KRDQN: An Interpretable Prediction Framework for Adverse Drug Reactions via Knowledge-Graph Reinforced Deep Q-Learning

The study introduces the KRDQN framework, which combines knowledge graphs and reinforcement learning to predict adverse drug reactions (ADRs) with high accuracy and interpretability. The framework outperformed existing models, achieving a recall of 0.8171 and an AUC of 0.8327, and effectively identified biological pathways linking drugs to ADRs, such as sunitinib-induced alopecia. Despite its strengths, including the ability to align predictions with clinical evidence, the model faces challenges like class imbalance and static databases, suggesting future improvements in sensitivity to rare ADRs and dynamic data updates. The study underscores the importance of multi-modal feature fusion and KRDQN's potential applications in pharmacovigilance and clinical decision-making.