DeepMet

Profiling 23 mouse tissues and biofluids by LC–MS/MS yielded 4,814 peaks that behaved like genuine metabolites; an in-house library of standards identified 250 of them, or 5.2%. The rest are the metabolome's chemical "dark matter" — real molecules, reproducibly measured, that no database can name. The bottleneck is candidacy rather than detection: a spectral search can only return a structure somebody has already written down.

DeepMet attacks that gap from the structure side. It is a chemical language model trained on the SMILES strings of metabolites already known to occur in human tissues and biofluids, on the premise that metabolites are assembled from a small pool of precursors through a limited repertoire of enzymatic transformations and therefore share a learnable biosynthetic grammar. Molecules sampled from the trained model cannot be reliably separated from real metabolites by a random forest classifier, and they recapitulate 77.5% of the one-step enzymatic transformations predicted for known metabolites by the rule-based BioTransformer platform, despite the model never being shown a reaction rule. It comes from the Rabinowitz and Skinnider groups at Princeton, with the Wishart lab at Alberta.

The scoring mechanism is what makes it usable. A structure has no unique SMILES spelling, and the model assigns very different likelihoods to different spellings of one molecule, so per-string likelihood is not a plausibility measure. Instead, one billion SMILES strings are drawn and the frequency of each unique structure tabulated; that aggregate sampling frequency is the model's belief that a structure belongs to the metabolome. An unrelated one-class model from Birmingham and Imperial also carries the name DeepMet, scoring metabolite-likeness of supplied structures rather than generating them.

#Key Features

• Frequency-based plausibility: Structures sampled more often are more similar to known metabolites and likelier to be plausible enzymatic products. Withheld metabolites separate from other generated molecules on sampling frequency alone at an AUC of 0.98.
• Zero-shot ranking from a mass: Generated structures matching a measured mass to ±10 ppm are ranked by sampling frequency. Supplying 176.0950 Da returns 2,818 unique candidates, of which serotonin — withheld from training — ranks first.
• MS/MS integration: Predicted tandem mass spectra are combined with the sampling frequency, so the model slots into conventional spectral-search workflows rather than replacing them.
• Calibrated confidence: Scores derived from sampling frequencies track the probability that an assignment is correct, so the most confident hypotheses are the right ones to synthesize.

#Technical Details

DeepMet is a recurrent chemical language model: a three-layer LSTM with 1,024-dimensional hidden states over a 128-dimensional character embedding, trained autoregressively to predict the next SMILES token. To offset the small size of the known metabolome it is first pretrained on drug-like structures from ChEMBL v28, then fine-tuned on 2,046 metabolites experimentally detected in human tissues or biofluids in HMDB 4.0, with lipids excluded and each structure augmented by 30 non-canonical SMILES enumerations. Of 313 metabolites added in HMDB 5.0 after that training set was fixed, DeepMet generated 252 (81%), and sampling frequency separated them from background generations at an AUC of 0.97. From accurate mass alone the top-ranked structure was correct for 29% of held-out metabolites; pairing the model with CFM-ID-predicted spectra raised top-1 accuracy to 52% in positive mode and 49% in negative mode. Predicted spectra for 2.4 million generated structures were searched against 29.1 million MS/MS spectra from 4,510 published analyses of human blood.

#Applications

The practical output is a ranked shortlist: of the 10,000 most frequently generated structures, 6,301 appear in no version of HMDB. Obtaining or synthesizing standards for the top-ranked ones and profiling them under a lab's own chromatographic conditions turns a prediction into a confirmed identification by retention time and MS/MS. That route yielded several dozen mammalian metabolites, among them N-carbamyl-proline, N-succinyl-tryptophan and N-lactoyl-glutamine, with isotope tracing in mice assigning individual compounds to host, dietary or microbial origin. The same database re-interrogates archived public data, where it surfaced an N-methylated imidazolelactic acid derivative that separated sepsis patients from healthy controls.

#Impact

DeepMet carries into metabolomics an idea that reshaped protein and genome analysis: a sequence model trained on what exists can propose what plausibly exists but has not been observed. An accurate mass stops being a lookup key and becomes a distribution over plausible biogenic structures, which helps even where MS1 data is all there is. The limits are concrete. Training on human metabolites means plant or bacterial biosynthesis will need its own models; compounds entering the body from drugs, food or the environment are generated less reliably; and spectrum and retention-time prediction from structure, on which the MS/MS workflow depends, are handled by external tools and remain open problems. No trained checkpoint is distributed, but the training and evaluation pipeline, a hosted web application, and the generated structures with their predicted spectra are all public.