Posterior predictive Bayesian phylogenetic model selection.

Area of Science:

• Evolutionary Biology
• Computational Biology
• Phylogenetics

Background:

• Bayesian phylogenetic model selection is crucial for understanding evolutionary relationships.
• Existing methods often rely solely on model fit to observed data, potentially overlooking predictive performance.
• New approaches are needed to evaluate models based on their predictive capabilities.

Purpose of the Study:

• To present and illustrate two novel posterior predictive approaches for Bayesian phylogenetic model selection: Gelfand-Ghosh (GG) and conditional predictive ordinate (CPO).
• To compare these methods with existing approaches, highlighting their unique contributions to assessing model fit.
• To demonstrate the application of GG and CPO using real-world phylogenetic datasets from green algae and flowering plants.

Main Methods:

• The Gelfand-Ghosh (GG) approach dissects model fit into posterior predictive variance (GGp) and goodness-of-fit (GGg) components.
• The conditional predictive ordinate (CPO) method provides site-specific model fit measures, combinable into the log pseudomarginal likelihood (LPML) for overall assessment.
• Both methods utilize samples from the posterior distribution, with CPO offering a computationally efficient cross-validation technique.

Main Results:

• The GG approach offers a nuanced view of model fit by separating variance and goodness-of-fit components.
• The CPO method provides valuable site-specific diagnostics and an efficient overall model fit measure (LPML).
• Both GG and CPO analyses were successfully applied to green algal cpDNA and flowering plant rDNA sequence data.

Conclusions:

• The GG and CPO methods provide novel perspectives on Bayesian phylogenetic model selection by focusing on predictive performance.
• These posterior predictive approaches complement traditional methods based on marginal likelihood and Bayes Factors.
• The presented methods enhance the evaluation of phylogenetic models, leading to more robust evolutionary inferences.