From Graph Neural Networks to Large Language Models: A Comprehensive Review of Deep Learning-Based Drug—Drug Interaction Prediction

Abstract

Drug-drug interactions (DDIs) are a major cause of adverse drug reactions, therapeutic failure, hospitalization, and mortality, particularly among elderly patients and individuals receiving multiple medications (polypharmacy). The mechanisms for the development of DDIs are either through pharmacokinetic effects, such as those related to absorption, distribution, metabolism, and excretion, or pharmacodynamic effects, such as physiological and receptor mediated effects. Common drugs that interact with each other clinically, including warfarin, macrolide antibiotics, ACE inhibitors, proton pump inhibitors, antifungal agents, non-steroidal anti-inflammatory drugs, and digoxin, can cause complications such as bleeding, hypoglycemia, hyperkalemia, cardiotoxicity, and organ toxicity. Traditional methods for identifying DDIs, including clinical trials, post-marketing surveillance, and laboratory studies, are expensive, time-consuming, and incapable of evaluating the enormous number of possible drug combinations. Consequently, computational approaches have become increasingly important for improving drug safety and supporting clinical decision-making. In recent years, deep learning has emerged as a transformative technology capable of learning complex biological, chemical, and pharmacological relationships extracted from vast biomedical data to predict DDI. This review primarily focuses on deep learning methods for drug-drug interaction (DDI) prediction. Related computational tasks, including molecular representation learning, drug-target interaction (DTI) prediction, and DDI extraction from biomedical literature, are discussed only where they have contributed to the development of modern DDI prediction frameworks based on GNNs, LSTMs, and Transformer-based large language models. The innovative architectures combine molecular graphs, SMILES sequences, biomedical knowledge graphs and multimodal clinical information to enhance prediction accuracy, interpretability and scalability. Overall, deep learning-enabled DDI prediction systems hold great promise for precision medicine, pharmacovigilance and drug discovery.