Scaffold-Aware Molecular Machine Learning for Drug Candidate Prioritization: A SARS-CoV-2 Main Protease Case Study

Abstract

Molecular machine learning can prioritize compounds before experimental follow-up, but its value depends on generalization to chemically distinct molecules rather than performance on randomly divided analogues. This study evaluated a scaffold-aware classification workflow using severe acute respiratory syndrome coronavirus 2 main protease (Mpro) as a case-study target. A harmonized multisource dataset assembled from ChEMBL-derived activities, the COVID Moonshot, and peer-reviewed biochemical screens contained 345 curated compounds accepted for modeling: 264 development compounds balanced between active and inactive/weak classes, and an 81-compound external benchmark with no Bemis-Murcko scaffolds shared with the development set. Molecules were represented by 21 RDKit physicochemical descriptors and a 1,024-bit Morgan fingerprint. A descriptor-only comparison of six algorithms showed that nonlinear models produced the three highest cross-validated F1-scores, led by Gradient Boosting (0.749). A prespecified Random Forest was subsequently tuned with grouped inner cross-validation using the combined 1,045-feature representation. On a development scaffold holdout, the model achieved accuracy 0.869, F1-score 0.852, and receiver operating characteristic area under the curve (ROC-AUC) 0.938. After refitting on all development compounds, performance on the external scaffold-disjoint benchmark was accuracy 0.852, F1-score 0.829, and ROC-AUC 0.925. Fingerprint bits and global size, surface-area, and complexity descriptors contributed to predictions. The results support scaffold-aware candidate prioritization while remaining insufficient to establish binding mechanisms or therapeutic activity.