Untangling the tree of life: Which partitioning strategies improve phylogenetic inference?

The rate of evolution is known to vary greatly between morphological characters, which complicates the inference of phylogeny. To accommodate this rate heterogeneity, we can group together characters that are expected to evolve at similar rates. Partitioning by codon position is commonplace in molecular phylogenetics – for morphological data, no such general rule exists. Rosa, Melo, and Barbeitos (2019) advocate partitioning characters according to their homoplasy index on a maximum parsimony tree. A possible concern with this method is that the choice of tree may influence results in a manner that is not detected by standard model-testing approaches. Such methods rely on a tree to infer a tree – but what if that first tree is unreliable? I tested homoplasy partitioning based on a spectrum of trees ranging from a published tree through to completely random trees, and found that the topology on which homoplasy is calculated does not affect the topology recovered by Bayesian analysis. I compared homoplasy partitioning to other partitioning strategies, including strategies informed by biological criteria: partitioning by character type (Sereno, 2007) and partitioning by anatomy. Using Steppingstone sampling (Xie et al., 2011) to estimate the fit of each partitioned model, I found that homoplasy is the only reliable approximator of evolutionary rate. Partitioning has the capacity to significantly increase model fit, but only homoplasy partitioning consistently produced better results than an unpartitioned model. No link between mosaic evolution or character type and evolutionary rates could be confirmed.