Machine learning coupled with quantum chemical and sensory validation: Mechanistic insight into flavor modulation of...

Cornelian cherry (Cornus officinalis) is a traditional Chinese medicinal herb that was first recorded in Shennong Ben Cao Jing approximately 1800 years ago. It has long been highly valued for its tonifying and astringent properties (such as nourishing the liver and kidneys and helping to retain semen and reduce urinary leakage). Although it possesses both medicinal and edible value, its flavor-which is characterized by strong astringency and slight bitterness-limits its use in food. Lactic acid bacteria can convert inorganic selenium into organic selenium nanoparticles (SeNPs), which not only improve the unpleasant taste but also increase the bioavailability of the resulting SeNPs. This study combines response surface methodology (RSM) and machine learning (ML) with quantum chemical calculations and sensory validation for predictive quality control, with the aim of elucidating the molecular basis of flavor improvement. The fermentation conditions optimized via RSM were 33 °C, a mass ratio of 1:3 between cornelian cherry and yacon, and an enzyme dosage of 0.6 g/L. HS-GC-IMS analysis revealed that esters with fruity odor characteristics continued to accumulate during fermentation, whereas aldehyde levels tended to decrease. Nontargeted metabolomics further revealed that the phenylpropanoid metabolic pathway and tannin-related metabolites were downregulated overall, providing a metabolic explanation for the reduction in astringency. Density functional theory (DFT) calculations indicated that esters with lower ΔE bind less strongly to the bitter taste receptor TAS2R14 than aldehydes with higher ΔE. Concentration-response, equimolar comparison, and spike-back experiments confirmed the sensory contribution of ester accumulation to bitterness reduction, while molecular docking suggested potential involvement of TAS2R14. Among the ML models evaluated, the TabPFN - a transformer-based model requiring no hyperparameter tuning achieved the highest predictive accuracy for the comprehensive quality score (R² = 0.928), outperforming XGBoost (R² = 0.912) and RSM (R² = 0.876). Offline multi-objective simulation suggested that batch-to-batch variability in DPPH scavenging could theoretically be reduced by 41.3% while maintaining a SeNP conversion of 89.4%. By dissecting the distinct mechanisms of astringency reduction (tannin degradation) and bitterness alleviation (low-ΔE ester/TAS2R14 modulation) and providing a deployable ML model, this work offers a transferable paradigm for the rational design of flavor-optimized functional fermented foods from medicinal and edible resources.