Best‐fit Maximum‐likelihood Models for Phylogenetic Inference: Empirical Tests With Known Phylogenies

Despite the proliferation of increasingly sophisticated models of DNA sequence evolution, choosing among models remains a major problem in phylogenetic reconstruction. The choice of appropriate models is thought to be especially important when there is large variation among branch lengths. We evaluated the ability of nested models to reconstruct experimentally generated, known phylogenies of bacteriophage T7 as we varied the terminal branch lengths. Then, for each phylogeny we determined the best-fit model by progressively adding parameters to simpler models. We found that in several cases the choice of best-fit model was affected by the parameter addition sequence. In terms of phylogenetic performance, there was little difference between models when the ratio of short: long terminal branches was 1:3 or less. However, under conditions of extreme terminal branch-length variation, there were not only dramatic differences among models, but best-fit models were always among the best at overcoming long-branch attraction. The performance of minimum-evolution-distance methods was generally lower than that of discrete maximum-likelihood methods, even if maximum-likelihood methods were used to generate distance matrices. Correcting for among-site rate variation was especially important for overcoming long-branch attraction. The generality of our conclusions is supported by earlier simulation studies and by a preliminary analysis of mitochondrial and nuclear sequences from a well-supported four-taxon amniote phylogeny.