Prediction of Adverse Drug Reactions to Anticancer Agents Based on Heterogeneous Graph Neural Networks

Objective Adverse reactions to anticancer drugs pose significant risks to treatment safety and efficacy, while current monitoring systems suffer from delayed detection and poor interpretability. To address these challenges, this study aimed to develop an interpretable heterogeneous graph neural network model for accurate prediction and risk backtracking of anticancer drug adverse reactions. Methods Multi-source data on drugs, proteins, diseases, and adverse drug reactions (ADRs) were integrated from DrugBank, STRING, CTD, and SIDER to construct a heterogeneous graph. A model named MP-HGNN based on short meta-path aggregation was proposed. It decomposes long meta-paths into semantically clear two-hop segments for local information fusion, aggregates them via learnable channel weights, and outputs drug-ADR association risks as conditional probabilities. Results Experimental results demonstrated that the MP-HGNN model achieved an area under the curve (AUC) of 0.9684 and an area under the precision-recall curve (AUPR) of 0.8386, outperforming 13 baseline models in both discriminative ability and calibration. Ablation studies confirmed the effectiveness of the short-path design and feature fusion strategy. The model enables interpretable backtracking of predictions through short evidence chains (e.g., drug-protein-disease-ADR). Literature verification confirmed that high-confidence predictions were consistent with established pharmacological mechanisms and clinical reports. Conclusions The "data harmonization and short-path aggregation" paradigm proposed in this study effectively mitigates semantic drift in long-path analysis. It balances prediction performance with interpretability and transferability, offering a practical technical solution for intelligent early warning and pre-medication risk assessment of anticancer drug adverse reactions.