Optimizing ChemBL-derived QSAR models for natural flavonoid screening: chemotype-specific predictive reliability in α-glucosidase inhibition
File(s)
Accepted version
Author(s)
Type
Journal Article
Abstract
α-Glucosidase inhibitors (AGIs) are therapeutic agents for postprandial glucose regulation, where flavonoids represent a major class of natural AGIs. Existing QSAR models often fail to generalize to natural compounds due to limited chemical space overlap. This study aimed to construct and evaluate Random Forest–based QSAR models trained on ChEMBL-derived α-glucosidase inhibitors and to determine their applicability natural flavonoids. All inhibitors from ChEMBL were curated into two datasets, one containing all scaffolds and another containing flavonoid-only. A custom RDKit-based classification pipeline automatically identified flavonoid chemotypes using SMARTS pattern recognition and fingerprint similarity, excluding nonphenolic compounds. Molecular descriptors were computed via Mordred, and regression and classification models were developed using 10-fold cross-validation. External validation employed 15 natural flavonoids from literature, categorized into four structural groups: (1) flavones/flavonols (aglycones), (2) flavonol glycosides (1–2 sugars), (3) isoflavones, and (4) flavan-3-ols. The general model trained on 563 diverse compounds achieved high regression (R2 = 0.837), while the flavonoid-specific model showed moderate regression performance (R2 = 0.564). The classification performance was higher for the general model than the flavonoid-specific model (accuracies of 0.915 and 0.880, respectively). When validated with external dataset, the flavonoid-specific model showed better predictive performance than the general model. The flavonoid-specific model is the most suitable for predicting pIC50 of catechins and aglycones (MAE = 0.182 and 0.233, respectively). The optimized QSAR model reliably predicts the inhibitory potential of natural flavones and flavonols (aglycones) but should be used cautiously for glycosides and isoflavones. The trained model is available online on https://github.com/bryangervais/QSAR-Predictor.
Date Acceptance
2026-01-03
Citation
Engineering, Technology & Applied Science Research
ISSN
2241-4487
Publisher
Engineering, Technology and Applied Science Research (ETASR)
Journal / Book Title
Engineering, Technology & Applied Science Research
Copyright Statement
Subject to copyright. This paper is embargoed until publication.
Publication Status
Accepted