Can AI identify an unknown molecule from 1D NMR alone?
An unreviewed system ranked the correct planar structure first for 82.3% of 2,031 benchmark molecules and helped resolve four new natural products. The four-compound demonstration is promising, not prospective proof of routine laboratory performance.
The short version
- What happened
- ChefNMR2 was trained on 2.52 million experimental and simulated molecule-spectrum pairs and evaluated on 2,031 molecules across five held-out or public experimental test sets.
- Who it affects
- Chemists could spend less time searching enormous candidate spaces and more time checking the few structures most consistent with the spectra. The risk is that a fast, calibrated-looking answer gains authority before independent measurements or expert review have excluded a plausible isomer.
- What remains unproven
- Unreviewed preprint; the complete frozen test sets and SpectraNP+ training collection were not yet all deposited for independent reproduction.
Keep in mind
Unreviewed preprint; the complete frozen test sets and SpectraNP+ training collection were not yet all deposited for independent reproduction.
The benchmark evaluates planar structure and does not establish stereochemistry, purity or performance in a complete laboratory workflow.
Only four novel natural products were used in the practical demonstration, and they were selected by the study team.
Human chemists had richer 2D NMR and mass-spectrometry evidence, so speed comparisons are not like-for-like.
Three authors are Edison Scientific employees, and the agent uses that company's Kosmos platform.
Source analysis · 1 linked source. Primary analysed source date: 31 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 5 min The evidence, methods, implications and limitations
A specialized generator beat both retrieval and general-purpose reasoning
A Princeton-led preprint reports that an AI system could infer small-molecule structures from a molecular formula and one-dimensional proton and carbon NMR peak lists. ChefNMR2 uses an atomic flow-matching model trained on 2.52 million molecules drawn from literature spectra, US-patent compounds and natural-product databases. The researchers tested it on 2,031 molecules across five sets: 500 each held out from three training sources, 315 from NMRShiftDB2 and 216 teaching compounds with experimental spectra.
Across those sets, the system ranked the correct planar structure first for 82.3% of molecules. The reported comparison was 60.1% for the retrieval system NMR-Solver, 52.0% for Claude Opus 5—the best of six tested general-purpose language models—and 35.4% for NMRTransformer, the best specialized de novo baseline. ChefNMR2's sampling frequency also tracked observed correctness with an expected calibration error of 1.3 percentage points. On the pooled benchmark, its median runtime was reported as much as 2,170 times lower and its median cost 360 times lower than the language models.[1]
Four new compounds tested whether the system could move beyond benchmarks
The more tangible demonstration involved four previously unreported natural products: two nocobactins from Nocardia transvalensis and two hybridamides from a Streptomyces strain. Human chemists first resolved the structures using one- and two-dimensional NMR and, where relevant, tandem mass spectrometry. The AI then received only the molecular formula and 1D peak lists. For nocobactin AI-1, the generator placed the correct structure fourth or fifth, while a reasoning agent used predicted shifts and literature context to rank it first in three of five runs. It kept nocobactin AI-2 first in all five runs.
For hybridamide B, the generator found the right components but reversed two units; the agent corrected the order. For hybridamide A, it moved an unusual prenyl group to the experimentally supported site, ranking the correct structure first in every run and within the top two in four of five. These examples show useful tool orchestration, but four hand-selected compounds are not a clinical-trial-style denominator. The human comparison also used richer measurements, so the reported 86- to 1,032-fold speed advantage is not a like-for-like replacement claim.[1]
The distinction that matters
What was tested
The paper combines a large retrospective benchmark with a much smaller real-discovery demonstration; they answer different questions.
Benchmark
2,031 molecules across five test sets
82.3% correct planar structure ranked first
Harder subset
35 large, dissimilar molecules
57.1% placed the correct structure in the top ten
Novel compounds
Four structures already resolved by human chemists
Agent recovered or corrected the reported structures from 1D inputs
The next test is blinded use on molecules the team did not select
Generalization remains the main uncertainty. The authors removed identical planar structures across their own train, validation and test splits, and reported top-ten accuracy of 84.8% when the nearest training molecule had Tanimoto similarity below 0.6. Performance fell to 57.1% among the 35 molecules that were both large and dissimilar. Some comparator tools had training overlap with test molecules, while the paper's own complete test sets are promised for Zenodo and one training collection is due only upon publication. Independent reproduction therefore cannot yet inspect every benchmark input.
The work is a bioRxiv preprint and has not been peer reviewed. Three authors work for Edison Scientific, whose agent platform was used; other funding included the NIH, Princeton programmes, Anthropic's AI for Science Program and foundations. What would change the assessment is a preregistered, blinded challenge on prospectively collected spectra, public frozen test sets, stereochemistry and impurity handling, and evidence that chemists save time without accepting confidently wrong structures. The immediate value is candidate ranking for expert review, not an autonomous certificate of molecular identity.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
Google DeepMind · 1 min 52 sec · Background explainer
Protein folding explained
YouTube loads only when you choose; playback starts when you press play.
A short introduction to why a protein’s shape matters. It provides background for understanding what AI structure prediction contributes to biological research.
Evidence check: This is the developer’s explanation. Predicting a structure does not by itself establish a treatment’s safety or benefit for patients.
Published 2020-11-30. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current researchContinue the story
Read next
Continue exploring ai for science: compare the finding and its evidence limits.
AI agents reproduced 13 of 93 scientific tasks—what failed?
PaperBenchX replayed agent submissions across 12 research areas and scored only regenerated evidence. The best configuration fully completed 13 tasks, while validation remained the weakest stage; 81 tasks and most reference workflows stay private.
Read the next reportExplore the full source and evidence record
What this means for people
- Chemists could spend less time searching enormous candidate spaces and more time checking the few structures most consistent with the spectra. The risk is that a fast, calibrated-looking answer gains authority before independent measurements or expert review have excluded a plausible isomer.
Global context
Automated interpretation could widen access to specialist chemical analysis in laboratories with limited expert capacity, but training data and reference spectra are uneven across compound classes. Prospective tests should include natural products and synthetic compounds from regions and institutions outside the development team.
What the evidence does not yet show
- Unreviewed preprint; the complete frozen test sets and SpectraNP+ training collection were not yet all deposited for independent reproduction.
- The benchmark evaluates planar structure and does not establish stereochemistry, purity or performance in a complete laboratory workflow.
- Only four novel natural products were used in the practical demonstration, and they were selected by the study team.
- Human chemists had richer 2D NMR and mass-spectrometry evidence, so speed comparisons are not like-for-like.
- Three authors are Edison Scientific employees, and the agent uses that company's Kosmos platform.
What to watch next
- Peer review and release of the frozen test sets
- Independent reproduction of the 82.3% benchmark result
- Blinded prospective structure-elucidation challenges
- Handling of stereochemistry, mixtures and noisy peak picking
- Chemist time saved and error interception in real workflows
- Performance on compounds far from the training distribution
Living evidence record
Impact record IAI-1YH4CST
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
1 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what bioRxiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 1 November 2026
Research question, topics and regions
AI-assisted small-molecule structure elucidation
Topics: AI for science · Chemistry · NMR · Molecular discovery · Scientific agents · Preprint
Regions: North America · United States
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.