Phylogenetic mixture models for proteins.

Area of Science:

• Computational Biology
• Evolutionary Biology
• Bioinformatics

Background:

• Standard protein substitution models utilize a single rate matrix, which fails to capture the heterogeneity of site evolution.
• Factors like genetic code, solvent exposure, and protein structure influence amino acid substitution patterns.

Purpose of the Study:

• To explore phylogenetic mixture models that combine multiple amino acid replacement matrices for improved protein evolution modeling.
• To compare unsupervised and supervised learning approaches for estimating these mixture models.

Main Methods:

• Developed maximum-likelihood phylogenetic mixture models combining several amino acid replacement matrices.
• Learned models from HSSP alignment database and tested on TREEBASE alignments.
• Compared unsupervised (site categories unknown) and supervised (site categories known) learning methods.
• Integrated gamma-distributed rates across sites.

Main Results:

• Mixture models achieved significant likelihood gains compared to single replacement matrices.
• Mixtures of matrices outperformed mixtures of profiles (e.g., CAT model).
• Unsupervised approach generally superior but sensitive to initial parameters; supervised approach useful for initialization.
• A three-matrix unsupervised model showed substantial AIC gains over LG, WAG, and JTT models.
• Inferred tree topologies differed when using mixture models, impacting evolutionary inference.

Conclusions:

• Phylogenetic mixture models provide a more accurate representation of protein evolution than single matrix models.
• The choice of model significantly impacts phylogenetic inference.
• Available models and PhyML implementation facilitate further research.