DeepMet
Summary¶
DeepMet is an LSTM chemical language model trained on the SMILES structures of 2,046 metabolites experimentally detected in human tissues or biofluids (pretrained first on ChEMBL drug-like structures) that learns to generate metabolite-like chemical structures. Its central innovation is using sampling frequency — how often a chemical structure recurs across a billion sampled SMILES strings, aggregating over the many valid SMILES that can represent the same structure — as a confidence signal for prioritizing candidate metabolites, since chemical structures (unlike protein or DNA sequences) lack a unique textual representation. Applying this to mass spectrometry data from human biofluids and 23 mouse tissues, the authors experimentally confirmed several dozen structurally diverse, previously unrecognized mammalian metabolites.[1]
Learning the Language of Metabolism¶
Evidence that DeepMet learned genuine metabolite-like chemistry, not generic drug-like chemistry: - Generated molecules overlapped extensively with known metabolites in UMAP-embedded chemical space. - A random forest classifier trained to separate generated molecules from metabolites withheld from training performed only marginally better than chance (AUC = 0.57). - DeepMet recapitulated 77.5% of one-step enzymatic transformations of known metabolites predicted by the rule-based tool BioTransformer, despite receiving no explicit reaction information during training, while also generating a much broader space of structures than BioTransformer's rule-based enumeration. - Generated molecules with a given molecular formula were more structurally similar to known metabolites than formula-matched molecules randomly sampled from PubChem or ChEMBL.
Anticipating Unrecognized Metabolites via Sampling Frequency¶
Because chemical structures have no unique SMILES representation, the authors sampled 1 billion SMILES strings from DeepMet and tabulated how often each unique underlying structure recurred. Structures generated more frequently were more structurally similar to known metabolites, more likely to correspond to plausible enzymatic transformations, more likely to share a scaffold with a known metabolite, and had more metabolite-like predicted biological activity (Fréchet ChemNet distance) — establishing sampling frequency as a metabolite-likeness signal.
- Retrospective validation: known metabolites withheld from training were reproduced disproportionately among the most frequently sampled structures (sampling frequency alone separated withheld metabolites from other generated molecules with AUC = 0.98).
- Prospective validation: of 313 metabolites added to HMDB version 5.0 after the training cutoff, DeepMet generated 252 (81%); most of the 61 not generated were exogenous (drug, food, microbiome, or environmental in origin) rather than endogenous. Sampling frequency again separated HMDB 5.0 metabolites from other generated molecules (AUC = 0.97), and the top 10,000 most frequently sampled structures were enriched ~1,500-fold for HMDB 5.0 metabolites over random expectation.
- Experimental discovery: of 80 synthesized/purchased chemical standards for top-10,000-ranked structures absent from HMDB, 17 were confirmed present in human biofluids by matching retention time and MS/MS against a large in-house urine/blood metabolomics bank (e.g., N-carbamyl-proline, N-succinyltryptophan, N-lactoyl-glutamine); literature review showed some were previously known metabolites simply missing from HMDB.
Prioritizing Structures from Mass Spectrometry Data¶
- From accurate mass alone: filtering DeepMet-generated structures to those matching a held-out metabolite's exact mass (±10 ppm) and ranking by sampling frequency correctly identified the metabolite as the top-ranked candidate in 29% of cases (e.g., serotonin was the single top-ranked structure among 2,818 unique candidates matching its mass), substantially outperforming the AddCarbon baseline of Renz et al. on top-k accuracy and structural similarity metrics.
- With MS/MS integration: CFM-ID was used to predict in silico MS/MS spectra for 2.4 million DeepMet-generated structures, creating a searchable spectral library from hypothetical metabolites. Combining DeepMet sampling frequency with MS/MS spectral matching correctly assigned the exact structure for 52% (positive mode) and 49% (negative mode) of held-out metabolites, with further improvements in top-3/top-10 accuracy and match quality; results held with an independent MS/MS dataset and alternative spectrum-prediction models.
- Confidence scores derived from sampling frequency correlated with the empirical probability that a structure assignment was correct, meaning DeepMet's most confident predictions are also its most reliable ones for experimental follow-up.
- Unlike "meta-score" approaches (citation counts, production volumes) that bias toward re-discovering well-studied compounds, DeepMet's sampling frequency is defined even for structures absent from any existing database, making it suited to genuine novel-metabolite discovery rather than re-annotation.
Availability¶
Interactive web application for structure prioritization from mass, formula, or MS/MS: deepmet.org. Snakemake pipeline for training/sampling/evaluating chemical language models on other metabolite classes: github.com/skinniderlab/clm (also archived on Zenodo). Generated structures and CFM-ID-predicted spectra: Zenodo. Experimental MS data: MassIVE accession MSV000097536.
See Also¶
- MetGenX — cites DeepMet as a deep generative approach to anticipating undiscovered metabolites, in contrast to its own template-guided structure-to-structure generation from experimental MS2 spectra.