Bayesian Computational Inference of Species Trees and Population Sizes

Abstract

Rapid advances in sequencing technology have allowed researchers to study many fundamental questions about Evolution and Life through the analysis of molecular sequence data. The computational analysis of sequence data can reveal the patterns and processes of speciation, domestication, epidemiology, evolution of traits, biogeography, and host/parasite co-evolution. Such analyses are also central to the systematic classification of life. All of these applications require a mathematical model that characterizes the random processes governing the evolution of DNA sequences over time, subject to the specific conditions of interest. It turns out that evolutionary models also share another common basis: the need to account for the ancestral relationships and sizes of the populations within which the genetic material is being propagated. Historically, those aspects were part of both the area of populations genetics and systematics - the study of biological diversity and its origins. Those two disciplines are experiencing a synthesis in the last few decades under the branch of Molecular Phylogenetics". In this thesis we develop two methods; one for estimating the effective size of a single population, the second for estimating the family (species) tree of several populations and their sizes. Both methods utilize multi-gene data from multiple individuals, a key ingredient for reducing the uncertainty of estimates. The methods are Bayesian, utilizing the generic Markov Chain Monte Carlo algorithm to compute posterior probabilities for high-dimensional and extremely complex models. Posterior probabilities are direct probabilities of the model under suitable prior assumptions, and uncertainty bounds are an integral part of the Bayesian approach. Uncertainty bounds are especially valuable in analysis of genetic data, which may contain an extremely weak signal. In this thesis we also explore some theoretical issues relating to the tree prior and model violations. Due to the complexity of the models, computer simulations play a major role in investigating model properties and performance as a function of the amount and quality of data. The methods are implemented as part of the software package BEAST and have already been used in several studies.