Detecting Selection in Multiple Populations by Modeling Ancestral Admixture Components

Abstract

One of the most powerful and commonly used approaches for detecting local adaptation in the genome is the identification of extreme allele frequency differences between populations. In this article, we present a new maximum likelihood method for finding regions under positive selection. It is based on a Gaussian approximation to allele frequency changes and it incorporates admixture between populations. The method can analyze multiple populations simultaneously and retains power to detect selection signatures specific to ancestry components that are not representative of any extant populations. Using simulated data, we compare our method to related approaches, and show that it is orders of magnitude faster than the state-of-the-art, while retaining similar or higher power for most simulation scenarios. We also apply it to human genomic data and identify loci with extreme genetic differentiation between major geographic groups. Many of the genes identified are previously known selected loci relating to hair pigmentation and morphology, skin, and eye pigmentation. We also identify new candidate regions, including various selected loci in the Native American component of admixed Mexican-Americans. These involve diverse biological functions, such as immunity, fat distribution, food intake, vision, and hair development.

Keywords: admixture; human evolution; population structure; positive selection; selective sweeps.

Fig. 1.

Simulation tests of Ohana performance and efficiency in detecting and mapping selected sites. (a–c) Power to detect selection relative to two comparable methods, BayPass and pcadapt. Error bars are 95% CIs; (d–f) efficacy of Ohana to fine-map the causal site; (g) computational efficiency compared with that of BayPass. Error bars are 5th to 95th percentiles.

Fig. 2.

Illustration of simulation models. (a) Model 1, a basic model of four-population split with no admixture. (b) Model 2, a four-population split with subsequent admixture. (c) Model 3, a four-population model mimicking human demographic models. Population size changes in model 3 are omitted from the visualization for simplicity. Selection is simulated to operate on the branch that has a larger width.

Fig. 3.

ROC curves for Ohana versus state-of-the-art methods, assessed using simulations with various values of the initial allele frequency at the beginning of selection (f) and different selection coefficients (s). Here the demographic model used was our human model (with selection in the Native American lineage), versus other demographic models considered (i.e., basic tree without and with admixture, supplementary figs. S1 and S2, Supplementary Material online, respectively).

Fig. 4.

Inferred unrooted tree of latent ancestry components for the analysis including the CHB, YRI, MXL, and GBR genomic panels. We label each component by the population in which it is maximized, but emphasize that the components and the populations are not equivalent entities.

Fig. 5.

Top 5 annotated peaks in each of the ancestry-specific selection studies. MXL-specific = scan for selection in Native American ancestry of MXL. GBR-specific = scan for selection in European ancestry of GBR. CHB-specific = scan for selection in CHB ancestry of CHB. YRI-specific = scan for selection in Yoruba African ancestry or ancestral non-African ancestry. We analyzed 5,601,710 variable sites across the autosomal genomes. We inferred genome-wide allele frequencies and covariances as described in Materials and Methods section. We applied a likelihood model for each SNP by rescaling all variances and covariances by a scalar multiplier α. Descriptions of each candidate region are in table 1. LLR, log-likelihood ratio score.

Fig. 6.

Selection hypotheses and their encodings as covariance matrices. In this example, the ancestry component E is assumed to be the potential target of selection. The entry E: E in the covariance matrix is therefore allowed to deviate from the globally estimated value.