Decoding the quantitative structure-activity relationship and astringency formation mechanism of oxygenated aromatic compounds

Astringency is a common sensory experience in the mouth, characterized by dryness, roughness, and puckering. Due to the inefficiency and expense of conventional astringency evaluation methods, the quantitative structureactivity relationship (QSAR) modeling correlates molecular structure with sensory feature, offering a scalable computational alternative. First, 54 oxygenated aromatic compounds were comprehensively collected, followed by molecular fingerprint similarity (MFS)-based hierarchical clustering for structural pattern classification. Subsequently, six machine learning regression models were constructed for predicting the astringency thresholds of the compounds, and the results indicated that the AdaBoost model performed the best, with an R2 of 0.778 and MSE of 0.058. Furthermore, the Shapley Additive exPlanations (SHAP) method was applied to interpret this model, revealing that BCUT2D_LOGPLOW and VSA_Estate1 were the most critical descriptors governing astringency thresholds. Two natural astringent oxygenated aromatic compounds were successfully identified through molecular fingerprint recognition and the Maximum Common Substructure (MCS) algorithm, and their astringency thresholds were predicted by the established model. The feasibility of the model was further validated through sensory experiments, where the predicted astringency thresholds closely matched the human astringency thresholds. The interaction mechanisms of astringent compounds were systematically investigated through turbidity measurements, zeta potential analysis, and molecular dynamics (MD) simulations. Results demonstrated that the protein-ligand complex aggregation was predominantly driven by hydrogen bonding and hydrophobic interactions. Therefore, the integration of QSAR and MD enables feature predictive frameworks to advance astringency-focused food development.