Gaolaboratory/MSConsensus-100M

MSConsensus

A hundred-million-scale, batch-effect-suppressed dataset and benchmark for proteomics machine learning.

MSConsensus holds 110,209,043 consensus MS/MS spectra. They were built from 1.01 PB of public raw mass-spectrometry data taken from 1,500 PRIDE repositories and covering all major Orbitrap and timsTOF platforms. A new consensus algorithm, Weighted-Score Binning (WSBIN), produces several consensus spectra per precursor instead of one representative spectrum. This suppresses instrument-, lab- and protocol-specific batch effects while keeping biologically meaningful intensity structure and intra-cluster variance.

This repository (Gaolaboratory/MSConsensus-100M) holds a 102,000,000-spectrum ML-ready release in Parquet: 100M train, 1M validation and 1M test spectra, about 190 GB compressed. Each row is one consensus spectrum with its peptide label and precursor charge. The full 110M corpus, with per-peak annotations, provenance metadata and msz files, is on Harvard Dataverse.

The full MSConsensus release ships with:

• msz, a binary container with random access. Files are 45% of mzML size and load at 25,378 random-access spectra/s (3.2× HDF5).
• MSDatasets, a drop-in PyTorch / JAX data loader.
• A three-task benchmark and public leaderboard covering spectrum embedding retrieval, fragment-intensity prediction and de novo peptide identification.

Primary hosting	Harvard Dataverse — doi:10.7910/DVN/NUCT4N
Leaderboard	msconsensus-leaderboard.netlify.app
Paper	MSConsensus: A Hundred-Million-Scale, Batch-Effect–Suppressed Dataset and Benchmark for Proteomics Machine Learning (under review, NeurIPS 2026)
License	CC BY 4.0

---

Dataset Summary

Proteomics machine learning faces three linked bottlenecks:

1. Noise and batch effects. Models trained on raw PRIDE spectra learn the instrument as easily as they learn the peptide, so cross-vendor generalization fails.
2. I/O bottlenecks. mzML (XML), MGF (text) and HDF5 were designed for analysis pipelines, not for shuffled GPU training.
3. Benchmark fragmentation. Each method ships its own splits, preprocessing and metrics, so results from different papers can't be compared.

MSConsensus addresses all three with one resource: a curated consensus corpus, an ML-native format and loader, and standardized held-out splits with a shared evaluation suite.

Comparison with existing resources

Resource	# Spectra	Batch-split	Per-peak annotation	PTM split	Data loader
MassIVE-KB v2	5.9 M	partial	✗	✗	✗
NIST Hybrid Libs	1.3 M	✗	partial	partial	✗
ProteomeTools	0.9 M	✗	✓	✓ (synth.)	✗
PRIDE Cluster	28 M	✗	✗	✗	✗
9-species (Casanovo)	28 M	✗	partial	✗	✗
AlphaPeptDeep train	~25 M	✗	✓	✓	✗
MSConsensus	110 M	✓ (WSBIN)	✓	✓ (8 PTMs)	✓ (msz)

MSConsensus is 19–55× larger than existing consensus libraries. It emits 5–100 consensus spectra per cluster, which keeps the intra-cluster variance that contrastive and representation-learning objectives need.

---

Supported Tasks and Leaderboards

The benchmark covers three core tasks, evaluated on identical held-out splits. The leaderboard lives at msconsensus-leaderboard.netlify.app. To submit a model, open a pull request to the leaderboard repository with a runnable evaluation script and a publicly hosted checkpoint. Submissions are merged once the reported numbers have been reproduced independently.

Task 1: Spectral embedding retrieval

Given a query spectrum, retrieve library spectra with the same (peptide, charge). Metrics: top-K accuracy at the (peptide, charge) level, and MRR.

Spec2Vec, with the library swapped from MassIVE-KB v2 to MSConsensus (held-out query set of 2,699 spectra):

Training library	Top-1	Top-5	Top-10	Top-20	Top-50
MassIVE-KB v2	39.20	55.39	60.65	64.32	70.17
MSConsensus	46.65	60.91	64.10	67.69	72.29
Δ (rel.)	+19.0%	+10.0%	+5.7%	+5.2%	+3.0%

On the consensus holdout (10,000 queries against a 100,000-spectrum library), GLEAMS beats Spec2Vec by +2.68 points top-1 and +0.0286 MRR.

Task 2: Fragment-intensity prediction

Predict fragment-ion intensities for a (peptide, charge, collision energy) input. Metrics: median spectral angle similarity (SAS, higher is better) and spectral angle (SA, lower is better) after L2 normalization.

Model	Paper holdout SAS ↑ / SA ↓	PXD053296 SAS ↑ / SA ↓	Consensus holdout SAS ↑ / SA ↓
Prosit	0.9147 / 0.0853	0.6964 / 0.3036	0.6748 / 0.3252
MS²PIP (HCD2021)	0.8255 / 0.1745	0.7954 / 0.2046	0.6714 / 0.3286

Prosit retrained on a matched 30M-spectrum MSConsensus subsample (consensus_30m_retrain), with the architecture and hyperparameters unchanged:

Variant	Paper holdout SAS	PXD053296 SAS	Consensus holdout SAS
Paper original	0.9147	0.6964	0.6748
consensus_30m_retrain	0.7425	0.8137	0.7317

PXD053296 has 30 raw files first released in 2025, so none of the evaluated models saw it during training.

Task 3: De novo peptide identification

Predict the peptide sequence directly from a spectrum. Metrics: amino-acid-level precision, recall and F1.

Model	AA Precision ↑	AA Recall ↑	AA F1 ↑
InstaNovo	0.8913	0.8923	0.8918
Casanovo	0.9418	0.6195	0.7474

---

Dataset Structure

Data instances

Each row is one consensus MS/MS spectrum made by WSBIN from a cluster of identified spectra that share a (peptide sequence, charge, fixed modifications) key. Here is an example from the validation split, with the arrays truncated:

{
  "mz":        [101.071, 102.055, 104.053, ...],        # 256 peaks, sorted ascending
  "intensity": [1513.7007, 601.9307, 3641.5234, ...],   # 256 values, aligned with mz
  "charge":    2,
  "peptide":   "MSIFGHSM[15.9949]GGHGALIC[57.0215]ALK"
}

Data fields

Field	Type	Description
mz	list<float32>	Consensus fragment-peak m/z values in Da, sorted ascending. Each value is the intensity-weighted mean m/z of the contributing peaks in a WSBIN bin.
intensity	list<float32>	Consensus peak intensities, the same length as mz and aligned with it. Each value is a trimmed-mean intensity scaled by the bin's weighted support. Values are not normalized, so they stay on the original detector scale (e.g., 10³–10⁶). Normalize them yourself, for example to unit max or with L2 or sqrt.
charge	int16	Precursor charge state. It ranges from 1 to 7; most spectra are 2+ or 3+ (mean 2.66).
peptide	string	Peptide sequence. Modifications are written inline as bracketed monoisotopic mass shifts, e.g. C[57.0215] for carbamidomethyl and M[15.9949] for oxidation. Sequences are 7–115 characters long, with a median of 17.

None of the columns are nullable. Spectra contain between 1 and 5,505 peaks, with a median of 207 and a mean of 257.

Available only in the full Dataverse release, not in this repository:

• Per-peak annotations: ion series (b / y / a / immonium / internal), fragment charge, chemical formula and confidence tier. T1 means the peak matches theoretical m/z within 5 ppm and is observed in ≥80% of contributing PSMs. T2 means it meets only one of these conditions. T3 is an unannotated peak kept above the WSBIN threshold τ.
• Provenance metadata: instrument and vendor, species, source PXD accession(s), per-project license header, and ethical-approval references.
• msz files, with MS level and retention time indexed in the footer.

Data splits

Split	Rows	Shards	Size on disk
train	100,000,000	400	186.7 GB
validation	1,000,000	4	1.8 GB
test	1,000,000	4	1.9 GB
Total	102,000,000	408	190.4 GB

Each shard is a ZSTD-compressed Parquet file with 250,000 rows. The splits were built from the 90 source consensus files (consensus_00.parquet to consensus_89.parquet). The build used seed 42, validation and test fractions of 1% each, 256 shuffle buckets for train and 4 for validation and test. manifest.json records the build parameters, a fingerprint of the source files, and the row count and SHA-256 of every shard, so you can verify downloads.

The paper also describes these held-out evaluation splits, which are distributed through MSDatasets and Dataverse:

• Held-out splits by instrument, species and modification (8 PTMs), plus vendor-stratified splits for cross-vendor evaluation (e.g., Orbitrap → timsTOF transfer).
• Peptide-level disjointness, enforced with 9-mer hashing.
• A consensus holdout of about 4M deduplicated spectra, used for the benchmark results reported above.

File formats

Format	Size relative to mzML	Sequential (spectra/s)	Random (spectra/s)
msz (MSDatasets)	0.45	29,134 ± 275	25,378 ± 126
mzML (MSDatasets)	1.00	9,307 ± 30	12,144 ± 411
Parquet (PyArrow)	0.46	1,672 ± 157	2,445 ± 218
MGF (Pyteomics)	2.04	1,518 ± 23	1,593 ± 9
HDF5 (h5py)	0.46	43,815 ± 463	7,896 ± 99

Benchmark setup: dual Intel Xeon Gold 5318Y, RAID 5 SSD array, batch size 128, 8 data-loader workers.

msz splits each file into three independently ZSTD-compressed block streams (XML metadata, m/z and intensity) behind a fixed header. A footer indexes spectrum offsets, MS levels and retention times. Encoding and decoding are lossless: all metadata and full floating-point precision are kept. Optional precision reduction and delta encoding are available for smaller files.

---

Usage

With 🤗 Datasets

from datasets import load_dataset

# Stream to avoid downloading about 190 GB up front
ds = load_dataset("Gaolaboratory/MSConsensus-100M", split="train", streaming=True)
ex = next(iter(ds))
print(ex["peptide"], ex["charge"], len(ex["mz"]))

# The evaluation splits are small enough to download in full
val = load_dataset("Gaolaboratory/MSConsensus-100M", split="validation")

---

Dataset Creation

Curation rationale

Public PRIDE data is dominated by per-instrument, per-lab and per-protocol systematic effects. Existing spectral libraries collapse each peptide-charge cluster into one representative spectrum. That suits search-engine identification, but it removes the intra-cluster variance that representation learning needs, and it doesn't quantify or isolate batch effects. MSConsensus is designed to suppress batch effects while keeping that variance.

Source data

1.01 PB of raw vendor files (Thermo .raw and Bruker timsTOF .d) from 1,500 PRIDE repositories. A repository was included only if it met all of these criteria:

1. At least 20 raw files per repository, so clusters can be averaged.
2. At least 100 proteins identified per file under a uniform Sage reanalysis.
3. Together, the repositories represent all major Orbitrap and timsTOF platforms (>90% of PRIDE submissions).

The exact PXD identifier list and MD5 manifests are released with the dataset.

Processing pipeline

1. Uniform reanalysis: a Sage + MSFragger union search at 1% PSM-level FDR.
2. Clustering by (peptide sequence, charge, fixed modifications).
3. WSBIN consensus aggregation. For each cluster C, draw K stochastic subsets of size ⌈f·|C|⌉. Bin the m/z axis at width w, then weight each bin by the quality-weighted count of contributing spectra. Compute the intensity-weighted mean m/z and the trimmed-mean intensity, and drop bins whose weight falls below the threshold τ (the explicit batch-effect suppressor). Each subset produces one consensus spectrum.
   • K ∈ {5, 20, 50, 100}, set by cluster size.
   • A tiered ratio regime runs from 100:1 for small clusters (keeps rare PTM evidence) to 5000:1 for the largest HeLa clusters (maximal denoising).
4. Per-peak annotation with ion series, fragment charge, chemical formula and confidence tier.
5. Encoding into msz with MSCompress, plus Parquet export.

Annotations: ground-truth criteria

A cluster's peptide-spectrum assignment is treated as ground truth only if all of the following hold:

• (a) At least 3 independent PSMs from at least 2 different PRIDE projects.
• (b) Sage and MSFragger agree on the peptide sequence at 1% FDR.
• (c) Precursor mass error is below 5 ppm in at least 80% of contributing PSMs.

Annotations are machine-generated by database search engines. No manual labeling was done.

Validation of batch-effect suppression

Five held-out PRIDE repositories were used: PXD003095, PXD009519, PXD009895, PXD010382 and PXD012715. They were searched with Comet, MetaMorpheus, MSFragger and Sage. Replacing raw spectra with consensus spectra from matching (peptide, charge) clusters had these effects:

• SNR rose by +3.86 dB on average (range +0.57 to +10.47 dB, p < 0.001).
• Protein, peptide and PSM identification counts stayed unchanged or slightly higher.
• In a sparse-PCA + UMAP projection, dataset-specific clusters collapsed into a largely overlapping distribution.

SNR is defined as 10·log₁₀(Σ_{b∈A} I_b² / Σ_{b∈N} I_b²) dB, where A is the set of annotated b/y-ion peaks and N is the set of all other peaks above the detector threshold.

---

Considerations for Using the Data

Known limitations and biases

• Vendor imbalance. Thermo Orbitrap instruments make up about 71% of the corpus, Bruker timsTOF about 22% and other vendors about 7%. Use the vendor-stratified splits in MSDatasets for fair cross-vendor evaluation.
• Consensus truth, not absolute truth. Labels come from database search engines (Sage, MSFragger), so they inherit those tools' limitations. Filtering was strict, but at this scale a marginal fraction of false-positive spectra may remain.
• In-distribution scores may drop. Models trained on MSConsensus can score lower on benchmarks built from their original training distribution (e.g., retrained Prosit fell from 0.9147 to 0.7425 SAS on its paper holdout) while generalizing better to unseen instruments. The authors attribute this to the removal of batch-effect shortcuts.
• Train/validation/test splits in this repository are random. They were drawn with 1% fractions and a fixed seed, and the build manifest doesn't record peptide-level disjointness. The same peptide may appear in more than one split. For generalization claims, use the held-out benchmark splits described in the paper.
• Instrument coverage follows what is deposited in PRIDE. Rare platforms and acquisition modes are underrepresented.

---

Licensing Information

MSConsensus is released under CC BY 4.0. Every ingested PRIDE submission is licensed CC BY 4.0 or CC0. Each project's license header is kept in the per-spectrum metadata, so you can build license-compliant subsets and attribute the original depositors.

Citation

@inproceedings{msconsensus2026,
  title     = {{MSConsensus}: A Hundred-Million-Scale, Batch-Effect--Suppressed Dataset and Benchmark for Proteomics Machine Learning},
  author    = {Anonymous},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
  note      = {Under review},
  doi       = {10.7910/DVN/NUCT4N}
}