Approximating model probabilities in Bayesian information criterion and decision-theoretic approaches to model...

Area of Science:

• Computational Biology
• Evolutionary Biology
• Bioinformatics

Background:

• Accurate phylogenetic model selection is crucial for inferring evolutionary relationships from molecular sequence data.
• Existing methods like the Bayesian information criterion (BIC) use approximations that may not always be reliable.
• The increasing complexity of evolutionary models necessitates robust model selection techniques.

Purpose of the Study:

• To evaluate the accuracy of the BIC approximation for phylogenetic model selection.
• To compare BIC-based model selection with a direct estimation method using reversible jump Markov chain Monte Carlo (RJ-MCMC).
• To assess the impact of model selection discrepancies on downstream phylogenetic analyses.

Main Methods:

• Extended the decision-theoretic (DT) approach using RJ-MCMC to directly estimate model probabilities.
• Evaluated model selection performance across an extensive set of 406 general time reversible + Γ models.
• Analyzed 250 diverse molecular sequence datasets to compare BIC approximation with direct estimation.

Main Results:

• Model selection differed between BIC approximation and direct estimation for 45% of datasets.
• Discrepancies in model choice led to significantly different phylogenetic trees in 26% of cases.
• The model with the lowest BIC score differed from the highest posterior probability model in 30% of datasets.

Conclusions:

• The BIC approximation is adequate only when data strongly supports a single model.
• A substantial proportion of biological datasets lead to the selection of underparameterized models using BIC.
• Direct estimation of model probabilities provides a more reliable approach for complex phylogenetic model selection.