ASPEN: A methodology for reconstructing protein evolutionwith improved accuracy using ensemble models

Abstract

Mapping the history of ancestor divergence can yield critical insight into protein function. However, reconstruction of this evolutionary history is subject to a great deal of uncertainty, as selection and alignment of extant descendants to represent each ancestor dramatically affects phylogenetic inference of all ancestral nodes. Although historically problematic and somewhat ignored, here we capitalize on that variability by subsampling descendants of each ancestor, generating an ensemble of reconstructions, and integrating over the resulting topological diversity to improve the accuracy of ancestor divergence reconstruction, above the level of subsampling. First, we demonstrate that reproducibility of an all-sequence, single-alignment reconstruction, measured by comparing topologies inferred from 90% subsamples, can be used to assess the likely accuracy of single-alignment reconstructions. We believe this relationship results from modulation of the likelihood landscape around the true topology induced by alignments of resampled sequences. Second, we hypothesize that topological features observed consistently across an ensemble of subsampled reconstructions best reflect true evolutionary history, despite the inaccuracy of each individual reconstruction due to the failure by substitution models to infer the correct sequence of mutations that produced the input sequences. We present a scoring function that evaluates topologies on the basis of their consistency with features extracted from an ensemble, and a reconstruction algorithm (ASPEN) that produces and ranks topologies according to that function. We test our methodology on 600 simulated protein families and the LacI transcription factor family, and find that ASPEN topologies are significantly more accurate than single-alignment reconstructions.