A multimodal feature fusion model for drug–drug interaction prediction

Background Accurate prediction of drug-drug interactions (DDIs) is essential for ensuring medication safety in clinical practice. However, existing methods have limitations in fully exploiting sequential features and effectively modeling dynamic interactions, while deep graph neural networks are susceptible to over-smoothing when extracting information from two-dimensional molecular graphs. Methods To address these limitations, a multimodal feature fusion framework, termed MFF-DDI, is proposed. MFF-DDI integrates three complementary modalities. In the one-dimensional sequence branch, a group enhancement module is incorporated with a dual information extraction and fusion module to dynamically capture intermolecular interaction signals between drug pairs. In the two-dimensional graph branch, a deep graph isomorphism network equipped with a SkipNode mechanism is designed to mitigate over-smoothing and extract higher-order topological features. Meanwhile, a pretrained graph attention network based on dual-view contrastive learning is employed to effectively integrate prior network topology information from the knowledge graph. Results MFF-DDI demonstrates strong performance across multiple evaluation metrics. For known-drug prediction, MFF-DDI achieves an accuracy of 0.9786, an AUC of 0.9970, and an AUPRC of 0.9969 on the DrugBank dataset, as well as an accuracy of 0.9634 and an AUC of 0.9894 on the KEGG dataset. For predictions involving completely unseen drugs, MFF-DDI achieves accuracies of 0.8686 and 0.8554 on the DrugBank and KEGG datasets, respectively, outperforming the baseline models. Conclusion Overall, MFF-DDI effectively improves the accuracy and generalization capability of DDI prediction through the deep extraction and synergistic fusion of multimodal features, demonstrating its potential for application in complex medication scenarios.