Comprehensive and Explainable Fragmentation: A Machine Learning Approach for Fast and Accurate Mass Spectrum Prediction.

Abstract

Mass spectrometry (MS) is a fundamental tool for chemical identification. The current in-silico prediction tools can handle broad instrument conditions, large molecular libraries or fragment structures only on a very limited level. In this work, we propose a dual-model machine learning strategy that can solve this problem by jointly a classification model for fragment identification and noise filtering, and a regression model for spectral prediction. With the help of attention mechanism, our method outperforms other algorithms in accuracy and efficiency, providing a deeper understanding of the molecular fragmentation behavior in mass spectra. Our method can facilitate the large-scale in-silico spectra calculations and the analysis of unknown molecular structures, which may promote wider applications for MS.