Spec2Class: Accurate Prediction of Plant Secondary Metabolite Class using Deep Learning

Mass spectrometry (MS)-based data is commonly used in studying metabolism and natural products, but typically requires domain-specific skill and experience to analyze. Existing computational tools for non-targeted metabolite analysis (i.e., metabolomics) mostly rely on comparison to reference MS spectral libraries for metabolite identification, limiting the annotation of metabolites for which reference spectra do not exist. This is the case in plant secondary metabolites, where most spectral features remain unidentified. Here, we developed Spec2Class , a deep-learning algorithm for the identification and classification of plant secondary metabolites from liquid chromatography (LC)-MS/MS spectra. We used the in-house spectral library of 7973 plant metabolite chemical standards, alongside publicly available data, to train Spec2Class to classify LC-MS/MS spectra to 43 common plant secondary metabolite classes. Tested on held out sets, our algorithm achieved an overall accuracy of 73%, outperforming state-of-the-art classification. We further established a prediction certainty parameter to set a threshold for low-confidence results. Applying this threshold, we reached an accuracy of 93% on an unseen dataset. We show a high robustness of our prediction to noise and to the data acquisition method. Spec2Class is publicly available and is anticipated to facilitate metabolite identification and accelerate natural product discovery.