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

Area of Science:

• Pharmacovigilance
• Computational Biology
• Drug Safety

Background:

• Adverse drug reactions (ADRs) present significant risks to patient safety and complicate clinical decisions.
• Current predictive models often lack interpretability in understanding drug-biological system interactions.

Purpose of the Study:

• To introduce the Knowledge Graph Reinforced Deep Q-Network (KRDQN) predictive framework for interpretable ADR prediction.
• To enhance the understanding of complex drug-induced biological effects.

Main Methods:

• Constructed a knowledge graph (KG) with drugs, targets, pathways, genes, and ADRs, enriched with node features.
• Employed a Deep Q-Network (DQN) within a reinforcement learning framework for ADR prediction.
• Validated KRDQN using five-fold cross-validation, reporting accuracy and AUC, and analyzed drug-drug and pathway similarities.

Main Results:

• KRDQN achieved superior performance over baseline methods, with a recall of 0.8171 and an AUC of 0.8327.
• The framework successfully predicted potential ADRs and their mechanistic pathways for sunitinib and indomethacin, aligning with clinical evidence.
• Identified biological pathways consistent with known drug effects, demonstrating practical utility.

Conclusions:

• The reinforcement learning-based KRDQN framework offers enhanced predictive performance for ADRs.
• KRDQN provides interpretable ADR predictions, making it a valuable tool for pharmacovigilance and clinical decision-making.
• This approach advances the prediction of drug safety and understanding of ADR mechanisms.