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

## TL;DR

This paper introduces KRDQN, a new framework that uses knowledge graphs and deep learning to predict adverse drug reactions with interpretable results.

## Contribution

The novel KRDQN framework combines knowledge graphs with reinforcement learning to improve ADR prediction and provide interpretable biological insights.

## Key findings

- KRDQN achieved a recall of 0.8171 and an AUC of 0.8327, outperforming existing methods.
- The framework successfully identified ADR mechanisms for sunitinib and indomethacin consistent with clinical evidence.

## Abstract

Background: Adverse drug reactions (ADR) pose substantial risks to patient safety and challenge clinical decision-making. However, traditional predictive approaches frequently fail to deliver interpretable insights into the complex interplay between pharmaceuticals and biological systems. Methods: We propose the KRDQN (Knowledge Graph Reinforced Deep Q-Network) predictive framework. First, a knowledge graph (KG) that encompasses five entity types—drug, target, pathway, gene, and adverse drug reaction (ADR)—is constructed, and each node is enriched with intrinsic attribute features. A Deep Q-Network (DQN) is subsequently deployed within a reinforcement learning paradigm to generate interpretable ADR predictions. Model performance is evaluated by five-fold cross-validation, with accuracy and AUC reported. Finally, the Spearman correlation coefficients between drug–drug similarity and path–path similarity are computed, and case studies are conducted to further assess the predictive capability of KRDQN. Results: We evaluated KRDQN on a comprehensive data set encompassing both drug–drug interactions and ADR records. Experimental results demonstrate that KRDQN surpasses state-of-the-art baselines, attaining a recall of 0.8171 and an AUC of 0.8327. Furthermore, to demonstrate the practical value of the KRDQN prediction framework, we applied it to predict potential ADRs and their mechanism pathways for the drugs sunitinib and indomethacin. The results indicated that the KRDQN framework could identify biological mechanism pathways consistent with clinical evidence. Conclusions: In this study, we developed the reinforcement learning-based KRDQN predictive framework, which outperforms existing baselines in predictive performance and yields interpretable adverse drug reaction (ADR) predictions, thereby serving as a powerful tool for pharmacovigilance and clinical decision-making.