Predicting Drug Sensitivity of Cancer Cell Lines via Collaborative Filtering with Contextual Attention
Proceedings of the 5th Machine Learning for Healthcare Conference, PMLR 126:660-684, 2020.
Accurate anti-cancer drug recommendations and the identification of essential biomarkers for this task are crucial to precision oncology. Large-scale drug response assays on cancer cell lines provide a potential way to understand the interplay of drugs and cancer cells. In this work, we present CADRE (Contextual Attention-based Drug REsponse), a model that accurately infers the response of cancer cell lines to a panel of candidate compounds based on the omics profiles, such as gene expressions, of cancer cells. CADRE builds on the framework of collaborative filtering, which provides robustness to the noise of biological data by leveraging similarities within drugs and cell lines. It utilizes the contextual attention mechanism to identify informative biomarkers of these cell lines, which boosts prediction accuracy and affords interpretability of results. In addition, CADRE incorporates external knowledge of drug target pathways and co-expression patterns of genes to further improve feature representations and model performance. Comprehensive evaluations of CADRE and competing models on two large-scale pharmacogenomic datasets show its superiority in both prediction performance and interpretability. CADRE identifies as vital biomarkers genes related to intracellular vesicles and signaling receptor binding, shedding light on its translational potential in the clinical practice of cancer treatment.