Insights
Every claim that a model generalizes to new chemistry rests on the test/train split: hide molecules that are structurally unlike anything in training, and test on those. For predicting structures from mass spectra, the field converged on one rule, a graph edit distance threshold called MCES-10, adopted by the MassSpecGym benchmark. That convention was set on 29,000 molecules. Public libraries now hold 227,690. Nobody had checked whether the rule survives that growth, because checking it means solving an NP-hard problem across 26 billion molecular pairs.
So we made it cheap enough to check. Most of those pairs never need solving: a few fast tests prove that two molecules are too far apart to matter, and once two molecules are already linked through others, the pair between them cannot change the answer. Skipping that work leaves 11.5 million problems instead of 26 billion, a 2,259x saving, and the result is provably identical. A split that cost tens of thousands of core hours now takes an afternoon.
At current library scale, the similarity graph percolates: a single connected cluster absorbs 95.5% of all known molecules, leaving at most 4.5% available for testing. A held-out compound is heavier than a training compound 88% of the time, because a fixed budget of ten edits is a large change to a small molecule and a trivial one to a large molecule. The benchmark no longer measures structural novelty alone. It measures extrapolation to bigger, harder-to-fragment chemistry, and it drifts further in that direction with every compound the community adds.

Before you can blame the models. In a Nature Metabolism Comment published this June, Regina Barzilay and Ling Min Serena Khoo examined why machine learning models for small-molecule structure elucidation fail to beat simple baselines, and closed with a call for meaningful benchmarks that can tell whether new models are actually better. We think our result is part of the answer to that call: before the field can measure whether models generalize, it needs a splitting criterion whose meaning is stable as libraries grow. Today’s is not.
The question raised in Barzilay and Khoo’s article is not an abstract one for us. Matterworks trains Large Spectral Models, foundation models for biochemical omics, and structure identification is one of the things they are asked to do. When a scientist asks whether a prediction can be trusted on chemistry no one has seen before, the evaluation behind that answer is the product.
This work was presented at MoML
MoML @ MIT took place on October 14, 2026 at the Luria Auditorium, Koch Institute, Cambridge. The paper is by Tomo Oga, Devesh Shah, Cailum Stienstra, Gabriel Asher, Antonio Fonseca, Niall O’Connor, and Michael Widrich.
Matterworks builds Large Spectral Models for biochemical omics, the study of small molecules, lipids, and peptides through mass spectrometry. Learn more at matterworks.ai.
