A Computational Pipeline for Graphical Modeling of Integrated Biomedical Data

Clinical decision making and biomedical research have the potential to be revolutionized by the abundance of readily available, multi-modal data. One of the main drivers of this wealth of data is next-generation sequencing technologies such as RNA-Seq and Single-cell RNASeq. These methods enable high-throughput measurements of the genome at a granular level. However, to truly understand the causes of disease and the effect of medical interventions, this data must be integrated with phenotypic, environmental, and behavioral data from individuals. In addition, effective modeling methods that can infer causal relationships from this data are required. This presents a host of modeling challenges such as 1) high dimensionality (low sample size and many variables), 2) redundancy among features, and 3) unmeasured variables that may confound the system under study. In addition, due to ethical concerns and cost, much of this data is observational. This means that no experimentally controlled perturbation was performed to measure the data. In this thesis, I present a pipeline to mine causal relationships from this integrated, observational biomedical data. The pipeline consists of three components: 1) Feature selection and clustering with prior knowledge, 2) Learning undirected graphical model structure to represent the joint distribution of mixed data, and 3) Learning causal graphical models with latent confounding. I demonstrate how this pipeline extracts useful knowledge via two cancer research applications: 1) Prediction of response to a prophylactic cancer vaccine and 2) Early detection of lung cancer from low-dose CT scans