A progressive screening strategy by combining UHPLCQTOF-MS and machine learning for distinguishing different varieties of Citri Reticulatae Pericarpium.

This study employed UHPLC-QTOF-MS-based untargeted metabolomics, integrated with chemometrics and machine learning, to discriminate between varieties of Citri Reticulatae Pericarpium (CRP). To identify specific markers between each pair of varieties more precisely, a progressive screening strategy was adopted. Initially, principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA) screened 46 universal differential metabolites across the four varieties. Subsequently, pairwise orthogonal partial least squares discriminant analysis (OPLS-DA) applied to these universal metabolites facilitated the detection of specific differential metabolites. LC-MS analysis identified a total of 268 metabolites. PCA revealed significant metabolic differences among the four CRP varieties, while PLS-DA screened 46 universal differential metabolites (VIP > 1). 9 machine learning discrimination models constructed from these 46 metabolites exhibited excellent performance (with accuracy, etc. > 0.9461) and successfully discriminated externally validated CRP samples from different varieties with a 100% recognition rate. Further, pairwise OPLS-DA based on the universal differential metabolites screened out 10 universal specific differential metabolites (VIP > 1), including 5-O-demethylnobiletin. Notably, nobiletin distinguished pharmacopoeia from non-pharmacopoeia CRP varieties. Finally, integrative analyses linked these findings to their biological significance, attributing differential flavonoid accumulation among CRP varieties to molecular evolutionary mechanisms driven by genetic background-dependent methylation and glycosylation processes. This study demonstrates the considerable potential of UHPLCQTOF-MS-based untargeted metabolomics for distinguishing CRP varieties, providing a useful reference for market quality control and clinical practice.