Leakage-controlled multi-assay antioxidant QSAR with class-conditional learning and evaluation of geometric–arithmetic connectivity indices

Reliable comparison of antioxidant potency is complicated by the strong dependence of measured activity on assay conditions. ABTS, DPPH, superoxide, lipid-peroxidation, and hydroxyl-radical tests capture different chemical environments and reaction mechanisms, making a single global model potentially misleading. We therefore evaluated a class-conditional quantitative structure-activity relationship framework for jointly predicting assay identity and antioxidant potency. The curated dataset comprised 1,059 compound-assay records from 898 standardized compounds and 330 Bemis-Murcko scaffolds. Models used 102 RDKit descriptors, 14 classical indices, and 14 geometric-arithmetic indices, with all preprocessing restricted to the training folds. Robustness was examined through repeated compound-grouped cross-validation, scaffold-grouped testing, descriptor-block ablation, architecture controls, response permutation, residual diagnostics, endpoint sensitivity, and applicability-domain analysis. Conditional XGBoost achieved a compound-grouped classification accuracy of \(0.718± 0.027\) and known-assay regression performance of \(R^2=0.718± 0.053\) . A Morgan ECFP4 random-forest comparator gave the best compound-grouped regression result ( \(R^2=0.739± 0.065\) ), but its performance declined under scaffold separation ( \(R^2=0.472± 0.303\) ). The connectivity-index blocks did not significantly improve the RDKit-only models after Holm adjustment. These findings support assay-aware modeling for related chemical structures, while showing that predictions for unfamiliar scaffolds require applicability-domain screening and prospective validation.