Machine Learning for Drug Response Prediction

Abstract Modeling cell type-specific responses to drugs is an important yet challenging topic in drug development and personalized medicine. Our research program has developed a suite of algorithms for predicting cell-line level drug responses, and seeks to address the next-step questions in improving the transferability of predicting cell type-specific and combinatorial drug responses. Our goal over the next five years is to tackle the following three major challenges that hinder the transferability of drug response models. The first challenge is delivering drug response models across datasets generated by different platforms and labs. The second challenge is transferring drug response predictions from in vitro systems to humans. The third challenge is the heterogeneity of the samples, which can hinder the effectiveness of targeted therapy due to the acquisition of mutations and the evolution of tumors to evade the immune system. To address these challenges, we will focus on improving the robustness of models across datasets by establishing coherence networks of drugs. We will leverage and fine-tune large language models to refine our collection of literature-based drug information and extract new features for drugs. We will also explore the possibility of using the heterogeneity of the samples to improve the accuracy and robustness of drug response models. Overall, we envision that our research program will contribute to personalized medicine and expedite the drug development process.