Abstract
Large-scale cancer drug sensitivity data have become available for a collection of cancer cell lines, but only limited drug response data from patients are available. Bridging the gap in pharmacogenomics knowledge between in vitro and in vivo datasets remains challenging. In this study, we trained a deep learning model, Scaden-CA, for deconvoluting tumor data into proportions of cancer-type-specific cell lines. Then, we developed a drug response prediction method using the deconvoluted proportions and the drug sensitivity data from cell lines. The Scaden-CA model showed excellent performance in terms of concordance correlation coefficients (>0.9 for model testing) and the correctly deconvoluted rate (>70% across most cancers) for model validation using Cancer Cell Line Encyclopedia (CCLE) bulk RNA data. We applied the model to tumors in The Cancer Genome Atlas (TCGA) dataset and examined associations between predicted cell viability and mutation status or gene expression levels to understand underlying mechanisms of potential value for drug repurposing.