Tread Lightly Interpreting Polygenic Tests of Selection

Abstract

In this issue of GENETICS, a new method for detecting natural selection on polygenic traits is developed and applied to several human examples (Racimo et al. 2018). By definition, many loci contribute to variation in polygenic traits, and a challenge for evolutionary geneticists has been that these traits can evolve by small, nearly undetectable shifts in allele frequencies across each of many, typically unknown, loci. Recently, a helpful remedy has arisen. Genome-wide association studies (GWAS) have been illuminating sets of loci that can be interrogated jointly for changes in allele frequencies. By aggregating small signals of change across many such loci, directional natural selection is now in principle detectable using genetic data, even for highly polygenic traits. This is an exciting arena of progress – with these methods, tests can be made for selection associated with traits, and we can now study selection in what may be its most prevalent mode. The continuing fast pace of GWAS publications suggest there will be many more polygenic tests of selection in the near future, as every new GWAS is an opportunity for an accompanying test of polygenic selection. However, it is important to be aware of complications that arise in interpretation, especially given that these studies may easily be misinterpreted both in and outside the evolutionary genetics community. Here, we provide context for understanding polygenic tests and urge caution regarding how these results are interpreted and reported upon more broadly. The foundations of polygenic tests of selection trace back to the very dawn of genetics, a century ago. An early challenge was to reconcile the inheritance of continuously distributed traits like height with the discretely inherited particulate genes of Mendelian inheritance. In a monumental advance in the history of genetics, R.A. Fisher reconciled this conflict by proposing that many genetic loci contribute to the variation of such traits – a polygenic model of inheritance – and by showing how to analyze their collective effect. In modern parlance, these traits are referred to as “quantitative,” “polygenic,” or “complex,” and each of the contributing loci are known as a quantitative trait locus (QTL). GWAS are revolutionary in allowing modern practitioners to uncover a fraction of the QTL that Fisher posited underlie most trait variation and to estimate effect sizes of each variant on a trait. Importantly for the discussion here, the polygenic model has evolutionary implications – most notably, the mean phenotypic value of a trait in a population can shift substantially and quickly via very subtle shifts in frequencies at the many QTL that underlie a trait. To see a signature of selection, modern polygenic tests use the QTL discovered by GWAS and detect shifts in their frequencies across time or across populations. The key advance has been to realize that while the impact of selection on any particular QTL cannot be reliably detected, the collective impact across many QTL can be. The first polygenic selection study to use human GWAS results was based on a commonly studied trait, human height. Turchin et al. (2012) found a difference between northern and southern Europeans in the genetic component of height that appears to have been driven by differential selection across populations. A subsequent foundational paper by Berg and Coop (2014) gave a more formal analysis, by establishing a null model for how polygenic scores should vary across populations. This approach generalized and modernized a traditional pre-GWAS approach [comparison of Qstvs. Fst, Whitlock and Guillaume 2009] and has itself been extended to consider multiple traits (Berg et al. 2017). These initial approaches focused on comparison across populations, but comparisons can also be made across time within single populations using either ancient DNA (Mathieson et al. 2015) or using haplotypic signatures of allele frequency change (Field et al. 2016). The new study published in this issue by Racimo et al. provides a novel methodology for comparing across populations by using a population tree with migration edges (an “admixture graph”). The new method allows one to isolate which branches in an admixture graph have experienced polygenic selection. In some sense then, Racimo et al. extend Berg and Coop’s (2014) approach to allow more detailed resolution of when and where directional selection took place. While these methodological contributions are important, the Racimo et al. study requires delicate handling. This is true for all polygenic selection studies, but all the more so here because of the comparison across populations for the complex behavioral trait of “educational attainment” (i.e., years of schooling). Genetic contributions to educational attainment have been of keen interest to social scientists and economists who have an eye on implications for policy (e.g. Conley and Fletcher 2017). The Racimo et al. study is not unique