CLDDI: A Novel Method for Predicting Drug-Drug Interaction Events Based on Graph Contrastive Learning

Adverse drug-drug interactions (DDIs) may occur when drugs are combined to treat complex or comorbid diseases, which can result in adverse drug events, injury, and even death. Therefore, accurate prediction of potential DDI events is critical. Recently, automated computational methods such as deep learning are widely used for DDI events prediction. However, most of these methods only consider single information about the drug or rely on a large amount of label data, which easily leads to insufficient robustness and generalization ability. Accordingly, we proposed a novel end-to-end graph contrastive learning model for predicting multi-relational DDI events (CLDDI). It comprehensively considers the rich biomedical information of the Knowledge Graph (KG) and the structural information of the drug network. Specifically, we first generate two graph views by randomly corrupting the original KG, and compute a contrastive loss to maximize the agreement of node representation in these two views. Then we extract the drug embeddings obtained by contrastive learning and aggregate their neighbor information in the multi-relational DDI network. Finally, we combine the contrastive and supervised loss to learn the feature representation of nodes in an end-to-end fashion. Extensive experiments on real datasets show that the performance of CLDDI is competitive with the best baselines. Experimental results on sparse datasets further demonstrate that CLDDI has strong generalization performance and robustness.