Combination of in silico prediction and convolutional neural network framework for targeted screening of metabolites from LC-HRMS fingerprints: A case study of “Pericarpium Citri Reticulatae - FructusAurantii”

Abstract

In this study, a novel approach is introduced, merging in silico prediction with a Convolutional Neural Network (CNN) framework for the targeted screening of in vivo metabolites in Liquid Chromatography-High Resolution Mass Spectrometry (LC-HRMS) fingerprints. Initially, three predictive tools, supplemented by literature, identify potential metabolites for target prototypes derived from Traditional Chinese Medicines (TCMs) or functional foods. Subsequently, a CNN is developed to minimize false positives from CWT-based peak detection. The Extracted Ion Chromatogram (EIC) peaks are then annotated using MS-FINDER across three levels of confidence. This methodology focuses on analyzing the metabolic fingerprints of rats administered with "Pericarpium Citri Reticulatae - Fructus Aurantii" (PCR-FA). Consequently, 384 peaks in positive mode and 282 in negative mode were identified as true peaks of probable metabolites. By contrasting these with “blank serum” data, EIC peaks of adequate intensity were chosen for MS/MS fragment analysis. Ultimately, 14 prototypes (including flavonoids and lactones) and 40 metabolites were precisely linked to their corresponding EIC peaks, thereby providing deeper insight into the pharmacological mechanism. This innovative strategy markedly enhances the chemical coverage in the targeted screening of LC-HRMS metabolic fingerprints.