GIL-DDI: Multi-View Graph Invariant Learning for Unknown Drug-Drug Interaction Prediction

Drug-Drug Interaction (DDI) prediction is essential for evaluating the side effects of a new drug and adverse interactions before the clinical application.The latest research applies multi-view data to enhance the generalization ability of models to predict new drug interactions, mainly unknown Drug-Drug Interaction (uDDI).However, a new drug's feature inevitably encounters the feature-shift problem; the trained models have not previously learned information about the new drug, significantly decreasing the uDDI prediction's accuracy.Thus, we proposed the GIL-DDI model that tries to extract the invariant features of known drugs, alleviating the impact of the feature-shift problem on the prediction of uDDI.In detail, the graph attention network(GAT) models embed multi-view drug graphs, including drug-chemical entities, drug substructures, drug-drug interactions, and molecular structures. Then, invariant features corresponding to the new drug are learned from the knowledge graph of the previous drugs.After that, a variant feature of the new drug is embedded through the GAT models and fused with learned invariant drug features to predict the DDI.Extensive experiments on real-world drug datasets indicate that the proposed method achieves new state-of-the-art records on new drug DDI prediction tasks. The source code is available at https://anonymous.4open.science/r/GIL-DDI-F701/README.md.