Teaching machines to read the language of molecules.

Somewhere in a lab right now, a mass spectrometer is producing a graph that almost nobody can read. It is a forest of peaks - thousands of molecular signals stacked on top of each other, recorded in a few seconds, and then quietly handed to a human who will spend the next several weeks trying to make sense of maybe a few dozen of them. The instrument is extraordinary. The bottleneck is everything that happens after it.

Matterworks exists in that gap. The company, founded in 2019 and run out of Somerville, Massachusetts, builds what it calls Large Spectral Models - foundation models that learn the structure of molecular data the same way a language model learns the structure of text. Feed them raw spectra and they read directly: this is the molecule, this is how much of it there is, here is what it might mean. No manual peak-picking. No per-analyte calibration curve drawn by hand. The forest becomes a list.

Mass spectrometry is a superpower nobody can scale

Here is the uncomfortable truth about one of biology's most powerful instruments: it generates more data than the field can actually use. Mass spectrometry can, in principle, measure thousands of molecules in a single sample - the metabolites, lipids, and proteins that describe what a cell is actually doing. In practice, most labs measure a short, pre-chosen list, because turning the raw signal into trustworthy numbers is slow, manual, and reserved for a small priesthood of analytical chemists.

The result is a strange kind of scarcity. The molecules are right there in the data. The answers are technically already recorded. They are just locked behind a workflow that does not scale - one calibration curve, one peak, one expert judgment at a time. Biology has spent years drowning in instruments that produce more than anyone can read.

A chemist and a computational biologist walk into a spectrum

The bet behind Matterworks is that spectra are a language, and languages can be learned. Jack Geremia, the CEO, brings the chemistry: a PhD in Chemistry from Princeton, a postdoc in control and dynamical systems at Caltech, and more than two decades building early-stage biotech - including Midori Animal Health, which was acquired by DSM. Mimoun Cadosch Delmar, the CTO, brings the machine learning, sharpened as a computational biologist at the Broad Institute of MIT and Harvard.

Their wager was self-supervision. Rather than hand-label data molecule by molecule - the very bottleneck they were trying to kill - they trained models on billions of raw spectra and let the models find the chemical and biological relationships themselves. It is the same trick that turned piles of unlabeled text into something that can hold a conversation, pointed instead at the unstructured molecular data biology had been throwing away.

Pyxis: a compass for molecular data

The models became a product called Pyxis - named, fittingly, after a small constellation that depicts a mariner's compass. Pyxis is a cloud platform that Matterworks describes as a predictive omics assistant. A scientist uploads raw LC-MS data; Pyxis performs untargeted absolute quantitation, returning identified molecules with concentrations attached. The per-analyte calibration, the peak selection, the data integration - the weeks of expert handwork - happen inside the model instead of on a researcher's desk.

"Untargeted" is the word doing the heavy lifting. Traditional workflows force you to decide what you are looking for before you look. Pyxis is designed to read what is actually there, which is a meaningfully different proposition for drug development, synthetic biology, microbiome research, toxicology, and biomanufacturing - anywhere the interesting molecule might be one you did not think to put on the list.

A data layer biology has been missing

Matterworks runs a model-as-a-service business: it sells access to its spectral models and to Pyxis, and customers point them at data from instruments they already own. The Series A money, the company says, goes toward growing that service and expanding its machine-learning and scientific teams. The framing is bigger than a software upgrade. If molecular data becomes something any biologist can read on demand, then biology gets a new layer to work with - measurements that were always being recorded but rarely used.

There is healthy skepticism to hold here. Foundation-model claims are easy to make and hard to verify, and "AI for biology" is a crowded, sometimes overpromised neighborhood. Matterworks' reported gains - 30% on isomers, 42% more correct identifications - are the company's own figures, and the real test is whether labs see them on their own messy samples, not in a press release. The endorsements from investors and partners are encouraging; they are not proof.

Back to that unreadable graph

Return to the lab from the opening, and the spectrometer humming through another sample. The forest of peaks is still there - thousands of signals, recorded in seconds. What has changed is what happens next. Instead of weeks of expert handwork that surfaces a few dozen molecules, the raw signal goes into a model and comes back as a list: identified, quantified, and ready to act on, in minutes rather than weeks.

That is the bet Matterworks is making - that the bottleneck in biology was never the instrument, but the reading. Whether the whole field comes to read molecules this way is still an open question. But the graph that almost nobody could read is, for a growing number of labs, starting to read back.