Abstract PR02: Towards a Cancer Dependency Map

The mapping of cancer genomes is rapidly approaching completion. The genomic information encoded by individual patients9 tumors should, in principle, provide a guide for predicting acquired cancer dependencies. Unfortunately, while the success of precision cancer genomics hinges on the decoding of such dependencies, we lack the ability to predict dependencies for most individual tumors. The challenge stems from the absence of clinical data relating genotypes with dependencies since most cancer mutations are rare and our arsenal of cancer drugs is incomplete. A comprehensive Cancer Dependency Map comprised of a catalog of genetic and small molecule vulnerabilities across a diverse set of cancers, along with robust statistical models able to predict these vulnerabilities from molecular and genomic features, would provide a roadmap of targets ripe for therapeutic development and would help reveal the mechanisms underlying the emergence of these vulnerabilities. Here, we report progress in creating a Cancer Dependency Map consisting of the following components: 1) Systematic genetic perturbation (RNAi/CRISPR) of over 600 cancer cell models representing a wide range of human cancers and cell lineages using massively parallel genome scale loss-of-function screens. 2) Computational segregation of on- from off-target effects of RNAi enabling the discovery of outlier dependencies. 3) Predictive modeling to discover biomarkers for each dependency. Our results demonstrate that our analytical approach (DEMETER) that models both gene and miRNA-based seed sequence effects effectively segregates on- from off-target effects of shRNAs. We discover 768 preferential dependencies whose suppression decreases viability at a level greater than six standard deviations in at least one of 503 cancer models and 105 such dependencies each present in at least 15 models. We find that 95% of the cancer models screened are strongly sensitive to the suppression of at least one of these dependencies, and that many models have common dependencies so that all models harbor at least one six-sigma dependency out of a set of only 76. Using a custom random forest based predictive modeling framework (ATLANTIS), we discover predictive biomarkers for hundreds of dependencies. These include known and novel vulnerabilities specified by somatic oncogenic alterations, overexpression of genes that specify lineage and differentiation, copy-number driven essentiality, and loss of functionally redundant paralogs. These observations provide a rigorous computational and experimental foundation for the creation of a comprehensive Cancer Dependency Map. Subsampling and projection analyses suggest that over 10,000 genomically characterized cancer cell models will be needed to achieve this important goal. This abstract is also being presented as Poster B43. Citation Format: Aviad Tsherniak, Francisca Vazquez, Barbara Weir, Philip Montgomery, Glenn Cowley, Stanley Gill, Gregory Kryukov, Sasha Pantel, Will Harrington, Mike Burger, Robin Meyers, Levi Ali, Amy Goodale, Yenarae Lee, Levi Garraway, Jesse Boehm, David Root, Todd Golub, William Hahn. Towards a Cancer Dependency Map. [abstract]. In: Proceedings of the AACR Precision Medicine Series: Targeting the Vulnerabilities of Cancer; May 16-19, 2016; Miami, FL. Philadelphia (PA): AACR; Clin Cancer Res 2017;23(1_Suppl):Abstract nr PR02.