About DeepMet

DeepMet is a chemical language model of the human metabolome. By representing chemical structures as short strings of text, DeepMet applies the same neural network architectures that are used to model the words in an English sentence to the atoms and bonds in a chemical structure. This paradigm allows DeepMet to learn the biosynthetic logic that is embedded within the structures of known human metabolites, and leverage this understanding to predict the existence of as-of-yet undiscovered metabolites.

The DeepMet web application offers multiple mechanisms to browse the structures of predicted metabolites and integrate DeepMet with mass spectrometry-based metabolomics data for metabolite discovery:

1. The Browse compounds page allows users to browse all 2.5 million chemical structures generated by DeepMet. The compounds are sorted in descending order by the frequency with which they were generated by the language model, which can be used to prioritize predicted metabolites for experimental discovery. Compounds can also be sorted by exact mass. The browsing interface additionally provides the chemical formula for each predicted metabolite and indicates whether it is a known metabolite found in the Human Metabolome Database (reproduced by the language model in cross-validation) and its annotation status in that database (detected, quantified, predicted, or expected).
2. The Mass search and Formula search pages allow users to filter the generated structures by exact mass or molecular formula, respectively, and thereby propose both known and unknown structures for an unidentified metabolite detected by mass spectrometry (at the MS1 level).
3. The MS/MS search page allows users to upload MS/MS spectra of unidentified metabolites, which are then searched against the predicted MS/MS spectra for all 2.5 million chemical structures generated by DeepMet, to propose both known and unknown structures for an unidentified metabolite detected by mass spectrometry (at the MS2 level).

The underlying data that powers the DeepMet web application, including all chemical structures generated by the language model and predicted MS/MS spectra for each generated structure, is available for download via Zenodo at http://doi.org/10.5281/zenodo.12999324.

We would love to hear your feedback, suggestions, or reports on errors or inconsistencies in the DeepMet web application. Please use the Wishart lab’s feedback form and we will direct your message to the appropriate expert. However, keep in mind we get a large volume of feedback, and although we read every message we cannot guarantee a turnaround time.

Questions / Feedback Form