Determining the chemical composition of unknown compounds plays a central role in chemical analysis. Mass spectral reference libraries, as the collections of chemical structures and their spectra, provide a means of identification for compounds. However, existing libraries suffer from a coverage problem, which can be partially solved by in-silico spectra augmentation. Alternative to reference library matching, we propose a direct spectra-to-molecule translation approach for structure prediction for electron-ionization mass spectra. This is the first method using neural machine translation to identify an unknown molecule from its mass spectrum. We evaluated our method with NIST main library, and the results indicate that our novel framework presents an effective compound identification by using Transformer architecture. Our method can work cooperatively with database spectral matching to improve identification accuracy. This proposed approach has a high potential for broad applications in the chemical analysis of unknown substances.

Compound identification in electron-ionization mass spectra via neural machine translation

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.