A Bayesian sampler for optimization of protein domain hierarchies

Area of Science:

• Computational Biology
• Bioinformatics
• Protein Domain Analysis

Background:

• Protein domain classification traditionally relies on manual curation.
• Automating the identification and hierarchical modeling of functionally divergent subgroups is a significant statistical and algorithmic challenge.

Purpose of the Study:

• To develop and validate an automated statistical and algorithmic approach for creating optimal hierarchical models of protein domain families.
• To model functionally divergent subgroups within protein domain classes and arrange them hierarchically.

Main Methods:

• Utilized Markov chain Monte Carlo (MCMC) sampling to create and optimize hierarchical structures from multiple sequence alignments.
• The MCMC sampler modifies hierarchies by adding/deleting nodes, moving subtrees, and redefining sequence patterns.
• The probability distribution models conserved and divergent patterns, linking them to protein function and biological properties.

Main Results:

• The MCMC sampler demonstrated convergence to similar, near-optimal hierarchical solutions across 60 diverse protein domains from multiple starting points.
• Achieved highly comparable log-likelihood ratio scores, indicating the method's ability to find optimal peaks in posterior probability distributions.
• Independent runs generated nearly optimal hierarchies, allowing for the distinction of robust versus uncertain features based on consensus.

Conclusions:

• The developed MCMC-based sampler provides an effective and automated solution for hierarchical protein domain modeling.
• The method's robustness and convergence suggest its potential for reliable identification of protein functional subgroups.
• Future applications include generating confidence measures for features within domain hierarchies, aiding in robust biological interpretation.