INTEGRATING MULTI-MODAL BIOLOGICAL DATA UNCOVERS VULNERABILITIES AND COMPLEX INTERACTIONS IN CANCER

Abstract

Although advancements in experimental technologies have enabled characterization of cancer genomes at an unprecedented scale and resolution, functional characterization of the genes that cancer depends on for proliferation remains a challenge due to their diversity and complex interactions. Exposing the molecular machinery behind genetic dependencies can expand potential therapeutic targets and improve development of cancer therapies. A promising avenue for this challenge is to combine the rich resources of genomic data across multiple modalities. Multi-modal data ranging from simple genotype such as somatic mutations to complex phenotype such as cell viability provide diverse, complementary information useful for understanding the complex molecular mechanisms underlying cancer. In this dissertation, I present computational methods to investigate the landscape of dependencies in cancer by combining multiple types of biological data.In Chapter 2, I present SuperDendrix, a computational method to identify a combination of genomic and cell-type biomarker features that are associated with cellular response to CRISPR-mediated gene knockout. SuperDendrix incorporates a principled statistical model and a practically efficient combinatorial algorithm, thus finding a higher number of accurate results over existing algorithms. The identified associations are consistent with the direction and type of regulatory interactions between genes in biological pathways, demonstrating the importance of pathway topology in the design of cancer therapeutic strategies. In Chapter 3, I present NetMix2, an algorithm to identify subnetworks that contain highly mutated genes in a biological interaction network. NetMix2 leverages global structure of the interaction network by incorporating principles of network propagation and outperforms existing algorithms at identifying genes implicated in cancer as well as several other diseases. I also demonstrate the importance of rigorous evaluation of network-based algorithms. Finally in Chapter 4, I present a computational framework to interrogate drug mechanism of action by integrating drug sensitivity measurements, CRISPR-mediated gene knockout responses, and genomic/cellular biomarker features. This framework extends SuperDendrix to investigate associations between heterogeneous variables of mixed datatypes. I demonstrate that the identified associations help discover both primary and alternative targets of drugs and the molecular context of drug sensitivity that collectively suggest a number of non-oncology drugs as prominent options for cancer treatment.