Characterization of complex genetic interactions in cancer cells

Abstract

Cancer dependencies are genes required for proliferation and survival in cancer cells, making them potential therapeutic targets. Cancer dependencies are often selective for subtypes of cancers, and are measured as conditional changes in fitness caused by genomic and molecular aberrations and are known as genetic interactions (GIs). However, cancer dependencies can be influenced by several factors, resulting in higher-order GIs, which are difficult to predict for any given cancer type. This has resulted in a varied success rate for the development and application of new targeted therapies. In order to systematically identify cancer dependencies and GIs, genome-wide CRISPR/Cas9 knock-out (KO) screens across pan-cancer libraries have been performed. Due to the limitations of large-scale screenings with multiple KOs, advanced computational strategies have to be used for inferring higher-order interactions and predicting cancer dependencies based on the molecular characteristics of the cancer cells.

In this study, we show a robust method to infer GIs from pan-cancer CRISPR screens based on the genetic and transcriptional background of the cancer cells. Pairwise GIs were inferred by predicting fitness change from CRISPR/Cas9-mediated gene deletion using multivariate penalised linear regression, combined with null-hypothesis testing. Furthermore, we developed an XGBoost approach, where regression tree structures were mined for variable interactions, to discover potential complex higher-order interactions from transcriptional changes between cancer cells. Novel GIs were subsequently mapped and analysed in a genome-wide GI network.

In conclusion, our study shows a robust framework for predicting complex GIs involved in the regulation of cancer fitness.