Proteomics · InstaDeepAI
winnow-ms-datasets
Tandem mass spectrometry datasets and peptide predictions for calibration and FDR estimation in de novo sequencing.
What winnow-ms-datasets is
winnow-ms-datasets is a collection of tandem mass spectrometry (MS/MS) datasets and corresponding peptide sequence predictions, published by InstaDeep. It is designed to train and benchmark Winnow, a tool for confidence calibration, rescoring, and label-free false discovery rate (FDR) control in de novo peptide sequencing.
De novo peptide sequencing predicts amino acid sequences directly from MS/MS spectra without relying on predefined protein reference databases. However, assessing the statistical reliability of these predictions is challenging without robust confidence calibration and FDR control. This resource provides curated spectrum metadata, experimental inputs, and aligned de novo sequencer outputs to standardise the evaluation of confidence scoring across diverse biological samples and organisms.
How it was made
The collection aggregates experimental spectra from public repositories (including PRIDE accessions and Figshare) covering diverse biological matrices, such as human cancer cell lines, plant tissues, snake venom, immunopeptidomes, and microbial samples. Datasets are partitioned into distinct search spaces: 'full' or 'raw' representing all experimental spectra, 'labelled' containing spectra validated via database search labels, 'raw_less_train' retaining unlabelled spectra alongside held-out labelled spectra, and 'unlabelled' isolating spectra lacking database identifications.
Spectra and associated metadata are stored in Apache Parquet format. Alongside the raw spectra, the repository supplies peptide predictions generated by the InstaNovo de novo sequencing model, as well as predictions from Casanovo (mzTab format) and pi-PrimeNovo (TSV format) for selected benchmarks. Reference FASTA proteomes for Homo sapiens, Arabidopsis thaliana, Caenorhabditis elegans, and Escherichia coli are also provided.
Key facts
- Record count
- Between 1 million and 10 million rows
- Primary data formats
- Parquet, CSV, TSV, mzTab, FASTA
- Data modalities
- Tabular, text
- Associated sequencers
- InstaNovo, Casanovo, pi-PrimeNovo
- Preprint identifier
- arXiv:2509.24952
- Digital Object Identifier
- 10.57967/hf/6610
What you can use it for
- Confidence score calibration
- Train machine learning models such as Winnow to map de novo peptide prediction scores into well-calibrated probabilities.
- Label-free FDR estimation
- Benchmark methods that control false discovery rates on de novo peptide sequencing outputs without traditional target-decoy database matches.
- Multi-engine de novo benchmarking
- Compare prediction outputs across different de novo algorithms—such as InstaNovo, Casanovo, and pi-PrimeNovo—on standardised HeLa single-shot datasets.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from InstaDeepAI/winnow-ms-datasets at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 384,643 rows of default/train
collision_energy
float- min
- 27
- median
- 27
- mean
- 27
- max
- 27
experiment_name
string_label- 20190601_QX6_JoMu_SA_uPac200cm_HepG2_f211%
- 20190601_QX6_JoMu_SA_uPac200cm_HepG2_f811%
- 20190601_QX6_JoMu_SA_uPac200cm_HepG2_f111%
frag_type
string_label- HCD100%
intensity_array
list- min
- 5
- median
- 157
- mean
- 185.69
- max
- 800
mz_array
list- min
- 5
- median
- 157
- mean
- 185.69
- max
- 800
precursor_charge
int- min
- 1
- median
- 2
- mean
- 2.55
- max
- 7
precursor_mz
float- min
- 299.49
- median
- 562.60
- mean
- 592.11
- max
- 1669.81
retention_time
float- min
- 0.08
- median
- 2687.81
- mean
- 3558.93
- max
- 10800.08
scan_number
int- min
- 4
- median
- 37,223
- mean
- 69071.96
- max
- 200k
sequence
string_text- min
- 6
- median
- 14
- mean
- 16.29
- max
- 112
source
string_label- hepg255%
- gluc29%
- tplantibodies5%
spectrum_id
string_text- min
- 28
- median
- 46
- mean
- 51.52
- max
- 75
Strengths and limits
Good at
- Covers a broad spectrum of biological systems, including bacterial, plant, human tissue, snake venom, and immunopeptidomic samples.
- Includes precomputed predictions from multiple de novo sequencing tools, facilitating direct comparative evaluations.
- Provides explicit data splits categorised by database labelling status to support supervised training and realistic open-search validation.
Watch out for
- Multi-sequencer predictions covering Casanovo and pi-PrimeNovo are provided for the HeLa single-shot dataset, while other benchmarks focus primarily on InstaNovo outputs.
- Requires high-throughput tabular data libraries such as Polars, Dask, or PyArrow to manage and query large Parquet partitions efficiently.
Questions
Can I look inside winnow-ms-datasets without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is winnow-ms-datasets under?
Its authors declare other. Read the terms on the original page before reusing the data.
What is the difference between the full, labelled, and unlabelled splits?
'Full' contains all spectra acquired in an experiment. 'Labelled' is restricted to spectra confirmed via database searching, whereas 'unlabelled' contains the remaining spectra where no database match was assigned.
Which sequencing models have predictions available in this dataset?
InstaNovo beam search predictions are provided across the training and evaluation splits, with additional Casanovo and pi-PrimeNovo predictions available for the HeLa single-shot (helaqc) dataset.
Which organisms are represented in the reference proteomes?
FASTA reference sequences are included for Arabidopsis thaliana, Caenorhabditis elegans, Escherichia coli (ecoli_zorya), and Homo sapiens.
How to cite it
Mabona, A., Daniel, J., Knudsen, H. S. J., Catzel, R., Eloff, K. M., Schoof, E. M., Lopez Carranza, N., Jenkins, T. P., Van Goey, J., and Kalogeropoulos, K. (2025). De novo peptide sequencing rescoring and FDR estimation with Winnow. arXiv preprint arXiv:2509.24952.
Written from winnow-ms-datasets's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.