Empirical studies of selection in admixed populations often look for regions of the genome that deviate from genome-wide patterns of genetic ancestry (Tang et al., 2007; Jin et al., 2012; Jeong et al., 2014; Rishishwar et al., 2015; Triska et al., 2015; Zhou et al., 2016; Busby et al., 2017; Laso-Jadart et al., 2017; Patin et al., 2017; Fernandes et al., 2019; Lopez et al., 2019; Norris et al., 2020; Vicuña et al., 2020). Regions of the genome with substantially higher ancestry from one source than present on average in the rest of the genome are hypothesized to be enriched for genes under selection. For an allele at different frequencies in the source populations, selection will increase the frequency of the ancestry on which the adaptive allele occurs at that locus. We tested if the DARC locus was an outlier within the genome for West African ancestry. We estimated local ancestry using the RFMix software (Maples et al., 2013) and calculated the proportion of individuals with West African ancestry at each SNP (see Materials and methods for details on local ancestry assignment). Figure 1B and Figure 1—figure supplement 1 show the distribution of West African ancestry for over ~880,000 SNPs by island, with the value for the Duffy-null SNP position marked in red. Within Santiago, this locus has the highest frequency of West African ancestry in the population, occurring at 0.834 frequency compared to mean West African ancestry across SNPs for individuals from Santiago of 0.730. In contrast, the DARC locus is not an outlier in its frequency of West African ancestry on the other island regions, occurring at the 75th and 65th percentiles for Fogo and the NW Cluster, respectively. High West African local ancestry proportion at the DARC locus in Santiago is consistent with the expectation that the Duffy-null allele rapidly increased in frequency following admixture, simultaneously increasing the proportion of individuals with West African ancestry at that locus relative to the genome-wide average.

Under a simple population genetic model, we expect the frequency of a neutral allele in an admixed population to be a linear combination of the allele frequencies in each source population and their relative ancestry contributions. The Duffy-null allele is nearly fixed in the West African source population and largely absent in the European source population; therefore, under neutrality, the expected frequency of the allele in each Cabo Verdean population is approximately equal to the West African ancestry contribution. Using the observed global ancestry proportion inferred with admixture (Alexander and Lange, 2011) as an estimate of the ancestry contribution from West Africa to each island, we found that the Duffy-null allele is at a higher frequency than expected under neutrality for the island of Santiago, but not the other regions of Cabo Verde (Figure 1—figure supplement 2, Table 1, binomial test, Santiago: p=2.193×10−5; Fogo: p=0.1915; NW Cluster: p=0.8172).

The distribution of the lengths of ancestry tracts spanning a selected locus can provide information for detecting and characterizing selection beyond single-locus outlier tests. In the case of recent admixture and strong selection, we might generally expect to see a parallel increase in local ancestry proportion in the regions surrounding the beneficial locus because the beneficial allele increases before recombination can break up large surrounding ancestry blocks. This is analogous to the increase in linkage disequilibrium and homozygosity in non-admixed populations (Sabeti et al., 2002; Kim and Nielsen, 2004; Voight et al., 2006). Figure 2A plots ancestry tracts that span the DARC locus for individuals from Santiago. As expected from source population allele frequencies, the Duffy-null allele is contained on all West African ancestry tracts spanning the locus and not found on any European ancestry tracts. Consistent with recent selection, West African ancestry tracts are longer and in higher frequency than European ancestry tracts covering the region. The median West African ancestry tract length spanning the locus is ~85 Mb, while the median European tract length is ~39 Mb.

In order to test if the observed local ancestry patterns are suggestive of selection beyond genome-wide ancestry proportion, we developed a summary statistic based on the length and frequency of the tract-length surrounding a locus. The integrated Decay in Ancestry Tract (iDAT) score compares the rate of decay of ancestry tract lengths as a function of distance from a site of interest (Figure 2B; see Materials and methods for details and performance of iDAT statistic under various demographic scenarios). The statistic is analogous to the commonly used integrated haplotype score (iHS) (Voight et al., 2006), but considers local ancestry tracts instead of haplotypes. Negative values for iDAT indicate longer West African ancestry tracts at higher frequencies compared to European ancestry tracts. Positive values indicate longer European ancestry tracts at higher frequencies compared to West African ancestry tracts. Windows that contain multiple extreme values of iDAT provide stronger evidence for recent selection. Figure 2C plots iDAT values along the genome for individuals from Santiago. Values are calculated by averaging over 20 Mb sliding windows (step size of 1 Mb) for 10,000 random standardized iDAT scores. Notably, in Santiago, the DARC gene is contained in the window with the lowest iDAT score in the genome (Figure 2C; window iDAT score = −2.602). iDAT scores near the DARC locus are not outliers in other island regions (Figure 2—figure supplement 1).

In order to estimate the expected distribution of ancestry within the population and test if the values of various summary statistics for the DARC locus on Santiago can be explained by drift alone, we conducted neutral simulations in SLiM (Haller and Messer, 2019) (Materials and methods). We calculated the following five summary statistics for each simulated population: the West African local ancestry proportion at DARC, the variance in the frequency of West African local ancestry across SNPs on chromosome 1, the median and mean West African ancestry tract length containing the Duffy-null allele, and the unstandardized iDAT score for the Duffy-null SNP. The variance in local ancestry along the chromosome provides a non-LD-based measure to capture high frequency and long tracts of West African ancestry using the population-wide measures of local ancestry for each SNP on the chromosome. Studies of demographic history and selection in recently admixed populations often assume constant population size and a single admixture event to simplify simulations. In order to confirm that assumptions about demographic history do not change our expectations, we considered multiple scenarios of population growth, differences in population size, and models of both constant contributions and single admixture events (Table 2). The values for the summary statistics for Santiago generally lie outside our expectations for all models, especially considered jointly (Figure 2—figure supplement 2).

Together, these summary statistics provide suggestive evidence that the DARC locus has been under positive selection on the island of Santiago since admixture started ~20 generations ago. To formally test this hypothesis, we extended the SWIF(r) framework developed by Sugden et al., 2018. SWIF(r) is a machine learning classification framework that explicitly learns the joint distributions for a set of features and returns a posterior probability of positive selection at a site of interest. It is particularly useful for handling summary statistics that are correlated, such as the length and frequency of ancestry tracts. We trained SWIF(r) using data simulated in SLiM and estimated the posterior probability of positive selection at the DARC locus using the five ancestry-based measures (Materials and methods). SWIF(r) returned a high posterior probability of positive selection at DARC on Santiago starting 20 generations ago (P>0.999).

The haplotype-based statistic, iHS, is often used to detect signatures of recent positive selection and partial selective sweeps (Voight et al., 2006), particularly in non-admixed populations. This statistic has been used as evidence of selection in recently admixed populations (Fernandes et al., 2019; Reynolds et al., 2019; Norris et al., 2020). However, the process of admixture results in the mixture of differentiated allele frequencies and diverged haplotypes, so interpretation of these statistics is difficult and applicability is limited. We demonstrate this by calculating iHS for all SNPs in our data set for each island region and performing the common standardization based on allele frequencies, using the software hapbin (Maclean et al., 2015). Figure 3 shows the distribution of absolute standardized iHS values along the genome for each island population, with the Duffy-null SNP indicated by the orange dot and flag. The absolute iHS values for all islands at the Duffy-null SNP are low. That is, the commonly used statistic iHS does not detect significant signatures of selection at the Duffy-null SNP position. This analysis, and other summaries of variation that do not account for the allele frequency and LD changes associated with admixture, may be detecting the high diversity in the African source populations rather than post-admixture selection. Without considering the process of admixture, we should be skeptical of the utility of these statistics in recently admixed populations. This emphasizes the importance of new methods that are admixture-aware.

Beyond identifying selection, inference of the strength of selection is informative about the evolutionary processes shaping human genomes. We used two complementary approaches to infer the strength of selection at the DARC locus. First, we considered a deterministic classical population-genetic model of selection based on the trajectory of allele frequencies over time on a grid of possible dominance and selection coefficients (Materials and methods). The estimate of the selection coefficient depends on dominance; past studies have modeled Duffy-null as recessive (Hodgson et al., 2014), dominant (Pierron et al., 2018), and additive (McManus et al., 2017) when estimating selection strength in other human populations. Figure 4A plots the selection strength (s) as a function of the dominance coefficient (h) of the Duffy-null allele for a set of three realistic initial frequencies, assuming 20 generations of constant selection strength. Functional studies suggest that heterozygotes have at least partial protection against P. vivax infection (Cavasini et al., 2007; Sousa et al., 2007; Gething et al., 2012; Kano et al., 2018); while not an exact correlate for population-genetic model parameters, this suggests that the Duffy-null allele is unlikely to be fully recessive or fully dominant. Taking the mean of selection coefficients for 0.2≤h≤0.8, we estimate the selection coefficient for each initial frequency, s0.65=0.106, s0.70=0.082, and s0.75=0.056, where spo is the inferred selection coefficient for initial allele frequency po.

Second, we used a simulation and rejection framework, approximate Bayesian computation (ABC), to jointly infer the selection coefficient and initial West African contribution while allowing for drift (Materials and methods). We used the five ancestry-based summary statistics described previously. We assumed an additive model, a single admixture event, and exponential growth in the population. Taking the median of the posterior distribution as the point estimate for selection coefficient, we estimated s=0.0795 (Figure 4B; see Figure 4—figure supplement 1 for estimates of s when modeling Duffy-null as either a dominant or a recessive mutation). This estimate of selection coefficient is consistent with those estimated under the deterministic population-genetic model.

Mean global ancestry proportion is often used as an estimate for initial ancestry contributions for admixed populations (Moreno-Estrada et al., 2013; Hellenthal et al., 2014; Bryc et al., 2015; Mathias et al., 2016; Laso-Jadart et al., 2017; Patin et al., 2017; Fernandes et al., 2019). However, our ABC estimates of the initial contributions from West Africa are lower than the mean ancestry currently observed in Santiago. The median of the posterior of initial contributions from West Africa is 0.690, with the middle 50 percentile of observed values in [0.682,0.697] (Figure 5A). In contrast, the observed mean ancestry in Santiago (Figure 1B) is 0.737. While this particular difference in observed and inferred founding contributions may be due to sampling biases or other neutral processes, it raises the more general question of how strong selection at a single locus may impact genome-wide ancestry patterns. We hypothesized that selection at DARC may have increased the genome-wide West African ancestry proportions in the current population of Santiago.

To test the genomic consequences of post-admixture selection at a single locus, we simulated whole human autosomes under a model of exponential growth and a single admixture event with selection at a single locus. We first considered a model based on the history of Santiago, using the posterior distributions of selection coefficient and initial West African ancestry contribution as parameters for the simulations (Materials and methods). Figure 5A plots the estimated posterior distribution of initial West African contribution in dark gray and the simulated distribution of global ancestry after 20 generations in white. The distribution of global ancestry in the populations simulated with selection (median 0.723, white) is noticeably higher than the initial contributions specified in the simulations (median 0.690, dark gray). This demonstrates that selection at a single locus is a plausible mechanism to increase mean global ancestry in an admixed population under a scenario similar to Santiago.

To explore the mechanism and relationship between selection strength at a single locus and genome-wide ancestry patterns, we simulated whole autosomes, assuming a single admixture event with initial West African ancestry contribution at 0.65 and selection coefficient varying from 0 to 0.2 at a single locus on chromosome 1. Figure 5B plots the mean ancestry for chromosome 1 and the other 21 autosomes in each simulated population after 20 generations as a function of the selection coefficient at a single locus. Perhaps surprisingly, mean ancestry on chromosomes 2–22 also increases with selection strength (gray), indicating that global ancestry increases beyond the contribution of higher ancestry on the selected chromosome alone (black). Together, this evidence suggests that strong selection at the DARC locus over 20 generations may have skewed global ancestry in Santiago and raises potential biases with a statistic that is often used to infer neutral demographic histories.

For this study, we used SNP array data from Beleza et al., 2013 which included 564 admixed individuals across the island regions of Cabo Verde. From this data set, we filtered individuals with greater than 5% missing calls overall or greater than 10% missing calls on a single chromosome. This resulted in removal of one individual with high missingness on chromosome 14 (11.36%). We merged genotypes for the remaining 563 individuals with genotypes from 107 IBS (Iberian Population in Spain) and 107 GWD (Gambian in Western Division – Mandinka) samples from high-coverage resequencing data released through the International Genome Sample Resource (Clarke et al., 2017; Fairley et al., 2020). Our analyses considered autosomal chromosomes only. We selected biallelic SNPs occurring in both the Cabo Verde samples and the reference samples. The final merged data set contained 884,656 autosomal SNPs. Average missingness by SNP was 0.0017.

Using the 884,656 autosomal SNP data set, we performed phasing with SHAPEIT2 using the Phase 3, NCBI build 37 (hg19) reference panel of haplotypes and associated genetic map in IMPUTE2 format (Delaneau et al., 2013). Following the SHAPEIT documentation, we first ran SHAPEIT –check to exclude sites not contained within the reference map, followed by SHAPEIT phasing to yield phased genotypes at 881,279 SNPs. For local ancestry inference, we ran RFMix v1.5 on the phased samples using a two-way admixture model (Maples et al., 2013). We used the RFMix PopPhased program with default window size, the --use-reference-panels-in-EM option, -e = 2 (2 EM iterations), and --forward-backward. Ancestry references for European and West African source populations, respectively, were IBS and GWD.

We observed a low overall proportion of regions within a given individual’s genome assigned as ‘unknown’ local ancestry by RFMix. The mean and median genome-wide proportion of unknown ancestry in our data set were 0.0089 and 0.0083, respectively. Typically, more than 99% of each individual’s genome could be assigned as either West African or European ancestry. Coupled with the fact that there is little historical or genetic evidence of migration from non-West African or non-European populations to Cabo Verde (Carreira, 1983; Verdu et al., 2017; Korunes et al., 2020), we find it is unlikely that individuals in our data set have high proportions of recent ancestry from a non-European or non-African source population.

To validate our phase imputation and RFMix local ancestry calls, we also performed local ancestry assignment using a second method, ELAI, which performs its own independent phasing prior to calling local ancestry (Guan, 2014). We ran ELAI under a two-way admixture model, again using IBS and GWD genotypes as references for the source populations. We set the parameters -mg (number of generations) to 20, -s (EM steps) to 30, -C (upper clusters) to 2, and -c (lower clusters) to 10, based on the 5 × C recommendation from the ELAI documentation. The estimates of each individual’s average genome-wide ancestry are highly correlated between ELAI and RFMix (Pearson’s R = 0.9964; p < 1 × 10⁻⁸).

Global ancestry inferred in admixture (Alexander and Lange, 2011) was averaged over 10 independent runs with randomly chosen seeds using supervised cluster mode with the GWD and IBS individuals specified as the reference populations. Our estimates of global ancestry by individual from admixture are consistent with those of RFMix (Pearson’s R = 0.9973; p < 1 × 10⁻⁸).

The island of Boa Vista was excluded from analyses due to our small sample from the region (26 individuals), leaving a final set of 537 individuals across the three island regions considered in this study (Santiago: 172, Fogo: 129, NW Cluster: 236; Figure 1A).

Local ancestry calls can be found at https://doi.org/10.5281/zenodo.4021277.

In order to account for global ancestry patterns that contribute to the ancestry tract-length distribution when identifying loci under selection post-admixture, we developed the iDAT score. This statistic quantifies the length and homozygosity of the tract-length distribution by comparing the decay of tract lengths from alternate ancestries at increasing distance from a site of interest.

First, we describe the Decay in Ancestry Tract (DAT) feature, which is calculated similar to expected haplotype homozygosity (EHH) (Sabeti et al., 2002), using local ancestry tracts rather than haplotypes. For each source population i we calculate:DATix= nxn 2,where nx is the number of ancestry tracts that extend some absolute distance x from a position of interest, and n is the total number of ancestry tracts (extending in either direction) that contain the site of interest. We calculate DAT at increasing distances from the site of interest. Similar to integrated haplotype homozygosity (iHH) (Voight et al., 2006), we then calculate the area under the curve for DAT as a function of distance from the position of interest, producing iDAT (Figure 2B). For this study, we calculated iDAT only for distances where DAT≥0.25, that is, where at least half of the ancestry tracts extend that absolute distance from the site of interest. We compare the difference in order of magnitude between iDAT for each ancestry, analogous to the integrated haplotype score (iHS), we have,iDATscore=ln⁡(iDAT2iDAT1),where iDAT_i is the iDAT calculated for source population i. The length of ancestry tracts, and therefore the iDAT score, will be influenced by relative ancestry contributions from each source population. So, when possible, we standardized the iDAT score using the empirical distribution of iDAT scores for 10,000 random positions across the genome. In this way, we deviate from the calculation of standardized iHS, which is standardized by the empirical distribution of SNPs with the same allele frequency because the length of haplotypes will be affected by the age of a variant (Voight et al., 2006). In the case of admixed populations, we instead need to account for the effect of global ancestry proportion on ancestry tract lengths, where the majority ancestry will tend to have longer contiguous tracts. We standardized by the genome-wide empirical distribution of iDAT scores, rather than iDAT scores for variants with the same local ancestry proportion, because the variance in local ancestry across the genome can be heavily influenced by drift. That is, to standardize the iDAT score, we calculateln⁡iDAT2iDAT1-Eln⁡iDAT2iDAT1SDln⁡iDAT2iDAT1.

By standardizing against a genome-wide distribution of iDAT values, we can account for demographic parameters, such as admixture proportions and timing since admixture, that may affect the global iDAT distribution. Single-locus deviations from the genome-wide expectations may then be indicative of selection at that site, and may warrant further study.

We used SLiM forward simulations with tree-sequence recording to track local ancestry (Haller et al., 2019; Haller and Messer, 2019). We considered eight different demographic scenarios: combinations of initial population size (N = 10,000 or N = 1000) with either a constant population size or exponential growth at a rate of 0.05 per generation, and either a single pulse of admixture at the start of simulation or continuous admixture at 1% total new migrants per generation (Table 2). The proportion of new migrants from each source population was weighted by the respective initial admixture contributions.

For each demographic scenario, we generated 1000 simulations of the human chromosome 1. For realistic recombination rates, we used the population-averaged human genetic map provided by IMPUTE2 (Delaneau et al., 2013) (https://mathgen.stats.ox.ac.uk/impute/1000GP_Phase3.html). We simulated admixture from two source populations to form a third admixed population (similar to recipe 17.5 in the SLiM manual) (Haller and Messer, 2016). One source population, representing the West African source population, was fixed for a neutral variant at the same position as the Duffy-null allele (chr1:159174683; GRCh37 coordinates in accordance with genetic map). Because the true West African ancestry contribution is unknown, for neutral simulations this parameter was drawn from a uniform distribution with lower and upper bounds at 0.65 and 0.75, respectively. We simulated the admixed population for 20 generations.

Using the tree-sequence files output from the SLiM simulations, we calculated the five ancestry-based summary statistics for each simulation: West African local ancestry proportion at DARC, variance in the distribution of West African local ancestry proportion across SNPs along the chromosome, mean and median West African ancestry tract length containing Duffy-null, and unstandardized iDAT score for the Duffy-null variant. We used the unstandardized iDAT score because there is no genome-wide distribution of iDAT scores for single-chromosome simulations. iDAT scores also could not be standardized using the distribution of simulated Duffy-null iDAT scores because each simulation had a different starting admixture proportion.

We sampled 172 individuals from each simulation and compared the simulated distribution to the observed values of the statistics for the 172 individuals from Santiago (Figure 2—figure supplement 2). The genetic map provided by IMPUTE2 is population-averaged. To determine whether population-specific differences in recombination rate may affect our ancestry-based statistics, we performed the same neutral simulations and comparison to empirical data using one of three genetic maps: GWD or IBS maps from Spence and Song, 2019, or an African American (AA) genetic map from Hinch et al., 2011. The expectations for mean and median tract length are affected by population-specific differences in recombination rate; however, because recombination rate affects both ancestries equally, the choice of genetic map does not change the expectation for iDAT score (Figure 2—figure supplement 3). Of note, the AA genetic map contains regions with extremely high recombination rate, resulting in the extreme differentiation in expectation of tract lengths between simulations using this and the other genetic maps, and, in particular, shorter expected tract lengths. Using this map would inflate our estimates of selection strength under the ABC framework; we chose to use the more conservative and general-use IMPUTE2 genetic map for population-averaged recombination rates.

Ancestry tracts extend over large proportions of the chromosome at the timescale of interest for this study (~20 generations). Therefore, in this case, fine-scale recombination rate differences are not expected to significantly affect our expectations for ancestry-based statistics.

The impact of different population size and migration scenarios on iDAT is summarized in the Materials and methods, under Single-chromosome simulations (Figure 2—figure supplement 2). These scenarios are relevant for the population history of Cabo Verde. Here, in order to understand the general behavior and applicability of iDAT for future analyses, we extend the scenarios considered beyond those likely to represent Cabo Verdean history. We consider combinations varying the generations since admixture, the selection coefficient, the initial contribution from the source populations, different chromosome lengths for the position of the adaptive allele, and different DAT cut-offs.

First, we considered the demographic history of the admixed population. Modifying the single-chromosome simulations described above, we conducted simulations setting the admixture timing to 10, 100, or 1000 generations in the past and admixture contribution from the source population with the adaptive allele to 0.1, 0.5, or 0.9. This source population was fixed for the variant at the Duffy-null position. We assumed a constant population size of N = 10,000 and a single-pulse of admixture. For each combination of admixture timing and admixture proportion, we generated 1000 simulations for each selection strength of s=0, s=0.01, s=0.05, or s=0.1. This resulted in 36 different scenarios and 36,000 simulations. For each simulation, we calculated iDAT for the variant at the Duffy-null position.

To interpret iDAT performance under these scenarios, we plot the proportion of Duffy-null iDAT values from the selection simulations that are in the bottom fifth percentile of the respective neutral Duffy-null iDAT distribution (Figure 2—figure supplement 4). We compare within demographic models because admixture proportion has a strong influence on the expectation and possible range of iDAT values. We also note that iDAT values cannot be calculated when an allele has been fixed in a population, as observed in the older admixture scenarios with high selection strength and high starting admixture proportion from the selected ancestry. As such, this statistic may be more useful under recent admixture and selection (i.e., fewer than 100 generations in the past) with substantial admixture contributions from both source populations of interest.

We next sought to assess how chromosome size and choice of DAT cut-off value affect the expected distribution of iDAT values. For this, we performed simulations of human chromosomes 1 (~250 Mb), 7 (~160 Mb), 15 (~100 Mb), and 22 (~50 Mb), to capture a range of reasonable chromosome sizes. We used the associated IMPUTE2 genetic map for each of these simulations. For consistency across simulations, one source population was fixed for a variant at a position at 80% of the physical length of the chromosome. We again assumed a constant population size at N = 10,000 and a single pulse of admixture. Based on the demographic history of Cabo Verde, we assumed the modern-day West African ancestry proportion to be the initial admixture contribution of 0.73 from the source population providing the variant of interest. We simulated the admixed population for 20 generations. We generated 1000 simulations each of neutral (s=0) and strong selection scenarios (s=0.05). For each simulation, we calculated iDAT for the tracked variant for distances where DAT≥0.25 (following our main analyses), DAT≥0.125, DAT≥0.0625, and DAT≥0.01. We show the proportion of iDAT values for the simulated variant under selection that are in the bottom fifth percentile of the neutral distribution of iDAT values, for each chromosome size and DAT cut-off (Figure 2—figure supplement 5).

Generally, on this timescale, selection strength, and admixture proportion, iDAT performs well across chromosome sizes and cut-off values. That said, we note that the cut-off of 0.25 works slightly better for the smallest human chromosome (chr 22), though we emphasize that its performance is not very different from the other cut-off values. Further, using lower cut-offs requires more computation time, and depending on the SNP density of the data set, this may be an important consideration. We encourage future studies looking to detect signatures of selection using this statistic on smaller chromosomes or different admixture histories, particularly for non-human populations, to consider testing iDAT performance under their specific model of interest.

We incorporated ancestry-based summary statistics and admixture simulations into the SWIF(r) framework developed by Sugden et al., 2018. For the simulated training set, we followed the single-chromosome simulation framework described above. We considered a single realistic demographic scenario for training: starting population size N = 10,000, exponential growth at a rate of 0.05 per generation, and single-pulse admixture with starting West African ancestry contribution randomly drawn from a uniform distribution from 0.65 to 0.75.

For simulations with selection at a single locus, we assumed an additive selection model (h=0.5). Selection coefficient was randomly drawn from a uniform distribution from 0 to 0.2. We calculated the five ancestry-based summary statistics for each simulation. Under this model, we generated 50,000 neutral simulations and 100 positive selection simulations for training; these training proportions are to reflect a prior probability of selection scenarios at 0.002. Since positive selection is relatively rare compared to neutral scenarios, SWIF(r) calibrates posterior probabilities according to this designated prior probability.

To validate this extension of SWIF(r), we generated a new set of 1000 neutral simulations and 1000 positive selection simulations. Figure 6 shows a precision-recall plot for a SWIF(r) implementation using the five ancestry-based statistics, a prior positive selection probability of 0.2% (reflecting the training set proportions above), and two classes (neutral or positive selection). Figure 6—figure supplement 1 demonstrates SWIF(r) performance for each class and across values of admixture contribution and selection coefficient. This SWIF(r) implementation had a low rate of false-negative classification of neutral simulations as positive selection scenarios.

To estimate the selection coefficient for the Duffy-null allele based on the dominance and the allele frequency trajectory over 20 generations, we used a deterministic population-genetic model of selection. We used the following recursion equation (Coop, 2020):pt+1=pt+pths+ qts1-h1-2ptqtsh-qt2sptqt,where pt is the frequency of Duffy-null in a given generation t, qt is the frequency of the alternate allele in generation t, s is the selection coefficient which is constant over time, and h is the dominance coefficient (0 if Duffy-null is dominant, 1 if recessive, and 0.5 in an additive selective model). We calculated the allele frequency over a grid of values for pairs of h∈[0,1] and s∈[0,0.2] in 0.005 and 0.001 increments, respectively. Figure 4A shows the combinations of h and s that produce |p20−pDuffy|<0.01, where p20 is the calculated frequency of the selected allele after 20 generations and pDuffy is the observed frequency of Duffy-null in Santiago (0.834). Because the true initial allele frequency po is not known, we performed the analysis for three reasonable starting allele frequencies: po=0.65, po=0.70, and po=0.75.

We used an ABC framework with nonlinear regression (neural network) adjustment to jointly estimate the posterior distributions for selection strength acting on the Duffy-null allele and the initial West African ancestry contribution (R package ‘abc’, with ‘neuralnet’ method) (Csilléry et al., 2012).

We followed the same simulation framework we previously described for our SWIF(r) positive selection simulations of chromosome 1, with the following demographic scenario for training: starting population size N = 10,000, exponential growth at a rate of 0.05 per generation, and single-pulse admixture. We assumed an additive selection model (h=0.5). The selection coefficient (s) and the initial West African ancestry contribution (m) were drawn from uniform prior distributions, s∼U(0,0.2) and m∼U(0.1,0.9). We generated 10,000 simulations for ABC inference and calculated the five ancestry-based summary statistics for each simulation.

We chose the tolerance and hidden layer sizes for ABC estimation based on the best combination of RMSE and R2 values for leave-one-out cross-validation under different combinations of these hyperparameters. We performed cross-validation for 1000 simulations. Estimation accuracy for this data set was similar under a variety of hyperparameter choices. We set the number of units in the hidden layer to ‘sizenet=2’ and the acceptance rate to ‘tol=0.05’. For the cross-validation, our estimates of the selection coefficient had an RMSE=0.0083 and R2=0.9785; our estimates of starting admixture proportion had an RMSE=0.0090 and R2=0.9985 (Figure 4—figure supplement 2).

To assess how selection at a single locus impacts genome-wide patterns of ancestry, we simulated changes in global ancestry over 20 generations (Figure 5). Whole autosome (22 chromosome) simulations with a realistic recombination map are computationally intensive. Instead, we took two complementary approaches.

First, we performed whole autosome simulations (22 independently segregating chromosomes) in SLiM by specifying a 'crossover rate' between chromosomes of 0.5 per generation. We used the total lengths of autosomal chromosomes from the Human Genome Assembly GRCh37.p13 (https://www.ncbi.nlm.nih.gov/grc/human/data?asm=GRCh37.p13). We used a uniform recombination rate within each chromosome of 1×10-8 crossovers per base position per generation. We considered the following demographic scenario: starting population size N = 10,000, exponential growth at a rate of 0.05 per generation, and single-pulse admixture. Global ancestry was calculated by taking the mean West African ancestry proportion across the autosome.

Second, we performed two-chromosome simulations (modeled based on human chromosomes 1 and 2) in order to incorporate a human genetic map for realistic recombination rates. We again used the genetic maps provided by IMPUTE2 for chromosomes 1 and 2. Two-chromosome simulations were performed under the same demographic model as whole autosome simulations. Global ancestry was calculated by taking a weighted mean of ancestry for the two chromosomes, where chromosome 2 represented the contribution for the ~92% of the genome that segregates independently from chromosome 1.

To test how selection at a single locus affects global ancestry, we considered both whole autosome simulations and two-chromosome simulations, using a West African ancestry contribution of 0.65 for all simulations in the founding generation. We varied the selection coefficient for the simulated Duffy-null variant on chromosome 1 from 0 to 0.2. We assumed an additive selection model (h=0.5). The results are similar across simulation methods (Figure 5—figure supplement 1).

To determine how our ABC estimates of initial West African ancestry contribution differ from final global ancestry after 20 generations of selection at the Duffy-null locus, we next simulated whole autosomes admixing for 20 generations drawing parameters from the previously inferred posterior distributions of the selection coefficient and initial ancestry contribution. Specifically, we used the paired estimates of selection coefficient and West African admixture contribution from each accepted simulation from the ABC analysis and passed those parameters as input for each whole autosome simulation; this produced a distribution of estimates of global ancestry for populations simulated with realistic selection and initial ancestry contributions (Figure 5A).