RADDI: A Retrieval Augmented Framework for Drug-Drug Interaction Prediction

Drug-drug interactions (DDIs) can significantly impact drug efficacy and safety, potentially leading to severe adverse effects. Existing works on DDI event prediction have typically relied on labels of specific events for supervision, neglecting the importance of mining textual descriptions. This limits their ability to address two challenges: (1) the lack of observable data for new drugs, hindering meaningful feature extraction; (2) the highly imbalanced event distribution, which causes models to overfit to common categories and struggle with rare interactions. To address these challenges, we propose RADDI, a retrieval-augmented DDI prediction method. This approach improves prediction accuracy and adapts to the dynamic nature of new drug discovery. Specifically, to solve the first challenge, RADDI introduces a collaborative prediction strategy that integrates general knowledge transfer with specialized knowledge retrieval. This approach uses pretrained language models to generate embeddings for drug descriptions at a coarse level, enabling broad interaction classification. At a finer level, RADDI incorporates retrieval augmentation, using drug pair descriptions as retrieval keys and interaction categories as retrieval targets, thereby enhancing semantic understanding. For the second challenge, we design a class-aware probability distribution strategy to mitigate class imbalance. By leveraging the prior distribution of event categories, RADDI adjusts the retrieval sample weights and normalizes category probabilities, thereby improving the prediction accuracy for rare-class interactions while reducing over-reliance on high-frequency categories. Experiments on benchmark datasets demonstrate that RADDI excels in zero-shot DDI prediction scenarios, effectively balancing generalization to new drugs and maintaining high accuracy across various interaction categories.