Rapid geographical authentication of Sauce-flavor baijiu using SICRIT-Q-TOF-MS and interpretable machine learning.

Geographical authentication of Sauce-flavor baijiu remains challenging due to its complex composition. A non-targeted metabolomics approach using soft ionization by chemical reaction in transfer coupled with quadrupole time-of-flight mass spectrometry (SICRIT-Q-TOF-MS) enabled rapid analysis of 134 samples from 10 provinces. 510 ion features were detected, and 202 discriminative features were screened (VIP > 1; FDR-adjusted p < 0.001; FC > 4). Orthogonal partial least squares discriminant analysis (OPLS-DA) showed internal cross-validated performance (R2Y = 0.900, Q2 = 0.896), with a bootstrap 95% confidence interval of 0.883-0.908 for Q2. Fifteen machine-learning algorithms were systematically screened, with Random Forest showing the most balanced internal test-set performance (accuracy = 0.796, AUC = 0.971). A 59-feature model maintained similar performance (accuracy = 0.805, AUC = 0.969). Shapley Additive Explanations (SHAP) indicated classification was driven by interacting features with nonlinear effects. This study provides an interpretable strategy for geographical origin discrimination of Sauce-flavor baijiu.