Comparative population genomics: power and principles for the inference of functionality

Abstract

The availability of sequenced genomes from multiple related organisms allows the detection and localization of functional genomic elements based on the idea that such elements evolve more slowly than neutral sequences. Although such comparative genomics methods have proven useful in discovering functional elements and ascertaining levels of functional constraint in the genome as a whole, here we outline limitations intrinsic to this approach that cannot be overcome by sequencing more species. We argue that it is essential to supplement comparative genomics with ultra-deep sampling of populations from closely related species to enable substantially more powerful genomic scans for functional elements. The convergence of sequencing technology and population genetics theory has made such projects feasible and has exciting implications for functional genomics.

Figure 1

Comparative population genomics. The power to detect purifying selection on either an individual functional element or any small collection of sites is improved greatly by the addition of multiple species. We model a functional element in which 80% of the sites are functional and maintained by a selective force of 4N_es = –50 across 1, 5, or 10 species (for details of 4N_es see Glossary: Effective selection). Each species has been sequenced to a depth of 501 individuals from a population. (A) The y axis is the power (in units of Δlog-likelihood) to distinguish between the true parameters (4N_es = –50, p = 0.8) and the null hypotheses H_{0, neutral} = (0,0). (B) The y axis is the power (in units of Δlog-likelihood) to distinguish between the true parameters (4N_es, p = 0.8) and the null hypotheses H_{0, lethal} = (–∞, p’). The dotted grey line represents 5% significance for the χ² test [–2Δlog(L) = 3.84]. We can see that as we increase the number of species, especially when those species have a higher level of polymorphism, we gain substantial power to detect shorter functional elements. When θ is on the order of 1%, and we go from a single population genomic dataset to a dataset from 10 species, we move from being able to detect purifying selection acting on an 80 bp element to being able to detect it acting on an 8 bp element. Similarly, panel (B) shows that, to distinguish finite selection from infinitely strong selection or mutation rate variation, 220 bp are needed if data from only a single species are available, versus 22 bp if polymorphisms from 10 species are used. Note that the increase in power (in log-likelihood space) is proportionate to the number of polymorphisms, such that increasing number of species, length, and θ all cause a proportional increase in power.