Frontier language models alone identified 2% of undocumented molecules. With Pyxis using our AI structure generator, LSM-Gen, it identified 38%, more than any state-of-the-art alternative.

In June, Anthropic reported that Claude could work out molecular structures from NMR spectra with no tools at all. We asked whether that carries over to tandem mass spectrometry (MS/MS), the dominant technique for identifying unknown molecules. That question matters. Most spectra in a typical sample are never matched to a structure. Yet identifying those molecules is what lets scientists check a new drug’s breakdown products for safety, or track a new pollutant in drinking water.

Separating literature recall from chemical reasoning. A language model can identify a well-studied compound by recalling it from its training data, without generalizable reasoning over the spectrum. That distinction matters because the hardest identifications are often novel molecules and scaffolds that no one has written about, exactly where an approach must generalize rather than recall. To distinguish recall from reasoning, we created a 100-molecule benchmark. Half the molecules are well documented, each with at least 100 linked PubMed papers and a common name. The other half are undocumented, with no linked papers and only a systematic chemical name. Each undocumented molecule is paired with a well-documented one whose spectrum is similar in mass, number of peaks, and ion type, so structures are comparably hard to derive in both halves. Claude Opus 5 and Pyxis each received the same spectrum and molecular formula and proposed five ranked structures. We scored each system's top-ranked answer.

Specialist tools enable Pyxis to identify undocumented molecules

Alone, frontier language models recalls the literature. On its own, Claude identified 52% of the documented molecules and only 2% of the undocumented ones, a single molecule out of 50. Its wrong answers on the undocumented half were structurally far from ground truth.

Specialist generators enable de novo identification. We then added specialist tools one at a time, keeping the model and its budget fixed. Tools that help Pyxis check its own ideas, by assigning a chemical formula to each peak or simulating the spectrum a candidate would produce, improved the documented half and left the undocumented half at 1 in 50. Only de novo generators, AI models that propose candidate structures directly from a spectrum, changed that. Paired with FRIGID and MS-GPT, the two leading public generators, Claude identified 12% and 26% of these molecules, respectively. Pyxis, using Matterworks’ structure generator, LSM-Gen, achieved 38%—the best result in the study, up from 32% for LSM-Gen alone.

LLMs excel at choosing, not inventing. The two parts of the system do different jobs. The generator proposes candidates; Pyxis weighs them against the spectrum. When the right structure was on LSM-Gen’s list, Pyxis ranked it first for 19 of 20 undocumented molecules. What frontier language models rarely did was find a correct structure the generator had missed, so on undocumented molecules the generator drives utility. Across every setup we tested, most misses happened because no tool and no model ever proposed the right structure.

Better answers at lower cost

On its own, LSM-Gen outperformed both state‑of‑the‑art generators, FRIGID and MS‑GPT, on well‑documented and undocumented molecules alike.

LSM-Gen cut the cost per correct identification across all 100 molecules from $4.70 to $2.14. Without a generator, frontier language models exhausted its reasoning budget before committing to an answer on 54% of undocumented spectra, listing and rejecting candidates from memory. With LSM-Gen, that fell to 6%.

What comes next

Better reasoning only helps if the right molecule makes the candidate list. Even Pyxis identified fewer than half of the undocumented molecules in this test. The next leap requires stronger structure generators, paired with agents that can put them to work. We generate our own spectra in an AI-native lab, giving us control over the data that raises the ceiling. We’re scaling LSM-Gen’s training data and model together, testing each gain against broader validation sets. Every AI system that reasons about chemistry will need a model that reads spectra. We’re building it.

Matterworks builds Large Spectral Models for biochemical omics, the study of small molecules, lipids, and peptides through mass spectrometry. Learn more at matterworks.ai.