Reconstructing substitution histories on phylogenies, with accuracy, precision, and coverage

Abstract

Ancestral sequences and substitution histories are usually averaged out in phylogenetic inference and are therefore not reported to the user. However, they can be recovered through ancestral sequence reconstruction (ASR) and stochastic mapping. By implementing and validating a new ASR and stochastic mapping package compatible with both single-locus and multispecies coalescent analysis, we show how reconstructing substitutions along tree branches provides a practical approach to phylogenetic inference that has a number of advantages, in terms of both efficiency and in the information gained about the tempo and mode of molecular evolution. Based on a range of simulated datasets, we observe that substitution histories are recovered more accurately and precisely on time trees with relaxed clocks compared with unconstrained substitution trees that lack temporal directionality (maximum likelihood and Bayesian). We show that codon-partition models with site-rate heterogeneity (i.e., with four nucleotide states) can effectively approximate synonymous and non-synonymous substitution histories while requiring far less runtime than computationally demanding 61-state codon models. In turn, this provides a low-bias estimator of dN/dS that can outperform existing stochastic-mapping methods. Lastly, we ground-truth the stochastic mapping approach by showing that it can recover expected patterns in molecular evolution and pathogen transmission in three different case studies: i) that smaller mammals tend to have faster sub-stitution rates than their larger relatives, ii) that the rate of change in 3Di structural-alphabet characters in the aminoacyl-tRNA synthetase anticodon binding domain is associated with amino acid substitution rate, and iii) that influenza A virus spread more frequently to nearby locations than distant ones, in a major H3N2 outbreak in New Zealand. These three effects were more pronounced when based on count estimates (via stochastic mapping) rather than evolutionary rates and branch lengths (the standard approach). Our open-source code comes with a user-friendly graphical interface, and is released as the BeastMap package for BEAST 2. The implementation is directly integrated into Bayesian phylogenetic analysis; supporting a wide range of clock, site, and tree models and data types, including insertions and deletions.