Bayesian nonparametric models for biomedical data analysis

Abstract

In this dissertation, we develop nonparametric Bayesian models for biomedical data analysis. In particular, we focus on inference for tumor heterogeneity and inference for missing data. First, we present a Bayesian feature allocation model for tumor subclone reconstruction using mutation pairs. The key innovation lies in the use of short reads mapped to pairs of proximal single nucleotide variants (SNVs). In contrast, most existing methods use only marginal reads for unpaired SNVs. In the same context of using mutation pairs, in order to recover the phylogenetic relationship of subclones, we then develop a Bayesian treed feature allocation model. In contrast to commonly used feature allocation models, we allow the latent features to be dependent, using a tree structure to introduce dependence. Finally, we propose a nonparametric Bayesian approach to monotone missing data in longitudinal studies with non-ignorable missingness. In contrast to most existing methods, our method allows for incorporating information from auxiliary covariates and is able to capture complex structures among the response, missingness and auxiliary covariates. Our models are validated through simulation studies and are applied to real-world biomedical datasets.