Bayesian Cross-Validation Comparison of Amino Acid Replacement Models: Contrasting Profile Mixtures, Pairwise...

Area of Science:

• Evolutionary biology
• Bioinformatics
• Computational phylogenetics

Background:

• Amino acid replacement models are fundamental to phylogenetic inference, especially for deep evolutionary history.
• Traditional models used single, empirically derived matrices to limit inference complexity.
• Modern phylogenetics increasingly uses complex models to account for data size and evolutionary heterogeneity.

Purpose of the Study:

• To systematically compare the predictive power of various amino acid replacement models.
• To quantify the contribution of different modeling aspects, such as mixtures and rate variation.
• To evaluate the performance of empirical vs. free mixtures and Dirichlet process models.

Main Methods:

• Bayesian cross-validation was employed to assess model performance.
• Different modeling components (e.g., single matrices, mixtures, rate variation) were systematically activated and deactivated.
• Model performance was evaluated on real biological data sets.

Main Results:

• Amino acid mixture models generally outperform single-matrix models, even those with gamma-distributed rates.
• Free finite mixture models consistently showed better predictive power than empirical finite mixtures.
• Dirichlet process-based infinite mixture models demonstrated the highest performance across most datasets.

Conclusions:

• Mixture models significantly enhance phylogenetic inference accuracy by capturing across-site heterogeneity.
• Free mixture models are preferable to empirical ones for modeling amino acid substitutions.
• Dirichlet process mixtures represent a powerful approach for complex evolutionary modeling in phylogenetics.