MS2Lipid: a lipid subclass prediction program using machine learning and curated tandem mass spectral data

Abstract

Untargeted lipidomics using collision-induced dissociation-based tandem mass spectrometry (CID-MS/MS) is essential for biological and clinical applications. However, annotation confidence is still guaranteed by manual curation by analytical chemists, although various software tools have been developed for automatic spectral processing based on rule-based fragment annotations. In this study, we provide a novel machine learning model, MS2Lipid, for the prediction of lipid subclasses from MS/MS queries to provide an orthogonal decision of lipidomics software programs to determine the lipid subclass of ion features, in which a new descriptor, MCH (mode of carbon and hydrogen), was designed to increase the specificity of lipid subclasses in nominal mass resolution MS data. The model trained with 5,224 and 5,408 manually curated MS/MS spectra for the positive- and negative-ion modes mapped the query into one or several categories of 97 lipid subclasses, with an accuracy of 95.5% queries in the test set. Our program outperformed the CANOPUS ontology prediction program, providing correct annotations for 38.7% of the same test set. The program was further validated using various datasets from different machines and curators, and the average accuracy exceeded 87.4 %. Furthermore, the function of MS2Lipid was showcased by the annotation of novel esterified bile acids, whose abundance was significantly increased in obese patients in a human cohort study, suggesting that the machine learning model provides an independent criterion for lipid subclass classification, in addition to an environment for annotating lipid metabolites that have been previously unknown.