Screening and prediction of quality markers of different origins of Lanxangia tsao-ko based on machine learning models

Lanxangia tsao-ko (LT) from various origins vary greatly in quality due to environmental factors, but there are no representative quality markers (QMs) for LT quality differentiation and evaluation. Five QMs were screened from five geographical sources, including (-)-Epicatechin, Astragalin, Isorhamnetin 3-O-glucoside, Narcissin, and Quercetin, were firstly identified by the data-generating model and the classical machine-learning model. After that, partial least squares established based on fourier transform near-infrared spectroscopy (FT-NIR) and Fourier transform infrared spectroscopy (ATR-FTIR) regression (PLSR) to predict the content of different QMs. The results show that the weighted correlation network (WCN) can match the “exclusive” bands of QMs and significantly improve the performance of the model compared with the traditional feature extraction method. • GAN efficiently amplifies metabolomic data. • Six machine learning models work together to identify five QMs. • ResNet model achieves 100% recognition of different LT origins. • WCN significantly improves PLSR prediction.