Composite measures of selection can improve the signal‐to‐noise ratio in genome scans

Abstract

Summary The growing wealth of genomic data is yielding new insights into the genetic basis of adaptation, but it also presents the challenge of extracting the relevant signal from multi‐dimensional datasets. Different statistical approaches vary in their power to detect selection depending on the demographic history, type of selection, genetic architecture and experimental design. Here, we develop and evaluate new approaches for combining results from multiple tests, including multivariate distance measures and methods for combining P‐values. We evaluate these methods on (i) simulated landscape genetic data analysed for differentiation outliers and genetic‐environment associations and (ii) empirical genomic data analysed for selective sweeps within dog breeds for loci known to be selected for during domestication. We also introduce and evaluate how robust statistical algorithms can be used for parameter estimation in statistical genomics. On the simulated data, many of the composite measures performed well and had decreased variation in outcomes across many sampling designs. On the empirical dataset, methods based on combining P‐values generally performed better with clearer signals of selection, higher significance of the signal, and in closer proximity to the known selected locus. Although robust algorithms could identify neutral loci in our simulations, they did not universally improve power to detect selection. Overall, a composite statistic that measured a robust multivariate distance from rank‐based P‐values performed the best. We found that composite measures of selection could improve the signal of selection in many cases, but they were not a panacea and their power is limited by the power of the univariate statistics they summarize. Since genome scans are widely used, improving inference for prioritizing candidate genes may be beneficial to medicine, agriculture, and breeding. Our results also have application to outlier detection in high‐dimensional datasets and to combining results in meta‐analyses in many disciplines. The compound measures we evaluate are implemented in the r package minotaur.

Genome-wide selection scans integrated with association mapping reveal mechanisms of physiological adaptation across a salinity gradient in killifish

“…F ST and π were combined using a multivariate outlier approach as implemented in the R package MINOTAUR [47]. Raw statistics were converted to rank based p-values based on a uniform distribution, reflecting quantile values from the empirical distribution [48]. To identify selection in fresh water, we based these p-value conversions on right-tailed and left-tailed expectations for F ST and π, respectively.…”

“…Importantly, this measure does not assume independence between statistics, though it does assume smooth dispersal from a centroid, making it necessary to perform the aforementioned p-value transformations. This approach follows the Md-rank-P method described in Lotterhos et al, [48]. From this distribution, we considered the top 1% of SNPs to be outliers, and therefore the most likely candidates as targets of natural selection.…”

How to measure natural selection

“…The paper by Lotterhos et al . () in this issue develops new methods of combining datasets to increase our power to detect selective sweeps.…”

“…Lotterhos et al . () provide clear methods for combining data across multiple outlier locus methods, either by combining the P ‐values from multiple tests or by combining the signal into a multivariate test for differentiation. Using a combination of simulations and analysis of empirical datasets, they provide much needed guidance for empiricists working in this area.…”

Next‐generation conservation genetics and biodiversity monitoring

“…Inappropriate choice of study design or analysis can lead to incorrect conclusions, and thus misguided interventions (Lotterhos & Whitlock, 2015;Meirmans, 2015). Moreover, there is recognition that currently available analyses do not make full use of large genomic datasets (Lotterhos et al, 2017;Villemereuil, Frichot, Bazin, François, & Gaggiotti, 2014), and that both informatic and theoretical advances are still needed. These improvements to genetic monitoring and analyses are the focus of this special issue.…”