GNNDRP: Graph Neural Network With Multi-Task Learning for Drug Response Prediction

Abstract

Using computational methods to personalize drug response prediction holds great promise to improve cancer therapy. Most existing methods use either biochemical information or response-related networks to predict drug response, nevertheless, the information they considered is not comprehensive. In this study, we present a novel end-to-end deep learning-based method Graph Neural Network with multi-task learning for Drug Response Prediction (GNNDRP). It leverages biochemical features as well as the hidden features from the heterogeneous network which incorporates the known drug-cell line responses, drug similarities, and cell line similarities, to complete the drug response prediction task. Moreover, GNNDRP designs a self-supervised task to enhance the representation capacity from the response network and further improve the model prediction performance. Extensive experiments show that GNNDRP outperforms existing state-of-the-art prediction methods under various experimental settings. The ablation analysis reveals that the biochemical characteristics, response-related network, and our self-supervised strategy can boost the predictive power. Additionally, case studies further validate the effectiveness of GNNDRP in identifying novel drug-cell line responses.