The work presented in this dissertation focuses on the development of robust deep learning models to predict in-vitro cancer drug responses, with utility in precision oncology research tasks such as drug repurposing and prioritization of disease-specific drug combinations. This work addresses two critical shortcomings in the current approaches to predicting cancer drug responses with deep learning: 1) data quality issues inherent in high-throughput drug screening datasets by developing algorithms for detection of atypical or low-quality data and 2) improved prediction and utility of perturbation biology models by developing algorithms that operate on mechanistic prior knowledge.