A good spectral match can still name the wrong molecule

The most instructive result in an automated structure-identification paper may be the assignment it gets wrong. A ranked molecular drawing can look definitive on a screen. The analytical question is whether the observed data distinguish that drawing from its alternatives, and whether the prediction machinery can represent the chemistry accurately enough to make the comparison useful.

A loop built around chemical shifts

Jin and colleagues' NMR-Solver, published in April 2026, combines candidate retrieval with fragment-based optimization guided by proton and carbon NMR chemical shifts. Its evaluation includes simulated spectra, curated literature spectra and laboratory examples.¹

In the 450-sample literature benchmark, with molecular formula supplied and both proton and carbon spectra, the reported first-ranked recall is 52.89%; recall within ten candidates is 67.33%. Those are different identification tasks, and neither result establishes a universal probability that the next proposed structure is correct.¹

A strained-system example reveals the boundary particularly clearly. An error in the forward chemical-shift predictor caused the procedure to favor an incorrect structure. The candidate matched the observations better according to the model, while the real product was disadvantaged by that model's prediction error.¹

Preserve the alternative explanation

Our interpretation is that the assignment record should carry more than the winning molecular representation. It should preserve the observations used, the constraints supplied and the alternatives that remained plausible. I would also want the reader to see which spectral features most strongly changed the ranking.

Imagine an illustrative case with two candidates that exchange places after a small change to how a peak is represented. I would treat that as a reason to inspect the interpretation, rather than to hide the earlier candidate once the software has settled. The important result would be the reason for choosing between the structures. Repeatedly obtaining the same answer from the same model would not answer that question by itself.

This is a proposed reporting discipline, not a new experiment conducted by AiChemEx. A practical follow-up could state the disputed assignments before collecting additional evidence, then show whether that evidence actually removed the ambiguity. The report should leave room for the answer that it did not.

Keep each source of information visible

The same discipline applies to known starting materials and a supplied molecular formula. In future coverage, I would label those inputs next to the output, so readers can distinguish a proposal aided by reaction context from one derived using fewer prior constraints. Comparisons would then begin with the information available to each method.

I would retain the unhelpful outputs as well: structures that failed basic checks, cases without a convincing leading candidate and assignments held for further analysis. Those outcomes can show where a useful assistant should stop claiming certainty.

NMR-Solver's explicit failure provides a concrete starting point for this standard.¹ The article's lesson is about the direction of the evidence: a proposed molecule should answer to the measurement, while the predicted spectrum remains a fallible bridge between them. A trustworthy analytical workflow would make a weak bridge visible before turning its preferred drawing into the permanent record of a reaction product.