Graph-based Approaches for Cancer Genomics, with Applications to Cancer Signaling and Dependencies

The era of widely applicable sequencing and genomic technologies led to the generation of many large-scale datasets exploring genomes, transcriptomes, or epigenomes of tumors. Availability of a wide range of datasets necessitated the development of new computational analysis approaches to generate novel insights from these datasets and improve our understanding of tumor development and progression.

This work focuses on a variety of graph-based approaches to evaluate their use in cancer genomics. We first focused on a graph-based semi-supervised learning approach called label propagation as a method to generate signaling networks from a gene set of interest. A distance metric based on the concept of maximal common subgraph was then established to quantify the degree of similarity observed across different networks. These two approaches were then combined to examine two separate cancer genomics datasets. Our first application focused on genes frequently altered across patients to build signaling networks that represent genes and pathways that are transcriptionally altered as a result of these mutations. These networks revealed the range of molecular events affected by each mutation and conserved changes observed across networks highlighted the critical signaling pathways tumors dysregulate through distinct alterations. The other area of focus for label propagation was the analysis of a set of melanoma samples resistant to BRAF inhibitors. Evaluation of networks of individual resistant samples revealed signaling changes shared across samples that have similar resistance mechanisms or originated from the same patient. Finally, drug response profiles of a large set of drugs were examined across cell lines belonging to eighteen different tumor types, by building bipartite graphs representing sensitivity patterns of drugs. These bipartite graphs were used to generate drug similarity graphs that revealed shared response profiles of drugs targeting distinct processes, which provided opportunities to refine the annotations of drug targets. Degree distributions of bipartite graphs also revealed drugs connected to exceptional responder cell lines, whose unique genomic profiles nominated potential markers of drug response. Collectively, the studies discussed here emphasize a variety of use cases for graph-based approaches in cancer genomics.