A0548
Title: A new variational family for Bayesian phylogenetics
Authors: Lloyd Elliott - Simon Fraser University (Canada) [presenting]
Evan Sidrow - Simon Fraser University (Canada)
Alexandre Bouchard - University of British Columbia (Canada)
Abstract: Bayesian phylogenetics aims to describe a posterior distribution over trees conditioned on a collection of genetic sequences (taxa) that arise from those trees. Modern MCMC can provide accurate posteriors but struggles to do so when the number of taxa is large. Variational Bayesian methods (in which the posterior is approximated by a variational family) are thus a natural choice for improving the scale of data for which Bayesian phylogenetics are viable. A new variational family is presented that factorizes over pairs of taxa. This family is parameterized by pairwise coalescent times and mapped to trees through single-linkage clustering. This factorization allows mean-field variational inference for the posterior, often using fewer parameters than other state-of-the-art variational phylogenetic methods based on subsplit Bayesian networks (SBN). Results are provided on COVID-19 RNA data. Compared to SBN methods, this method VIPR (Variational phylogenetic Inference with PRoducts over bipartitions) tends to have better initial conditions, fewer parameters and reaches posteriors with high likelihood faster. Gradient estimation is explored using REINFORCE, VIMCO, and the reparameterization trick. Autodiff is used for gradient calculations. The factorized nature of the variational family allows for parameter interpretation, extensions to unclocked trees, tip-dates, or further parameterization with low-dimensional latent states.
