Novel machine learning model for predicting cancer drugs' susceptibilities and discovering novel treatments

Background and Objective: Timely treatment is crucial for cancer patients, so it's important to administer the appropriate treatment as soon as possible. Because individuals can respond differently to a given drug due to their unique genomic profiles, we aim to use their genomic information to predict how various drugs will affect them and determine the best course of treatment. Methods: We present Kernelized Residual Stacking (KRS), a new multi-task learning approach, and use it to predict the responses to anti-cancer drugs based on genomic data. We demonstrate the superior predictive performance of KRS, outperforming popular competitors, by utilizing the Genomics of Drug Sensitivity in Cancer (GDSC) study and the Cancer Cell Line Encyclopedia (CCLE) study. Downstream analysis of feature genes selected by KRS is conducted to discover novel therapies. Results: We used two genomic studies to show that KRS outperforms a few popular competitors in predicting drugs' susceptibilities. Through downstream analysis of feature genes selected by KRS, we found that the PI3KAkt pathway could alter drugs' susceptibilities, and its expression correlated positively with the hub gene ERBB2. We discovered eight novel small molecules based on these feature genes, which could be developed into novel combination therapies with anti-cancer drugs. Conclusions: KRS outperforms competitors in prediction performance and selects feature genes highly correlated with drugs' susceptibilities. Novel biological results are found by investigating KRS's feature genes.