MS2Prop: A machine learning model that directly predicts chemical properties from mass spectrometry data for novel compounds

Mass spectrometry is a key analytical tool for the study of complex small molecule mixtures. The contents of these mixtures are the subject of untargeted metabolomics applications ranging from understanding metabolism, disease, biomarkers, and environmental contaminants to natural products based drug discovery. Yet identifying from mixtures the compounds or their properties from mass spectrometry data remains very challenging for most small molecules. For most compounds there will not be an annotation, and nearly all annotation techniques rely on partial matches to spectral and structural databases with limited coverage. However, property prediction of unknowns in untargeted metabolomics relies heavily on those annotations. Here we introduce MS2Prop, a complement to compound identification, that directly predicts chemically relevant properties of compounds for drug discovery and other applications from mass spectrometry data for any mass spectrometry feature, regardless of whether the corresponding compound is in an existing database. On compounds excluded from the training set MS2Prop has an average R^2 = 0.73 across ten properties, including synthetic accessibility and quantitative drug likeness properties, and R^2 = 0.96 for compounds in the training set, but with disjoint spectra. For compounds excluded from the training set, MS2Prop outperforms predictions based on compound identification by over a factor of three, setting the stage for future use of computational prioritization of compounds for diagnostic and drug discovery applications. ### Competing Interest Statement PCD is a Scientific Advisor to Cybele and a Scientific Co-founder and advisor to Ometa and Enveda with prior approval by UC-San Diego. All other authors are employees of Enveda.