Graphical models for de novo and pathway-based network prediction \ud over multi-modal high-throughput biological data

It is now a standard practice in the study of complex disease to perform many high-throughput -omic experiments (genome wide SNP, copy number, mRNA and miRNA expression) on the same set of patient samples. These multi-modal data should allow researchers to form a more complete, systems-level picture of a sample, but this is only possible if they have a suitable model for integrating the data. Due to the variety of data modalities and possible combinations of data, general, flexible integration methods that will be widely applicable in many settings are desirable. In this dissertation I will present my work using graphical models for de novo structure learning of both undirected and directed sparse graphs over a mixture of Gaussian and categorical variables. Using synthetic and biological data I will show that these models are useful for both variable selection and inference. Selecting the regularization parameters is an important challenge for these models so I will also cover stability based methods for efficiently setting these parameters, and for controlling the false discovery rate of edge predictions. I will also show results from a biological application to data from metastatic melanoma patients where our methods identified a PARP1 slice site variant that is predictive of response to chemotherapy. Finally, I present work incorporating miRNA into a pathway based graphical model called PARADIGM. This extension of the model allows us to study patient-specific changes in miRNA induced silencing in cancer