1 Abstract
A powerful way to detect selection in a population is by modeling local allele frequency changes in a particular region of the genome under scenarios of selection and neutrality, and finding which model is most compatible with the data. Chen et al. [1] developed a composite likelihood method called XP-CLR that uses an outgroup population to detect departures from neutrality which could be compatible with hard or soft sweeps, at linked sites near a beneficial allele. However, this method is most sensitive to recent selection and may miss selective events that happened a long time ago. To overcome this, we developed an extension of XP-CLR that jointly models the behavior of a selected allele in a three-population tree. Our method - called 3P-CLR - outperforms XP-CLR when testing for selection that occurred before two populations split from each other, and can distinguish between those events and events that occurred specifically in each of the populations after the split. We applied our new test to population genomic data from the 1000 Genomes Project, to search for selective sweeps that occurred before the split of Africans and Eurasians, but after their split from Neanderthals, and that could have presumably led to the fixation of modern-human-specific phenotypes. We also searched for sweep events that occurred in East Asians, Europeans and the ancestors of both populations, after their split from Africans.
2 Introduction
Genetic hitchhiking will distort allele frequency patterns at regions of the genome linked to a beneficial allele that is rising in frequency [2]. This is known as a selective sweep. If the sweep is restricted to a particular population and does not affect other closely related populations, one can detect such an event by looking for extreme patterns of localized population differentation, like high values of Fst at a specific locus [3]. This and other related statistics have in fact been used to scan the genomes of present-day humans from different populations, so as to detect signals of recent positive selection [4-7].
Once it became possible to sequence entire genomes of archaic humans (like Neanderthals) [8-10], researchers also began to search for selective sweeps that occurred in the ancestral population of all present-day humans. For example, ref. [8] searched for genomic regions with a depletion of derived alleles in a low-coverage Neanderthal genome, relative to what would be expected given the derived allele frequency in present-day humans. This is a pattern that would be consistent with a sweep in present-day humans. Later on, ref. [10] developed a hidden Markov model (HMM) that could identify regions where Neanderthals fall outside of all present-day human variation (also called “external regions”), and are therefore likely to have been affected by ancient sweeps in early modern humans. They applied their method to a high-coverage Neanderthal genome. Then, they ranked these regions by their genetic length, to find segments that were extremely long, and therefore highly compatible with a selective sweep. Finally, ref. [11] used summary statistics calculated in the neighborhood of sites that were ancestral in archaic humans but fixed derived in all or almost all present-day humans, to test if any of these sites could be compatible with a selective sweep model. While these methods harnessed different summaries of the patterns of differentiation left by sweeps, they did not attempt to explicitly model the process by which these patterns are generated over time.
Chen et al. [1] developed a method called XP-CLR, which is designed to test for selection in one population after its split from a second, outgroup, population tAB generations ago. It does so by modeling the evolutionary trajectory of an allele under linked selection and under neutrality, and then comparing the likelihood of the data under each of the two models. The method detects local allele frequency differences that are compatible with the linked selection model [2], along windows of the genome.
XP-CLR is a powerful test for detecting selective events restricted to one population. However, it provides little information about when these events happened, as it models all sweeps as if they had immediately occurred in the present generation. Additionally, if one is interested in selective sweeps that took place before two populations a and b split from each other, one would have to run XP-CLR separately on each population, with a third outgroup population c that split from the ancestor of a and b tABC generations ago (with tABC > tAB). Then, one would need to check that the signal of selection appears in both tests. This may miss important information about correlated allele frequency changes shared by a and b, but not by c, limiting the power to detect ancient events.
To overcome this, we developed an extension of XP-CLR that jointly models the behavior of an allele in all 3 populations, to detect selective events that occurred before or after the closest two populations split from each other. Below we briefly review the modeling framework of XP-CLR and describe our new test, which we call 3P-CLR. In the Results, we show this method outperforms XP-CLR when testing for selection that occurred before the split of two populations, and can distinguish between those events and events that occurred after the split, unlike XP-CLR. We then apply the method to population genomic data from the 1000 Genomes Project [12], to search for selective sweep events that occurred before the split of Africans and Eurasians, but after their split from Neanderthals. We also use it to search for selective sweeps that occurred in the Eurasian ancestral population, and to distinguish those from events that occurred specifically in East Asians or specifically in Europeans.
3 Methods
3.1 XP-CLR
First, we review the procedure used by XP-CLR to model the evolution of allele frequency changes of two populations a and b that split from each other tAB generations ago (Figure 1.A). For neutral SNPs, Chen et al. [1] use an approximation to the Wright-Fisher diffusion dynamics [13]. Namely, the frequency of a SNP in a population a (pA) in the present is treated as a random variable governed by a normal distribution with mean equal to the frequency in the ancestral population (β) and variance proportional to the drift time ω from the ancestral to the present population: where ω = tAB/ (2Ne) and Ne is the effective size of population A.
If a SNP is segregating in both populations - i.e. has not hit the boundaries of fixation or extinction - this process is time-reversible. Thus, one can model the frequency of the SNP in population a with a normal distribution having mean equal to the frequency in population b and variance proportional to the sum of the drift time (ω) between a and the ancestral population, and the drift time between b and the ancestral population (Ψ):
For SNPs that are linked to a beneficial allele that has undergone a sweep in population a only, Chen et al. [1] model the allele as evolving neutrally until the present and then apply a transformation to the normal distribution that depends on the distance to the selected allele r and the strength of selection s [14,15]. Let where q0 is the frequency of the beneficial allele in population A before the sweep begins. The frequency of a neutral allele is expected to increase from p to l — c + cp if the allele is linked to the beneficial allele, and this occurs with probability equal to the neutral allele (p) before the sweep begins. Otherwise, the frequency of the neutral allele is expected to decrease from p to cp. This leads to the following transformation of the normal distribution: where σ2=(ω+ψ)pb(1-pb)and I[x,y] (z) is 1 on the interval [x,y] and 0 otherwise.
For s → 0 or r>>s, this distribution converges to the neutral case. Let v be the vector of all drift times that are relevant to the scenario we are studying. In this case, it will be equal to (ω, Ψ) but in more complex cases below, it may include additional drift times. Let r be the vector of recombination fractions between the beneficial alleles and each of the SNPs within a window of arbitrary size. We can then calculate the product of likelihoods over all k SNPs in that window for either the neutral or the linked selection model, after binomial sampling of alleles from the population frequency and conditioning on the event that the allele is segregating in the population:
We note that the denominator in the above equation is not explicitly stated in ref. [1] for ease of notation, but appears in the published online implementation of the method. Because we are ignoring the correlation in frequencies produced by linkage, this is a composite likelihood [16,17]. Finally, we obtain a composite likelihood ratio statistic SXP-CLR of the hypothesis of linked selection over the hypothesis of neutrality:
For ease of computation, Chen et al. [1] assume that r is given (via a recombination map) and we will do so too. Furthermore, they empirically estimate v using F2 statistics [18] calculated over the whole genome, and assume selection is not strong or frequent enough to affect their genome-wide values. Because we are interested in selection over long time scales, the new methods we will present below are optimally run using drift times calculated from population split times and effective population sizes estimated using model-based demographic inference methods, like ∂a∂i [19] or fastsimcoal2 [20].
3.2 3P-CLR
We are interested in the case where a selective event occurred more anciently than the split of two populations (a and b) from each other, but more recently than their split from a third population c (Figure 1.B). We begin by modeling pA and pB as evolving from an unknown common ancestral frequency β:
Let χ be the drift time separating the most recent common ancestor of a and b from the most recent common ancestor of a, b and c. Additionally, let v be the drift time separating population c in the present from the most recent common ancestor of a, b and c. Given these parameters, we can treat β as an additional random variable that either evolves neutrally or is linked to a selected allele that swept immediately more anciently than the split of a and b. In both cases, the distribution of β will depend on the frequency of the allele in population c (pC) in the present. In the neutral case:
In the linked selection case: where K2 = (v + χ)pc (1 - pc)
The frequencies in a and b given the frequency in c can be obtained by integrating β out. This leads to a density function that models selection in the ancestral population of a and b.
Additionally, formula 10 can be modified to test for selection that occurred specifically in one of the terminal branches that lead to a or b (Figures 1.C and 1.D), rather than in the ancestral population of a and b. For example, the density of frequencies for a scenario of selection in the branch leading to a can be written as:
We will henceforth refer to the version of 3P-CLR that is tailored to detect selection in the internal branch that is ancestral to a and b as 3P-CLR(Int). In turn, the versions of 3P-CLR that are designed to detect selection in each of the daughter populations will be designated as 3P-CLR(A) and 3P-CLR(B).
We can now calculate the probability density of specific allele frequencies in populations a and b, given that we observe mC derived alleles in a sample of size nC from population c:
Where B(x,y) is the Beta function and
Conditioning on the event that the site is segregating in the population, we can then calculate the probability of observing mA and mB derived alleles in a sample of size nA from population a and a sample of size nB from population b, respectively, given that we observe mC derived alleles in a sample of size nC from population c, using binomial sampling:
Where and
This allows us to calculate a composite likelihood of the derived allele counts in a and b given the derived allele counts in c:
As before, we can use this composite likelihood to produce a composite likelihood ratio statistic that can be calculated over regions of the genome to test the hypothesis of linked selection centered on a particular locus against the hypothesis of neutrality. Due to computational costs in numerical integration, we skip the sampling step for population c (formula 13) in our implementation of 3P-CLR. In other words, we assume pC= mC/nC, but this is also assumed in XP-CLR when computing its corresponding outgroup frequency. We implemented our method in a freely available C++ program that can be downloaded from here:
https://github.com/ferracimo [WILL POST IT AFTER PUBLICATION]
4 Results
4.1 Simulations
We generated simulations in SLiM [21] to test the performance of XP-CLR and 3P-CLR in a three-population scenario. We focused specifically on the performance of 3P-CLR(Int) in detecting ancient selective events that occurred in the ancestral branch of two sister populations. We assumed that the population history had been correctly estimated by the researcher (i.e. the drift parameters and population topology were known). First, we simulated scenarios in which a beneficial mutation arose in the ancestor of populations a and b, before their split from each other but after their split from c (Table 1). Although both XP-CLR and 3P-CLR are sensitive to partial or soft sweeps (as they do not rely on extended patterns of homozygosity[1]), we required the allele to have fixed before the split (at time tab) to ensure that the allele had not been lost before it, and also to ensure that the sweep was restricted to the internal branch of the tree. We fixed the effective size of all three populations at Ne = 10,000. Each simulation consisted in a 5 cM region and the beneficial mutation occurred in the center of this region. The mutation rate was set at 2.5 *10-8 per generation and the recombination rate was set at 10-8 per generation.
To make a fair comparison to 3P-CLR(Int), and given that XP-CLR is a two-population test, we applied XP-CLR in two ways. First, we pretended population b was not sampled, and so the “test” panel consisted of individuals from a only, while the “outgroup” consisted of individuals from c. In the second implementation (which we call “XP-CLR-avg”), we used the same outgroup panel, but pretended that individuals from a and b were pooled into a single panel, and this pooled panel was the “test”. The window size was set at 0.5 cM and the space between the center of each window was set at 600 SNPs. To speed up computation, and because we are largely interested in comparing the relative performance of the three tests under different scenarios, we used only 20 randomly chosen SNPs per window in all tests. We note, however, that the performance of all three tests can be improved by using more SNPs per window.
Figure 2 shows receiver operating characteristic (ROC) curves comparing the sensitivity and specificity of 3P-CLR(Int), XP-CLR and XP-CLR-avg in the first six demographic scenarios described in Table 1. Each ROC curve was made from 100 simulations under selection (with s = 0.1 for the central mutation) and 100 simulations under neutrality (with s = 0 and no fixation required). In each simulation, 100 individuals were sampled from population a, 100 from population b and 10 from the outgroup population c. This emulates a situation in which only a few individuals have been sequenced from the outgroup, while large numbers of sequences are available from the tests (e.g. two populations of present-day humans). For each simulation, we took the maximum value at a region in the neighborhood of the central mutation (+/- 0.5 cM) and used those values to compute ROC curves under the two models.
When the split times are recent or moderately ancient (models A to D), 3P-CLR(Int) outperforms the two versions of XP-CLR. When the split times are very ancient (models E and F), none of the tests perform well. The root mean squared error (RMSE) of the genetic distance between the true selected site and the highest scored window is comparable across tests in all six scenarios (Figure S2). Finally, Figures S1 and S3 show the ROC curves and RMSE plots, respectively, for a case in which 100 individuals were sampled from all three populations (including the outgroup), with similar results.
Importantly, the usefulness of 3P-CLR(Int) resides not just in its performance at detecting selective sweeps in the ancestral population, but in its specific sensitivity to that particular type of events. Because the test relies on correlated allele frequency differences in both population a and population b (relative to the outgroup), selective sweeps that are specific to only one of the populations will not lead to high 3P-CLR(Int) scores. Figure 3 shows ROC curves in two scenarios in which a selective sweep event occurred only in population a (Models I and J in Table 1), using 100 sampled individuals from each of the 3 populations. Here, XP-CLR performs well, but 3P-CLR(Int) shows almost no sensitivity to the recent sweep, under reasonable specificity cutoffs. For example, in Model I, at a specificity of 95%, XP-CLR has 80% sensitivity, while at the same specificity, 3P-CLR(Int) only has 14% sensitivity. One can compare this to the same demographic scenario but with selection occurring in the ancestral population (Model C, Figure S1), where at 95% specificity, XP-CLR has 69% sensitivity, while 3P-CLR has 83% sensitivity.
4.2 Selection in Eurasians
We first applied 3P-CLR to modern human data from the 1000 Genomes Project [12]. We used the African-American recombination map [22] to convert physical distances into genetic distances. We focused on two populations (Europeans and East Asians), using Africans as the outgroup population (Figure S4.A). We randomly sampled 100 individuals from each population and obtained sample derived allele frequencies every 10 SNPs in the genome. We then calculated likelihood ratio statistics by a sliding window approach, where we sampled a “central SNP” once every 20 SNPs. The central SNP in each window was the candidate beneficial SNP for that window. We set the window size to 0.25 cM, and randomly sampled 100 SNPs from each window, centered around the candidate beneficial SNP. In each window, we calculated 3P-CLR to test for selection at three different branches of the population tree: the terminal branch leading to Europeans (3P-CLR Europe), the terminal branch leading to East Asians (3P-CLR East Asia) and the ancestral branch of Europeans and East Asians (3P-CLR Eurasia). Results are shown in Figure 4. For each scan, we selected the windows in the top 99.9% quantile of scores and merged them together if they were contiguous. Tables 2,3 and 4 show the top hits for Europeans, East Asians and the ancestral Eurasian branch, respectively
We observe several genes that have been identified in previous selection scans. In the East Asian branch, one of the top hits is EDAR. This gene codes for a protein involved in hair thickness and incisor tooth morphology [23,24]. It has been repeatedly identified in earlier selections scans as having undergone a sweep in East Asians [25,26].
Furthermore, 3P-CLR allows us to narrow down on the specific time at which selection occurred in the history of particular populations. For example, ref. [1] performed a scan of the genomes of East Asians using XP-CLR with Africans as the outgroup, and identified a number of genes as being under selection [1]. 3P-CLR confirms this signal in several of these loci when looking specifically at the East Asian branch: CYP26B1, EMX1, SPR, SFXN5, SLC30A9, PPARA, PKDREJ, GTSE1, TRMU, CELSR1, PINX1, XKR6, CD226, ACD, PARD6A, GFOD2, RANBP10, TSNAXIP1, CENPT, THAP11, NUTF2, CDH16, RRAD, FAM96B, CES2, CBFB, C16orf70, TRADD, FBXL8, HSF4, NOL3, EXOC3L1, E2F4, ELMOS, LRRC29, FHOD1, SLC9A5, PLEKHG4, LRRCS6, ZDHHC1, HSD11B2, AtP6V0D1, AGRP, FAM65A, CTCF and RLTPR. However, when applied to the ancestral Eurasian branch, 3P-CLR finds some genes that were previously found in the XP-CLR analysis of East Asians, but that are not among the top hits in 3P-CLR applied to the East Asian branch: COMMD3, BMI1, SPAG6, CD226, SLCS0A9,LONP2, SIAH1, ABCC11 and ABCC12. This suggests selection in these regions occurred earlier, i.e. before the European-East Asian split. Figure 5 shows a comparison between the 3P-CLR scores for the three branches in the region containing genes BMI1 (a proto-oncogene [27]) and SPAG6 (involved in sperm motility [28]). In that figure, the score within each window was standardized using its chromosome-wide mean and standard deviation, to make a fair comparison. One can observe that the signal of Eurasia-specific selection is evidently stronger than the other two signals.
Other selective events that 3P-CLR infers to have occurred in Eurasians include the region containing HERC2 and OCA2, which are major determinants of eye color [29-31]. There is also evidence that these genes underwent selection more recently in the history of Europeans [32], which could suggest an extended period of selection - perhaps influenced by migrations between Asia and Europe - or repeated selective events at the same locus.
When running 3P-CLR to look for selection specific to Europe, we find that TYRP1 and MYO5A, which play a role in human skin pigmentation [33-36], are among the top hits. Both of these genes have been previously found to be under strong selection in Europe [37], using a statistic called iHS, which measures extended patterns of homozygosity that are characteristic of selective sweeps. Interestingly, a change in the gene TYRP1 has also been found to cause a blonde hair phenotype in Melanesians [38].
4.3 Selection in ancestral modern humans
We applied 3P-CLR to modern human data combined with recently sequenced archaic human data [10]. We sought to find selective events that occurred in modern humans after their spit from archaic groups. We used the combined Neanderthal and Denisovan high-coverage genomes [9,10] as the outgroup population, and, for our two test populations, we randomly sampled 100 Eurasian genomes and 100 African genomes from the 1000 Genomes data (Figure S4.B). We used previously estimated drift times as fixed parameters [10], and tested for selective events that occurred more anciently than the split of Africans and Eurasians, but more recently than the split from Neanderthals. We run 3P-CLR using 0.25 cM windows as above, but also verified that the density of scores was robust to the choice of window size and spacing (Figure S5). As before, we selected the top 99.9% windows and merged them together if they were contiguous. Table 5 and Figure S6 show the top hits. To find putative candidates for the beneficial variants in each region, we queried the catalogs of modern human-specific high-frequency or fixed derived changes that are ancestral in the Neanderthal and/or the Denisova genomes [10,39].
We observe several genes that have been identified in previous scans that looked for selection in modern humans after their split from archaic groups [8,10]: SIPA1L1, ANAPC10, ABCE1, RASA1, CCNH, KCNJ3, HBP1, COG5, GPR22, DUS4L, BCAP29, CALDPS2, RNF133, RNF148, FAM172A, POU5F2, FGF7, RABGAP1, GPR21, STRBP, SMURF1, GABRA2, ALMS1, PVRL3, EHBP1, VPS54, OTX1, UGP2, HCN1, GTDC1, ZEB2, OIT3, USP54, MYOZ1 and DPYD. One of our strongest candidate genes among these is ANAPC10. This gene is a core subunit of the cyclosome, is involved in progression through the cell cycle [40], and may play a role in oocyte maturation and human T-lymphotropic virus infection (KEGG pathway [41]). ANAPC10 is noteworthy because it was found to be significantly differentially expressed in humans compared to other great apes and macaques: it is up-regulated in the testes [42]. The gene also contains two intronic changes that are fixed derived in modern humans, ancestral in both Neanderthals and Denisovans and that have evidence for being highly disruptive, based on a composite score that combines conservation and regulatory data (PHRED-scaled C-scores > 11 [10,43]). The changes, however, appear not to lie in any obvious regulatory region [44,45].
We also find ADSL among the list of candidates. This gene is known to contain a nonsynonymous change that is fixed in all present-day humans but homozygous ancestral in the Neanderthal genome, the Denisova genome and two Neanderthal exomes [39] (Figure 6.A). It was previously identified as lying in a region with strong support for positive selection in modern humans, using summary statistics Implemented in an ABC method[11]. The gene is interesting because it is one of the members of the Human Phenotype ontology category “aggression hyperactivity” which is enriched for nonsynonymous changes that occurred in the modern human lineage after the split from archaic humans [39,46]. ADSL codes for adenylosuccinase, an enzyme involved in purine metabolism [47]. A deficiency of adenylosuccinase can lead to apraxia, speech deficits, delays in development and abnormal behavioral features, like hyperactivity and excessive laughter [48]. The nonsynonymous mutation (A429V) is in the C-terminal domain of the protein (Figure6.B) and lies in a highly conserved position (primate PhastCons = 0.953; GERP score = 5.67 [43.49.50). The ancestral amino acid is conserved across the tetrapod phylogeny, and the mutation is only three residues away from the most common causative SNP for severe adeny-losuccinase deficiency[51_55]. The change has the highest probability of being disruptive to protein function, out of all the nonsynonymous modern-human-specific changes that lie in the top-scoring regions (C-score = 17.69). While ADSL is an interesting candidate and lies in the center of the inferred selected region (Figure 6.A), there are other genes in the region too, including TNRC6B and MKL1. TNRC6B may be involved in miRNA-guided gene silencing [56], while MKL1 may play a role in smooth muscle differentiation[57], and has been associated with acute megakaryocytic leukemia [58].
RASA1 was also a top hit in a previous scan for selection [8], and was additionally inferred to have a high Bayes factor in favor of selection in ref. [11]. The gene codes for a protein involved in the control of cellular differentiation [59]. Human diseases associated with RASA1 include basal cell carcinoma [60] and ateriovenous malformation [61,62].
The GABAA gene cluster in chromosome 4p12 is also among the top regions. The genes within the putatively selected region code for three of the subunits of the GABAA receptor (GABRA2, GABRA4, GABRB1), which codes for a ligand-gated ion channel that plays a key role in synaptic inhibtion in the central nervous system (see review by ref. [63]). GABRA2 is significantly associated with the risk of alcohol dependence in humans [64], perception of pain [65] and asthma [66]. In turn, GABRA4 is associated with autism risk [67,68].
Two other candidate genes that may be involved in brain development are FOXG1 and CADPS2. FOXG1 was not identified in any of the previous selection scans, and codes for a protein called forkhead box G1, which plays an important role during brain development. Mutations in this gene have been associated with a slow-down in brain growth during childhood resulting in microcephaly, which in turn causes various intellectual disabilities [69,70]. CADPS2 was identified in [8] as a candidate for selection, and has been associated with autism [71]. The gene has been suggested to be specifically important in the evolution of all modern humans, as it was not found to be selected earlier in great apes or later in particular modern human populations [72].
Finally, we find a signal of selection in a region containing the gene EHBP1 and OTX1. This region was identified in both of the two previous scans for modern human selection [8,10]. EHBP1 codes for a protein involved in endocytic trafficking [73] and has been associated with prostate cancer [74]. OTX1 is a homeobox family gene that may play a role in brain development [75]. Interestingly, EHBP1 contains a single-nucleotide intronic change (chr2:63206488) that is almost fixed in all present-day humans and homozygous ancestral in Neanderthal and Denisova [10]. This change is also predicted to be highly disruptive (C-score = 13.1) and lies in a position that is extremely conserved across primates (PhastCons = 0.942), mammals (PhastCons = 1) and vertebrates (PhastCons = 1). The change is 18 bp away from the nearest splice site and overlaps a VISTA conserved enhancer region (element 1874) [76], which suggests a putative regulatory role for the change.
4.4 Modern human-specific high-frequency changes in GWAS catalog
We overlapped the genome-wide association studies (GWAS) database [77,78] with the list of fixed or high-frequency modern human-specific changes that are ancestral in archaic humans [10] and that are located within our top putatively selected regions in modern humans (Table 6). None of the resulting SNPs are completely fixed derived, because GWAS can only yield associations from sites that are segregating. Among these SNPs, the one with the highest probability of being disruptive (rs10003958, C-score = 16.58, Gerp score = 6.07) is located in a highly-conserved regulatory (“strong enhancer”) region in the RAB28 gene [44,45] (Primate PhastCons = 0.951), and is significantly associated with obesity [79] (Figure 7.A). Interestingly, the region containing RAB28 is inferred to have been under positive selection in both the modern human and the Eurasian ancestral branches (Tables 4,5). In line with this evidence, the derived allele of rs10003958 is absent in archaic humans, at very high frequencies in Eurasians (> 94%), and only at moderately high frequencies in Africans (74%) (Figure 7.B).
We also find a highly disruptive SNP (rs10171434, C-score = 8.358) associated with urinary metabolites [80] and suicidal behavior in patients with mood disorders [81]. The SNP is located in an enhancer regulatory freature [44,45] located between genes PELI1 and VPS54, in the same putatively selected region as genes EHBP1 and OTX1 (see above). Finally, there is a highly disruptive SNP (rs731108, C-score = 10.31) that is associated with renal cell carcinoma [82]. This SNP is also located in an enhancer regulatory feature [44,45], in an intron of ZEB2. In this last case, though, only the Neanderthal genome has the ancestral state, while the Denisova genome carries the modern human variant.
5 Discussion
We have developed a new method called 3P-CLR, which allows us to detect positive selection along the genome. The method is based on an earlier test (XP-CLR [1]) that uses linked allele frequency differences between two populations to detect population-specific selection. However, 3P-CLR can allow us to distinguish between selective events that occurred before and after the split of two populations. Our method also has some similiarities to an earlier method developed by [83], which used an Fst-like score to detect selection ancestral to two populations. In that case, though, the authors used summary statistics and did not explicitly model the process leading to allele frequency differentiation.
We used our method to confirm previously found candidate genes in particular human populations, like EDAR, TYRP1 and HERC2, and find some novel candidates too (Tables 2,3,4). Additionally, we can infer that certain genes, which were previously known to have been under selection in East Asians (like SPAG6), are more likely to have undergone a sweep in the population ancestral to both Europeans and East Asians than in East Asians only.
We also used 3P-CLR to detect selective events that occurred in the ancestors of modern humans, after their split from Neanderthals and Denisovans (Table 5). These events could perhaps have led to the spread of phenotypes that set modern humans apart from other hominin groups. We find several intersting candidates, like SIPA1L1, ADSL, RASA1, OTX1, EHBP1, FOXG1, RAB28 and ANAPC10, some of which were previously detected using other types of methods [8,10,11].
An advantage of differentiation-based tests like XP-CLR and 3P-CLR is that, unlike other patterns detected by tests of neutrality (like extended haplotype homozygostiy, [84]) that are exclusive to hard sweeps, the patterns that both XP-CLR and 3P-CLR are tailored to find are based on regional allele frequency differences between populations. These patterns can also be produced by soft sweeps from standing variation or by partial sweeps [1], and there is some evidence that the latter phenomena may have been more important than classic sweeps during human evolutionary history [85].
Another advantage of both XP-CLR and 3P-CLR is that they do not rely on an arbitrary division of genomic space. Unlike other methods which require the partition of the genome into small windows of fixed size, our composite likelihood ratios can theoretically be computed over windows that are as big as each chromosome, while only switching the central candidate site at each window. This is because the likelihood ratios use the genetic distance to the central SNP as input. SNPs that are very far away from the central SNP will not contribute much to the likelihood function of both the neutral and the selection models, while those that are close to it will. While we heuristically limit the window size in our implementation in the interest of speed, this can be arbitrarily adjusted by the user as needed. The use of genetic distance in the likelihood function also allows us to take advantage of the spatial distribution of SNPs as an additional source of information, rather than only relying on patterns of population differentiation restricted to tightly linked SNPs.
3P-CLR also has an advantage over HMM-based selection methods, like the one implemented in ref. [10]. The likelihood ratio scores obtained from 3P-CLR can provide an idea of how credible a selection model is for a particular region, relative to the rest of the genome. The HMM-based method previously used to scan for selection in modern humans [10] can only rank putatively selected regions by genetic distance, but cannot output a statistical measure that may indicate how likely each region is to have been selected in ancient times. In contrast, 3P-CLR provides a composite likelihood ratio score, which allows for a statistically rigorous way to compare the neutral model and a specific selection model (for example, recent or ancient selection). The score also gives an idea of how much fainter the signal of ancient selection in modern humans is, relative to recent selection specific to a particular present-day population. For example, the outliers from Figure 4 have much higher scores (relative to the rest of the genome) than the outliers from Figure S6. This may be due to both the difference in time scales in the two sets of tests and to the uncertainty that comes from estimating outgroup allele frequencies using only two archaic genomes. This pattern can also be observed in Figure S7, where the densities of the scores looking for patterns of ancient selection have much shorter tails than the densities of scores looking for patterns of recent selection.
Like XP-CLR, 3P-CLR is largely robust to the underlying population history, even when this is wrongly specified, as it relies on looking for extreme allele frequency differences that are restricted to a particular region. We have noticed, however, that these types of tests may not be robust to admixture events from the outgroup population used. For example, we observe that 3P-CLR finds evidence for selection in the region containing HYAL2 (involved in the cellular response to ultraviolet radiation), when run in the East Asian branch. This makes sense, as a variant of this gene is known to have been pushed to high frequencies by selection specifically in East Asians. However, this variant likely came from Neanderthals via introgression [86,87]. While we do not observe that the HYAL2 region is a top hit in either the European or the Eurasian ancestral branches, we do observe it as a top hit in the modern human ancestral branch. This is puzzling, given that the selected haplotype should not have been introduced into modern humans until after the split of Africans and non-Africans. One explanation for the appearance of this region in both the East Asian and the modern human top hits is that the introgression event could perhaps confound the signal that 3P-CLR targets, as we are assuming no admixture in our demographic model. Another possibility is that the region has suffered multiple episodes of repeated selection. Incorporating admixture into the modeling procedure may help to disentangle this pattern better, but we leave this to a future work.
A further limitation of composite likelihood ratio tests is that the composite likelihood calculated for each model under comparison is obtained from a product of individual likelihoods at each site, and so it underestimates the correlation that exists between SNPs due to linkage effects [1,16,17,88]. One way to mitigate this problem is by using corrective weights based on linkage disequilibrium (LD) statistics calculated on the outgroup population [1]. Our implementation of 3P-CLR allows the user to incorporate such weights, if appropriate LD statistics are available from the outgroup. However, in cases where these are unreliable, it may not be possible to fully correct for this (for example, when only a few unphased genomes are available, as in the case of the Neanderthal and Denisova genomes).
While 3P-CLR relies on integrating over the possible allele frequencies in the ancestors of populations a and b (formula 10), one could envision using ancient DNA to avoid this step. Thus, if enough genomes could be sampled from that ancestral population that existed in the past, one could use the sample frequency in the ancient set of genomes as a proxy for the ancestral population frequency. This may soon be possible, as several early modern human genomes have already been sequenced in recent years [89-91].
Though we have limited ourselves to a three-population model in this manuscript, it should be straightforward to expand our model to a larger number of populations, albeit with additional costs in terms of speed and memory. Our method relies on a similar framework to the demographic inference method implemented in TreeMix [92], which can estimate complex population trees that include migration events, using genome-wide data. With a more complex modeling framework, it may be possible to estimate the time and strength of selective events with better resolution, and to incorporate additional demographic forces, like continuous migration between populations or pulses of admixture.
Acknowledgments
We thank Montgomery Slatkin, Rasmus Nielsen, Joshua Schraiber, Nicolas Duforet-Frebourg, Emilia Huerta-Sanchez, Hua Chen, Nick Patterson, David Reich, Joachim Hermisson, Graham Coop and members of the Slatkin and Nielsen labs for helpful advice and discussions. This work was supported by NIH grant R01-GM40282 to Montgomery Slatkin.
Footnotes
↵* Email: fernandoracimo{at}gmail.com