MultiFG: integrating molecular fingerprints and graph embeddings via attention mechanisms for robust drug side effect prediction.

Accurate prediction of drug side effect frequencies is critical for drug safety assessment but remains challenging due to the high cost of clinical trials and the limited generalizability of existing models. We propose Multi Fingerprint and Graph Embedding model (MultiFG), a novel deep learning framework that integrates diverse molecular fingerprint types, graph-based embeddings, and similarity features of drug-side effect pairs. MultiFG incorporates attention-enhanced convolutional networks and utilizes the recently developed Kolmogorov-Arnold Networks (KAN) as the prediction layer to effectively capture complex relationships. In the task of predicting side effect associations for approved drugs, MultiFG achieved an AUC of 0.929, precision@15 of 0.206, and recall@15 of 0.642, outperforming the previous state-of-the-art by 0.7% points, 7.8%, and 30.2%, respectively. For side effect frequency prediction, MultiFG attained an RMSE of 0.631 and an MAE of 0.471, representing improvements of 0.413 and 0.293 over the best existing model. Moreover, MultiFG demonstrated strong generalization performance when predicting side effects for novel drugs. Overall, MultiFG offers a significant advancement in both side effect association and frequency prediction tasks, providing a practical and powerful tool for risk assessment across both marketed and investigational drugs.