Complecto errorem: Embracing Uncertainty in Molecular Phylogenetic Analysis.

Abstract

Model-based inference is ubiquitous to myriad fields of science, but perhaps has not elsewhere effected such a boon as when applied to the inherently noisy data involved in studies of ecology and evolutionary biology. Today, model-based inference in both maximum likelihood and Bayesian frameworks is ubiquitous in phylogenetic analysis. However, despite the advances made possible through this general approach, sensitivity of inferences to underlying analytical assumptions have not been well explored. In chapter 1, I review the molecular clock literature, critically examining the methods that are currently available. Throughout I comment on how the adoption of individual relaxed clock methods has contributed to our ‘phylochronological’ understanding of early avian diversification. I then discuss potential drawbacks with current molecular clock inference, and suggest directions to take to make these models more general and realistic. In chapters 2 and 3 I apply various molecular clock approaches to the analysis of independent avian nuclear and mitochondrial DNA matrices, respectively. I show that, despite the various methods making very different assumptions about how molecular substitution rates change across the tree, both genomes strongly support an early Cretaceous origin of most of the major extant avian lineages. These results support previous molecular studies, and suggest that a literal reading of the fossil record may be biased for this particular clade. In chapter 4, I seek an optimal partitioning strategy (model) for phylogenetic analysis of a mixed nuclear/mitochondrial matrix for New World Vultures (Cathartidae). Although the specific model chosen depended somewhat on the statistical criterion employed, in general the model selected partitioned nuclear data by gene, and mitochondrial data by codon position. The tree topology inferred assuming this optimal model is incompatible with trees inferred assuming simple (unpartitioned) models. Finally, in chapter 5 I apply a posterior predictive approach to arbitrating amongst candidate partitioned models for example primate and avian mitochondrial DNA alignments. The discriminating power of the summary statistic employed (a multinomial test statistic summarizing the shape of the alignment column frequency spectrum) is generally low. However, the approach is easily extendable through constructing more sufficient summary statistics that more fully gauge model predictive ability.