Interpretable machine learning-augmented quantitative targeted flavoromics for quality grade prediction of Jiangxiangxing baijiu.

The quality of Jiangxiangxing (JXX) baijiu depends on its sensory and flavor characteristics, which traditional methods struggle to evaluate accurately. This study employed quantitative targeted flavoromics and interpretable machine learning (ML) to analyze 578 baijiu samples across different quality grades. A total of 106 flavor compounds were quantified using GC-FID, LLME-GC-MS, and UPLC-HRMS analyses. Among these, 50 compounds were identified as odor-active (OAV >1) and 50 were screened as potential differential markers (VIP >1) through PLS-DA analyses. By integrating the results of analysis of variance (ANOVA), 23 flavor quality markers were further confirmed to have robust features for effectively distinguishing quality grades. Subsequently, eight ML algorithms were applied to construct quality prediction models. Of which, XGBoost displayed optimal performance in the discovery dataset (DISC), as determined via a comprehensive evaluation using eight metrics (all >0.93), alongside calibration and decision curve analyses. The model's robustness was further validated across three external independent validation datasets (VALI 1-VALI 3), achieving macro-average AUC values of 0.89, 0.92, and 0.91, respectively. SHapley Additive exPlanations (SHAP) analysis revealed the contributions of each marker, revealing that furfural (roasted aroma), ethyl decanoate (fruity aroma), ethyl 2-methylbutanoate (fruity aroma), and isobutyric acid (acidic aroma) demonstrated the greatest impact on predictions and thereby played key roles in discriminating the quality grades of baijiu. Higher concentrations of these compounds were associated with enhanced positive flavor attributes in JXX baijiu. Collectively, this study provides a reliable, data-driven approach for Baijiu quality evaluation and control, offering substantive support for liquor blending and process optimization.