Fruit wines classification enabled by combing machine learning with comprehensive volatiles profiles of GC-TOF/MS and GC-IMS

Abstract

Fruit wines, produced through the fermentation of various fruits, are well-documented for their distinct flavor profiles. Intelligent sensory analysis, GC-TOF/MS and GC-IMS were used for the analysis of the volatile profile of eight types of fruit wines including 5 grape wine (SJ, LS, HY, TJ, FT), 1 fermented plum wine (FZ), 1 blueberry wine (HZ), as well as 1 configured plum wine (LM). A total of 281 compounds were identified through GC-TOF/MS, with esters and acids constituting over 80% of all samples. GC-IMS identified 60 compounds, predominantly including 16 esters, 11 alcohols, and 6 ketones, and 7 sulfur-containing compounds. This observation leads to the assumption that the IMS and MS data contain different information about the composition of the volatile profile. 37 and 18 differential compounds for TOF/MS data and IMS data were obtained, respectively. Three ranking algorithms combined with five machine learning models Neural Networks (NN), Random Forests (RF), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Logistic Regression (LR) applied and identified both 58 key features from volatiles. LR and KNN achieved an overall classification of 0.95 and an F1 score greater than 0.9. For the IMS data, NN, LR, and KNN models exhibited accuracies and F1 scores greater than 0.9. This study advances fruit wine classification, benefiting the beverage industry and food chemistry research.