A mass spectrum still leaves a choice

An unknown sample produces a spectrum. A model returns a molecular drawing. The important question is whether the drawing identifies the sample, or offers one plausible candidate among several. DiffMS, presented at ICML in 2025, tackles candidate generation from mass spectra while keeping the molecular formula as an explicit input.¹

Fix what is known

The model combines a spectrum encoder with a graph diffusion decoder constrained by the heavy-atom composition of a supplied formula.¹ Its illustrated workflow assumes that external tools determine that formula.¹ The task is therefore more specific than recovering every aspect of a molecular identity from a raw, otherwise uninterpreted measurement.

For our reading, the useful distinction is between evidence provided to the model and evidence inferred by it. A candidate matching an input formula has satisfied a constraint, not independently verified that the formula was correct. We would keep the provenance and confidence of that upstream input beside the generated candidates.

A better benchmark result can still be a hard problem

In the paper's main comparison, DiffMS achieves 8.34% top-one exact accuracy on NPLIB1 and 2.30% on MassSpecGym.¹ Those numbers describe the first-ranked candidate matching the benchmark structure. They make it difficult to confuse improved performance with routine, unambiguous identification.

We would want an evaluation to report the practical cost of examining the remaining candidates too. Does a shortlist help an analyst choose a discriminating next measurement? Does it contain the relevant alternative, and how far down the list? These questions could connect a benchmark advance to an actual identification workflow without pretending that the connection has already been demonstrated here.

The paper motivates its task with cases in which different structures have very similar mass spectra.¹ A computational preference cannot create experimental information that the measurement did not distinguish. Our interpretation is that ambiguity belongs in the output record rather than being hidden by the convenience of displaying one structure.

Separate pretraining from measurement evidence

The decoder is pretrained using fingerprint–structure pairs before the spectrum-conditioned task.¹ This is an important use of structural information, but it is not a collection of new measured spectra. The history of the candidate generator and the evidence about the sample should remain distinguishable.

When assessing another implementation, we would ask which candidate ranking was fixed before the true identity was known, what data were excluded during evaluation and whether the ranking changed after additional measurements. Such a record would help a reader identify where the model actually contributed useful information.

Keep the output provisional

The practical role we would investigate is candidate assistance: proposing structures that can be challenged with further evidence. An informative report would retain unsuccessful proposals and record why a later observation ruled them out. It would not substitute a persuasive molecular image for a completed structure assignment.

AiChemEx inspected the official proceedings paper but did not rerun DiffMS or identify a laboratory sample. The study supports a computational method and benchmark comparison. Its remaining error rate is part of the result, and essential context for deciding what a generated molecular candidate is allowed to mean.