A model-free method for genealogical inference without phasing and its application for topology weighting

Abstract

Recent advances in methods to infer and analyse ancestral recombination graphs (ARGs) are providing powerful new insights in evolutionary biology and beyond. Existing inference approaches tend to be designed for use with fully-phased datasets, and some rely on model assumptions about demography and recombination rate. Here I describe a simple model-free approach for genealogical inference along the genome from unphased genotype data called Sequential Tree Inference by Collecting Compatible Sites (sticcs). sticcs applies a heuristic algorithm based on the perfect phylogeny principle to reconstruct a local genealogy at each variant site in the genome, using a collecting procedure to import information from other nearby sites. Using simulations, I show that sticcs is accurate for ARG inference, but only when the sample size is small. However, I also describe how it can be applied for the purpose of topology weighting by stacking tree sequences inferred for multiple subsets of the data, removing the sample size restriction. Topology weights estimated in this way from unphased data tend to be more accurate than those computed with full ARGs inferred from perfectly phased data using several popular tools. The methods presented therefore have promise for analysis of relatedness and introgression in non-model species, including polyploids. The new methods are implemented in two Python packages, sticcs (for ARG inference) and twisst2 (for topology weighting using the stacking procedure), which both interface with the tskit library for analysis of tree sequence objects.