Benchmarking Experiments with External Data, Domain-Specific Features, and Learning Strategies for Modeling ADME and Potency of Coronavirus Main Protease Inhibitors

For ADME modeling, ensemble gradient boosting models trained on our bespoke feature space consistently outperformed both broader tree-based approaches and GNN variants across the five ADME end points in blind evaluation. By a factorial design of experiments and statistical benchmarking, we evaluated the performance of the optimal XGBoost model for each end point. For modeling potency, we explored varied pretraining paradigms and ensemble fusion strategies, and found that a stacking ensemble combining docking-enhanced XGBoost with a stereochemistry-aware AttentiveFP Graph-Transformer model yielded the strongest significant generalization on the Challenge test set, suggesting that orthogonal model families reduced epistemic uncertainty attributable to model selection in data-limited, multitarget settings.

Beyond benchmarking, our experiments provided a testbed for evaluating critical assumptions in the field, encompassing data augmentation strategies, splitting protocols, and ensemble integration approaches. Our analysis revealed that a consistently optimal model is elusive, and that several standard industry practices fail to withstand close inspection. Overall, our work provides both practical benchmarked modeling strategies for ADME and potency prediction and a reproducible computational foundation for AI-driven antiviral drug discovery efforts.