Lab Newsletter — September 4, 2026: Naming the Unknowns

Of all the molecular layers, the metabolome sits closest to the action. Genes say what a cell could do, RNA what it’s transcribing, proteins what machinery it’s built — but metabolites are the small molecules of the actual chemistry: the sugars being burned, the lipids being assembled, the signals being passed. As the team behind CSI:FingerID (Dührkop … Böcker, PNAS, 2015) puts it in a single clean sentence, “metabolites provide a direct functional signature of cellular state.” The instrument that reads them, mass spectrometry, is dazzlingly sensitive — and yet the field carries an uncomfortable open secret, stated in the same abstract: “untargeted metabolomics experiments usually rely on tandem MS to identify the thousands of compounds in a biological sample. Today, the vast majority of metabolites remain unknown.” A typical run lights up thousands of peaks; only a small fraction can be named. This is the dark metabolome, and machine learning is what’s finally lighting it up.

🧩 The core trick: from a spectrum to a fingerprint

A tandem mass spectrum is a molecule shattered into fragments — an indirect, cryptic sketch of a structure. The breakthrough idea in CSI:FingerID is to not guess the structure directly but to guess a fingerprint first. The method “computes a fragmentation tree that best explains the fragmentation spectrum of an unknown molecule,” then uses “the fragmentation tree to predict the molecular structure fingerprint of the unknown compound using machine learning.” That predicted fingerprint — a checklist of chemical substructures — becomes the search key against databases like PubChem. The approach, the authors report, “improve[s] on the competing methods for computational metabolite identification by a considerable margin.” It reframed a chemistry problem as a learning problem, and the rest of the field followed.

The tool that put it in everyone’s hands was SIRIUS 4 (Dührkop … Böcker, Nature Methods, 2019). Its abstract is refreshingly blunt about the stakes — “mass spectrometry is a predominant experimental technique in metabolomics and related fields, but metabolite structural elucidation remains highly challenging” — and its contribution is to make the CSI:FingerID pipeline fast and usable: SIRIUS 4 “integrates CSI:FingerID for searching in molecular structure databases,” and with it the authors “achieved identification rates of more than 70% on challenging metabolomics datasets.” Seventy percent, on the hard cases, for a problem that used to strand most peaks unnamed.

🗂️ When the molecule is in no database

Fingerprint search only works if the answer is somewhere in a database. For genuinely novel chemistry — a natural product no one has catalogued — there’s nothing to match against. CANOPUS (Dührkop … Böcker, Nature Biotechnology, 2021) takes the humbler-but-powerful route: if you can’t name the molecule, name its kind. The authors note plainly that “structural molecule annotation is limited to structures present in libraries or databases, restricting analysis and interpretation of experimental data,” and answer with a deep network that “predict[s] 2,497 compound classes from fragmentation spectra, including all biologically relevant classes.” Crucially, CANOPUS “explicitly targets compounds for which neither spectral nor structural reference data are available” — the truly dark peaks — and still reaches “an average accuracy of 99.7% in cross-validation.” Knowing a mystery peak is a particular class of lipid or alkaloid is often enough to turn a dead end into a lead.

✍️ Writing the structure from scratch

The boldest step is to skip the database entirely and draw the molecule. MSNovelist (Stravs … Zamboni, Nature Methods, 2022) does exactly that. Noting that “current methods for structure elucidation of small molecules rely on finding similarity with spectra of known compounds, but do not predict structures de novo for unknown compound classes,” it “combines fingerprint prediction with an encoder-decoder neural network to generate structures de novo solely from tandem mass spectrometry (MS2) spectra.” Tested on “3,863 MS2 spectra from the Global Natural Product Social Molecular Networking site, MSNovelist predicted 25% of structures correctly on first rank, retrieved 45% of structures overall … without having ever seen the structure in the training phase.” It’s the small-molecule cousin of the de novo peptide sequencing we covered last month — a generative model that writes an answer instead of looking one up — and, like the best of these tools, it’s framed as a complement: “ideally suited to complement library-based annotation in the case of poorly represented analyte classes and novel compounds.”

🌐 Better similarity, and an open commons

Two more pieces make the ecosystem work. Spec2Vec (Huber … van der Hooft, PLOS Computational Biology, 2021) rethinks how we compare spectra at all. Since “spectral similarity is used as a proxy for structural similarity” in library matching and molecular networking, it borrows from language modeling — “a novel spectral similarity score inspired by a natural language processing algorithm — Word2Vec” — treating fragment peaks like words and learning embeddings that “correlate better with structural similarity than cosine-based scores,” and do so fast enough for “structural analogue searches in large databases within seconds.” And GNPS (Wang … Bandeira, Nature Biotechnology, 2016) supplies the thing every learning system needs — shared data — as “an open-access knowledge base for community-wide organization and sharing of raw, processed or identified tandem mass (MS/MS) spectrometry data,” where “crowdsourced curation” improves annotations and the data is treated as “’living data’ through continuous reanalysis.” Molecular networking on GNPS lets a confident annotation propagate to its unknown neighbors — one named peak lighting up the dark ones around it. The honest scorekeeper for all of this is the community CASMI contest, the prove-it standard MSNovelist reports against.

🧭 Why it’s our kind of problem

It lands right on the lab’s map. If a serious virtual cell or Human Cell Simulator is going to be more than a catalog of genes and proteins, it needs metabolism — the small-molecule chemistry that is, quite literally, “a direct functional signature of cellular state.” Metabolomics is the omics layer that our multi-omics integration story is still mostly missing, the phenotype-end partner to the proteome. And the shape of the challenge is one we keep returning to: a dark unknown to be annotated — the same discipline as the dark proteome, aimed at a different class of molecule. The way forward is the ethos this digest keeps championing: open, callable, benchmarked tools — SIRIUS and GNPS’s “living data,” scored against a public contest like CASMI — the same publish-the-model-and-the-test spirit behind the BioImage Model Zoo and BioEngine. Teach a machine to read a shattered spectrum and name the molecule that made it, and thousands of anonymous peaks per sample start becoming biology you can reason about.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.