Hypergraph Representation Learning for Cancer Drug Response Prediction

Accurately predicting drug response is crucial for personalized cancer treatment. Current graph neural network methods mainly focus on pairwise relationships between cell lines and drugs, neglecting higher-order interactions. In this work, we propose a novel computational method called HRLCDR based on Hypergraph representation learning for predicting cancer drug responses. Firstly, HRLCDR constructs hypergraphs for cell lines and drugs based on their similarities, respectively. Subsequently, HRLCDR applies low-pass and high-pass hypergraph convolutions on these hypergraphs to capture the common and different features from the complex high-order interactions between nodes. Then it builds a heterogeneous graph based on the known cell line responses to drugs and employs parallel heterogeneous graph convolutions to learn primary interaction features of drugs and cell lines from known cell line-drug associations. Finally, HRLCDR integrates the cell line features and drug features learned from the hypergraphs and the heterogeneous graph to reconstruct the relationships between cell lines and drugs. We tested our model on the Cancer Drug Sensitivity Data (GDSC) and the Cancer Cell Line Encyclopedia (CCLE) databases. The results indicate that HRLCDR exhibits superior performance in predicting cancer drug responses compared with other state-of-the-art methods. The source code is available at https://github.com/weiba/HRLCDR .