Detecting Convergence of Bayesian Searches in Computational Phylogeny

Computational phylogeny attempts to use descriptions of various taxa to generate an evolutionary tree of life. When generating phylogenic trees using Bayesian methods, it is difficult for the researcher to determine the run time necessary in order to get a good sample from the target distribution. Less than adequate run times can miss important features, resulting in an inaccurate representation of the data. Current methods to detect whether the analysis must continue require the extra overhead of multiple runs and are subject to researcher interpretation. We propose a set of convergence diagnostics that reduce overhead and provide intuitive information about convergence. First, a consensus tree is built during runtime of rejected tree proposals. When this reject consensus tree matches a consensus tree built from accepted trees, a local maximum has been reached. Second, the set of proposed trees is analyzed by taking the Robinson-Foulds distance between each proposals and a set of random trees. This metric evaluates the variety of search propositions used by reporting back the percentage of random trees that shared bipartitions with a sample. This set of metrics was integrated into the MrBayes software package and tested on a variety of real data sets. The metrics were found to be both predictive and conservative, demonstrating their effectiveness for testing search convergence.