EPISTASIS AND EVOLUTION OF DISEASE TRAJECTORIES IN MULTI-DIMENSIONAL STUDY OF GENOMIC AND PHENOMIC INTERACTIONS

Clinical care today is largely fragmented and reactive in nature, with presentation of symptoms a pre-requisite for further testing. Without a priori knowledge about a patient’s future disease risks, lines of treatment are often generic. A holistic approach evaluating predispositions and future comorbidities is needed to widen the clinical intervention window and tailor treatments to the patient. Using genomic and phenomic approaches, we examine inter-individual differences in complex disease risk and disease progression in captured and generated health data from diverse patient populations. In this dissertation, we first develop a framework to evaluate the role of complex genetic interactions in disease risk. We then develop methodology to holistically identify early indicators of complex disease outcomes in temporal patient data and model disease trajectories. In our genomic analysis, we identify 4 replicating pairwise epistasis models in long-range and high linkage disequilibrium. These models support the hypothesis that epistasis regulates evolutionarily-conserved essential gene functions. We identify epistasis as a “fine-tuning” mechanism that could explain phenotypic heterogeneity observed among patients with the same underlying disease. For example, we identify pleiotropic interactions between FLRT2 (chr14q31.3) and PDE4D (chr5q11.2) and RBFOX1 (chr16p13.3) in association with psoriasis and type 2 diabetes, respectively, which may explain a previously unknown shared genetic etiology between both conditions. In our phenomic analysis, we develop the DETECT feature extraction algorithm to generate curated patient trajectories for predictive modeling. The hypertension population in Penn Medicine was used to identify common disease paths from hypertension to 15 diverse related comorbidities such as renal failure, stroke, and heart attack. Clinical labs, medications, and procedural data were integrated to develop a prediction model for identifying which diagnosis came next for a patient. Our recurrent neural network model predicting stroke, renal failure, and retinopathy performed robustly with an ROCAUC of 0.713-0.837. This work employs a multi-dimensional framework to better understand disease risk and the longitudinal progression of disease at the patient-level, generating precision medicine insight and demonstrating how a preventative care paradigm can be implemented.