Scalable Empirical Mixture Models That Account for Across-Site Compositional Heterogeneity

Area of Science:

• Computational biology
• Evolutionary biology
• Bioinformatics

Background:

• Biochemical constraints create variations in amino acid usage across protein sites.
• Ignoring this heterogeneity in phylogenetic models causes errors like long-branch attraction.

Purpose of the Study:

• To introduce EDCluster, a scalable method for building empirical distribution mixture models.
• To develop universal distribution mixture (UDM) models for improved phylogenetic inference.

Main Methods:

• EDCluster uses cluster analysis with coordinate transformations to identify specialized amino acid distributions.
• UDM models were created using HOGENOM and HSSP databases, with up to 4,096 components.
• The method was implemented in IQ-TREE, Phylobayes, and RevBayes software.

Main Results:

• UDM models effectively removed long-branch attraction artifacts.
• The new models demonstrated superior performance compared to existing C10-C60 models.
• EDCluster provides a computationally efficient approach to model compositional heterogeneity.

Conclusions:

• EDCluster offers a scalable and effective solution for modeling amino acid compositional heterogeneity.
• UDM models enhance phylogenetic accuracy by accounting for site-specific amino acid preferences.
• The provided implementations facilitate the adoption of these advanced models in evolutionary research.