UC San Diego

Metabolomics, or the study of compounds essential for cellular function, is a field withincreasing application in the research of organisms and organic systems; one component of this is the automatic identification of metabolites from the spectral analysis of samples. This can be challenging, however, due to chemical shifts on account of laboratory conditions, as well as noise arising from the experiment itself. In this work, we develop an approach to this task using a convolutional neural network (CNN) on 1H-13C nuclear magnetic resonance (NMR) spectra. With very limited experimental data, we synthesized our own set of metabolic mixture data and trained the neural network to achieve good performance compared to other automatic metabolite identification methods predicting from NMR data.