WinnowNet: A Deep Learning-Based Filtering Framework for Metaproteomics

This repository contains the datasets, scripts, and instructions required to reproduce the experiments presented in our paper, particularly the sections on Training WinnowNet, Benchmark Dataset Descriptions, Performance Comparison to State-of-the-art Filtering Algorithms, and Performance Evaluation of WinnowNet-Integrated Protein Identification Pipelines.

WinnowNet is a deep-learning-based PSM filtering algorithm for metaproteomics that demonstrates superior performance over existing statistical and machine learning tools. This repository is organized to facilitate easy access and reproducibility of results.

Repository Structure

WinnowNet4Review/
│
├── Training WinnowNet/
│   └── Scripts and configuration files to train both the self-attention-based and CNN-based WinnowNet models.
│
├── Benchmark Dataset Descriptions/
│   └── Descriptions of the twelve benchmark datasets (Synthetic, P1–P3, Marine 1–3, Soil 1–3, Human Gut, and Human Gut TimsTOF).
│       Includes guidelines for data preprocessing and database construction with entrapment proteins.
│
├── Performance Comparison to State-of-the-art Filtering Algorithms/
│   └── Commands to benchmark WinnowNet against Percolator, Q-ranker, PeptideProphet, iProphet, MS²Rescore, and DeepFilter.
│       Includes results and plotting scripts for comparisons on PSM, peptide, and protein levels.
│
├── Performance Evaluation of WinnowNet-Integrated Protein Identification Pipelines/
│   └── Instructions for integrating WinnowNet with Sipros-Ensemble, FragPipe, Peaks, and AlphaPept pipelines.
│       Includes evaluation scripts and performance metrics (e.g., FDR control and identification gain).
│
└── README.md

Reproducibility Overview

Each subdirectory contains:

• Step-by-step instructions for reproducing the results reported in the paper.
• Dependencies and environment setup details (e.g., specific software versions like ProteoWizard 3.0.11841).
• Preprocessed or synthetic datasets, where applicable.
• Output files and scripts for evaluation, including FDR calculations.

Highlights

• Entrapment-based FDR estimation methods (including paired and combined strategies).
• Self-contained training scripts for WinnowNet, enabling retraining on new datasets.
• Pipeline integration and benchmarking across diverse platforms and sample complexities.

Citation

If you find this repository helpful, please cite our paper:

Citation coming soon.

License

MIT License.

Contact

For any questions or issues, please submit here