Bayesian nonparametric models of genetic variation

We will develop three new Bayesian nonparametric models for genetic variation. These models are all dynamic-clustering approximations of the ancestral recombination graph (or ARG), a structure that fully describes the genetic history of a population. Due to its complexity, efficient inference for the ARG is not possible. However, different aspects of the ARG can be captured by the approximations discussed in our work. The ARG can be described by a tree valued HMM where the trees vary along the genetic sequence. Many modern models of genetic variation proceed by approximating these trees with (often finite) clusterings. We will consider Bayesian nonparametric priors for the clustering, thereby providing nonparametric generalizations of these models and avoiding problems with model selection and label switching. Further, we will compare the performance of these models on a wide selection of inference problems in genetics such as phasing, imputation, genome wide association and admixture or bottleneck discovery. These experiments should provide a common testing ground on which the different approximations inherent in modern genetic models can be compared. The results of these experiments should shed light on the nature of the approximations and guide future application of these models.