In modern medical systems, drug therapy remains one of the most essential approaches for disease prevention and treatment. In particular, combination therapy has become a common strategy for improving therapeutic efficacy and reducing drug resistance in the treatment of complex diseases. However, the use of multiple drugs simultaneously significantly increases the risk of drug–drug interactions (DDIs), which may alter the pharmacological effects of drugs or even lead to adverse reactions, posing a threat to patient safety. Therefore, accurate prediction of potential DDIs at the early stages of drug development and clinical application is of great importance. With the advancement of computational methods, DDI type prediction models based on machine learning and deep learning have gradually become a research focus and have improved prediction efficiency and accuracy to some extent. Nevertheless, existing approaches still have room for improvement. First, most graph neural networks tend to focus on node features when processing molecular graphs, while the utilization of edge features remains relatively insufficient. Second, current models often fail to fully capture the latent relationships among different types of features. In addition, with the rapid growth of pharmaceutical data, how to effectively integrate multi-source information to further improve prediction performance remains a key challenge. Therefore, developing DDI type prediction methods that can effectively integrate multi-dimensional features and model complex relationships is of significant importance for advancing drug safety evaluation and the development of precision medicine.

Recently, the research group led by Gao Jie from Jiangnan University, China published an article titled "MFCN-DDI: Capsule network based on multimodal feature for multitype drug–drug interaction prediction" in Quantitative Biology, proposing an algorithm called MFCN-DDI for predicting DDI types based on multimodal features and capsule networks. By extracting multimodal features of drugs and utilizing capsule networks to achieve efficient feature fusion, the algorithm accurately captures the complex relationships among different features, thereby enabling precise prediction of DDI types.

As shown in Figure 1, the research team develops a DDI type prediction method named MFCN-DDI. This approach fully exploits edge features in drug molecules during molecular representation learning to enhance the expressive capability of molecular graphs. It further integrates knowledge graph features, molecular fingerprint features, and pharmacological features to characterize drug properties from multiple perspectives. On this basis, a capsule network-based feature fusion module is introduced, which dynamically adjusts information transmission weights according to feature relevance and contribution, enabling more effective capture of complex relationships among different modalities and thereby improving prediction accuracy. To comprehensively evaluate model performance, MFCN-DDI is assessed on both multiclass and multilabel prediction tasks. Ablation studies and case studies are conducted to validate the importance of various features and key modules, as well as the effectiveness and applicability of the model in real-world scenarios. The results demonstrate that MFCN-DDI achieves significant improvements in DDI type prediction accuracy, providing strong support for safe clinical medication and facilitating the optimization of combination therapy and patient health. In the future, the research will further focus on practical application needs by enhancing model interpretability to improve the clinical understandability and utility of predictions, exploring richer drug feature representations and learning methods to build more comprehensive drug profiles, and incorporating more real-world clinical DDI data to further improve the model’s generalization ability and practical value in complex clinical scenarios.
DOI
10.1002/qub2.70021