UltraMS MSnLib benchmark: full-library formula-level split and strict-adduct molecule-level library-search split

Two frozen, independent benchmark splits of the MSnLib tandem mass spectral library, released by UltraMS. Split A — UltraMS MSnLib formula-level split. The complete library (560,084 spectra, all adducts) is split by molecular formula into 16,885 / 2,110 / 2,112 formulas and 446,667 / 55,980 / 57,437 spectra for train / validation / test (21,107 unique formulas, none in more than one fold). The test fold has 57,437 spectra, 2,112 formulas, 5,003 distinct MSnLib source SMILES and 4,889 distinct RDKit canonical isomeric SMILES. This split is the population for single-spectrum candidate ranking, multiple-spectrum voting, label-blind grouping, fingerprint prediction and embedding analyses (Figure 2b/2j, 3a/3b/3d/3g, 3a_sp_4, 3a_sp_5 and the single-spectrum baselines). The official multiple-spectrum voting units are the 5,002 MSnLib source-SMILES groups covering 57,436 spectra (the grouping of the published voting task, n_molecules = 5,002); an alternative canonical-SMILES view (4,888 groups, same 57,436 spectra) is also provided. The split-A directory also contains a compact per-spectrum retention-time table (msnlib_rt.csv.gz, row_index, rt in seconds, 560,084 rows) aligned 1:1 to MSnLib.csv and exported from MSnLib_with_rt.csv (755,984,910 bytes, SHA-256 a5b7c38228287c5c1de401050445319b331ee9a5f17993d6e104e9a734fe2964) after a row-by-row alignment check (560,084 rows, 0 mismatches); it supplies the retention times used by the label-blind grouping analysis (Figure 3d). Split B — strict-adduct molecule-level split. The strict [M+H]+ and [M-H]- subsets (447,956 source rows) are re-split 70/15/15 at the level of RDKit canonical isomeric SMILES, separately per polarity, with frozen queries and libraries for spectrum-to-spectrum library search. [M+H]+: 41,348 compounds and 301,636 preprocessed spectra, split into 28,944 / 6,202 / 6,202 compounds, with 6,202 test queries against a 295,434-spectrum test library (6,201 queries have an exact-adduct library positive) and 6,202 validation queries against a 250,424-spectrum library. [M-H]-: 21,293 compounds and 140,193 spectra, split into 14,907 / 3,193 / 3,193 compounds, with 3,193 test queries against 137,000 spectra (3,160 with a positive) and 3,193 validation queries against 116,135 spectra. This split is the population for Figure 3e/3f/3h, 3h_sp_1 and the cross-collision-energy task (6,161 positive / 2,999 negative test queries). Split seeds are 42 ([M+H]+) and 43 ([M-H]-); one query per validation/test compound is drawn with seed 119. Disjointness. Split A is formula-disjoint; split B is molecule-disjoint within each polarity. The two splits must not be combined into a global-disjointness claim: about 41% of split-B compounds lie in a different split-A fold, and a spectrum can be a split-A test spectrum while its compound is a split-B train compound. The deposit documents the full cross-tabulation. The deposit contains the split-A formula split, row-level formula/fold map, test and validation spectra tables and two voting tables; the split-B compound split, query and library tables (plain and gzipped CSV); the frozen split-B protocol arrays and manifest; row metadata; the producer scripts for exact rebuilding; a self-contained verifier; checksums of every file; verification reports with the complete cross-check evidence; the figure-to-split mapping; and provenance/licensing notices. Mass spectra are not part of the split tables: every row is joined to the spectra through source-row pointers into MSnLib.csv, which is hosted with the dataset (560,084 data rows, 752,707,651 bytes, SHA-256 be54b33039cf18675751ea3c6354b592177c899a0cd5df303ae616330133ec07; the UltraMS CSV conversion of the MSnLib original release, Zenodo 10.5281/zenodo.11163380, CC BY 4.0). The dataset also hosts MSnLib_candidates.json (1,368,264,026 bytes, SHA-256 125a1cdcc57a3fcef8cc76d5322ac0c434cd09b0bfb9e0c9d28bddf052f45fcc), the candidate SMILES lists behind the single-spectrum candidate-ranking task (9,886 query entries). The package content is released under CC BY 4.0 with attribution to MSnLib; the two UltraMS producer scripts are Apache-2.0.