Using environmental correlations to identify loci underlying local adaptation

Abstract

Loci involved in local adaptation can potentially be identified by an unusual correlation between allele frequencies and important ecological variables or by extreme allele frequency differences between geographic regions. However, such comparisons are complicated by differences in sample sizes and the neutral correlation of allele frequencies across populations due to shared history and gene flow. To overcome these difficulties, we have developed a Bayesian method that estimates the empirical pattern of covariance in allele frequencies between populations from a set of markers and then uses this as a null model for a test at individual SNPs. In our model the sample frequencies of an allele across populations are drawn from a set of underlying population frequencies; a transform of these population frequencies is assumed to follow a multivariate normal distribution. We first estimate the covariance matrix of this multivariate normal across loci using a Monte Carlo Markov chain. At each SNP, we then provide a measure of the support, a Bayes factor, for a model where an environmental variable has a linear effect on the transformed allele frequencies compared to a model given by the covariance matrix alone. This test is shown through power simulations to outperform existing correlation tests. We also demonstrate that our method can be used to identify SNPs with unusually large allele frequency differentiation and offers a powerful alternative to tests based on pairwise or global F(ST). Software is available at http://www.eve.ucdavis.edu/gmcoop/.

Figure 1.—

The distance from the equator for each of 52 human populations, plotted against sample allele frequencies for the SNP AGT M235T in each population. The points are colored according to the geographic region each population belongs to, following region definitions of Rosenberg et al. (2002). The data were generated using HGDP samples by Thompson et al. (2004) and are replotted on the basis of a figure in that article.

Figure 2.—

(A) A single draw from the posterior of the covariance matrix estimated for the HGDP SNPs of Conrad et al. (2006). (B) The correlation matrix calculated from the covariance matrix shown in A. The matrices are displayed as heat maps with lighter colors corresponding to higher values. The rows and columns of these matrices have been arranged by broad geographic label.

Figure 3.—

The power of various methods to detect a correlation between latitude and allele frequency.

Figure 4.—

The power of various methods to detect a correlation between summer precipitation and allele frequency.

Figure 5.—

The power of various methods to detect a “European effect” on allele frequency.

Figure 6.—

A plot of the log₁₀ Bayes factor for each SNP along the human genome for (A) a European effect and (B) a western Eurasian effect. Bayes factors <1 are not plotted. The numbers on the x-axis indicate chromosome number, with SNPs on different chromosomes colored alternately in red and black. We list the name of the gene that is nearest to each of the six highest-ranking SNPs in each plot (considering only the peak SNP in each cluster of high Bayes factors).