Soft Sweeps Are the Dominant Mode of Adaptation in the Human Genome

Abstract The degree to which adaptation in recent human evolution shapes genetic variation remains controversial. This is in part due to the limited evidence in humans for classic “hard selective sweeps”, wherein a novel beneficial mutation rapidly sweeps through a population to fixation. However, positive selection may often proceed via “soft sweeps” acting on mutations already present within a population. Here, we examine recent positive selection across six human populations using a powerful machine learning approach that is sensitive to both hard and soft sweeps. We found evidence that soft sweeps are widespread and account for the vast majority of recent human adaptation. Surprisingly, our results also suggest that linked positive selection affects patterns of variation across much of the genome, and may increase the frequencies of deleterious mutations. Our results also reveal insights into the role of sexual selection, cancer risk, and central nervous system development in recent human evolution.


Introduction
Spurred by the ongoing revolution in DNA sequencing capacity, human population genetic datasets have grown exponentially in size over the past 5 years (Auton et al. 2015;UK10K Consortium 2015). Such growth enables insight into the evolutionary histories of human populations with hitherto unrivaled precision. A central question in the study of human evolution is the extent to which adaptation has driven recent evolution and affected patterns of genetic diversity (Akey 2009). This can be addressed by scanning genomic data for evidence of selective sweeps, wherein a beneficial mutation is favored by natural selection and therefore rapidly increases in frequency within a population. Such selective sweeps leave a characteristic footprint in variation; they create a valley of diversity around the selected site (Maynard Smith and Haigh 1974;Kaplan et al. 1989;Stephan et al. 1992), a deficit of both low-and high-frequency derived alleles at linked sites (Fay and Wu 2000), and an increase in linkage disequilibrium in flanking regions (Kim and Nielsen 2004). Thus there are multiple population genetic signals to exploit. Accordingly numerous theoretical and methodological advances (Kaplan et al. 1989;Stephan et al. 1992;Fu 1997;Kim and Stephan 2002;Nielsen, Williamson, et al. 2005;Voight et al. 2006) in the study of selective sweeps have given researchers the ability to uncover the genetic basis of adaptation on a genome-wide scale.
There are two complimentary approaches to studying the impact of adaptive evolution on genetic variation. The first approach aims to infer genome-wide rates of adaptive evolution by estimating the mean effects of selective sweeps across the genome (Wiehe and Stephan 1993;Kern et al. 2002;Andolfatto 2007;Jensen et al. 2008;Hernandez et al. 2011;Sattath et al. 2011). Such approaches may estimate the rates of sweeps or their effects with respect to the genomic background, but do not focus on the targets of sweeps themselves. An alternative approach is to focus on finding individual selective sweeps throughout the genome, and in so doing characterize specific cases of adaptation with hopes of gaining general insight into the adaptive process (Sabeti et al. 2002;Voight et al. 2006;Williamson et al. 2007). The search for selective sweeps has shed light into the recent evolutionary histories of natural populations, and has shown a pervasive impact of adaptive evolution on polymorphism in some species such as Drosophila melanogaster (Begun et al. 2007;Macpherson et al. 2007;Langley et al. 2012;Lee et al. 2014;Garud et al. 2015). In humans, the picture remains less clear: while scans for selective sweeps have discovered numerous compelling candidates for strong positive selection (Ruwende et al. 1995;Stephens et al. 1998;Tishkoff et al. 2007;Bryk et al. 2008;Huerta-S anchez et al. 2014), some recent studies have suggested that the impact of adaptation on patterns of variation genome-wide is quite limited (Hernandez et al. 2011;Lohmueller et al. 2011). Conversely, Enard et al. (2014) argue that the genome-wide reduction in diversity around substitutions is driven in part by positive selection.
One possible explanation for the difficulty in characterizing the contributions of adaptive and nonadaptive forces in human populations is that genetic hitchhiking effects may be muted by human demographic history. Many human populations appear to have experienced bottlenecks and/or recent growth (Marth et al. 2004;Fagundes et al. 2007;Gravel et al. 2011;Auton et al. 2015), which cause much of the genome to resemble selective sweeps (Nielsen, Williamson, et al. 2005). Moreover, positive selection has historically been modeled as the process of a de novo beneficial mutation rapidly sweeping to fixation, a process now referred to as a hard sweep. However selection may act on previously segregating neutral or weakly deleterious variants (Orr and Betancourt 2001;Innan and Kim 2004). Selection on standing variation will produce qualitatively different skews in linkage disequilibrium and allele frequencies, along with a shallower valley in diversity (Hermisson and Pennings 2005;Przeworski et al. 2005;Berg and Coop 2015;-such an event is thus referred to as a soft sweep. If selection typically proceeds through soft sweeps, as may be the case in Drosophila (Garud et al. 2015), then many sweeps may have been missed by previous scans that were designed to detect signatures produced under a hard sweep model.
We sought to address the controversy over the impact of adaptation on human genomic variation by conducting a genome-wide scan for both hard and soft selective sweeps across human populations. We previously developed S/HIC (Soft/Hard Inference through Classification), a machine learning method capable of detecting completed sweeps and inferring their mode of selection with unparalleled accuracy and robustness to nonequilibrium demography . Here we apply S/HIC to uncover hard and soft sweeps in six population samples from the 1000 Genomes Project (Auton et al. 2015), thereby performing the most comprehensive investigation of completed selective sweeps in humans to date. Surprisingly, our results suggest that patterns of polymorphism across much of the human genome may be affected by linked positive selection-primarily soft sweeps. Moreover, we find evidence that the mode of selection differs substantially across populations, with non-African populations adapting via hard sweeps to a much greater extent than African populations. Finally, we investigate the biological targets of selection in recent human evolution, with particular processes such as immunity, cancer, and sexual reproduction playing outsized roles.

Results
We set out to detect completed hard and soft selective sweeps in six populations from Phase 3 of the 1000 Genomes Project: two West-African populations (YRI and GWD from Yoruba and The Gambia, respectively), one East-African population (LWK from Kenya), one European population (CEU, from UT), one East Asian population (JPT from Japan), and one from the Americas (PEL from Peru). For each population we trained and applied a S/HIC classifier to identify hard and soft selective sweeps across the genome (see Materials and Methods), distinguishing them from neutrally evolving regions as well as those linked to sweeps . Briefly, S/HIC is a machine learning method that leverages spatial patterns of a variety of statistics across a large genomic window in order to infer the mode of evolution at the center of the window. We previously showed that S/HIC is exceptionally robust to the confounding effect of linked selection (e.g., the "soft shoulder" effect where regions linked to hard sweeps resemble soft sweeps; , as well as nonequilibrium demographic histories, making it well suited for a survey of positive selection in humans. We also assessed the accuracy of our classifiers on simulated test data with the same demographic history used to generate training data, finding that S/HIC achieved good power for each demographic history, with somewhat higher accuracy for histories inferred from the African than non-African populations (supplementary fig. S1, Supplementary Material online).
We also performed forward simulations under the GWD and JPT models (see Materials and Methods) in order to assess whether purifying selection and its effect on variation at linked unselected sites (i.e., background selection Charlesworth et al. 1993) could result in false sweep calls. The results of these simulations suggest that S/HIC's false positive rate is essentially unaffected by these forces (supplementary fig. S1, Supplementary Material online). Note that we exposed each classifier to a wide range of mutation and recombination rates (see Materials and Methods) during training (and testing) in order to improve (and assess) our robustness to variation in these rates across the genome. We also examined values of Garud et al.'s (2015) H 12 and H 2 /H 1 within windows classified by S/HIC as hard, soft, or neutral, noting that as expected, H 12 is higher in sweeps than neutral regions, while H 2 /H 1 is higher for soft sweeps than hard sweeps (supplementary fig. S2, Supplementary Material online). Below, we begin with a brief overview of the broad patterns of adaptation we observe across populations, before discussing genomic features and biological pathways with a strong enrichment of selective sweeps, as well as compelling novel candidates for recently completed selective sweeps.

The Majority of Sweeps in Humans Resemble Selection on Standing Variation
We found a total of 1,927 distinct selective sweeps merged across all six populations (see Materials and Methods). About 190 (9.9%) of these are present in all populations, 59 (3.1%) are shared among the African populations, 71 (3.7%) are shared among the non-African populations, and 701 (36.4%) are population-specific (supplementary table S1, Supplementary Material online). The remaining 906 (47.0%) sweeps were present in more than one population but do not fit into any of the categories above. We observe that across populations, the vast majority (1,776, or 92.2%) of sweeps were classified as soft, and note that this trend does not change qualitatively as we impose increasingly strict posterior probability thresholds before assigning a class label to a given window (supplementary table S2, Supplementary Material online; see Materials and Methods). These events may represent soft sweeps on standing genetic variants that our classifier was trained to detect, but we note that a similar signature can be created by a soft sweep resulting from recurrent origination of the adaptive allele(s), or by a de novo mutation that has been placed onto multiple haplotypes by allelic gene conversion events (see Discussion). Schrider and Kern . doi:10.1093/molbev/msx154 MBE Although hard sweeps appear to be quite rare globally, the fraction of hard sweeps is significantly higher in non-African than African populations (table 1). For example, when comparing PEL to GWD, we observe a significantly higher fraction of hard sweeps in PEL (4.7% versus 1.6%; P ¼ 0.05). For each other African versus non-African comparison we see an even greater (and more significant) disparity. Further, we observe a suggestive correlation between the fraction of sweeps in a population that were classified as soft and the harmonic mean of its population size within the last 4N generations (Pearson's q¼À0.96; see Materials and Methods). Though taken at face value this correlation appears to be highly significant, we note that due to the six populations' shared evolutionary history a statistical test of this correlation would be invalid.
Comparing our results to those of previous scans we find that 519 of S/HIC's sweep calls (26.9%) have previously been identified according to dbPSHP, a database of candidate regions for recent positive selection across human populations (Li et al. 2014). This accounts for 10.9% of the loci in the dbPSHP set (ignoring regions not classified by S/HIC). The remaining 1,408 sweeps called by S/HIC (73.1% of calls) represent potentially novel selective sweeps. There are several possible explanations for the modest overlap between our set of sweep candidates and those in dbPSHP. First, the sweep candidates in dbPSHP have been identified by a variety of methods, some of which are designed to detect selective scenarios other than completed sweeps (e.g., partial sweeps, spatially varying selection). Second, when comparing results from methods designed to detect the same type of sweeps, the intersection between studies is often fairly small (Akey 2009). Although most scans undoubtedly recover a large number of true selective sweeps, different methods may produce different false positives and false negatives, resulting in imperfect concordance between scans.

Selective Sweeps Preferentially Target Genes Involved in Cancer and Viral Infection
Examining the locations of selective sweeps across the genome, we find that regions classified as selective sweeps are significantly overrepresented for both coding sequence and untranslated regions (q < 0.05 in several populations for hard sweeps, and each population for putative soft sweeps; fig. 1A and B; supplementary table S3, Supplementary Material online), relative to data sets with permuted classifications (see Materials and Methods). Enrichment for transcription factor binding sites was less pronounced, and only significant in soft sweeps for the three African populations along with PEL.
The most striking result we observed was a dramatic enrichment of sweep windows for mutations in the COSMIC data set of somatic mutations that have been observed in cancer cells (Forbes et al. 2015) and may therefore play a role in tumor suppression/progression. Averaged across populations, the number of COSMIC mutations found in soft sweeps represents a 3.7-fold increase relative to that observed in permuted data sets; this enrichment was significant in each population, and peaked at 4.5-fold in PEL. For hard sweeps, this enrichment was 12-fold on an average, reaching as high as 21-fold in CEU, though this was the only population for which the enrichment was statistically significant. We also observed a sizeable overrepresentation of genes encoding virusinteracting proteins (VIPs) curated by (Enard et al. 2016) in soft sweeps, with a 1.9-fold increase relative to permuted sets (averaged across populations). VIPs show a similar magnitude of enrichment in hard sweeps for some populations, but does not achieve significance at q < 0.05.

Selective Sweeps Increase Linked Deleterious Variation
Because S/HIC not only detects selective sweeps, but also attempts to identify regions of the genome that appear to be linked to recent sweeps, our classifications allow us to examine the effect of linked selection in a principled way. We found that while a minority of genomic windows were classified as selective sweeps (7.6% on an average across all populations), a large fraction of windows were classified as linked to a completed selected sweep, either hard or soft (56.4% on an average). These estimates range from 41.5% in JPT to 74.0% in GWD ( fig. 2).
We also asked whether selective sweeps have a detectable impact on linked deleterious variation. As beneficial alleles increase in frequency in a population, they may carry along with them linked deleterious polymorphisms as hitchhikers, potentially increasing the frequency of deleterious variants over what would be expected given mutation-selectiondrift equilibrium (Birky and Walsh 1988;Hartfield and Otto 2011). To this end we asked whether relatively common candidate deleterious mutations were enriched in regions classified as either hard-linked or soft-linked. Indeed, we observed a fairly subtle but significant overrepresentation of SNPs with derived allele frequencies of at least 0.01 but predicted to be damaging by SIFT (Kumar et al. 2009) in both the hard-linked (mean enrichment across populations: 1.3-fold) and softlinked (mean enrichment: 1.1-fold) classes for most populations ( fig. 1C and D; supplementary table S3, Supplementary Material online). We find a similar enrichment in these MBE sweep-linked classes of common SNPs in regions inferred to be conserved across primates according to phastCons (Siepel et al. 2005). Phenotype-associated variants from the GWAS catalogue (Welter et al. 2014) were also significantly overrepresented sweep-linked regions in several populations ( fig. 1C and D).
Sexual Reproduction, the Central Nervous System, and Immunity Are Targets of Recent Sweeps In order to determine if positive selection preferentially acts on particular organismal functions, we asked which Gene Ontology (GO) terms were enriched in our sweep calls relative to the permuted data (see Materials and Methods). In soft sweeps, we found a sizeable and significant enrichment (q < 0.05) of terms related to sperm development, structure, and function. For example, "spermatogenesis" (4.4-fold enrichment averaged across populations), and "sperm-egg recognition" (3.9-fold enrichment on an average) were enriched in soft sweeps in several populations. We also observed an overrepresentation of genes involved in the "glutamate receptor signaling pathway" in our soft sweep sets for each population (4-fold mean enrichment). Glutamate receptors are the primary excitatory neurotransmitter in the central nervous system, and important for both proper brain development and function (Luj an et al. 2005). Indeed, soft sweeps are enriched for "central nervous system development" in multiple populations (1.6-fold mean enrichment). Numerous GO terms related to immune response, especially adaptive immunity, as well as KEGG pathways related to immunity and cancer progression/tumor suppression were also significantly enriched among soft sweeps (see supplementary  table S4, Supplementary Material online, for full list).

Positive Selection on Interacting Gene Pairs
We examined three types of gene interaction networks: protein-protein interactions (PPIs), transcription factor-target gene interactions, and genetic interactions where one gene modifies the effect of another (see Materials and Methods). Interestingly, we observed a dramatic enrichment of sweeps in genes that encode proteins that physically interact with one another ( fig. 3A and B): if a gene overlapped a window classified as a soft sweep, genes that interact with this gene were on an average 3.3 times more likely to overlap a putative soft sweep than expected by chance (P < 0.0001 for each population; fig.  3B). Despite the smaller number of candidate regions, we found a significant enrichment for PPIs in hard sweeps, though this was only significant in for non-African populations (4.0-fold enrichment averaged across populations; P < 0.05 in CEU, JPT, YRI; fig. 3A). For transcription factor-target interactions, we observe no overrepresentation of soft sweeps, but a significant enrichment of hard sweeps in non-African populations (P < 0.05 for each; 8.5-fold enrichment on an average; fig. 3C and D). There were no populations exhibiting an overrepresentation of pairs of genes with genetic interactions and experiencing sweeps of either type ( fig. 3E and F).
, who did not find any evidence of selection in Africa. However, S/HIC infers that this region has experienced a soft sweep that is found in GWD and YRI, as well as the non-African populations. Consistent with this, Europeans, Asians, and African populations show a reduction in p, a trough in Tajima's D (Tajima 1989), and a peak in Nielsen et al.'s SweepFinder composite likelihood ratio (CLR) test statistic, which captures regions that appear to be at the epicenter of the spatial skew in the SFS expected around sweeps (Nielsen, Williamson, et al. 2005). Intriguingly, GRIA2 interacts with the GRID2 glutamate receptor gene (Kohda et al. 2003), which itself is classified as a soft sweep in CEU, LWK, PEL, and GWD. The remaining glutamate receptors overlapping identified sweeps are GRIA4, GRID1, GRIK1, GRIK3, GRM2, and GRM7. Of these genes, GRIA4 and GRID2 were shown by Liu et al. (2012) to have evolved a human-specific developmental expression profile. Figure 4 shows a region on chromosome 9 that exhibits strong evidence of a previously undetected hard sweep in each of our six populations. This region contains several members of the spermatogenesis associated 31 gene family: SPATA31B1, SPATA31D1, SPATA31D3, and SPATA31D4. Across populations this region shows dramatic valleys in p and Tajima's D, as well as an elevated CLR near the center of the sweep window. These genes are highly testis-specific according to data from the GTEx project (Lonsdale et al. 2013), and male mice are infertile when lacking Spata31, another member of these gene family (Wu et al. 2015). Figure 4 also shows that each of these genes overlaps a cluster of nonrepetitive piRNAs (data from piRBase; Zhang et al. 2014). Also near this region is DDX10P2, which GENCODE annotates as a processed pseudogene (Pei et al. 2012). DDX10P2, which is located at the center of the CLR peak for CEU, is expressed with a high degree of testis-specificity according to GTEx data, similar to the neighboring SPATA31 genes. A BLAT search (Kent 2002) revealed that this putative pseudogene exhibits 99.5% sequence identity to the orthologous sequence in chimpanzees. The parent gene of DDX10P2, DDX10, is expressed in many tissues, but shows highest expression in the testis.
On chromosome 11 we detected what appear to be several novel soft sweeps present in and upstream of CADM1 (cell adhesion molecule 1; fig. 5), one of which is present in each population. This gene is essential for spermatogenesis in mice (Van Der Weyden et al. 2006), and is also a tumor suppressor that is hypermethelated in various cancers (Kuramochi et al. 2001;Allinen et al. 2002;Fukuhara et al. 2002), as it works with the adaptive immune system to suppress metastasis (Faraji et al. 2012). CADM1 is also active in the brain where it is involved in synaptic adhesion and has been linked to autism (Zhiling et al. 2008;Fujita et al. 2010). CADM1 forms a complex with two other genes: the GABA receptor GABBR2, which has a soft sweep in YRI, and MUPP1, which has a soft sweep found in each population; this complex appears to localize to Purkinje cell dendrites (Fujita et al. 2012). Thus, this example encompasses many of the functions that we find are highly enriched across our sweep sets:

Discussion
Understanding the history of human adaptation at the genetic level is a central goal of population genomics and human evolutionary biology. Accordingly, since the completion of the human genome assembly (Lander et al. 2001) and subsequent proliferation of population genomic data, numerous genome-wide scans for selection have been conducted using differing methodologies (Sabeti et al. 2002(Sabeti et al. , 2007Voight et al. 2006;Pickrell et al. 2009;Field et al. 2016). The majority of these studies searched primarily for partial selective sweepsthe signature of a beneficial mutation currently sweeping through a population (see Williamson et al. 2007 for a notable exception)-and rightly so, as these sweeps can reveal the targets of ongoing adaptation in human populations. However, because the sojourn of an adaptive mutation to fixation should be rapid (e.g., on the order of 400 generations, assuming N ¼ 10 4 and a moderately strong selection coefficient of s ¼ 0.05, and 4,000 generations for s ¼ 0.005), the success of efforts to detect ongoing selection implies the presence of a larger number of recently completed sweeps. We have therefore focused on completed sweeps in order to complement previous studies and to construct a more comprehensive catalogue of the loci underpinning recent human adaptation. Using a powerful and robust machine learning method that we have recently introduced (S/HIC; Schrider and Kern 2016) for finding completed selective sweeps, we performed a genome-wide search for the targets of recent positive selection in six human populations. Furthermore, we sought to determine the mode of positive selection, distinguishing between selection on de novo mutations and on previously standing variation.

Soft Sweeps Dominate Human Adaptation
Perhaps our most consequential result is the finding that the majority of our candidate sweeps resemble soft sweeps on standing variation. This result implies that adaptation in Adaptation via Soft Sweeps Is Prevalent in Humans . doi:10.1093/molbev/msx154 MBE humans may not be mutation-limited (Gillespie 1991;Karasov et al. 2010): rather than waiting for a novel mutation to arise, human populations may often be able to respond via selection on previously segregating polymorphisms, thereby more rapidly responding to novel environmental challenges. This may be surprising given the apparently small effective population size and low nucleotide diversity levels in humans. However, if the mutational target for the trait to be selected on is fairly large, then the probability of a population harboring a mutation affecting that trait may be appreciable.
While soft sweeps appear to be the dominant mode of selection globally, there is a significant increase in the proportion of putative hard sweeps in non-African populations relative to African populations. This is consistent with theoretical expectations, as larger populations have more standing variation for selection to act on (Hermisson and Pennings 2005). Moreover, the human migration out of Africa was associated with a severe population bottleneck (Marth et al. 2004;Fagundes et al. 2007). Soft selective sweeps may be "hardened" by a reduction in population size, which can result in the stochastic loss of some genetic backgrounds harboring the adaptive allele so that only a single haplotype reaches fixation (Wilson et al. 2014). Thus, though one might expect selection on segregating neutral or nearly neutral variation when a population enters a new environment with novel selective pressures, if the migration event is accompanied by a bottleneck then the population may experience a somewhat counterintuitive increase in the proportion of hard sweeps. Moreover, the causal relationship between population size and mode of adaptation may not be unidirectional. As Orr and Unckless (2014) have shown in the context of evolutionary rescue, when faced with a changing environment, a population which does not harbor standing variation that is beneficial may experience a more protracted decline in size while it waits for an adaptive de novo mutation.
Our genome-wide results amplify results of earlier studies that by design have tried to infer the mode of adaptation in a smaller number of targeted loci. For instance Peter et al.  (Peter et al. 2012) attempted to infer the mode of adaptation among 7 loci previously identified to be under selection in human populations. They report that half of the loci that they could confidently classify supported selection on standing variation. In Drosophila melanogaster, when looking among strong outliers of haplotype homozgosity, Garud et al. (2015) found that patterns of variation in those regions were consistent with recent soft selective sweeps. Our finding, that the vast majority of sweeps in human populations are soft sweeps, thus underscores the ubiquity of selection from standing variation in natural populations. Indeed it seems plausible that adaptation from standing variation might be the rule, rather than the exception. There are two caveats affecting our ability to discriminate between selection on standing variation and on de novo mutations. First, while we have trained our classifier to detect soft sweeps on previously segregating mutations, soft sweeps may also occur via recurrent mutation to the adaptive allele Hermisson 2006a, 2006b). Though there are some qualitative differences between these two models of soft sweeps (Berg and Coop 2015;, these are fairly subtle in comparison to the differences between the other models we consider. Thus, our classifiers may have sensitivity to both types of sweeps. If this is so, then some of the soft sweeps that we detect may result from recurrent mutation. Additionally, gene conversion during a sweep can transfer the adaptive mutation on to new genetic backgrounds (Jones and Wakeley 2008), thereby "softening" the sweep . This implies that selection on a single de novo mutation could sometimes appear to be a soft sweep in our classification. In any case, our finding that most sweeps in humans do not appear to be hard sweeps underscores the importance of using methods that are sensitive to soft sweeps.

Extensive Impact of Linked Positive Selection
Our analysis demonstrates that the impact of linked positive selection on genetic variation is considerable, with roughly half of the genome classified by S/HIC as being influenced by a nearby sweep. This result has important implications for efforts to infer demographic histories from patterns of genetic polymorphism, as most inference methods hinge on the assumption of neutrality. Indeed, we have recently shown that linked positive selection has the potential to severely confound demographic inferences . Similarly, Ewing and Jensen (2016) have found that background selection (Charlesworth et al. 1993) can also bias demographic estimates. One strategy is to use only those polymorphisms that are distant from genes and conserved noncoding elements to mitigate these effects (Gazave et al. 2014). One could further supplement such an approach by using S/HIC to directly ask which intergenic regions are unaffected by hitchhiking in order to further diminish the bias introduced by linked selection. We note that the putatively neutrally evolving regions found in this study can be obtained from our raw classification output (available at https://github.com/ kern-lab/shIC/tree/master/humanScanResults).
If linked positive selection affects much of the genome, then that implies that the frequencies of many neutral or weakly deleterious mutations may be altered by genetic draft (Gillespie 2000). That is to say, deleterious mutations that happen to reside on chromosomes that begin to sweep may be able to reach higher frequencies than expected from mutation-selection-drift equilibrium. Consistent with this, we observe a slight but significant excess of potentially deleterious polymorphisms in windows classified as linked to selective sweeps. Previously, Chun and Fay (2011) found evidence that the ratio of deleterious to neutral polymorphisms is elevated in sweep regions, concluding that hitchhiking carries linked deleterious variants to higher frequencies. Our finding that SNPs from the GWAS catalogue are also enriched regions linked to selective sweeps lends further support to this hypothesis. Indeed, several compelling examples of hitchhiking mutations known or suspected of causing disease have been described in the literature (Helgason et al. 2007;Chun and Fay 2011;Huff et al. 2012). Moreover it seems that the phenomenon of deleterious alleles hitchhiking along with strongly beneficial alleles is not restricted to humans: a recent study also uncovered evidence that selection during domestication increased the frequency of deleterious polymorphisms in dogs (Marsden et al. 2016).

Targets of Recent Human Selective Sweeps
Our catalogue of sweep candidates allowed us to characterize the biological functions that are overrepresented in sweeps. Notably, we found a strong excess of spermatogenesis genes within sweep regions, a phenomenon previously observed by Voight et al. (2006). This signature may be a result of sexual selection, sexual conflict, and/or sperm competition (Swanson and Vacquier 2002). We also observed a significant enrichment of cancer-related genes among our sweep candidates.  found a similar enrichment of candidate genes under selection related to cancer when examining protein divergence between humans and chimpanzees. These authors found that some of these genes are also involved in spermatogenesis (much like our CADM1 example), and concluded that genomic conflict between tumor suppression and the advantage of avoiding apoptosis during spermatogenesis may explain the selection on cancer genes. An alternative (and nonmutually exclusive) explanation is that the increase in longevity along the human lineage has created an immense selective pressure to reduce the rate of cancer progression by orders of magnitude (Nunney and Muir 2015).
We also observed a significant excess of glutamate receptor genes targeted by sweeps, suggesting that these loci may underlie some of the dramatic neurological changes that have occurred along the human lineage. Consistent with this, we previously found evidence suggesting some of these glutamate receptor genes (along with other neurotransmitters) may have recently gained novel regulatory elements in humans Meyer et al. 2017). The most striking examples of glutamate receptors experiencing sweeps are GRIA2 and GRID2, which show strong signatures of selection in multiple populations and physically interact with one Adaptation via Soft Sweeps Is Prevalent in Humans . doi:10.1093/molbev/msx154 MBE another. The action of positive selection on multiple members of the protein complex appears to be a general phenomenon ( fig. 3). For a more in-depth examination of positive selection in the PPI network, see Qian et al. (2015), who found that genes in candidate regions for positive selection were more likely to lie close together in the PPI network.

Conclusions
Our investigation has revealed several valuable insights into the adaptive process in human populations. The success of our approach exemplifies the potential of machine learning methods to elucidate the adaptive process in humans and other species (Fan et al. 2016). To date several machine learning methods have been devised to detect selective sweeps (Pavlidis et al. 2010;Lin et al. 2011;Ronen et al. 2013;Pybus et al. 2015;Sheehan and Song 2016), and they tend to substantially outperform more traditional approaches (see . We suspect that machine learning could be used to make important inroads in answering a variety of evolutionary questions. Finally, Hernandez et al. (2011) argued that hard selective sweeps might be rare in human populations, and instead suggested that the majority of adaptation might be a consequence of selection on standing variation or selection on polygenic traits. We here find direct evidence that indeed this is the case-the vast bulk of human adaptation is occurring as a consequence of soft sweeps. Our observation thus reconciles Hernandez et al.'s findings with those of Enard et al., who conclude that the reduction in diversity around amino acid substitutions is caused by widespread selective sweeps (Enard et al. 2014). Moreover, while our scan leveraged a method that performs very well in detecting both hard and soft sweeps, it was not trained to detect cases of polygenic selection (Berg and Coop 2014). It is fair to assume that a large majority of phenotypes are determined by multiple loci, thus polygenic selection should be expected to be common. If that were the case, then it could very well be that an even larger portion of genetic variation is influenced by natural selection and its linked effects throughout the genome.

Sequence and Annotation Data
We downloaded phased genotype data from Phase 3 of the 1000 Genomes Project (Auton et al. 2015). This data set consists of 26 population samples from Africa, East Asia, South Asia, Europe, and the Americas. We wished to include only populations where the influence of admixture/migration on genetic variation appeared to be minimal, while still allowing us to characterize selection across multiple continents. We therefore chose to scan the following populations for selective sweeps: the GWD (Gambians in Western Divisions in The Gambia) and YRI (Yoruba in Ibadan, Nigeria) populations from West Africa, LWK (Luhya in Webuye, Kenya) from East Africa, JPT (Japanese in Tokyo, Japan) from Asia, CEU (Utah residents with Northern and Western European Ancestry) from Europe, and PEL (Peruvians from Lima, Peru) from the Americas. Examining Auton et al.'s results from running ADMIXTURE (Alexander et al. 2009), we see that for most values of K, each of these populations appears to correspond primarily to a single ancestral population rather than displaying multiple clusters of ancestry (see Extended Data Figure 5 from Auton et al. 2015). One exception may be the PEL population, but among the highly admixed American samples it appears to exhibit the smallest amount of possible mixed ancestry (for most values of K), so we retained this population in order to have some representation from the Americas. We opted not to examine any South Asian population, as for each of these samples ADMIXTURE inferred evidence of ancestry from three or more ancestral populations.
We downloaded numerous annotation data sets containing genomic features to test for enrichment/depletion of selective sweeps and perform other downstream analyses. These included GENCODE gene model release 19 (Harrow et al. 2012) including pseudogenes (Pei et al. 2012), virusinteracting proteins from Enard et al. (2016), enhancers gained or along the human lineage since diverging from Old World monkeys (Cotney et al. 2013), and SIFT's (Kumar et al. 2009) predictions of damaging amino acid polymorphisms from dbNSFP version 3.2a (Liu et al. 2016). We obtained Gene Ontology (GO) annotations from ENSEMBL release 75 (Yates et al. 2016). We also downloaded coordinates of previously identified selective sweeps from dbPSHP (Li et al. 2014).
We used the UCSC Table Browser (Karolchik et al. 2004) to obtain the following data sets: phenotype-associated SNPs  Kanehisa et al. 2015). For each of these data sets we used GRCh37/hg19 coordinates.
In order to examine the prevalence of selective sweeps within interacting gene networks, we downloaded physical and genetic interactions from BioGRID version 3.4.136 (Chatr-Aryamontri et al. 2015). Our set of genetic interactions consisted of those annotated as "synthetic genetic interaction defined by inequality", "suppressive genetic interaction defined by inequality", or "additive genetic interaction defined by inequality". Physical interactions included those annotated as "direct interaction", "association", or "physical association". We extracted transcription factor-target interactions from ORegAnno (accessed Dec 22, 2015;Griffith et al. 2008), retaining only interacting pairs where the ENSEMBL gene identifier were provided for both genes in order to avoid ambiguity.

Building Classifiers to Detect Selective Sweeps
To detect sweeps we used S/HIC (https://github.com/kernlab/shIC), a machine learning approach we previously Schrider and Kern . doi:10.1093/molbev/msx154 MBE described and showed to be remarkably powerful and robust to nonequilibrium demography . Briefly, the S/HIC machine learning approach leverages spatial patterns (along a genome) of a variety of population genetic summary statistics to classify genomic windows as being the target of a completed hard sweep (hard), being closely linked to a hard sweep (hard-linked), a completed soft sweep (soft), linked to a soft sweep (soft-linked), or evolving neutrally (neutral). While this classification approach allows inference when considering a large number of features jointly, it necessitates training from a large number of data instances known to belong to each class. Because the number of genomic windows known to belong to each our five classes is limited, we must rely on simulation to generate our training data. To this end we used the program discoal  to simulate large chromosomal regions, subdivided into 11 subwindows. Training examples for the hard class experienced a hard sweep in the center of the central sub-window (i.e., the 6th window), while examples for the hard-linked class experienced a hard sweep in the center of one of the remaining sub-windows (selected randomly). Analogous simulations with soft sweeps were generated for the soft and softlinked classes, respectively. Finally, neutrally evolving examples did not experience any selective sweep.
We sought to train a classifier for each population under a demographic model that offers a better approximation to the population size history than the standard neutral model. For this we used Auton et al.'s (2015) population histories inferred by PSMC (Li and Durbin 2011). The 1000 Genomes Project's PSMC output did not contain estimates of h, the population mutation rate parameter. Thus for each population we conducted a grid search by simulating genomic windows with the appropriate sample size under each demographic model with varying values of h ¼ 4NuL (where L is the length of the locus, which we set to 100 kb); the grid of h values raged from 10 to 250, examining multiples of 10. For each value of h, we compared the values of p (Nei and Li 1979),ĥ w (Watterson 1975), h H (Fay and Wu 2000), H 2 /H 1 (Garud et al. 2015), and Z nS (Kelly 1997) from 1,000 simulations to those from 1,000 randomly selected genomic loci (calculated as described below), calculating the mean of each statistic in the real and simulated datasets. We chose as the final values of h that for which the sum of the percent deviations of the simulated from the observed means of each statistic was minimized. This estimate of h allowed us to calculate estimated population sizes and times scaled by the number generations for each time point in the history inferred by PSMC. The harmonic mean of each population's size was calculated by taking the estimated population size for each of the last 4N generations. We note that these models may not accurately capture the demographic histories of the populations we examined due to the confounding effects of positive  and negative (Ewing and Jensen 2016) selection. However, because of S/HIC's robustness to demographic misspecification, we do not expect this to severely impact our analysis .
For each population we simulated a total of 2,000 regions for each of our five classes. For simulations involving sweeps, we drew the selection coefficient from U(0.005, 0.1), the sweep completion time from U(0, 2000), the initial selected frequency for soft sweeps from U(1/N, 0.2). We drew values of h uniformly from a range spanning exactly one order of magnitude, specified so that the mean value of h was equal to that estimated for the population as described earlier. We drew recombination rates from an exponential distribution with mean 1 Â 10 À8 , truncated at triple the mean due to memory constraints. The simulation program discoal requires some of these parameters to be scaled by the present-day effective population size; we did this by taking the mean value of h and dividing by 4uL, where u was set to 1.2 Â 10 À8 (Kong et al. 2012). The full command lines we used to generate 1.1 Mb regions (to be subdivided into 11 windows each 100 kb in length) for each population are shown in supplementary table S5, Supplementary Material online. We also simulated 1,000 test examples for each population in the same manner as for the training data.
In order to address the potential for purifying and background selection to confound our classifiers, we simulated additional test sets of 1,000 genomic windows 1.1 Mb in length with varying arrangements of selected sites. In order to mimic patters of purifying/background selection expected in the human genome as closely as possible, for each of our 1,000 replicates we randomly selected a 1.1-Mb window from the human genome and asked which sites were found within either a GENCODE exon (Harrow et al. 2012) or within a phastCons (Siepel et al. 2005) conserved element from the UCSC Genome Browser's 100-way vertebrate alignment . Sites in the simulated chromosome corresponding to these functional elements in the human genome were labeled as "selected" in the simulations. In "selected" regions, 25% of all new mutations had no fitness effect, while the remaining 75% had a selection coefficient drawn from a gamma distribution with mean of À0.0294 and a shape parameter of 0.184 (the African model from Boyko et al. 2008). We limit fitness effects of new mutations to 75% in an effort to mimic coding regions of the genome. We note that this percentage may not be accurate for noncoding functional regions, though it is likely that some fraction of mutations in these regions is effectively neutral. All mutations outside of the selected regions were fitness-neutral. These simulations were performed for both our GWD and JPT demographic models using version 0.0.1 of the fwdpy11 (https://github. com/molpopgen/fwdpy11) forward population genetic simulator (Thornton 2014), using the same mutation rates, recombination rates, and history of instantaneous population size changes as used in our coalescent simulations described in supplementary table S5, Supplementary Material online. Feature vectors were then generated for each of these simulated test examples in the same manner as for our coalescent simulations. We also tested each population's classifier against test sets generated by discoal with different fixed values of h (but otherwise with the same parameterizations shown in supplementary table S5, Supplementary Material online) in order to ensure that our approach was robust to uncertainty in the estimate of this parameter (supplementary fig. S4, Supplementary Material online).
Adaptation via Soft Sweeps Is Prevalent in Humans . doi:10.1093/molbev/msx154 MBE Our feature vector for each simulated region examined the spatial patters (following Schrider and Kern 2016) of each of the following statistics: p (Nei and Li 1979),ĥ w (Watterson 1975),ĥ H (Fay and Wu 2000), the number of distinct haplotypes, average haplotype homozygosity, Garud et al.'s (2015) H 12 and H 2 /H 1 statistics, Z nS , x (Kim and Nielsen 2004), and the maximum frequency of derived mutations (Li 2011). Before calculating these summary statistics we masked a number of sites within each simulation by randomly selecting a 1.1-Mb region from our empirical windows sampled throughout the genome and masking the same regions in the simulated window as were masked in the genomic window (see below). Thus our simulated windows exhibit the same distribution of regions of missing data as the windows to which we applied our classifiers. We then used S/HIC to train extra-trees classifiers (Geurts et al. 2006), one for each population.

Classifying Genomic Windows in Each Population
Having trained our classifiers, we then applied them to genomic data from the corresponding population. We inferred ancestral states of polymorphisms and masked inaccessible sites (whether polymorphic or not) in the same manner as described previously . We then used S/HIC to classify the central 100-kb sub-window of 1.1-Mb windows across the autosomes, while taking the stringent approach of omitting those for which any sub-window was <25% accessible, before sliding 100-kb downstream to examine the next window. We also removed windows where any of the three central sub-windows had a mean recombination rate of zero (using data from Kong et al. 2010). Importantly, for each retained 1.1-Mb window, we recorded the locations of all sites deemed inaccessible for use in masking our training data (see above). In total we classified 13,968 windows, accounting for 48.5% of the assembled autosomes. For our classifications we simply took the class that S/HIC's classifier inferred to be the most likely one, but we also used S/HIC's posterior class membership probability estimates in order to experiment with different confidence thresholds (results shown in supplementary table S2, Supplementary Material online). For a given threshold, we required the sum of a windows' hard and soft sweep posterior probabilities to be greater than or equal to the threshold before labeling the window as a sweep; the mode of the sweep was that corresponding to the greater posterior probability among the hard and soft sweep classes.
In order to count the number of distinct sweep candidates found within our set of populations, we simply merged all 100-kb windows classified as a sweep of either type that were located either at the exact same coordinates or adjacent to one another, repeating this until no more sweep regions could be merged. If all constituent windows were classified as soft, we counted the sweep as soft; otherwise we counted it as a hard sweep. We used a similar approach but examining classifications from only one population at a time in order to count the number of sweeps of each type in that population. If a gene found within a sweep window identified by S/HIC was not found in an entry of dbPSHP (Li et al. 2014), we referred to it as a novel sweep. Visualization of sweep candidates was performed using the UCSC Genome Browser , along with custom tracks showing values of various population genetic summary statistics and selection scan scores for the CEU, YRI, and JPT populations from the Human Positive Selection Browser (Pybus et al. 2014). Our classification results are available at https://github.com/kern-lab/shIC/ tree/master/humanScanResults.

Permutation Tests for Enrichment of Annotation Features in Sweeps
To determine whether certain annotation features were enriched within any of our five classes, we carefully designed a permutation test to account for the subset of the genome that we examined with S/HIC, as well as the spatial correlation of S/HIC's classifications (i.e., adjacent windows are especially likely to receive the same classification). Briefly, the permutation algorithm begins by examining our classification results for a given population and keeping track of the length of runs of consecutive windows assigned to each class. The permutation algorithm then selects a chromosome, and begins at its first classified window (i.e., not removed by data filtering). A run length and associated class assignment is then randomly drawn without replacement. This process continues until the end of the chromosome, and then another chromosome is selected until the end of the final chromosome is reached, at which point the permutation has been completed. We then repeated this permutation procedure 10,000 times for each population. Note that this process preserves the run length distribution of our classifications while permuting them across the set of genomic windows that had enough unmasked data to be included in our scan.
After constructing our permuted data sets, we conducted one-sided enrichment tests by counting the number of base pairs in the intersection between the S/HIC class of interest and the annotation feature of interest, and comparing this number to its distribution among the permuted data sets. The fraction of permuted data sets where this intersect was greater than or equal to that observed for the real data is the P value. Because we tested each of S/HIC's five classes for enrichment of a fairly large number of genomic features (supplementary table S3, Supplementary Material online), we corrected for multiple testing using false discovery rate q values following Storey (2002). When testing GO terms and KEGG pathways for enrichment, we considered only the hard and soft sweep classes, corrected for calculating q values separately for each class.
We also asked whether the number of pairs of interacting genes both overlapping windows classified as sweeps was greater than in our permuted data sets. To ensure that our results were not inflated by the spatial clustering of interacting genes, we only counted interacting pairs overlapping sweep windows if they were separated by at least 10 Mb or on separate chromosomes. In addition, if we observed an interaction between two genes, A and B, that each overlapped sweeps, and a third sweep candidate gene, C, was found, to avoid redundancy we counted at most one interaction between A and C and B and C, even if C was found interact with Schrider and Kern . doi:10.1093/molbev/msx154 MBE both other genes. As with GO and KEGG terms, we only searched the hard and soft classes for enrichments before calculating one-sided q values as described earlier.

Supplementary Material
Supplementary data are available at Molecular Biology and Evolution online.