Deep Generative Models for Cellular Representation Learning and Drug Sensitivity Prediction

The idea of precision oncology with drug sensitivity prediction was first introduced in the 1950s. With the emergence and promotion of in vitro cytotoxicity assays and high-through cell profiling techniques in the past three decades, precision oncology has advanced into an active research topic. The introduction of quantitative and computational methods has further boosted the growth of this area. Some impressive achievements have been made through the advancements, and yet we still have a long way to go towards the goal of precision oncology. Most previous studies were focused on using traditional statistical models to quantize the correlations between a small set of genetic features and drug responses. The models were often limited to a specific cancer type, and could only predict responses for a small number of drugs. These limitations prevent deploying computational approaches into the standard clinical practice. To further promote precision oncology, we need to embrace the tremendous amount of genomic features and utilize the comprehensive information they provide about the state of cellular signaling systems to build versatile computational tools that can predict sensitivity for various cancer drugs. In this dissertation project, we explore machine learning techniques, with a special focus on deep generative models for learning cellular state representations from omics data. We hypothesize that such representations can be used to replace traditional clinical and genetic features to significantly improve drug sensitivity prediction accuracy. Learning latent representations from raw input features and tuning the representations for downstream tasks has been successful in a number of deep learning application areas, including computer vision and natural language processing. Such strategies used to be impractical in systems biology due to the limited amount of data. With large systematic perturbation datasets like TCGA, LINCS, and GDSC that are now available, there is an unprecedented opportunity for introducing representation learning into the study of drug sensitivity prediction. We believe that the integration of deep learning representations presented in this dissertation will help advance the practice of pre-clinical drug-response prediction and contribute to a new age of precision and personalized cancer therapy