Accurate Detection of Incomplete Lineage Sorting via Supervised Machine Learning

Abstract

Gene tree discordance due to incomplete lineage sorting or introgression has been described in numerous genomic datasets. Among distantly related taxa, however, it is difficult to differentiate these biological sources of discordance from discordance due to errors in gene tree reconstruction, even when supervised machine learning techniques are used to infer individual gene trees. Here, rather than applying machine learning to the problem of inferring single tree topologies, we develop a model to infer important properties of a particular internal branch of the species tree via genome-scale summary statistics extracted from individual alignments and inferred gene trees. We show that our model can effectively predict the presence/absence of discordance, estimate the probability of discordance, and infer the correct species tree topology in the presence of multiple, common sources of error. While gene tree topology counts are the most salient predictors of discordance at short time scales, other genomic features become relevant for distantly related species. We validate our approach through simulation, and apply it to data from the deepest splits among metazoans. Our results suggest that the base of Metazoa experienced significant gene tree discordance, implying that discordant traits among current taxa can be explained without invoking homoplasy. In addition, we find support for Porifera as the sister clade to the rest of Metazoa. Overall, these results demonstrate how machine learning can be used to answer important phylogenetic questions, while marginalizing over individual gene tree--and even species tree--topologies.