Significant Advances in Research on Drug-Drug Interactions in the Past Five Years
Yun Zuo, Jiayi Ji, Yuwen Li, Bulanni Xiong, Dandan Qiao, Jiankang ChenIntroduction:
Drug–Drug Interaction (DDI) prediction is pivotal in pharmaceutical research and clinical practice, enhancing medication safety and optimizing therapeutic strategies. Traditional experimental approaches, constrained by high costs and time consumption, are insufficient for large-scale screening. Consequently, computational prediction methods have garnered increasing attention.
Methods:
This review systematically analyzes key challenges in DDI prediction: balancing model performance with interpretability, integrating heterogeneous features, ensuring sample set reliability, and optimizing algorithmic architectures. Database resources are categorized into drug interaction databases (e.g., DrugBank), compound feature databases (e.g., PubChem), and protein-related databases (e.g., UniProt).
Results:
Four primary feature encoding strategies are examined: pre-trained language model representations, multimodal fusion encoding, chemical structure-based encoding, and knowledge graphbased encoding. Mainstream DDI prediction approaches are classified into five paradigms: traditional machine learning, deep learning, graph neural networks, knowledge graph-based methods, and contrastive learning approaches.
Discussion:
Empirical results from benchmark datasets demonstrate that deep learning models integrating multi-scale features and attention mechanisms (e.g., MSDF) achieve superior predictive accuracy, while graph neural networks (e.g., GCN-BMP) excel at capturing complex relational structures.
Conclusion:
This review provides a robust reference for methodological selection and outlines promising directions for future innovation in DDI prediction, offering valuable guidance for advancing drug safety evaluation in precision medicine.