Network-based interpretation of the regulatory role of non-coding mutations and prediction of vulnerabilities using deep learning in cancer

Abstract

Cancer research has been actively conducted for centuries, and many progress has been made to increase the cancer survival rate. However, cancer genome exhibits extensive heterogeneity which poses a fundamental problem in understanding cancer biology. Therefore, I have attempted to identify the role of non-coding mutations and to discover the essential genes for novel therapeutic approach through network-based analysis. Cancer driving genes have been identified as recurrently affected by variants that alter protein-coding sequences. However, a majority of cancer variants arise in noncoding regions, and some of them are thought to play a critical role through transcriptional perturbation. Here I identified putative transcriptional driver genes based on combinatorial variant recurrence in cis-regulatory regions. The identified genes showed high connectivity in the cancer type-specific transcription regulatory network, with high outdegree and many downstream genes, highlighting their causative role during tumorigenesis. In the protein interactome, the identified transcriptional drivers were not as highly connected as coding driver genes but appeared to form a network module centered on the coding drivers. The coding and regulatory variants associated via these interactions between the coding and transcriptional drivers showed exclusive and complementary occurrence patterns across tumor samples. Transcriptional cancer drivers may act through an extensive perturbation of the regulatory network and by altering protein network modules through interactions with coding driver genes. Systematic in vitro loss-of-function screens provide valuable resources that facilitate the understanding of cancer vulnerabilities. Here, I developed deep learning-based methods to predict tumour-specific vulnerabilities in patient samples by leveraging a wealth of cell line screening data. Acquired dependencies of tumours were inferred in cases in which one allele was disrupted by inactivating mutations or in association with oncogenic mutations. Nucleocytoplasmic transport by Ran GTPase was identified as a common vulnerability in triple-negative or Her2-positive breast cancers. Vulnerability to loss of Ku70/80 was predicted for tumours that are defective in homologous recombination and rely on nonhomologous end joining for DNA repair. my experimental validation for Ran, Ku70/80, and a proteasome subunit using patient-derived cells showed that they can be targeted specifically in particular tumours that are predicted to be dependent on them. This approach can be applied to facilitate the discovery of novel therapeutic targets for different types of cancers.