Machine learning prediction for constructing a universal multidimensional information library of Panax saponins (ginsenosides)

Abstract

Accurate characterization of Panax herb ginsenosides is challenging because of the isomers and lack of sufficient reference compounds. More structural information could help differentiate ginsenosides and their isomers, enabling more accurate identification. Based on the Vion^TM ion-mobility high-resolution LC-MS platform, a multidimensional information library for ginsenosides, namely GinMIL, was established by predicting retention time (t_R) and collision cross section (CCS) through machine learning. Robustness validation experiments proved t_R and CCS were suitable for database construction. Among three machine learning models we attempted, gradient boosting machine (GBM) exhibited the best prediction performance. GinMIL included the multidimensional information (m/z, molecular formula, t_R, CCS, and some MS/MS fragments) for 579 known ginsenosides. Accuracy in identifying ginsenosides from diverse ginseng products was greatly improved by a unique LC-MS approach and searching GinMIL, demonstrating a universal Panax saponins library constructed based on hierarchical design. GinMIL could improve the accuracy of isomers identification by approximately 88%.

Keywords: Collision cross section; Hierarchical design; Machine learning; Multidimensional information library; Retention time.