Abstract 5323: Integrated computational cell-line modeling of drug sensitivity and high-throughput siRNA screening reveals novel molecular biomarkers for conventional chemotherapy

Abstract

Abstract Introduction: Identifying molecular aberrations that drive carcinogenesis remains a daunting problem. Uncovering important combinations of such aberrations, e.g. overexpression of a gene in concordance with copy number amplification, pose additional challenges. Publicly available large-scale cell line panels profiled for drug response often do not reflect the true tumor heterogeneity and therefore predictions made solely based on such data may be difficult to translate into therapeutic targets. Integrating predicted associations with drug sensitivity (resistance) on one hand, with siRNA screens with the same drug (patient-derived samples and cell lines) on the other hand, should allow for the identification of therapeutically relevant molecular traits. Methods: We propose a novel Bayesian model for multi-task learning with a gene-wise prior, which accounts for multiple aberrations in a given gene (expression, mutation and copy number). Our method allows for the simultaneous learning of drug response for multiple drugs in a given panel (N_Sanger=138, N_CCLE=24 drugs). Our model aims to identifying sets of important aberrations, which individually occur at low frequencies. To train and evaluate the performance of our model we use the CCLE and Sanger datasets. Further, we use the Cisplatin and Doxorubicin drug responses from the Sanger dataset to rank genome-wise aberrations to find features best predictive of response to the chemotherapeutic agents. We then consider our findings in the context of high-throughput siRNA synthetic lethal screens in head and neck, ovarian and breast cancer with Cisplatin and Doxorubicin in patient-derived samples and cell lines. Results: Our approach improves the quality of drug sensitivity prediction (Pearson correlation in five-fold cross validation) over methods that model individual drug responses (e.g. Elastic Net). It further allows for the subtyping of drugs by target/mechanism of action commonalities, which we recover closely for drugs with known mechanism of action. Our model recovers reasonably the targets of consensus drug interactions (e.g. Erlotinib and EGFR) and in addition identifies other known associations currently undergoing preclinical/clinical validation (i.e. Cisplatin and JAK2 mutation, and RUNX3 expression and Doxorubicin). Integrated analysis of the results from our predictive model in conjunction with the siRNA screens are currently under investigation. Conclusions: We present an integrated approach that combines a novel Bayesian multi-task learning model with high-throughput siRNA screens. Our approach aims to uncover sets of important aberrations and allows for the subtyping of drugs based on similarities in targets and mechanisms of action. We integrate our results with high-throughput RNAi experiments to identify synthetic lethal events in specific therapeutic context. Citation Format: Olga H. Nikolova, Mehmet Gönen, Rodrigo Dienstmann, In Sock Jang, Russell Moser, Silvia Cermelli, Chang Xu, Ryan M. Mitchell, Eduardo Mendez, Carla Grandori, Christopher Kemp, Stephen Friend, Justin Guinney, Adam Margolin. Integrated computational cell-line modeling of drug sensitivity and high-throughput siRNA screening reveals novel molecular biomarkers for conventional chemotherapy. [abstract]. In: Proceedings of the 105th Annual Meeting of the American Association for Cancer Research; 2014 Apr 5-9; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2014;74(19 Suppl):Abstract nr 5323. doi:10.1158/1538-7445.AM2014-5323