Morgan fingerprints outperform physicochemical descriptors for scaffold-aware machine learning classification of β3-adrenergic receptor agonists

Abstract The β3-adrenergic receptor (ADRB3) is a relevant but challenging target because public bioactivity data are heterogeneous and ligands overlap structurally with other β-adrenergic receptor subtypes. We developed a curated ligand-based machine-learning workflow for ADRB3 agonist prediction and benchmarked the effect of molecular representation under scaffold-aware validation. Bioactivity records from ChEMBL were standardized and converted into binary labels using a two-level consensus strategy. Confirmed ADRB1/ADRB2 agonists without ADRB3 agonist evidence were added as subtype-selectivity decoys to strengthen the control class, and near-duplicate conflicts were removed. The final dataset comprised 2147 curated compounds represented as ECFP4 fingerprints, ECFP6 fingerprints, RDKit physicochemical descriptors, and two hybrid feature sets. Six classifier variants were evaluated using grouped scaffold cross-validation and a held-out scaffold-grouped test set. Random Forest trained on ECFP6 fingerprints performed best, reaching ROC-AUC 0.976 and MCC 0.880 on out-of-fold predictions, and ROC-AUC 0.990 and MCC 0.941 on the held-out test set (ECFP4-based models performed closely comparably). Test sensitivity and specificity were 0.976 and 0.965, respectively. Y-randomization showed clear separation between real and permuted models, and SHAP. These results show that ADRB3 agonist activity can be predicted robustly from scaffold-aware ligand-based models and that ECFP4 fingerprints alone provide the most informative representation in this benchmark. Fingerprint-based models outperformed descriptor-only models, and adding descriptors did not provide a consistent benefit. Scientific contribution This study provides a reproducible, scaffold-aware ADRB3 agonist benchmark built from curated ChEMBL data and subtype-selectivity decoy controls. It directly compares fingerprint-only, descriptor-only, and hybrid feature spaces across multiple classifiers under the same validation framework. The results show that local structural environments encoded by ECFP4 fingerprints dominate predictive performance for ADRB3 agonist classification.