De novo peptide sequencing rescoring and FDR estimation with Winnow

preprint · arXiv · 2025

preprint · arXiv · 2025. Amandla Mabona et al. Machine learning has markedly advanced de novo peptide sequencing (DNS) for mass spectrometry-based…
Date 2025-09-29
Type preprint
Venue arXiv
Publisher arXiv
Contribution post-processor
DOI 10.48550/arXiv.2509.24952
Citations (OpenAlex) 0

Abstract

Machine learning has markedly advanced de novo peptide sequencing (DNS) for mass spectrometry-based proteomics. DNS tools offer a reliable way to identify peptides without relying on reference databases, extending proteomic analysis and unlocking applications into less-charted regions of the proteome. However, they still face a key limitation. DNS tools lack principled methods for estimating false discovery rates (FDR) and instead rely on model-specific confidence scores that are often miscalibrated. This limits trust in results, hinders cross-model comparisons and reduces validation success. Here we present Winnow, a model-agnostic framework for estimating FDR from calibrated DNS outputs. Winnow maps raw model scores to calibrated confidences using a neural network trained on peptide-spectrum match (PSM)-derived features. From these calibrated scores, Winnow computes PSM-specific error metrics and an experiment-wide FDR estimate using a novel decoy-free FDR estimator. It supports both zero-shot and dataset-specific calibration, enabling flexible application via direct inference, fine-tuning, or training a custom model. We demonstrate that, when applied to InstaNovo predictions, Winnow’s calibrator improves recall at fixed FDR thresholds, and its FDR estimator tracks true error rates when benchmarked against reference proteomes and database search. Winnow ensures accurate FDR control across datasets, helping unlock the full potential of DNS.

Authors

  1. Amandla Mabona · InstaDeep Ltd
  2. Jemma Daniel · InstaDeep Ltd
  3. Henrik Servais Janssen Knudsen · Technical University of Denmark
  4. Rachel Catzel · InstaDeep Ltd
  5. Kevin Eloff · InstaDeep Ltd
  6. Erwin M. Schoof · Technical University of Denmark
  7. Nicolas Lopez Carranza · InstaDeep Ltd
  8. Timothy P. Jenkins · Technical University of Denmark
  9. Jeroen Van Goey · InstaDeep Ltd
  10. Konstantinos Kalogeropoulos · Delft University of Technology, Kavli Institute of Nanoscience, Technical University of Denmark

Methods and tools

  • Winnow: NN rescoring + decoy-free FDR

Data used

  • Analysis outputs (as deposited) · 10.6084/m9.figshare.30147601
  • High-throughput MS-based immunopeptidomics (as deposited) · PXD006939
  • High-throughput proteomics of the 26 medically most important elapids and vipers from sub-Saharan Africa (as deposited) · PXD036161
  • Supplementary Data for “Comprehensive evaluation of peptide de novo sequencing tools for monoclonal antibody assembly” (as deposited) · 10.6084/m9.figshare.21394143
  • Systematic HLA Epitope Ranking Pan Algorithm (SHERPA) (as deposited) · PXD023064
  • The Proteome Landscape of the Kingdoms of Life (as deposited) · PXD014877, PXD019483
  • Winnow MS datasets (version not stated)

Cited by (4)

Seen in the charts

Back to the full map

Back to top