Multi-task regression learning for prediction of response against a panel of anti-cancer drugs in personalized medicine

The main goal of medicine research is to precisely provide treatment for each patient. In particular, clinicians want to use right drugs and doses for each patient according to their biological characteristics to maximize treatment's efficiency. With support of high-throughput technologies, a large amount of -omics and drug response data has been generated for human's tumors and cell lines. This wealth data facilitates the research for predicting drug response for each patient before selecting drugs for intake. Many computational methods have been also proposed for this purpose using -omics data of cell lines to predict the response on untested drugs. Previous methods have assumed a linear/non-linear relationship between -omics data and drug response for individual drugs. However, it should be noted that drugs can be functionally related to others since they may share some chemical structures. In addition, clinicians are also interested in knowing which biomarkers affect the response. Therefore, in this study, we assumed a linear relation and proposed a method based on multi-task regression learning to predict response for not only one but a panel of drugs. Comparing with previous single-task linear regression method, we found that our method achieves better performance in terms of correlation coefficient for major number of drugs.