Mass-spec structure splits drift as libraries grow: one component swallows 95.5% of 227,690 molecules, leaving a 4.5% holdout

Splitting a mass-spectrum structure-prediction benchmark by single-linkage clustering at an MCES-10 threshold yields a holdout whose size depends on how large the library is, rather than a fixed benchmark. The rule was fixed at a library of about 29,000 molecules, while public mass-spectrometry libraries now hold 227,690. On the 227,690-molecule public union the largest component holds 95.5% of the library, capping the achievable holdout at 4.5%; on Enveda-180 alone it falls to 0.84%.

The holdout does not merely shrink: the compounds left in it are systematically heavier, because a fixed absolute edit budget is a tighter relative tolerance on a larger molecule. The split had been treated as a construction that guarantees structural separation without introducing another difference. The probability that a holdout compound outweighs a training compound rises monotonically with library size, reaching 0.881 on the union; on Enveda-180 the training median mass is 329.2 Da against 384.3 Da in the holdout.

To make the criterion computable at today's library sizes, the authors reduce the MCES distance computation so that only a tiny fraction of molecular pairs reach an integer program. MCES is NP-hard per pair and the number of pairs grows quadratically with library size, so the criterion had never been evaluated at this scale. Vectorized bounds certify 87% of pairs as far, a union-find skip removes a further 97.8%, and 11,476,315 pairs reach the integer program, 0.04% of all pairs.

The bias is a property of the criterion rather than of one library's composition, since the same trends reappear on a library assembled from a different source. The two experimental libraries have markedly different mass profiles, with Enveda-180 dominating the shape of the union while SpectraVerse contributes more high-mass compounds. On a third library combining SpectraVerse with COCONUT, the achievable holdout falls monotonically from 40.6% to 14.6% and the mass AUC rises monotonically from 0.647 to 0.726.

The scope is how MCES-10 structure-disjoint splits behave as mass-spectrometry structure-prediction libraries grow, for researchers who build or use such benchmarks. The authors measure on the public union, on Enveda-180, and on a third library combining SpectraVerse with COCONUT, at the threshold of 10 that MassSpecGym adopts, and rerun the union at a threshold of 5 in an appendix. They state that structure-disjoint benchmarks must adopt criteria whose meaning is stable under scale.

Both experimental libraries carry meaningful biases in their compound-mass distributions; the authors consider benchmarking on more regular mass distributions preferable but keep the current libraries to reflect realistic deployment conditions. Only 19.4% of the third library has spectra, and its ladder stops at a smaller size because of compute constraints. The integer program runs under a five-second budget and timed-out pairs are conservatively banked near, so reported giant fractions are upper bounds; resolving 3,265 of the union's 55,478 timeouts to completion moves the giant from 95.4983% to 95.3441% and the mass AUC from 0.8810 to 0.8853.