Edison Scientific introduced ChefNMR2, "a novel AI system for rapid, automated de novo structure elucidation from simple one-dimensional (1D) NMR spectra," built with Princeton's Ellen Zhong, assistant professor of computer science, and Mo Seyedsayamdost, professor of chemistry. Nuclear magnetic resonance spectroscopy is the gold standard for working out a molecule's structure, but reading the spectra is slow and needs specialist expertise. Small molecules account for more than half of FDA-approved medicines.
Two components
- ChefNMR2: a generative model that frames structure elucidation as generating atomic coordinates from a molecular formula and 1D 1H and 13C peak lists. It combines a peak encoder with a multimodal diffusion transformer with joint attention between peak and atom tokens, and was trained on 2.52 million molecules with experimental and simulated spectra. At inference it turns noise into coordinates, assigns bonds to get candidate structures, and uses how often each distinct planar structure appears across repeated samples as its confidence.
- ChefNMR2-Agent: an agent with a natural-language interface that adds chemical reasoning from an LLM, chemical-shift prediction and literature search. It uses ChefNMR2's confidence to decide how much to sample and analyze, returning confident answers quickly and digging further into uncertain ones.
The results
Across five test sets totalling 2,031 molecules, ChefNMR2 ranks the correct planar structure first for 82.4% of molecules in the announcement (82.3% in the full write-up). The comparisons in the write-up are 52.0% for Claude Opus 5, the best LLM tested (51.9% in the announcement), 60.1% for the retrieval method NMR-Solver and 35.4% for the best specialized de novo model, NMRTransformer.
- Generalization: the correct structure is in the top ten for 87.5% of molecules with more than 30 heavy atoms and 84.8% of molecules whose nearest training neighbor has a Tanimoto similarity below 0.6.
- Calibration: sampling frequency tracks accuracy, with an expected calibration error of 1.3 percentage points, and the confidence of NMRPeak-G-Raw and two LLMs strays further from observed accuracy.
- Speed and cost: after 100 initial samples the model adapts its budget to confidence, running two to three orders of magnitude faster than frontier LLMs, with up to 2,170 times lower median runtime and 360 times lower median cost per molecule.
Natural products
On nine well-known natural products, including vancomycin aglycone, rapamycin and taxol, ChefNMR2 gets all nine right at rank one. NMRTransformer gets three, NMRPeak-G-Raw none, and the frontier LLMs GPT-5.6 Sol and Claude Opus 5 get eight and seven, though Edison notes that for such famous molecules it is hard to tell inference from recall. On seven closely related taxane diterpenoids from a taxol biosynthesis study, ChefNMR2 has the right structure in its top ten for five. NMRTransformer, NMRPeak-G-Raw and four of six LLMs identify none, while GPT-5.6 Sol and Claude Opus 5 each identify four, taking 6 to 25 minutes per molecule. Opus 5 solves taxol in about a minute, which Edison reads as familiarity.
Unpublished molecules
ChefNMR2 was built into Edison's AI scientist, Kosmos, to elucidate four unpublished natural products isolated in the Seyedsayamdost lab, matching human-chemist accuracy "in a fraction of the time" from 1D NMR spectra and molecular formulae only.
- Nocobactin AI-1: ChefNMR2 ranks the right structure fourth or fifth. With chemical-shift prediction, spectral interpretation and literature search, the agent ranks it first in three of five runs. Without ChefNMR2 the agent usually fails, and every specialized baseline and LLM tested fails.
- Nocobactin AI-2: solved from automatically picked peak lists.
- Hybridamides A and B: both solved, including an unprecedented prenyl substitution at the Nε-nitrogen of arginine in hybridamide A.
Availability
ChefNMR2 is available in Kosmos for Edison's partners. A web app and MCP tool are on Hugging Face, and the project page is hosted at Princeton. All results are Edison's own. For other AI work on molecules, see Hugging Face's Carbon-X DNA model.