Reliable Machine Learning for DPPH Antioxidant QSAR: Scaffold- and Study-Aware Validation with an Out-of-Domain Sulfonamide-Tyr-Gly Case Study
- 1Department of Industrial Chemistry, Faculty of Applied Physical Sciences
- 2Enugu State University of Science and Technology (ESUT
Abstract
Machine-learning QSAR models can prioritize antioxidant compounds, but apparent accuracy depends strongly on how chemical data are partitioned for validation. This study develops a reproducible DPPH antioxidant QSAR workflow using 1,891 curated molecules from a public 30 min DPPH dataset. Molecular structures were represented by nine RDKit physicochemical descriptors and 2,048-bit Morgan fingerprints (radius 2), giving 2,057 combined features.Random Forest regression was evaluated as the primary model,with Extra Trees as an independent ensemble comparator.Random Forest achieved mean R-squared = 0.712 and MAE = 0.297 pIC50 units under random five-fold cross-validation, but performance decreased to R-squared = 0.447 and MAE = 0.433 under Bemis-Murcko scaffold-aware validation, and to R-squared = 0.124 and MAE = 0.545 under DOI-grouped validation. In an internal held-out set, molecules with scaffolds represented during training performed better than those with unseen scaffolds(R-squared=0.824 versus 0.580).Nearest-neighbour Morgan Tanimoto analysis further separated structurally supported predictions from extrapolation: inside-domain molecules achieved R-squared = 0.830, whereas outside-domain performance was R- squared =-0.031. The workflow was then applied to two experimentally investigated sulfonamide-Tyr-Gly derivatives. Both had low maximum Tanimoto similarity(0.364 and 0.333), unseen scaffolds,and were overpredicted by approximately one pIC50 unit.The results show that random-split performance alone is insufficient for judging DPPH-QSAR reliability and that chemically informed validation and applicability-domain analysis are necessary when predictions are extended to structurally novel compounds.
Related
- SimSJSAlert: A Similarity-Augmented Multi-View Learning Framework with Scaffold Alerts for Drug-Induced Stevens-Johnson Syndrome Risk Assessment
- Reliable Structure-Processing-Performance Learning in Organic Photovoltaics: Chemical Generalization, Processing-Aware Prediction and Uncertainty under Domain Shift
- 1,3,4-Oxadiazole-sulfonamide hybrids: recent advances in biological activities, structure-activity relationships, and therapeutic applications.
- Photoisomerisable zero-field Co(II) single-ion magnets with azo-functionalized sulfonamide ligands: synthesis, crystal structure, magnetism and light-induced switching
- 4-Thiazolidinone/2,4-Thiazolidinedione-Sulfonamide Hybrids: A Concise Review of Their Design and Pharmacological Significance.
BibTeX
@misc{orji2026reliable,
title = {{Reliable Machine Learning for DPPH Antioxidant QSAR: Scaffold- and Study-Aware Validation with an Out-of-Domain Sulfonamide-Tyr-Gly Case Study}},
author = {Dennis Obinna Orji and Ike Ozoemena Christian},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008549/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008549/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Reliable Machine Learning for DPPH Antioxidant QSAR: Scaffold- and Study-Aware Validation with an Out-of-Domain Sulfonamide-Tyr-Gly Case Study
AU - Dennis Obinna Orji
AU - Ike Ozoemena Christian
PY - 2026
DA - 2026/09/09
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008549/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008549/v1
AB - Machine-learning QSAR models can prioritize antioxidant compounds, but apparent accuracy depends strongly on how chemical data are partitioned for validation. This study develops a reproducible DPPH antioxidant QSAR workflow using 1,891 curated molecules from a public 30 min DPPH dataset. Molecular structures were represented by nine RDKit physicochemical descriptors and 2,048-bit Morgan fingerprints (radius 2), giving 2,057 combined features.Random Forest regression was evaluated as the primary model,with Extra Trees as an independent ensemble comparator.Random Forest achieved mean R-squared = 0.712 and MAE = 0.297 pIC50 units under random five-fold cross-validation, but performance decreased to R-squared = 0.447 and MAE = 0.433 under Bemis-Murcko scaffold-aware validation, and to R-squared = 0.124 and MAE = 0.545 under DOI-grouped validation. In an internal held-out set, molecules with scaffolds represented during training performed better than those with unseen scaffolds(R-squared=0.824 versus 0.580).Nearest-neighbour Morgan Tanimoto analysis further separated structurally supported predictions from extrapolation: inside-domain molecules achieved R-squared = 0.830, whereas outside-domain performance was R- squared =-0.031. The workflow was then applied to two experimentally investigated sulfonamide-Tyr-Gly derivatives. Both had low maximum Tanimoto similarity(0.364 and 0.333), unseen scaffolds,and were overpredicted by approximately one pIC50 unit.The results show that random-split performance alone is insufficient for judging DPPH-QSAR reliability and that chemically informed validation and applicability-domain analysis are necessary when predictions are extended to structurally novel compounds.
ER -