Datasets:

dsadd4
/

UltraMS-MSnLib-Benchmark

# Figure → split mapping

This package contains the two frozen MSnLib splits behind the UltraMS manuscript. They are
different objects and answer different questions:

* **Split A — UltraMS MSnLib formula-level split** (`split_A_full_msnlib/`): the complete
  MSnLib library (560,084 spectra, all adducts) split by molecular formula into 16,885 /
  2,110 / 2,112 formulas and 446,667 / 55,980 / 57,437 spectra (train / validation / test).
  Formula-disjoint, not molecule-disjoint. Test fold: 57,437 spectra, 2,112 formulas,
  5,003 source SMILES / 4,889 canonical molecules.
* **Split B — strict-adduct molecule-level split** (`splits/`, `queries/`, `libraries/`):
  the strict `[M+H]+` and `[M-H]-` subsets (447,956 source rows) 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. Molecule-disjoint within each
  polarity. Test tasks: 6,202 positive and 3,193 negative queries against 295,434 and
  137,000 library spectra.

## 1. Panel mapping

| figure panel | split | evaluation population / task |
| --- | --- | --- |
| Figure 2b — ionization-polarity recognition | **A** | polarity prediction over the split-A evaluation population (A test fold; A train / validation used for fitting and reference). |
| Figure 2j — peak-token embedding chemical atlas | **A** | peak-token embedding visualization/analysis over the split-A test fold with split-A reference spectra; colours are chemical properties, not a model output. |
| Figure 3a — single-spectrum candidate ranking (Hit@1/5/20) | **A** | 57,437 split-A test spectra as queries, ranked against the per-formula candidate tables used by the manuscript (candidate ranking tables are part of the UltraMS repository, not of this split package). |
| Figure 3b — score/rank density and cumulative Top-k | **A** | same 57,437 split-A test queries and their candidate rankings. |
| Figure 3d — multiple-spectrum voting | **A** | 57,437 split-A test spectra grouped into the **official voting units: 5,002 MSnLib source-SMILES groups covering 57,436 spectra** (`split_A_full_msnlib/test_molecules_voting.csv`); the alternative canonical-isomeric-SMILES view (4,888 groups, same 57,436 spectra) is `split_A_full_msnlib/test_molecules_voting_canonical_view.csv`; the retention times used by this analysis are `split_A_full_msnlib/msnlib_rt.csv.gz` (`row_index, rt`). See the parent `README.md` §4.3/§4.5. |
| Figure 3g — empirical rank CDF | **A** | 57,437 split-A test queries. |
| Figure 3a_sp_4 — label-blind consensus grouping | **A** | split-A test fold, label-blind grouping of same-molecule spectra. |
| Figure 3a_sp_5 — promoted MSnLib fingerprint analysis | **A** | split-A fingerprint-prediction evaluation population (split-A train / validation for fitting). |
| baselines — single-spectrum Top-1 table | **A** | 57,437 split-A test queries, one ranking per method. |
| Figure 3e — strict-adduct library search, FDR–recall | **B** | strict `[M+H]+`: 6,202 test queries vs 295,434 library spectra (recall denominator 6,201); strict `[M-H]-`: 3,193 vs 137,000 (denominator 3,160). Validation tasks: 6,202 / 250,424 and 3,193 / 116,135. |
| Figure 3f — cross-polarity molecular retrieval | **B** | molecule-level split-B compounds; one polarity's spectra query the other polarity's library. |
| Figure 3h — seven-method strict-adduct FDR–recall (unused figure) | **B** | same split-B query / library sets as Figure 3e. |
| Figure 3h_sp_1 — FDR threshold transfer | **B** | split-B validation/test query identities; 200 paired molecule-disjoint calibration/held-out assignments over those identities. |
| cross-collision-energy search | **B** | split-B test queries that have at least two collision energies: **6,161** positive and **2,999** negative cross-CE queries (from 6,202 / 3,193 split-B test queries). |
| NPLIB1 and MassSpecGym comparison panels | **not this dataset** | NPLIB1 (3,699 test molecules) and MassSpecGym are independent external benchmarks with their own splits and libraries; they are not derived from MSnLib and are not covered by this package. |
| Figure 3c / Figure 3i case studies | **neither** | case-study panels use specific MSnLib spectra and MAGMa peak annotations; they are not defined by either split table. |

## 2. Cross-split inconsistency: split A (formulas) vs split B (compounds)

The two splits are built on different units (formula vs canonical molecule) and different
populations (all adducts vs strict `[M+H]+`/`[M-H]-`), so their fold labels disagree for a
large share of compounds. The tables below are computed by mapping every split-B compound
to its molecular formula through `msnlib_csv_data_row` → `split_A_full_msnlib/row_formula_fold.csv`.
Compound counts are split-B compounds (a compound belongs to one polarity and one split-B
fold).

### `[M+H]+` — 41,348 compounds

| split B ↓ / split A → | train | validation | test | total |
| --- | ---: | ---: | ---: | ---: |
| **train** | 23,059 | 2,901 | 2,984 | 28,944 |
| **validation** | 4,924 | 647 | 631 | 6,202 |
| **test** | 4,922 | 626 | **654** | 6,202 |
| total | 32,905 | 4,174 | 4,269 | 41,348 |

Same-fold compounds: **24,360 / 41,348 (58.9 %)**.
Of the 6,202 split-B positive test compounds only **654 (10.5 %)** are split-A test
compounds; 4,922 (79.4 %) are split-A train compounds.

### `[M-H]-` — 21,293 compounds

| split B ↓ / split A → | train | validation | test | total |
| --- | ---: | ---: | ---: | ---: |
| **train** | 12,035 | 1,447 | 1,425 | 14,907 |
| **validation** | 2,552 | 307 | 334 | 3,193 |
| **test** | 2,541 | 349 | **303** | 3,193 |
| total | 17,128 | 2,103 | 2,062 | 21,293 |

Same-fold compounds: **12,645 / 21,293 (59.4 %)**.
Of the 3,193 split-B negative test compounds only **303 (9.5 %)** are split-A test
compounds.

### Spectra-level view

The split-B tables cover all 441,829 preprocessed strict-adduct spectra (301,636 positive +
140,193 negative); each maps to exactly one split-A fold:

`[M+H]+` — 301,636 preprocessed spectra:

| split B ↓ / split A → | train | validation | test | total |
| --- | ---: | ---: | ---: | ---: |
| **train** | 168,804 | 20,917 | 22,024 | 211,745 |
| **validation** | 35,535 | 4,604 | 4,742 | 44,881 |
| **test** | 35,716 | 4,490 | **4,804** | 45,010 |
| total | 240,055 | 30,011 | 31,570 | 301,636 |

`[M-H]-` — 140,193 preprocessed spectra:

| split B ↓ / split A → | train | validation | test | total |
| --- | ---: | ---: | ---: | ---: |
| **train** | 79,644 | 9,523 | 9,234 | 98,401 |
| **validation** | 16,760 | 1,981 | 2,186 | 20,927 |
| **test** | 16,609 | 2,263 | **1,993** | 20,865 |
| total | 113,013 | 13,767 | 13,413 | 140,193 |

The spectra-level matrix is computed from the frozen `protocol/` row metadata (each
preprocessed row → its compound → its split-B fold) and `split_A_full_msnlib/row_formula_fold.csv`;
the per-split totals equal the split-B `split_spectra_counts` of the frozen manifest
(211,745 / 44,881 / 45,010 positive and 98,401 / 20,927 / 20,865 negative).

(Recompute with `rebuild/verify_release.py`, which rebuilds both matrices from the shipped
tables.)

### Relation of the split-A test fold to split B

* 57,437 split-A test spectra in total;
* **31,570 (55.0 %)** are strict `[M+H]+` spectra of the split-B tables, **13,413 (23.3 %)**
  are strict `[M-H]-` spectra; 44,983 (78.3 %) therefore lie inside split B;
* the remaining 12,454 split-A test spectra carry other adducts (`[M+FA]-`, `[M+Cl]-`,
  `[M+Na]+`, `[M+NH4]+`, `[M+H-H2O]+`, …) and never occur in split B.

### What this means — and what must not be claimed

* Split A is **formula-disjoint only**; split B is **compound-disjoint within each
  polarity**. Each property holds only inside its own split.
* The two splits **cannot be merged into a global-disjointness claim**: about 41 % of the
  split-B compounds are in a different split-A fold, and the same spectrum can be a *test*
  spectrum in split A while belonging to a *train* compound in split B (e.g. 4,922 of the
  6,202 positive split-B test compounds are split-A train compounds).
* A score computed on an split-A test spectrum whose compound is split-B train must not be
  reported as a "held-out molecule" result, and vice versa. Train and evaluate within one
  split only; if a cross-split experiment is unavoidable, state explicitly which split
  defines the training population and which defines the test population.

	# Figure → split mapping

	This package contains the two frozen MSnLib splits behind the UltraMS manuscript. They are
	different objects and answer different questions:

	* Split A — UltraMS MSnLib formula-level split (`split_A_full_msnlib/`): the complete
	MSnLib library (560,084 spectra, all adducts) split by molecular formula into 16,885 /
	2,110 / 2,112 formulas and 446,667 / 55,980 / 57,437 spectra (train / validation / test).
	Formula-disjoint, not molecule-disjoint. Test fold: 57,437 spectra, 2,112 formulas,
	5,003 source SMILES / 4,889 canonical molecules.
	* Split B — strict-adduct molecule-level split (`splits/`, `queries/`, `libraries/`):
	the strict `[M+H]+` and `[M-H]-` subsets (447,956 source rows) 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. Molecule-disjoint within each
	polarity. Test tasks: 6,202 positive and 3,193 negative queries against 295,434 and
	137,000 library spectra.

	## 1. Panel mapping

	\| figure panel \| split \| evaluation population / task \|
	\| --- \| --- \| --- \|
	\| Figure 2b — ionization-polarity recognition \| A \| polarity prediction over the split-A evaluation population (A test fold; A train / validation used for fitting and reference). \|
	\| Figure 2j — peak-token embedding chemical atlas \| A \| peak-token embedding visualization/analysis over the split-A test fold with split-A reference spectra; colours are chemical properties, not a model output. \|
	\| Figure 3a — single-spectrum candidate ranking (Hit@1/5/20) \| A \| 57,437 split-A test spectra as queries, ranked against the per-formula candidate tables used by the manuscript (candidate ranking tables are part of the UltraMS repository, not of this split package). \|
	\| Figure 3b — score/rank density and cumulative Top-k \| A \| same 57,437 split-A test queries and their candidate rankings. \|
	\| Figure 3d — multiple-spectrum voting \| A \| 57,437 split-A test spectra grouped into the official voting units: 5,002 MSnLib source-SMILES groups covering 57,436 spectra (`split_A_full_msnlib/test_molecules_voting.csv`); the alternative canonical-isomeric-SMILES view (4,888 groups, same 57,436 spectra) is `split_A_full_msnlib/test_molecules_voting_canonical_view.csv`; the retention times used by this analysis are `split_A_full_msnlib/msnlib_rt.csv.gz` (`row_index, rt`). See the parent `README.md` §4.3/§4.5. \|
	\| Figure 3g — empirical rank CDF \| A \| 57,437 split-A test queries. \|
	\| Figure 3a_sp_4 — label-blind consensus grouping \| A \| split-A test fold, label-blind grouping of same-molecule spectra. \|
	\| Figure 3a_sp_5 — promoted MSnLib fingerprint analysis \| A \| split-A fingerprint-prediction evaluation population (split-A train / validation for fitting). \|
	\| baselines — single-spectrum Top-1 table \| A \| 57,437 split-A test queries, one ranking per method. \|
	\| Figure 3e — strict-adduct library search, FDR–recall \| B \| strict `[M+H]+`: 6,202 test queries vs 295,434 library spectra (recall denominator 6,201); strict `[M-H]-`: 3,193 vs 137,000 (denominator 3,160). Validation tasks: 6,202 / 250,424 and 3,193 / 116,135. \|
	\| Figure 3f — cross-polarity molecular retrieval \| B \| molecule-level split-B compounds; one polarity's spectra query the other polarity's library. \|
	\| Figure 3h — seven-method strict-adduct FDR–recall (unused figure) \| B \| same split-B query / library sets as Figure 3e. \|
	\| Figure 3h_sp_1 — FDR threshold transfer \| B \| split-B validation/test query identities; 200 paired molecule-disjoint calibration/held-out assignments over those identities. \|
	\| cross-collision-energy search \| B \| split-B test queries that have at least two collision energies: 6,161 positive and 2,999 negative cross-CE queries (from 6,202 / 3,193 split-B test queries). \|
	\| NPLIB1 and MassSpecGym comparison panels \| not this dataset \| NPLIB1 (3,699 test molecules) and MassSpecGym are independent external benchmarks with their own splits and libraries; they are not derived from MSnLib and are not covered by this package. \|
	\| Figure 3c / Figure 3i case studies \| neither \| case-study panels use specific MSnLib spectra and MAGMa peak annotations; they are not defined by either split table. \|

	## 2. Cross-split inconsistency: split A (formulas) vs split B (compounds)

	The two splits are built on different units (formula vs canonical molecule) and different
	populations (all adducts vs strict `[M+H]+`/`[M-H]-`), so their fold labels disagree for a
	large share of compounds. The tables below are computed by mapping every split-B compound
	to its molecular formula through `msnlib_csv_data_row` → `split_A_full_msnlib/row_formula_fold.csv`.
	Compound counts are split-B compounds (a compound belongs to one polarity and one split-B
	fold).

	### `[M+H]+` — 41,348 compounds

	\| split B ↓ / split A → \| train \| validation \| test \| total \|
	\| --- \| ---: \| ---: \| ---: \| ---: \|
	\| train \| 23,059 \| 2,901 \| 2,984 \| 28,944 \|
	\| validation \| 4,924 \| 647 \| 631 \| 6,202 \|
	\| test \| 4,922 \| 626 \| 654 \| 6,202 \|
	\| total \| 32,905 \| 4,174 \| 4,269 \| 41,348 \|

	Same-fold compounds: 24,360 / 41,348 (58.9 %).
	Of the 6,202 split-B positive test compounds only 654 (10.5 %) are split-A test
	compounds; 4,922 (79.4 %) are split-A train compounds.

	### `[M-H]-` — 21,293 compounds

	\| split B ↓ / split A → \| train \| validation \| test \| total \|
	\| --- \| ---: \| ---: \| ---: \| ---: \|
	\| train \| 12,035 \| 1,447 \| 1,425 \| 14,907 \|
	\| validation \| 2,552 \| 307 \| 334 \| 3,193 \|
	\| test \| 2,541 \| 349 \| 303 \| 3,193 \|
	\| total \| 17,128 \| 2,103 \| 2,062 \| 21,293 \|

	Same-fold compounds: 12,645 / 21,293 (59.4 %).
	Of the 3,193 split-B negative test compounds only 303 (9.5 %) are split-A test
	compounds.

	### Spectra-level view

	The split-B tables cover all 441,829 preprocessed strict-adduct spectra (301,636 positive +
	140,193 negative); each maps to exactly one split-A fold:

	`[M+H]+` — 301,636 preprocessed spectra:

	\| split B ↓ / split A → \| train \| validation \| test \| total \|
	\| --- \| ---: \| ---: \| ---: \| ---: \|
	\| train \| 168,804 \| 20,917 \| 22,024 \| 211,745 \|
	\| validation \| 35,535 \| 4,604 \| 4,742 \| 44,881 \|
	\| test \| 35,716 \| 4,490 \| 4,804 \| 45,010 \|
	\| total \| 240,055 \| 30,011 \| 31,570 \| 301,636 \|

	`[M-H]-` — 140,193 preprocessed spectra:

	\| split B ↓ / split A → \| train \| validation \| test \| total \|
	\| --- \| ---: \| ---: \| ---: \| ---: \|
	\| train \| 79,644 \| 9,523 \| 9,234 \| 98,401 \|
	\| validation \| 16,760 \| 1,981 \| 2,186 \| 20,927 \|
	\| test \| 16,609 \| 2,263 \| 1,993 \| 20,865 \|
	\| total \| 113,013 \| 13,767 \| 13,413 \| 140,193 \|

	The spectra-level matrix is computed from the frozen `protocol/` row metadata (each
	preprocessed row → its compound → its split-B fold) and `split_A_full_msnlib/row_formula_fold.csv`;
	the per-split totals equal the split-B `split_spectra_counts` of the frozen manifest
	(211,745 / 44,881 / 45,010 positive and 98,401 / 20,927 / 20,865 negative).

	(Recompute with `rebuild/verify_release.py`, which rebuilds both matrices from the shipped
	tables.)

	### Relation of the split-A test fold to split B

	* 57,437 split-A test spectra in total;
	* 31,570 (55.0 %) are strict `[M+H]+` spectra of the split-B tables, 13,413 (23.3 %)
	are strict `[M-H]-` spectra; 44,983 (78.3 %) therefore lie inside split B;
	* the remaining 12,454 split-A test spectra carry other adducts (`[M+FA]-`, `[M+Cl]-`,
	`[M+Na]+`, `[M+NH4]+`, `[M+H-H2O]+`, …) and never occur in split B.

	### What this means — and what must not be claimed

	* Split A is formula-disjoint only; split B is **compound-disjoint within each
	polarity**. Each property holds only inside its own split.
	* The two splits cannot be merged into a global-disjointness claim: about 41 % of the
	split-B compounds are in a different split-A fold, and the same spectrum can be a test
	spectrum in split A while belonging to a train compound in split B (e.g. 4,922 of the
	6,202 positive split-B test compounds are split-A train compounds).
	* A score computed on an split-A test spectrum whose compound is split-B train must not be
	reported as a "held-out molecule" result, and vice versa. Train and evaluate within one
	split only; if a cross-split experiment is unavoidable, state explicitly which split
	defines the training population and which defines the test population.