Data-specific substitution models improve protein-based phylogenetics

Calculating amino-acid substitution models that are specific for individual protein data sets is often difficult due to the computational burden of estimating large numbers of rate parameters. In this study, we tested the computational efficiency and accuracy of five methods used to estimate substitution models, namely Codeml, FastMG, IQ-TREE, P4 (maximum likelihood), and P4 (Bayesian inference). Data-specific substitution models were estimated from simulated alignments (with different lengths) that were generated from a known simulation model and simulation tree. Each of the resulting data-specific substitution models was used to calculate the maximum likelihood score of the simulation tree and simulated data that was used to calculate the model, and compared with the maximum likelihood scores of the known simulation model and simulation tree on the same simulated data. Additionally, the commonly-used empirical models, cpREV and WAG, were assessed similarly. Data-specific models performed better than the empirical models, which under-fitted the simulated alignments, had the highest difference to the simulation model maximum-likelihood score, clustered further from the simulation model in principal component analysis ordination, and inferred less accurate trees. Data-specific models and the simulation model shared statistically indistinguishable maximum-likelihood scores, indicating that the five methods were reasonably accurate at estimating substitution models by this measure. Nevertheless, tree statistics showed differences between optimal maximum likelihood trees. Unlike other model estimating methods, trees inferred using data-specific models generated with IQ-TREE and P4 (maximum likelihood) were not significantly different from the trees derived from the simulation model in each analysis, indicating that these two methods alone were the most accurate at estimating data-specific models. To show the benefits of using data-specific protein models several published data sets were reanalysed using IQ-TREE-estimated models. These newly estimated models were a better fit to the data than the empirical models that were used by the original authors, often inferred longer trees, and resulted in different tree topologies in more than half of the re-analysed data sets. The results of this study show that software availability and high computation burden are not limitations to generating better-fitting data-specific amino-acid substitution models for phylogenetic analyses.