Multiscale Computational Investigation of Antioxidant Mechanisms in Natural Products: DFT-Based Reaction Pathways, Solvent Effects and Explainable Machine Learning

Abstract

Oxidative stress arising from an imbalance between reactive oxygen/nitrogen species and endogenous defences underlies the pathophysiology of numerous chronic diseases, motivating sustained interest in plant-derived antioxidants. This work presents an integrated, multiscale computational framework that couple’s density functional theory (DFT) with explainable machine learning (ML) to rationalise and predict the radical-scavenging behaviour of natural phenolic compounds. Density functional calculations at the M06-2X/6-311++G(d,p) level, combined with the SMD implicit solvation model, were used to evaluate the thermodynamics (bond dissociation enthalpy, ionization potential, proton affinity, electron transfer enthalpy) and kinetics (transition-state geometries, activation barriers) of the hydrogen-atom transfer (HAT), single-electron-transfer-proton-transfer (SET-PT) and sequential-proton-loss-electron-transfer (SPLET) pathways for a curated set of natural antioxidants with experimentally reported DPPH/ABTS/FRAP activities. Conceptual-DFT descriptors (Fukui functions, electrophilicity, global hardness) and structural features (intramolecular hydrogen bonding, protonation state) were subsequently used, together with the quantum-derived descriptors, to train and benchmark random forest, gradient-boosted tree and neural-network regressors. SHAP-based explainability was applied to connect model predictions back to the underlying electronic and thermodynamic descriptors. Gradient-boosted trees achieved the best generalisation (test R² = 0.88, RMSE = 0.53), and Shapley attributions confirmed that BDE, PA and the dual-descriptor-based radical-attack index were the dominant predictors, in agreement with the DFT-derived mechanistic picture. The proposed multiscale pipeline reproduces the solvent-dependent switch from HAT-dominated behaviour in the gas phase to SPLET dominance in aqueous/protic media reported in the literature, while offering an interpretable, low-cost surrogate for large-scale virtual screening of natural antioxidants. The results demonstrate that combining first-principles reaction-pathway analysis with explainable ML narrows the gap between mechanistic understanding and predictive throughput in antioxidant discovery.