Inference of selective sweep parameters through supervised learning

Abstract

A selective sweep occurs when positive selection drives an initially rare allele to high population frequency. In nature, the precise parameters of a sweep are seldom known: How strong was positive selection? Did the sweep involve only a single adaptive allele (hard sweep) or were multiple adaptive alleles at the locus sweeping at the same time (soft sweep)? If the sweep was soft, did these alleles originate from recurrent new mutations (RNM) or from standing genetic variation (SGV)? Here, we present a method based on supervised machine learning to infer such parameters from the patterns of genetic variation observed around a given sweep locus. Our method is trained on sweep data simulated with SLiM, a fast and flexible framework that allows us to generate training data across a wide spectrum of evolutionary scenarios and can be tailored towards the specific population of interest. Inferences are based on summary statistics describing patterns of nucleotide diversity, haplotype structure, and linkage disequilibrium, which are estimated across systematically varying genomic window sizes to capture sweeps across a wide range of selection strengths. We show that our method can accurately infer selection coefficients in the range 0.01 < s < 100 and classify sweep types between hard sweeps, RNM soft sweeps, and SGV soft sweeps with accuracy 69 % to 95 % depending on sweep strength. We also show that the method infers the correct sweep types at three empirical loci known to be associated with the recent evolution of pesticide resistance in Drosophila melanogaster. Our study demonstrates the power of machine learning for inferring sweep parameters from present-day genotyping samples, opening the door to a better understanding of the modes of adaptive evolution in nature. Author summaryAdaptation often involves the rapid spread of a beneficial genetic variant through the population in a process called a selective sweep. Here, we develop a method based on machine learning that can infer the strength of selection driving such a sweep, and distinguish whether it involved only a single adaptive variant (a so-called hard sweep) or several adaptive variants of independent origin that were simultaneously rising in frequency at the same genomic position (a so-called soft selective sweep). Our machine learning method is trained on simulated data and only requires data sampled from a single population at a single point in time. To address the challenge of simulating realistic datasets for training, we explore the behavior of the method under a variety of testing scenarios, including scenarios where the history of the population of interest was misspecified. Finally, to illustrate the accuracy of our method, we apply it to three known sweep loci that have contributed to the evolution of pesticide resistance in Drosophila melanogaster.