Abstract 2099: Assessing cancer drug response prediction from gene expression

Abstract

Abstract Customizing treatments targeted to a patient specific disease landscape remains an open challenge in precision medicine. Gene expression data, generated through next generation sequencing technologies, has shown significant power for supervised machine learning approaches. Here, we evaluated neural network performance at predicting cancer cell-line specific drug sensitivity. Gene expression data for 17,737 genes across 1014 human cancer cell-lines with IC50 concentrations for 251 anti-cancer drugs were obtained from the Genomics of Drug Sensitivity in Cancer Project. Cell-lines were considered sensitive if IC50 concentration was below maximum drug concentration. For neural network input, gene expression values were scaled to a range of -0.5 to 0.5 through log transformation, z-scoring, subtracting the minimum and dividing by the range. A hyperparameter optimized neural network model consisting of two hidden layers of size 1024 and 512 respectively resulted in a test set accuracy ranging from 60.2% to 67.8% across drugs compared to an average baseline performance of 53.4%. Dimension reduction through clustering and Nearest Neighbor Networks did not result in improved performance. We found that classical machine learning classification methods including support vector machines showed comparable performance to our neural network approach. However, neural network cross-training on drugs sharing the same putative targets (e.g. AURKA) led to a 5.03% improvement in neural network prediction accuracy. Overall, our analysis shows that utilizing gene expression profiles independent of other -omics data for cancer drug response prediction through machine learning frameworks offers modest predictive capabilities. To increase performance, we suggest augmenting training size through shared pathway cross-training, optimizing feature encoding to maximize neural network predictive capabilities, and incorporating other -omics data. Citation Format: James Talwar, Hannah Carter. Assessing cancer drug response prediction from gene expression [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2099.