BAYESIAN FRAMEWORKS FOR PARSIMONIOUS MODELING OF MOLECULAR CANCER DATA

In this era of precision medicine, clinicians and researchers critically need the assistance of computational models that can accurately predict various clinical events and outcomes (e.g,, diagnosis of disease, determining the stage of the disease, or molecular subtyping). Typically, statistics and machine learning are applied to ‘omic’ datasets, yielding computational models that can be used for prediction. In cancer research there is still a critical need for computational models that have high classification performance but are also parsimonious in the number of variables they use. Some models are very good at performing their intended classification task, but are too complex for human researchers and clinicians to understand, due to the large number of variables they use. In contrast, some models are specifically built with a small number of variables, but may lack excellent predictive performance. This dissertation proposes a novel framework, called Junction to Knowledge (J2K), for the construction of parsimonious computational models. The J2K framework consists of four steps: filtering (discretization and variable selection), Bayesian network generation, Junction tree generation, and clique evaluation. The outcome of applying J2K to a particular dataset is a parsimonious Bayesian network model with high predictive performance, but also that is composed of a small number of variables. Not only does J2K find parsimonious gene cliques, but also provides the ability to create multi-omic models that can further improve the classification performance. These multi-omic models have the potential to accelerate biomedical discovery, followed by translation of their results into clinical practice