Xu Shengjie ; Xie Lingxi

Abstract

Antibody-drug conjugates (ADCs) have emerged as a promising class of targetedcancer therapeutics, but the design and optimization of their cytotoxicpayloads remain challenging. This study introduces DumplingGNN, a novel hybridGraph Neural Network architecture specifically designed for predicting ADCpayload activity based on chemical structure. By integrating Message PassingNeural Networks (MPNN), Graph Attention Networks (GAT), and GraphSAGE layers,DumplingGNN effectively captures multi-scale molecular features and leveragesboth 2D topological and 3D structural information. We evaluate DumplingGNN on acomprehensive ADC payload dataset focusing on DNA Topoisomerase I inhibitors,as well as on multiple public benchmarks from MoleculeNet. DumplingGNN achievesstate-of-the-art performance across several datasets, including BBBP (96.4\%ROC-AUC), ToxCast (78.2\% ROC-AUC), and PCBA (88.87\% ROC-AUC). On ourspecialized ADC payload dataset, it demonstrates exceptional accuracy(91.48\%), sensitivity (95.08\%), and specificity (97.54\%). Ablation studiesconfirm the synergistic effects of the hybrid architecture and the criticalrole of 3D structural information in enhancing predictive accuracy. The model'sstrong interpretability, enabled by attention mechanisms, provides valuableinsights into structure-activity relationships. DumplingGNN represents asignificant advancement in molecular property prediction, with particularpromise for accelerating the design and optimization of ADC payloads intargeted cancer therapy development.