Computational approaches for drug–drug interaction prediction: A survey

Drug–Drug Interactions (DDIs) continue to pose a significant risk to clinical safety, especially in cases of polypharmacy, where unforeseen interactions can cause adverse effects and financial burdens on the healthcare system. Existing experimental and rule-based methods for detection are limited in both scalability and coverage. This has created a need for methods that are capable of screening potential interactions automatically on a large scale. This review traces the methodological evolution of computational DDI prediction, from descriptor-based Quantitative Structure–Activity Relationship (QSAR) models to graph neural networks, transformer architectures, and multimodal learning frameworks. Rather than only summarizing model categories, it explores how dataset construction, feature choice, and evaluation strategy influence published results and model performance generalization. The review addresses existing methodological issues, such as dataset bias, negative sample generation, benchmarking inconsistency, restricted interpretability, and the challenge of predicting interactions involving unseen drugs or higher-order drug combinations. Recent progress in hierarchical graph learning, multimodal integration, and clinically oriented DDI prediction is analyzed to highlight both current opportunities and remaining barriers to robust and clinically useful prediction systems.