Using decision trees to study the convergence of phylogenetic analyses

Abstract

In this paper, we explore the novel use of decision trees to study the convergence properties of phylogenetic analyses. A decision learning tree is constructed from the evolutionary relationships (or bipartitions) found in the evolutionary trees returned from a phylogenetic analysis. We treat evolutionary trees returned from multiple runs of a phylogenetic analysis as different classes. Then, we use the depth of a decision tree as a technique to measure how distinct the runs are from each other. Decision trees with shallow depth reflect non-convergence since the evolutionary trees can be classified with little information. Deep decision tree depths reflect convergence. We study Bayesian and maximum parsimony phylogenetic analyses consisting of thousands of trees. For some datasets studied here, a single distinguishing bipartition can classify the entire tree collection suggesting non-convergence of the underlying phylogenetic analysis. Thus, we believe that decision trees lead to new insights with the potential for helping biologists reconstruct more robust evolutionary trees.