Dumpling GNN: Hybrid GNN Enables Better ADC Payload Activity Prediction Based on the Chemical Structure

Abstract

Antibody-drug conjugates (ADCs) are promising cancer therapeutics, but optimizing their cytotoxic payloads remains challenging. We present DumplingGNN, a novel hybrid Graph Neural Network architecture for predicting ADC payload activity and toxicity. Integrating MPNN, GAT, and GraphSAGE layers, DumplingGNN captures multi-scale molecular features using both 2D and 3D structural information. Evaluated on a comprehensive ADC payload dataset and MoleculeNet benchmarks, DumplingGNN achieves state-of-the-art performance, including BBBP (96.4% ROC-AUC), ToxCast (78.2% ROC-AUC), and PCBA (88.87% ROC-AUC). On our specialized ADC payload dataset, it demonstrates 91.48% accuracy, 95.08% sensitivity, and 97.54% specificity. Ablation studies confirm the hybrid architecture's synergy and the importance of 3D information. The model's interpretability provides insights into structure-activity relationships. DumplingGNN's robust toxicity prediction capabilities make it valuable for early safety evaluation and biomedical regulation. As a research prototype, DumplingGNN is being considered for integration into Omni Medical, an AI-driven drug discovery platform currently under development, demonstrating its potential for future practical applications. This advancement promises to accelerate ADC payload design, particularly for Topoisomerase I inhibitor-based payloads, and improve early-stage drug safety assessment in targeted cancer therapy development.