Deep transfer learning for drug response prediction

The goal of precision oncology is to make accurate predictions for cancer patients via some omics data types of individual patients. Major challenges of computational methods for drug response prediction are that labeled clinical data is very limited, not publicly available, or has drug response for one or two drugs. These challenges have been addressed by generating large-scale pre-clinical datasets such as cancer cell lines or patient-derived xenografts (PDX). These pre-clinical datasets have multi-omics characterization of samples and are often screened with hundreds of drugs which makes them viable resources for precision oncology. However, they raise new questions: how can we integrate different data types? how can we handle data discrepancy between pre-clinical and clinical datasets that exist due to basic biological differences? and how can we make the best use of unlabeled samples in drug response prediction where labeling is extra challenging? In this thesis, we propose methods based on deep neural networks to answer these questions. First, we propose a method of multi-omics integration. Second, we propose a transfer learning method to address data discrepancy between cell lines, patients, and PDX models in the input and output space. Finally, we proposed a semi-supervised method of out-of-distribution generalization to predict drug response using labeled and unlabeled samples. The proposed methods have promising performance when compared to the state-of-the-art and may guide precision oncology more accurately.