The structure is in the spectra

A packet of ordinary spectra is enough for a frozen procedure to name the constitution and the relative configuration of a small molecule that has never been published. A person may start the procedure. A person may not choose or repair the structure after seeing the packet.

Structure assignment sits between a routine measurement and every decision that follows it: which isomer the reaction made, which impurity is in the batch, which metabolite is in the extract. A laboratory can already record the packet in an afternoon. What it cannot yet do is trust a single, unedited output the way it trusts the mass measurement. Solving this would turn structure elucidation into a check a second laboratory can grade, including for a model.

Published top-1 rates on structure-from-spectrum benchmarks do not clear the grade below.

Xue and co-workers report a top-1 accuracy of 92.07 percent across the input conditions of their own benchmark, and 85.10 percent when the inputs are one- and two-dimensional NMR without a molecular formula or fragments (Analytical Chemistry, 2025, doi:10.1021/acs.analchem.5c03783 (opens in a new tab)). Their abstract also says the model was used on six previously uncharacterized natural products from Magnolia officinalis and on six unexpected synthetic products. Fang and co-workers, re-evaluating named models on an experimental, scaffold-split benchmark, report a top-1 sequence accuracy for that model of 15.49 percent with the molecular formula supplied and 11.75 percent without it (arXiv:2601.15763, 2026, HTML (opens in a new tab)). Their Table 1 marks the large Mind-Gaussian set associated with the same work as non-experimental. Neither number is a relative-configuration score on unpublished molecules.

Hu, Chen, Rotskoff, Kanan, and Markland report that a multitask model finds the exact molecule 69.6 percent of the time within the first 15 predictions, on simulated proton and carbon spectra, for molecules of up to 19 heavy atoms, with no molecular formula given (ACS Central Science, 2024, doi:10.1021/acscentsci.4c01132 (opens in a new tab)). That figure is a top-15 rate on simulated spectra.

Chen and co-workers report that a diffusion model places 38 of 45 structures first when experimental chemical shifts are supplied and the two-dimensional spectra are reconstructed from those shifts together with the molecular structures (arXiv:2510.26231, PDF (opens in a new tab)). Reconstructing the two-dimensional spectra from the answer is a different task from reading a raw packet. On eleven natural products with raw experimental data, the same preprint places three structures first.

Tandem mass spectrometry is a different packet. Nowatzky and co-workers state that in silico methods in the 2016 CASMI challenge reached a recall of up to 34 percent on previously unknown compounds, and that identification rates in CASMI 2022 were below 30 percent (Nature Communications, 2025, doi:10.1038/s41467-025-57422-4 (opens in a new tab)). CASMI scores compound identification from mass spectra. It does not score relative configuration from a nuclear-magnetic-resonance packet.

An open doubt: the 92 percent author-reported figure and the 15 percent independent figure cannot both describe the same experiment. Until the spectra, the split, and the stereochemical rule are the ones in the specification below, neither figure is a solution of this problem. The Fiehn Lab results page did not load when this file was checked (a DNS error); the “below 30 percent” sentence is taken from Nowatzky and co-workers, who cite that page.

An independent laboratory prepares 20 compounds. Each has a molecular weight between 150 and 500 and at most three stereocenters. At least 10 of the 20 have at least one stereocenter, so a method that only names planar constitutions cannot fill the set. None has a published spectrum or a published structure on the day of the challenge; the laboratory records the search. Before the packet is released, a second chemist, who does not know the synthetic route, solves the structure from the packet alone and writes it down. A packet that this second chemist cannot solve is replaced. That chemist’s solution is a screen for an ambiguous packet. It is not an entry, and it is not shown to the attempt.

Each packet contains: the molecular formula from high-resolution mass spectrometry, an infrared spectrum, a proton spectrum, a proton-decoupled carbon spectrum, a COSY, an HSQC, and an HMBC. If the molecule has stereocenters, a NOESY or ROESY is included. Solvent and spectrometer frequency are stated. The packet does not contain a name, a starting material, a yield, or a drawing. The two-dimensional spectra are the spectra that were recorded, not spectra rebuilt from a candidate structure.

The attempt registers the procedure before the packets are opened: the program, the weights, and every setting that can change the output. It then returns one structure per packet within 24 hours, as the molfile or SMILES string the program writes, with relative stereo. Absolute configuration is not required. Nobody edits that file after the program stops. No further experiment is allowed. Searching a spectral database with the packet is not allowed; the compounds are unpublished so that a match would be a failure of the set, and the set would be replaced.

Grading compares the returned structure with the structure the laboratory made. Constitution is compared by the non-stereo layers of a standard InChI. Relative configuration is compared separately, at every stereocenter the packet’s NOESY or ROESY can support. A correct constitution with a wrong relative configuration does not get full credit.

At least 16 of the 20 structures are correct in constitution and in relative configuration at every stereocenter the packet’s NOESY or ROESY can support. At least 8 of those 16 are among the compounds that have a stereocenter.

Sixteen of twenty, and 8 of those 16, are pass marks defined here. It is high enough that a few misses do not pass, and it asks for a single committed structure rather than a rank inside a long list. The second-chemist screen keeps the set solvable from the packet. The locked program is what keeps a person from supplying the answer the screen already showed a person can find.

The second chemist’s solution is archived before the attempt is scored and is the key, together with the synthetic characterization the laboratory already holds. The grader does not consult the attempt’s reasoning, only the structure in the answer file. The grader also checks that the answer file matches the program’s output and was not edited.

A packet whose InChI matches more than one constitution consistent with the formula is not used. Two graders score disagreements. A disagreement that is not resolved from the packet counts the structure wrong.

Any program, model, or ranking rule that was registered before the packets were opened and that writes the answer file itself. Training on published spectra is allowed. The compounds in the challenge are unpublished.

A person assigning the structure, or choosing among the program’s candidates, after the packet is opened. Editing the answer file. The synthetic procedure, the reagents, or a photograph of the sample. A database lookup of the spectrum or of the molecular formula against a structure collection. Returning a list and asking to be graded on the best entry. Rebuilding a two-dimensional spectrum from a proposed structure and treating that rebuilt spectrum as data. Absolute-configuration claims are ignored, unless they contradict a relative configuration that was required.

At least 16 of the 20 structures are constitutionally correct, with relative configuration ignored. That partial bar is the same count, on an easier question. It is not a solution. A method that gets the skeleton and misses every stereocenter has not finished the job the packet was built to ask.

The set size, the 16-of-20 marks, the molecular-weight window, and the 24-hour clock are pass marks defined here.

Cited measurements, each opened for this draft:

• Xue et al., Analytical Chemistry, 2025, 97, 22603–22614, doi:10.1021/acs.analchem.5c03783 (opens in a new tab). Author-reported top-1 rates of 92.07 percent and 85.10 percent.
• Fang et al., arXiv:2601.15763, 2026, https://arxiv.org/html/2601.15763 (opens in a new tab). Independent experimental top-1 sequence rates of 15.49 percent and 11.75 percent. The version opened has a placeholder DOI.
• Hu, Chen, Rotskoff, Kanan, and Markland, ACS Central Science, 2024, 10, 2162–2170, doi:10.1021/acscentsci.4c01132 (opens in a new tab). 69.6 percent within the first 15 predictions, on simulated spectra.
• Chen et al., arXiv:2510.26231, https://arxiv.org/pdf/2510.26231 (opens in a new tab). 38 of 45 with two-dimensional spectra reconstructed using the structures; 3 of 11 natural products from raw experimental data placed first.
• Nowatzky et al., Nature Communications, 2025, doi:10.1038/s41467-025-57422-4 (opens in a new tab). CASMI 2016 recall up to 34 percent; CASMI 2022 identification rates below 30 percent.