Winnow

Winnow maps raw de novo peptide sequencing model scores to calibrated confidences using a neural network trained on peptide-spectrum match-derived features, enabling false discovery rate estimation without decoy databases.

Key Features:

• Confidence Calibration: Winnow maps raw model scores to calibrated confidences using a neural network trained on peptide-spectrum match-derived features.
• Decoy-Free FDR Estimation: Winnow computes peptide-spectrum match-specific error metrics and an experiment-wide false discovery rate estimate using a novel decoy-free FDR estimator.
• Flexible Calibration Modes: Winnow supports both zero-shot and dataset-specific calibration, enabling application via direct inference, fine-tuning, or training a custom model.
• Model-Agnostic Framework: Winnow operates independently of specific de novo peptide sequencing tools and can be applied across different models.

Scientific Applications:

• De Novo Peptide Sequencing Validation: Winnow improves recall at fixed false discovery rate thresholds when applied to predictions from tools such as InstaNovo.
• Cross-Model Comparison: Winnow enables comparison of results across different de novo peptide sequencing tools by providing calibrated, comparable confidence scores.
• Proteome Analysis of Uncharted Regions: Winnow supports accurate false discovery rate control for mass spectrometry-based proteomics in regions lacking reference databases.

Methodology:

Winnow uses a neural network trained on peptide-spectrum match-derived features to calibrate raw model scores, then applies a novel decoy-free FDR estimator to compute error metrics and experiment-wide FDR.