitly model both divergence and admixture in an easily interpretable framework [22, 23]. This makes it difficult to understand the signatures of selection when working with sets of multiple populations that may be related to each other in complex ways. Here, we introduce a method to efficiently detect selective sweeps across the genome when using many populations that are related via an arbitrarily complex history of population splits and mergers. We modified the QB statistic [24] which was originally meant to detect polygenic adaptation using admixture graphs. Unlike QB, our new statistic—which we call SB—does not need gene-trait association data and works with allele frequency data alone. It can be used to both scan the genome for regions under strong single-locus positive selection, and to pinpoint where in the population graph the selective event most likely took place. We demonstrate the usefulness of this statistic by performing selection scans on human, bovine and codfish data, recovering existing and new candidate loci, while obtaining a clear picture of which populations were most affected by positive selection in the past.
Methods
Theory
We modified the previously-developed QB statistic [24] to detect strong branch-specific deviations in single-locus allele frequencies, but without having to use effect size estimates from an association study. We assume that the topology of the admixture graph relating a set of populations is known and that we have allele frequency data for all the populations we are studying. For a single SNP, let p be the vector of allele frequencies across populations. We then make a multivariate normal approximation to obtain a distribution with which we can model these frequencies [14, 17, 18, 25]: where F is the neutral covariance matrix and e is the ancestral allele frequency of all populations. We use the genome-wide covariance matrix as an estimate of the neutral covariance matrix. In practice, the ancestral allele frequency is unknown, so the mean allele frequency among populations can be used as an approximate stand-in. We then obtain a mean-centered version of the vector p, which we call y: where 1 is a column vector of as many ones as there are populations. For an arbitrarily-defined, mean-centered vector b with the same number of elements as there are populations:
Our test statistic—which we call SB—is then defined as:
The key is to choose a vector b that represents a specific branch of our graph. Essentially, for a branch j, the elements of its corresponding branch vector bj are the ancestry contributions of that branch to each of the populations in the leaves of the graph. For a more detailed description of how to construct this vector, see Racimo, Berg and Pickrell [24].
If we choose b to be the vector corresponding to branch j when computing the statistic in equation 4, then significant values of the statistic SB(j) will capture deviations from neutrality in the graph that are attributable to a disruption that occurred along branch j.
If we only have a few genomes per population, the true population allele frequencies will be poorly estimated by our sample allele frequencies, potentially decreasing power. However, we can increase power at the cost of spatial genomic resolution and rigorous statistical interpretation, by combining information from several SNPs into windows, as was done, for example, in Skoglund et al. [26]. We can compute the average χ2 statistic over all SNPs in each window and provide a new P-value for that averaged statistic. As the chi-squared distributional assumption only holds for low amounts of drift, it may be useful to standardize the scores using the mean and variance of the genome-wide distribution, especially when working with populations that diverged from each other a long time ago.
Implementation
We implemented the SB statistic in a program called GRoSS (Graph-aware Retrieval of Selective Sweeps), which is freely available on GitHub: https://github.com/FerRacimo/GRoSS. The program runs on R and makes use of the admixturegraph R library [27]. We also wrote a module that allows one to input a file specifying the admixture graph topology directly.
Figure 1 shows a schematic workflow for GRoSS. The user begins by estimating an admixture graph using genome-wide data, via a program like TreeMix [23], MixMapper [28] or qpGraph [22]. Then, the user writes the topology of the graph to a text file. The format of this file can be either the dot-format or the input file format for qpGraph, so it can be skipped if the initial step was run using qpGraph. Then, the user inputs the graph topology, and a file with major/minor allele counts for each SNP into GRoSS. The allele counts can also be polarized as ancestral/derived or reference/alternative. GRoSS will compute the genome-wide covariance matrix and the b vectors for each branch, and then calculate the SB scores and corresponding P-values, which can then be plotted.
Simulations
We used SLiM 2 [29] to simulate genomic data and test how our method performs at detecting positive selection, with sample sizes of 100, 50, 25 and 4 diploid genomes per population (Figures 2, S1, S2 and S3, respectively). We simulated a genomic region of length 10Mb, a constant effective population size (Ne) of 10,000, a mutation rate of 10−8 per base-pair per generation and a uniform recombination rate of 10−8 per base-pair per generation. We placed the beneficial mutation in the middle of the region, at position 5Mb. We used a burn-in period of 100,000 generations to generate steady-state neutral variation. For each demographic scenario that we tested, we simulated under neutrality and two selective regimes, with selection coefficients (s) of 0.1 and 0.01. We considered two types of selection scenarios for each demographic scenario: one in which we condition on the beneficial mutation reaching > 1% frequency at the final generation of the branch in which we simulated the positive selection event, and one in which we condition the mutation reaching > 5% frequency. We discarded simulations that did not fulfill these conditions. We set the time intervals between population splits at 2,000 generations for all branches of the population graph in the 3-population and 6-population graphs, and at 500 generations in the 16-population graph. To speed up the simulations, we scaled the values of the population size and of time by a factor of 1/10 and, consequently, the mutation rate, recombination rate and selection coefficients by a factor of 10 [29].
Human data
We used data from Phase 3 of the 1000 Genomes Project [30] and a SNP chip dataset of present-day humans from 203 populations genotyped with the Human Origins array [22, 31]. The SNP chip dataset was imputed using SHAPEIT [32] on the Michigan Imputation Server [33] with the 1000 Genomes data as the reference panel [24]. We used inferred admixture graphs that were fitted to this panel using MixMapper (v1.02) [28] in a previous publication [24]. For the 1000 Genomes dataset, the inferred graph was a tree where the leaves are composed of panels from 7 populations: Southern Han (CDX), Han Chinese from Beijing (CHB), Japanese from Tuscany (JPT), Toscani (TSI), Utah Residents (CEPH) with Northern and Western European Ancestry (CEU), Mende from Sierra Leone (MSL) and Esan from Nigeria (ESN) (Figure 3). For the Human Origins dataset, the inferred graph was a 7-leaf admixture graph that includes Native Americans, East Asians, Oceanians, Mandenka, Yoruba, Sardinians and Europeans with high ancient-steppe (Yamnaya) ancestry (Figure 3). This graph contains an admixture event from a sister branch to Native Americans and a sister branch to Sardinians into Europeans, representing the ancient steppe ancestry known to be present in almost all present-day Europeans (but largely absent in present-day Sardinians).
We removed sites with < 1% minor allele frequency or where at least one population had no coverage. We then ran GRoSS on the resulting SNPs in each of the two datasets (Tables S1 and S2). We selected SNPs with –log10(P) larger than 7, and merged SNPs into regions if they were within 100kb of each other. Finally, we retrieved all HGNC protein-coding genes that overlap each region, using biomaRt [34].
Bovine data
We assembled a population genomic dataset (Table S3) containing different breeds of Bos taurus using: a) SNP array data from ref. [35], corresponding to the Illumina BovineHD Genotyping BeadChip (http://dx.doi.org/10.5061/dryad.f2d1q); b) whole-genome shotgun data from 10 individuals from the indigenous African breed NĎama [36] (Bioproject ID: PRJNA312138); c) shotgun data from two commercial cattle breeds (Holstein and Jersey; Bioproject IDs: PRJNA210521 and PRJNA318089, respectively); d) shotgun data for 8 Iberian cattle breeds [37].
We used TreeMix [23] to infer an admixture graph (Figure S4) using allele counts for 512,358 SNPs in positions that were unambiguously assigned to the autosomes in the cattle reference genome version UMD_3.1.1 [38] using SNPchiMp [39]. For shotgun data, allele counts were obtained from allele frequencies calculated in ANGSD [40] for positions covered in at least three individuals. We removed SNPs for which at least one panel had no coverage or in which the minor allele frequency was less than 1%.
We applied the statistic to the TreeMix-fitted graph model in Figure 4. We performed the scan in two ways: in one we computed a per-SNP chi-squared statistic, from which we obtained a P-value (Table S4), and in the other we combined the chi-squared statistics in windows of 10 SNPs (Table S5), with a step size of 1 SNP, obtaining a P-value for a particular window using its average SB score (Figure 4). We used this windowing scheme because of concerns about small sample sizes in some of the populations, and aimed to pool information across SNPs within a region. After both scans, we combined windows that were within 100kb of each other into larger regions, and retrieved HGNC and VGNC genes within a +/- 100kb window around the boundaries of each region using biomaRt [34] with the April 2018 version of Ensembl.
Codfish data
Codfish were randomly sampled from a large tissue sample database [41] and the J. Mork collections from populations covering a wide distribution from the western Atlantic to the northern and eastern Atlantic (Figure S5 and Table S6). The populations differ in various life-history and other biological traits [42, 43], and their local environment ranges from shallow coastal water (e.g. western Atlantic and North Sea) to waters of great depth (e.g. parts of Iceland and Barents Sea). They also differ in temperature and salinity (e.g. brackish water in the Baltic).
We isolated genomic DNA from gill tissue and fin clips stored in ethanol using the E.Z.N.A.® Tissue DNA Kit (Omega biotek) following the manufacturer’s protocol. Libraries were prepared and individually indexed for sequencing using the Nextera® DNA Library Preparation Kit (Illumina, FC-121–1031). Pooled libraries were sequenced on the HiSeq 2500 in rapid run mode (paired-end, 2×250 cycles) at the Bauer Core Facility at Harvard University. We aligned fastq files to the gadmor2 assembly [44] using bwa mem [45], merged and deduplicated using Picard (http://broadinstitute.github.io/picard) and GATK [46] in accordance with GATK best practices [47]. Details of the molecular and analysis methods are given in [48, 49]. We ran ANGSD [40] on the genome sequences from all populations, computed base-alignment quality [50], adjusted mapping quality for excessive mismatches, and filtered for mapping quality (≥ 30) and base quality (≥ 20). We then estimated the allele frequencies in each population at segregating sites using the –sites option of ANGSD.
We applied the statistic to the graph model in Figure 6, estimated using TreeMix [23] allowing for 3 migration events (Figure S6). We removed SNPs where at least 1 panel had no coverage or in which the minor allele frequency was less than 1%, and we only selected sites in which all panels had 2 or more diploid individuals covered. We performed the scan by combining the per-SNP chi-squared statistics in windows of 10 SNPs, with a step size of 5 SNPs, obtaining a P-value for a particular window using its average SB score (Figures S7, S8, S9, S10). In a preliminary analysis, we identified 4 large regions of high differentiation related to structural variants, which span several mega-bases (see Results and Discussion below). In our final analysis, we excluded sites lying within linkage groups that contain these regions from the TreeMix-fitting and covariance matrix estimation, so as to prevent them from biasing our null genome-wide model.
Selection of candidate regions
Given the myriad of plausible violations of our null multivariate-Normal model (see Discussion), we do not expect the P-values of the SB statistic to truly reflect the probability one has of rejecting a neutral model of evolution. We therefore see these P-values as a guideline for selecting regions as candidates for positive selection, rather than a way for rigorously determining the probability that a region has been evolving neutrally. In all applications below, we used arbitrary P-value cutoffs to select the top candidate regions based on visual inspection of the Manhattan plots of the genome-wide distribution of SB P-values. These empirical cutoffs vary across study species and also depend on the specific scheme we use to calculate the SB statistic (per-SNP or averaged over a window) and we do not claim these cutoffs to have any statistical motivation beyond being convenient ways to separate regions that lie at the tails of our empirical distribution.
Alternative approaches could involve using a randomization scheme or generating simulations based on a fitted demographic model to obtain a neutral distribution of loci and derive a P-value from that. While any of those approaches could be pursued with the SB framework, we do not pursue any of those approaches in this paper. We think that the chosen mode of randomization or the fitted demographic parameters will also necessarily rely on assumptions about unknown or unmodeled parameters, and may provide unmerited confidence to the cutoff that we could end up choosing. Instead, we recommend that the reader take our chi-squared-distributed P-values with a grain of salt, and merely use them as a way to prioritize regions for more extensive downstream modeling and validation (for example, using methods like those described in refs. [51–53]).
Results
Simulations
We performed simulations on SLiM 2 [29], and used ROC and precision-recall curves to evaluate the performance of our method under different demographic scenarios, and to compare the behavior of our scores under selection and neutrality. For each demographic scenario, we tested four selective sweep modes: comparing simulations under two different selection coefficients (s=0.1 and s=0.01) against a set of neutral simulations (s=0), and, for each of these, conditioning on establishment of the beneficial mutation at more than 5% frequency or at more than 1% frequency. Each branch of each graph had a diploid population size (Ne) of 10,000.
First, we simulated an episode of positive selection occurring on a branch of a three-population tree with no admixture. Each branch of the tree lasted for 2,000 generations. We sampled 100 individuals from each population. Unsurprisingly, the performance of the method under both selection coefficients is higher when we condition on a higher frequency of establishment of the beneficial allele, and is also better under strong selection (Figure 2). We also kept track of which branch in each simulation had the highest score in a region of 100kb centered on the beneficial mutation. As shown in Figure 2, the highest values typically correspond to the population in which selection was simulated.
We then simulated more complex demographic histories including a 6-population graph with admixture. Each branch of the graph lasted for 2,000 generations. We explored two different selection scenarios. In one scenario, the selective sweep was introduced in one of the internal branches whereas in another scenario, it was introduced in one of the external branches. Interestingly, the performance under this graph appears to improve relative to the three-population scenario (Figure 2). The reason is that the SB statistic depends on having an accurate estimate of the ancestral allele frequency (e). This estimate is calculated by taking the average of all allele frequencies in the leaf populations, so the more leaf populations in a well-balanced graph we have, the more accurate this estimate will be.
In Figure 2, we also show the maximum scores for each branch in a 100kb region around the beneficial allele under both scenarios. We find that the branch where the selective sweep was simulated also tends to have the highest SB scores. We also explored a larger population tree with sixteen leaf populations. In this case, each branch of the graph lasted for 500 generations. ROC and recall curves show a similar performance to the ones from the 6-population admixture graph (Figure 2).
Finally, we explored the performance of the method when the number of diploid individuals per population was smaller than 100. Figure S1 shows the performance of the method with 50 diploid individuals per population, figure S2 shows the performance with 20 individuals and figure S3 shows the performance with 4 individuals. Even when the number of individuals is this small, we can still recover most of the simulated sweeps, especially when selection is strong.
Positive selection in human populations across the world
Applying our method to both the 1000 Genomes and the Human Origins population graphs, we observe many candidate loci that have been identified in previous world-wide positive selection scans (some of them due to archaic adaptive introgression). Previously reported selection candidates that we recover include the LCT/MCM6, BNC2, OCA2/HERC2, TLR and SLC24A5 regions in northern Europeans, the CHMP1A/ZNF276/FANCA, ABCC11 and POU2F3 regions in East Asians, and the SLC45A2 and SLC12A1 genes in an ancestral European population [1, 11, 12, 54–59] (Tables S1, S2).
We find that the IGH immune gene cluster (also containing gene KIAA0125) is the strongest candidate for selection in the 1000 Genomes scan, and the signal is concentrated on the Chinese Dai branch. This cluster has been recently reported as being under selection in a large Chinese cohort of over 140,000 genomes [60]. Our results suggest that the selective pressures may have existed somewhere in southern China, as we do not see such a strong signal in other parts of the East Asian portion of the graph.
A region containing TARBP1 was the strongest candidate for selection in the Human Origins scan (East Asian terminal branch). The gene codes for an HIV-binding protein and has been previously reported to be under balancing selection [61]. The top SNP (rs2175591) lies in an H3K27Ac regulatory mark upstream of the gene. The derived allele at this SNP is at more than 50% frequency in all 1000 Genomes East Asian panels but at less than 2% frequency in all the other worldwide panels, except for South Asians where it reaches frequencies of around 10%. Interestingly, the TARBP1 gene has been identified as a target for positive selection in milk-producing cattle [62] and in sheep breeds [63, 64]. It has also been associated with resistance to gastrointestinal nematodes in sheep [65]. Our results suggest it may have also played an important role during human evolution in eastern Eurasia, possibly as a response to local pathogens.
Another candidate for selection is the NFAM1 gene in East Asians, which codes for a membrane receptor that is involved in development and signaling of B-cells [66]. This gene was also found to be under positive selection in the Sheko cattle of Ethiopia, along with other genes related to immunity [67].
In the Native American terminal branch of the Human Origins scan, we find a candidate region containing two genes: GPR156 and GSK3B. GPR156 codes for a G protein-coupled receptor, while GSK3B codes for a kinase that plays important roles in neuronal development, energy metabolism, and body pattern formation [68]. We also find a candidate region in the same branch in the protamine gene cluster (PRM1, PRM2, PRM3, TNP2), involved in spermatogenesis [69, 70], and another region overlapping MDGA2, which is specifically expressed in the nervous system [71].
Cattle breeding: morphology, tameness and milk yield
We performed two scans on the bovine data, one in which computed the SB statistic per SNP (Table S4), and one in which we computed it in 10-SNP windows (Table S5). The window-based scan retrieved 12 top candidate regions, 10 of which overlap with regions previously detected to be under selection in cattle (reviewed in [72]). Additionally, 28 of the 43 top candidate SNPs from the single-SNP scan are also in regions that have been previously reported as selection candidates.
The two top scoring regions from the 10-SNP scan are both in chromosome 7, and include regions that have been detected to be under selection in both taurine and zebu cattle (the latter not represented in our sample), and are potentially associated with cattle traits of general interest in domesticated species. The region located between 30 and 33 Mb of chromosome 7 appears as a top candidate in both the 1-SNP and 10-SNP scan, on the terminal branch leading to the Maremmana (MA) breed (Figure 5). It includes genes related to variation in body shape, such as CEP120 whose mutations have been linked to a type of skeletal dysplasia that results in thoracic cage and extra-skeletal abnormalities [73], and PRDM6, a histone modifier that can induce various smooth muscle phenotypes [74].
The region located between 50 and 55 Mb contains members of the three protocadherin (Pcdha, Pcdhb and Pcdhg) gene clusters (Figure 5). It is identified by GRoSS to be under selection in Romanian Grey cattle (RO), which is well-known for its docile disposition. Protocadherins are cell-adhesion molecules that are differentially expressed in individual neurons [75]. They have been implicated in mental retardation and epilepsy in humans [76] and in fear-conditioning and memory in mice [77], and have also been shown to be under selection in cats [78]. Genes of the protocadherin family have also been detected to have expression and allele frequency differences consistent with adaptation in an analysis of tame and aggressive foxes [79].
The largest window (4.4 Mb) detected by GRoSS corresponds to the branch leading to the Holstein (HOL) breed (Figure 5). This window overlaps regions found to be under selection in Holstein using various tests (reviewed in [72]). Some of the outlier genes that were also identified in an earlier XP-EHH scan [80] include VPS18, implicated in neurodegeneration [81] and CAPN3, associated with muscle dystrophy. The window also contains genes that are differentially expressed between high and low milk yield cows (PLCB2 and CCDC9B). [82].
Large regions of extreme differentiation in Atlantic codfish
When running GRoSS on the Atlantic codfish data, we found 4 large genomic regions of high differentiation spanning several mega-bases, on 4 different linkage groups: LG01, LG02, LG07, and LG12 (Figure 6, Table S7). These regions were previously detected by pairwise FST analyses [48, 83, 84]. They are associated with inversions that suppress recombination in heterozygous individuals, and have thereby favored dramatic increases in differentiation between haplotypes. The signals in the LG01, LG02 and LG07 regions are strongest among north Atlantic populations. The LG01 signal is particularly concentrated in the terminal branches leading to the Icelandic and Barents Sea branches. The LG02 signal is concentrated in the Icelandic terminal branch and the parent branches of the east Atlantic / north European. This region also contains a low-differentiation region inside it, suggesting it may be composed of two contiguous structural variants, as the LG01 region is known to be [85]. The LG07 signal is concentrated in nearly the same branches as the LG02 signal, and also in the Faroe plateau terminal branch. In contrast, the highly differentiated region in LG12 is concentrated among more downstream branches of the east Atlantic / north European populations, including the Celtic Sea terminal branch (Figure 6). Notably, none of the highly differentiated regions appears to show strong signs of high differentiation in the west Atlantic / North American populations.
Discussion
We have developed a method for detecting positive selection when working with species with complex histories. The method is fast—it only took 486 seconds to run the bovine scan (including 512,358 SNPs and 36 populations) on a MacBook Air with a 1.8 GHz Intel Core i5 processor and 8 Gb of memory. Assuming a null model of genetic drift based on a multivariate Normal distribution, the SB statistic is chi-squared distributed with 1 degree of freedom. This is accurate as long as the graph topology is accurate and the branches in the graph do not contain high amounts of drift. In an admixture graph with K branches, there are K possible versions of the SB statistic. If the differences in allele frequencies at a SNP can be explained by an allele frequency shift that occurred along branch k, then SB (k) will be large, and a P-value based on the null drift model can be calculated from it. By design, branches whose parent are the root node and branches that have the same descendant nodes have the same SB scores, so selective events on these branches are not distinguishable from each other under this scheme.
The SB statistic is most accurate when a large number of individuals have been sampled from each population. If this is not the case, then one can average the scores over windows of SNPs to obtain power from correlated allele frequency shifts in a region (for example, as in ref. [26]), at the expense of losing spatial resolution across the genome due to larger test regions, as we did here in the bovine dataset example. The statistic, however, does not account for the structure of linkage disequilibrium within or between windows.
We have found the method performs best when there are many leaves in a graph because it uses a population-averaged allele frequency to estimate the ancestral allele frequency in the graph. We therefore recommend using this method when working with more than a few populations at a time, to make this estimate as accurate as possible. A possible future improvement of the method could be the incorporation of a model-based ancestral allele frequency estimation scheme, to address this issue. For example, to more appropriately account for the population covariance structure, we could also use the unbiased linear estimate of e with minimum variance, ê = (1TF−1p)/(1TF−11) [14]. However, we found this estimate generates an excess of significant P-values when working with allele frequencies at the boundaries of fixation and extinction, due to poor modelling of frequency dynamics by the multivariate normal distribution, and that this effect is ameliorated when using the simpler population-averaged allele frequency.
Another critical issue is that the more branches one tests, the more of a multiple-testing burden there will be when defining significance cutoffs. In a way, our method improves on previous approaches to this problem, because—given an inferred admixture graph—one does not need to perform a test for all possible triplets or pairs of populations, as one would need to do when applying the PBS statistic or pairwise FST methods, respectively. Instead, our method performs one test per branch. For example, if the graph is a rooted tree with m leaves and no admixture, the number of branches will be equal to 2m – 2, while the number of possible triplets will be equal to , and the number of pairs will be equal to , both of which grow much faster with larger m than does 2m – 2.
While the SB statistic is fast and easy to compute, it is not as principled as other approaches for multi-population selection that rely on explicit models of positive selection (e.g. [86]). This means that it only detects significant deviations from a neutral null model and does not provide likelihoods or posterior probabilities supporting specific selection models. We recommend that, once a locus with high SB has been detected in a particular branch of a graph, biologists should perform further work to disentangle exactly what type of phenomenon would cause this value to be so high and test among competing selection hypotheses.
Among the genes that emerge when applying our method to human data, we found several known candidates, like LCT/MCM6, SLC45A2, SLC24A5, POU2F3, OCA2/HERC2 and BNC2. We also found several new candidate regions, containing genes involved in the immune response, like the TARBP1 and NFAM1 genes in East Asians. Additionally, we found new candidate regions in Native Americans, like GSK3B and the protamine gene cluster.
Analysis of the bovine dataset yielded numerous regions that may be implicated in the breeding process. One of the strongest candidate regions contains genes involved in musculoskeletal morphology, including CEP120 and PRDM6, and GRoSS narrows this signal down to the branch leading to the Maremmana breed. This is an Italian beef cattle breed that inhabits the Maremma region in Central Italy, and has evolved a massive body structure well adapted to draft use in the marshy land that characterizes the region [87]. Interestingly, when comparing muscle samples between Maremmana and the closely-related Chianina breed (CH), gene ontology categories related to muscle structural proteins and regulation of muscle contraction have been reported to be enriched for differentially expressed genes. Additionally, the Maremmana is enriched for over-expressed genes related to hypertrophic cardiomyopathy pathways [87].
Another strong candidate region is the protocadherin gene cluster, associated with neuronal functions in humans and mice [75–77], and shown to be under positive selection in domesticated cats and foxes [78, 79]. GRoSS identifies this region as under selection in the Romanian Grey breed terminal branch. Given that this breed is popularly known to be very docile, it is plausible that this gene cluster might have been a target for selection on behavior during the recent breeding process.
Additionally, GRoSS detects a very large 4.4-Mb region as a selection candidate in the Holstein breed, currently the world’s highest-production dairy animal. This region overlaps several candidate genes earlier identified to be under selection in Holstein using other methods (see [72] for an extensive review). These genes are related to several traits usually targeted by breeding practices, such as behaviour, muscle development and milk yield.
Our method also recovered previously reported regions of high differentiation among Atlantic codfish populations and served to pinpoint where in the history of this species the inversions may have arisen, or at least where they have most strongly undergone the process of differentiation between haplotypes. The largest of these regions is in LG01 and is composed of two adjacent inversions covering 17.4 Mb [85], which suppress recombination in heterozygous individuals and promote differentiation between haplotypes. The inversions effectively lock together a super-gene of alleles at multiple loci [85]. Two behavioral ecotypes—a deep-sea frontal (migratory) ecotype and a shallow-water coastal (stationary) ecotype—have been associated with inversion alleles in the region [88–90]. Several putative candidate selected genes are located within the LG01 inversions [85, 91, 92] that may be of adaptive value for deep sea as well as long-distance migration.
Similarly, the other large inversions observed on linkage groups LG02, LG07 and LG12 (5, 9.5, and 13 Mb respectively) also suppress recombination [93, 94]. Allele frequency differences observed between individuals living offshore and inshore environments are suggestive of ecological adaptation driving differentiation in these regions [93–95]. Previously, a pairwise FST outlier analysis of populations in the north (Greenland, Iceland, and Barents Sea localities combined) vs. populations in the south (Faroe Islands, North Sea, and Celtic Sea combined) showed clear evidence of selection in these regions [48]. However, in comparisons of West (Sable Bank, Western Bank, Trinity Bay, and Southern Grand Banks combined) with either North or South localities only some of these regions displayed signatures of high differentiation [48], indicating these inversions had different spatiotemporal origins. By modeling all these populations together in a single framework, our method provides a way to more rigorously determine in which parts of the graph these inversions may have originated (Figure 6), and suggests they were largely restricted to East Atlantic populations.
In conclusion, GRoSS is a freely-available, fast and intuitive approach to testing for positive selection when the populations under study are related via a history of multiple population splits and admixture events. It can identify signals of adaptation in a species by accounting for the complexity of this history, while also providing a readily interpretable score. This method will help evolutionary biologists and ecologists pinpoint when and where adaptive events occurred in the past, facilitating the study of natural selection and its biological consequences.
Acknowledgments
We thank Jeremy Berg, Anders Albrechtsen and Kathleen Lotterhos for helpful advice and discussions. FR thanks the Natural History Museum of Denmark, the Danish National Research Foundation and the Lundbeck Foundation for their support. EÁ and KH were supported by a grant from Svala Árnadóttir’s private funds, by a grant from the University of Iceland Research Fund, by institutional funds from R.C. Lewontin, and by a grant from the Icelandic Research Fund (no. 185151–051). R.R.F. thanks the Danish National Research Foundation for its support of the Center for Macroecology, Evolution, and Climate (grant DNRF96). The authors declare no competing interests.
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.
- 16.↵
- 17.↵
- 18.↵
- 19.
- 20.
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.↵
- 27.↵
- 28.↵
- 29.↵
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.↵
- 36.↵
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.↵
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.
- 53.↵
- 54.↵
- 55.
- 56.
- 57.
- 58.
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.↵
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵
- 82.↵
- 83.↵
- 84.↵
- 85.↵
- 86.↵
- 87.↵
- 88.↵
- 89.
- 90.↵
- 91.↵
- 92.↵
- 93.↵
- 94.↵
- 95.↵
- 96.↵