High-throughput targeted quantification and interpretable machine learning for differentiating Baijiu aroma types.

Abstract

The quantitative analysis of volatile flavor compounds in Baijiu presents challenges owing to the extensive variety and differing concentrations of these compounds. This research developed a quantitative method for 249 compounds using gas chromatography flame ionization detection (GC-FID) and gas chromatography tandem mass spectrometry (GC-MS/MS). Chemometrics with an interpretable machine learning model were employed to analyze the flavor profiles of Baijiu. The results indicated that the strategy demonstrated excellent linearity (R² > 0.99), precision (RSD < 7.18%), accuracy (recovery of 82.18%-111.03%), and an extensive analytical working range (0.0025-4500 mg/L). The variable importance in projection (VIP), p-value, and fold changes were integrated to identify the differential volatile compounds among Baijiu aroma types. Following that, machine learning models were introduced, with Random Forest (RF) demonstrating optimal classification performance, achieving 100% accuracy in the test set. The contributions of individual volatile compounds to the classification of Baijiu aroma types by the RF model were quantified using Shapley Additive Explanations (SHAP) analysis.