Supervised topic modeling for predicting molecular substructure from mass spectrometry [version 1; peer review: 2 approved with reservations]

• Published in: F1000 research Vol. 10; p. 403
• Main Authors: Reder, Gabriel K, Young, Adamo, Altosaar, Jaan, Rajniak, Jakub, Elhadad, Noémie, Fischbach, Michael, Holmes, Susan
• Format: Journal Article
• Language: English
• Published: 2021
• Subjects: Metabolomics; Tandem Mass Spectrometry; Structure Identification; LC-MS; Mass Spectrometry; MS/MS; Machine Learning
• ISSN: 2046-1402; 2046-1402
• DOI: 10.12688/f1000research.52549.1

Small-molecule metabolites are principal actors in myriad phenomena across biochemistry and serve as an important source of biomarkers and drug candidates. Given a sample of unknown composition, identifying the metabolites present is difficult given the large number of small molecules both known and yet to be discovered. Even for biofluids such as human blood, building reliable ways of identifying biomarkers is challenging. A workhorse method for characterizing individual molecules in such untargeted metabolomics studies is tandem mass spectrometry (MS/MS). MS/MS spectra provide rich information about chemical composition. However, structural characterization from spectra corresponding to unknown molecules remains a bottleneck in metabolomics. Current methods often rely on matching to pre-existing databases in one form or another.  Here we develop a preprocessing scheme and supervised topic modeling approach to identify modular groups of spectrum fragments and neutral losses corresponding to chemical substructures using labeled latent Dirichlet allocation (LLDA) to map spectrum features to known chemical structures. These structures appear in new unknown spectra and can be predicted. We find that LLDA is an interpretable and reliable method for structure prediction from MS/MS spectra. Specifically, the LLDA approach has the following advantages: (a) molecular topics are interpretable; (b) A practitioner can select any set of chemical structure labels relevant to their problem; (c ) LLDA performs well and can exceed the performance of other methods in predicting substructures in novel contexts.