Deep learning for population genetic inference
Given genomic variation data from multiple individuals, computing the likelihood of complex population genetic models is often infeasible. To circumvent this problem, we introduce here a novel likelihood-free inference framework by applying deep learning, a powerful modern technique in machine learning. In contrast to Approximate Bayesian Computation, another likelihood-free approach widely used in population genetics and other fields, deep learning does not require a distance function on summary statistics or a rejection step, and it is robust to the addition of uninformative statistics. To demonstrate that deep learning can be effectively employed to estimate population genetic parameters and learn informative features of data, we focus on the challenging problem of jointly inferring natural selection and demography (in the form of a population size change history). Our method is able to separate the global nature of demography from the local nature of selection, without sequential steps for these two factors. Studying demography and selection jointly is motivated by Drosophila, where pervasive selection confounds demographic analysis. We apply our method to 197 African Drosophila melanogaster genomes from Zambia to infer both their overall demography, and regions of their genome under selection. We find many regions of the genome that have experienced hard sweeps, and fewer under selection on standing variation (soft sweep) or balancing selection. Interestingly, we find that soft sweeps and balancing selection occur more frequently closer to the centromere of each chromosome. In addition, our demographic inference suggests that previously estimated bottlenecks for African Drosophila melanogaster are too extreme, likely due in part to the unaccounted impact of selection.