Making Claude a Chemist
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic published a white paper comparing three Claude models against dedicated NMR software ChemDraw and MestReNova on 20 novel compounds drawn from ChemRxiv preprints posted after the models' training cutoff to avoid selection bias.
- Opus 4.7 posted an average hydrogen-spectrum error of ±0.079 ppm—well under half the ±0.20 ppm tolerance window—while on carbon it effectively tied with MestReNova at ±1.37 vs ±1.48 ppm.
- Claude models predicted sub-peak spacing to within half a hertz roughly 80% of the time, against 26–35% for ChemDraw and MestReNova, and Opus 4.7 matched experimentally reported splitting patterns more often than any other tool tested.
- On inverse structure elucidation, Opus 4.7 recovered all eight simpler target structures on every attempt using only the molecular formula and 1D NMR spectra, and on seven harder targets with a starting-material hint, returned the correct structure on all three runs for four of them and on two of three runs for the rest.
- Anthropic notes the inverse elucidation task—proposing a molecule's structure from its spectrum—is the harder direction, and one that existing dedicated software typically requires 2D NMR, specialized training, and licensed tools to attempt.
- Anthropic acknowledges limitations: the 20-compound assessment spans only four structural families, and a stronger evaluation would cover several hundred compounds across 20–30 scaffold classes with at least 15 per class.
Why it matters: NMR structure elucidation is one of the most time-consuming steps in synthetic chemistry, and the fact that a general-purpose model like Claude can now match or beat ChemDraw and MestReNova on forward prediction—and handle 1D inverse elucidation without specialized training or 2D spectra—means academic and small-lab chemists who don't currently use retrosynthesis or AI tools gain a low-setup path to a task that the article says is still done largely by hand.
Ask SkimNews
