Bayesian Variable Selection Approaches to Identifying Selection in Evolutionary Datasets

Abstract

We consider the problem of identifying individual sites and characters under selection in phylogenetic datasets. We introduce an approach based on the application of Bayesian variable selection to mutation-selection models, which enables both sites and selection coefficients to be explicitly excluded from the model in order to control parameter complexity. We provide an accelerated Gibbs sampling algorithm for model selection and parameter estimation, which enables our approach to scale to realistic sequence lengths and dataset sizes. We compare this approach to previous models that do not produce sparse solutions but attempt to control complexity through Dirichlet-process (DP) mixture model prior distributions. We demonstrate that the Bayesian variable selection approach outperforms the DP approach for a wide range of settings and selection signal strengths, while offering additional advantages of sparsity and interpretability. We also identify a problem of the DP prior producing "phantom selection" artifacts, on both simulated and real datasets, which the Bayesian variable selection approach avoids.