Predicting locus phylogenetic utility using machine learning

Abstract

Disentangling evolutionary signal from noise in genomic datasets is essential to building phylogenies. The efficiency of current sequencing platforms and workflows has resulted in a plethora of large-scale phylogenomic datasets where, if signal is weak, it can be easily overwhelmed with non-phylogenetic signal and noise. However, the nature of the latter is not well understood. Although certain factors have been investigated and verified as impacting the accuracy of phylogenetic reconstructions, many others (as well as interactions among different factors) remain understudied. Here we use a large simulation-based dataset and machine learning to better understand the factors, and their interactions, that contribute to species tree error. We trained Random Forest regression models on the features extracted from simulated alignments under known phylogenies to predict the phylogenetic utility of the loci. Loci with the worst utility were then filtered out, resulting in an improved signal-to-noise ratio across the dataset. We investigated the relative importance of different features used by the model, as well as how they correspond to the originally simulated properties. We further used the model on several diverse empirical datasets to predict and subset the least reliable loci and re-infer the phylogenies. We measure the impacts of the subsetting on the overall topologies, difficult nodes identified in the original studies, as well as branch length distribution. Our results suggest that subsetting based on the utility predicted by the model can improve the topological accuracy of the trees and their average statistical support, and limits paralogy and its effects. Although the topology generated from the filtered datasets may not always be dramatically different from that generated from unfiltered data, the worst loci consistently yielded different topologies and worst statistical support, indicating that our protocol identified phylogenetic noise in the empirical data.