Predicted MS/MS spectra for DeepMet-generated structures in cross-validation

This record contains predicted MS/MS spectra for hypothetical metabolites sampled 5 or more times by DeepMet, a chemical language model trained on the known metabolome. Spectra were predicted by FraGNNet (Young*, Wang* et al., arXiv 2024) in both positive and negative ionization modes, at collision energies of 10, 20 and 40 eV, and then merged to produce a single spectrum per compound in each ionization mode. FraGNNet was trained in five-fold cross-validation, and these models were used to predict MS/MS spectra for known metabolites in their respective test sets, such that there was no overlap between the compounds in the training and test sets of each model (i.e., no MS/MS spectrum was predicted by a model that had seen any spectrum for that compound in its training dataset). A sixth model was trained on the entire training dataset and used to predict MS/MS spectra for DeepMet-generated metabolites whose reference spectra did not appear in the training dataset. For more information, see:Qiang, H. et al. Language model-guided anticipation and discovery of mammalian metabolites. Nature 651, 211–220 (2026). doi: 10.1038/s41586-025-09969-xYoung, A. et al. FraGNNet: a deep probabilistic model for tandem mass spectrum prediction. Trans. Machine Learning. Res (2025). https://openreview.net/pdf?id=UsqeHx9Mbx