Nucleotide Substitution Model Selection Is Not Necessary for Bayesian Inference of Phylogeny With Well-Behaved Priors.

This article investigates whether substitution model selection is necessary in Bayesian phylogenetic inference or if using the most complex substitution model, specifically the general time reversible model with gamma-distributed rate variation and a proportion of invariant sites (GTR+Γ+I), suffices without biasing results. Through extensive simulations under the simplest substitution model (Jukes–Cantor) and inference with over-parameterized models and various prior distributions—including defaults from MrBayes, BEAST2, RevBayes, and a newly proposed "Tame" prior—the study finds that Bayesian phylogenetic inference is robust to substitution and partition model over-parameterization when well-behaved priors like the Tame prior are applied. While over-parameterization does not affect the accuracy of tree topology estimates, it can bias estimates of tree length under default priors, an issue mitigated by the Tame prior. The authors conclude that substitution and partition model selection steps can be omitted in Bayesian pipelines if appropriate priors are used, and recommend focusing efforts on developing more biologically realistic substitution models.