Origin discrimination and screening of potential geographical origin markers for American ginseng based on metabolomics and machine learning
Аннотация
American ginseng is highly valued worldwide as both a traditional medicine and a premium ingredient in functional foods and dietary supplements. While the United States and Canada are recognized as its native and primary geographical production areas, its extensive introduction and cultivation in China have complicated the differentiation of quality based on origin in the global market. This study proposes a traceability strategy to decipher metabolomic disparities across different countries. Utilizing UHPLC-MS/MS-based non-targeted metabolomics, we characterized the ginsenoside profiles of ginseng from three distinct geographical origins and annotated 37 key compounds via the GNPS platform. Six machine learning classifiers—L1- and L2-regularized logistic regression, multilayer perceptron, random forest, gradient boosting decision tree, and support vector machine—were systematically compared under stratified cross-validation. The L2-regularized logistic regression (LR-L2) model delivered the best balanced performance, with a cross-validated accuracy of 88.6% and a macro-averaged F1 score of 88.0%; on the independent test set, it correctly classified 14 of 15 samples (93.3%), achieving a macro-averaged F1 score of 0.933, and was therefore selected to identify origin-discriminating saponins. To enhance model transparency, Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) were applied to interpret the LR-L2 model, and the resulting markers were further corroborated by univariate statistical analysis of the LC-MS peak areas. This work not only enriches the quality assessment framework for American ginseng but also provides a robust theoretical foundation for food authenticity screening, cross-border identification, and the targeted development of ginseng-based functional foods.