Abstract: Adverse drug reactions (ADRs) to anticancer drugs severely impact patients’ quality of life and treatment efficacy, while existing studies lack both predictive accuracy and mechanistic interpretation. This study aims to develop a novel computational model to predict drug-ADR associations and uncover their molecular mechanisms. A dual-channel graph autoencoder (CombinedGAE) is proposed integrating graph convolutional networks (GCN) and GraphSAGE, trained on a dataset of 85 anticancer drugs and 1,827 ADRs. Protein-protein interaction (PPI) networks are further incorporated to identify key ADR-related proteins via network diffusion analysis. CombinedGAE achieves superior prediction performance (AUC = 0.92±0.03) compared with baseline methods. Network analysis reveals core proteins (e.g., SH2B2, NRP2) and pathways implicated in drug toxicity. This study provides a high-accuracy tool for drug safety evaluation and mechanistic insights for clinical decision-making. Future work may extend to polypharmacy scenarios to advance personalized therapy.