Machine learning applications for identify the geographical origin, variety and processing of black tea using 1H NMR chemical fingerprinting

The geographical origin of black tea can affect commercial value and is highly susceptible to food fraud. In this study, nuclear magnetic resonance (NMR) spectroscopy was used for untargeted metabolomics analysis of 219 black tea samples from seven major black tea producing regions in China (Anhui, Yunnan, Fujian, and Guangdong), India (Darjeeling and Assam) and Sri Lanka (Kandy). Black tea from different geographical origins can be distinguished according to the variety and processing, among which caffeine and alanine were identified as the main differential metabolites of the variety, theaflavin 3, 3′-digallate and succinic acid were identified as the main differential metabolites of the processing. Several machine learning algorithms were used to identify the origin of black tea, and the test set accuracy results showed that the nonlinear model random forest (92.7%) and support vector machine (91.8%) algorithms were better than the linear model linear discriminant analysis (86.3%) and K-nearest neighbor (86.3%). The random forest model screened 14 black tea geographical origin marker metabolites, such as caffeine, malic acid, lysine and β-glucose, and based on these marker metabolites, the chemical fingerprint pattern of origin was drawn. Black tea origin marker metabolites proved that variety contributed more to the origin metabolite fingerprint than processing. The results support that 1H NMR metabolomics combined with machine learning can be used as an effective tool for the construction of black tea chemical fingerprints for quality assessment and fraud detection.