HyperCPI: A Novel Method Based on Hypergraph for Compound Protein Interaction Prediction with Good Generalization Ability

Identifying Compound Protein Interaction (CPI) through experiments is expensive, making CPI prediction with deep learning crucial in drug discovery. However, existing deep learning approaches face a challenge due to the lack of representations for non-pairwise relations and substructures in compounds, leading to limited performance and poor generalization ability. To address this challenge, a novel method named HyperCPI is proposed in this study. HyperCPI employs hypergraphs to represent compounds, where non-pairwise relations and substructures are represented as hyperedges. A hypergraph attention network is then utilized to extract high-order information from compounds. Additionally, an attention pooling layer is introduced to enhance the modeling of complex interactions between compounds and proteins. We conducted extensive experiments to evaluate HyperCPI. The results demonstrated that HyperCPI outperformed state-of-the-art (SOTA) methods in out-of-distribution (OOD) settings, underscoring its strong generalization ability. In typical experimental settings, HyperCPI achieved comparable performance with SOTA methods. Furthermore, we visualized HyperCPI to identify key atoms in compounds, and the findings align closely with real-world experiments. This demonstrates that HyperCPI offers excellent interpretability, a crucial feature for its application in drug discovery.