Small molecule structure elucidation from NMR spectra with generative and reasoning models

ChefNMR2-Agent enables rapid de novo small-molecule structure elucidation from molecular formulae plus 1H and 13C NMR spectra, substantially outperforming existing tools with 2–3 orders of magnitude faster inference than frontier LLMs.

• Why it matters: NMR is a gold standard for structure elucidation, but interpreting spectra is labor-intensive and expert-dependent, limiting speed through the large space of possible structures.
• What they did: The system combines ChefNMR2, an atomic flow-matching generative model, with ChefNMR2-Agent that uses LLM-based reasoning to refine candidate structures from formulae and 1D 1H/13C NMR.
• The result: Across diverse benchmarks it provides calibrated confidence estimates and, when applied to four novel natural products, matches human-chemist accuracy in a fraction of the time.

AI summary of the abstract · may contain errors · not medical advice · read the paper.