Winnow General Probability Calibrator

Winnow recalibrates confidence scores and provides FDR control for de novo peptide sequencing (DNS) workflows.
This repository hosts a pretrained, general-purpose calibrator that maps raw InstaNovo model confidences and complementary features (mass error, retention time, beam features, fragment matching features) to well-calibrated probabilities.

Intended inputs: spectrum input data and corresponding MS/MS PSM results produced by InstaNovo
Outputs: rescored and calibrated per-PSM probabilities in calibrated_confidence with de novo FDR control.

What’s inside

model.safetensors: trained classifier
config.json: classifier hyperparameter settings

How to use

Python

from pathlib import Path
from huggingface_hub import snapshot_download
from winnow.calibration.calibrator import ProbabilityCalibrator
from winnow.datasets.data_loaders import InstaNovoDatasetLoader
from winnow.scripts.main import filter_dataset
from winnow.fdr.nonparametric import NonParametricFDRControl

# 1) Download model files
general_model = Path("general_model")
snapshot_download(
    repo_id="InstaDeepAI/winnow-general-model",
    repo_type="model",
    local_dir=general_model,
)

# 2) Load calibrator
calibrator = ProbabilityCalibrator.load(pretrained_model_name_or_path=general_model)

# 3) Load your dataset (InstaNovo-style config)
dataset = InstaNovoDatasetLoader().load(
    data_path="path_to_spectrum_data.parquet",
    predictions_path="path_to_instanovo_predictions.csv",
)
dataset = filter_dataset(dataset)  # standard Winnow filtering

# 4) Predict calibrated confidences
calibrator.predict(dataset)  # adds dataset.metadata["calibrated_confidence"]

# 5) Optional: FDR control on calibrated confidence
fdr = NonParametricFDRControl()
fdr.fit(dataset.metadata["calibrated_confidence"])
cutoff = fdr.get_confidence_cutoff(0.05)  # 5% FDR cutoff
dataset.metadata["keep@5%"] = dataset.metadata["calibrated_confidence"] >= cutoff

CLI

# After `pip install winnow`
winnow predict \
  data_loader=instanovo \
  dataset.spectrum_path_or_directory=my_data.parquet \
  dataset.predictions_path=my_preds.csv \
  calibrator.pretrained_model_name_or_path=config_with_dataset_paths.yaml \
  fdr_control.fdr_threshold=0.05 \
  output_folder=outputs

Inputs and outputs

Required columns for calibration:

Spectrum data (parquet\ipc\mgf)
- spectrum_id (string): unique spectrum identifier
- experiment_name (string): MS run identifier
- retention_time (float): retention time (seconds)
- precursor_charge (float): charge of the precursor ion (from MS1)
- precursor_mz (float): mass-to-charge of the precursor ion (from MS1)
- mz_array (list[float]): mass-to-charge values of the MS2 spectrum
- intensity_array (list[float]): intensity values of the MS2 spectrum
Beam predictions (csv)
- spectrum_id (string)
- predictions (string): top prediction, untokenised sequence
- predictions_tokenised (string): comma‐separated tokens for the top prediction
- log_probability (float): top prediction log probability
- token_log_probabilities (list[float]): per-token log-probabilities for the top prediction
- predictions_beam_k (string): untokenised sequence for beam k (k≥0)
- log_probability_beam_k (float)
- token_log_probabilities_k (string/list-encoded)

Outputs:

metadata.csv: spectrum metadata and computed features. Contains everything except the prediction and FDR columns, i.e.:
- spectrum_id, experiment_name, precursor_mz, precursor_charge, retention_time, etc. (all pass-through spectrum columns) All computed feature columns, including intermediate results (mass_error_da, irt_error, ion_matches, margin, etc.)
preds_and_fdr_metrics.csv: predictions and FDR results. Always contains:
- spectrum_id
- prediction
- calibrated_confidence: calibrated probability
- psm_fdr
- psm_q_value
- Optional: psm_pep

Training data

The general model was trained on a pooled, labelled set spanning multiple public datasets to encourage cross-dataset generalisation:
- HeLa single-shot (PXD044934)
- Candidatus Scalindua Brodae (PXD044934)
- Wound exudates (PXD025748)
- HepG2 (PXD019483)
- Immunopeptidomics (PXD006939)
- HeLa degradome (PXD044934)
- Snake venoms (PXD036161)
- Therapeutic nanobodies (PXD044934)
This model uses fragment match features, iRT features, beam features, token score features and the mass error feature.
Predictions were obtained using InstaNovo v1.2.0 with beam search set to 5 beams.

Citation

If you use winnow in your research, please cite our preprint: De novo peptide sequencing rescoring and FDR estimation with Winnow

@article{mabona2025novopeptidesequencingrescoring,
      title         = {De novo peptide sequencing rescoring and FDR estimation with Winnow},
      author        = {Amandla Mabona and Jemma Daniel and Henrik Servais Janssen Knudsen and
                       Rachel Catzel and Kevin Michael Eloff and Erwin M. Schoof and Nicolas
                       Lopez Carranza and Timothy P. Jenkins and Jeroen Van Goey and
                       Konstantinos Kalogeropoulos},
      year          = {2025},
      eprint        = {2509.24952},
      archivePrefix = {arXiv},
      primaryClass  = {q-bio.QM},
      url           = {https://arxiv.org/abs/2509.24952},
}

If you use this pretrained calibrator, please cite:

@misc{instadeep_ltd_2025,
    author       = { InstaDeep Ltd },
    title        = { winnow-general-model (Revision ef3002daf2254369d04095731c76a022553ba63a) },
    year         = 2026,
    url          = { https://huggingface.co/InstaDeepAI/winnow-general-model },
    doi          = { 10.57967/hf/6611 },
    publisher    = { Hugging Face }
}

If you use the InstaNovo model to generate predictions, please also cite: InstaNovo enables diffusion-powered de novo peptide sequencing in large-scale proteomics experiments

@article{eloff_kalogeropoulos_2025_instanovo,
        title        = {InstaNovo enables diffusion-powered de novo peptide sequencing in large-scale
                        proteomics experiments},
        author       = {Eloff, Kevin and Kalogeropoulos, Konstantinos and Mabona, Amandla and Morell,
                        Oliver and Catzel, Rachel and Rivera-de-Torre, Esperanza and Berg Jespersen,
                        Jakob and Williams, Wesley and van Beljouw, Sam P. B. and Skwark, Marcin J.
                        and Laustsen, Andreas Hougaard and Brouns, Stan J. J. and Ljungars,
                        Anne and Schoof, Erwin M. and Van Goey, Jeroen and auf dem Keller, Ulrich and
                        Beguir, Karim and Lopez Carranza, Nicolas and Jenkins, Timothy P.},
        year         = 2025,
        month        = {Mar},
        day          = 31,
        journal      = {Nature Machine Intelligence},
        doi          = {10.1038/s42256-025-01019-5},
        issn         = {2522-5839},
        url          = {https://doi.org/10.1038/s42256-025-01019-5}
}