Learning to Fragment Molecular Graphs for Mass Spectrometry Prediction

Abstract

Mass spectrometry is a common technique used to identify metabolites, small molecules that are yielding important new insights into a number of important biological and physiological processes, including organ function, nutrient sensing, and gut physiology. However, identifying these metabolites is a difficult task. Most methods for metabolite identification compare the spectrum of an unknown compound against a database containing referential spectra of known compounds. However, this approach fails if the true compound is not in the reference database, motivating approaches to generate a larger standard library by directly learning to simulate the forward fragmentation process. In this thesis, I propose a machine learning framework that learns the fragmentation pathways for each molecule to predict its mass spectrum, drawing upon both advances in machine learning and domain insights from analytical chemistry.