WinnowNet

This algorithm was implemented and tested on Ubuntu 20.04.6 LTS (GNU/Linux 5.15.0-84-generic, x86_64).

Note:

This repository contains the development version of WinnowNet. For the code used to reproduce the experiments in the paper, please refer to the following repository: https://github.com/Biocomputing-Research-Group/WinnowNet4Review

Overview

WinnowNet is designed for advanced processing of mass spectrometry data with two core methods: a CNN-based approach and a self-attention-based approach. The repository includes scripts for feature extraction, model training, prediction (inference), and evaluation. A toy example is included to help users get started.

Table of Contents

• Setup and Installation
• Requirements
• Downloading Required Files
• Feature Extraction
• Training
  • CNN-based WinnowNet
  • Self-Attention-based WinnowNet
• Inference
  • PSM Re-scoring
• Evaluation
• Contact and Support

Setup and installation

1. Create a new conda environment and activate it.

It is recommended to use Conda for dependency management. Run the following commands in your terminal:

conda create --name WinnowNet python=3.8
conda activate WinnowNet

2. Install dependencies:

CUDA version 11.8 Pytorch GPU version is compatible with corresponding cuda version

pip install -r ./requirements.txt

Requirements

• Operation system: Linux
• GPU Memory
  • Inference Mode: At least 8 GB (adjust batch size if necessary)
  • Training Mode: At least 20 GB

Download Required Files

• Pre-trained model can be downloaded via:
  • CNN-based WinnowNet: cnn_pytorch.pt
  • Self-Attention-based WinnowNet: marine_att.pt
• A toy example is provided in this repository.
• Sample Input Datasets Mass spectra data for Benchmark datasets Other raw files benchmark datasets can be downloaded via: PXD007587, PXD006118, PXD013386, PXD023217, PXD035759

Input pre-processing

Extract fragment ion matching features along with 11 additional features derived from both theoretical and experimental spectra. The PSM (peptide-spectrum match) candidate information should be provided in a tab-delimited file (e.g., a TSV file output from Percolator).

python SpectraFeatures.py -i <tsv_file> -s <ms2_file> -o spectra.pkl -t 48 -f cnn

• Replace <tsv_file> with the path to your PSM candidates file.
• Replace <ms2_file> with the path to your experimental spectra file.
• The -t 48 option sets the number of threads (adjust this value as needed).
• Use -f cnn when preparing input for the CNN-based architecture or -f att for the self-attention-based model.

Training WinnowNet Models

This folder contains scripts, datasets, and instructions for training two variants of the WinnowNet deep learning model: a self-attention-based model and a CNN-based model. Training is carried out in two phases to enable curriculum learning from synthetic (easy) to real-world metaproteomic (difficult) datasets.

Requirements

• Python 3.7+
• PyTorch
• NumPy, Pandas, scikit-learn

Datasets

• Prosit_train.zip (Phase 1 training set): https://figshare.com/articles/dataset/Datasets/25511770?file=55257041
• marine1_train.zip (Phase 2 training set): https://figshare.com/articles/dataset/Datasets/25511770?file=55257035

---

Self-Attention-Based WinnowNet

Phase 1: Training on Easy Tasks (Synthetic Data)

python SpectraFeatures_training.py -i filename.tsv -s filename.ms2 -o spectra_feature.pkl -t 20 -f att
python WinnowNet_Att.py -i spectra_feature_directory -m prosit_att.pt

Explanation of options:

• -i: Input tab-delimited file with PSMs, including labels and weights.
• -s: Corresponding MS2 file (filename should match TSV).
• -o: Output file to store extracted features as a pkl file.
• -t: Number of threads for parallel processing.
• -f: Feature type (att for self-attention model).
• -m: Filename to save the trained model.
• A for-loop is needed to convert all tsv files to pkl files.

Phase 2: Training on Difficult Tasks (Real Data)

python SpectraFeatures_training.py -i filename.tsv -s filename.ms2 -o spectra_feature.pkl -t 20 -f att
python WinnowNet_Att.py -i spectra_feature_directory -m marine_att.pt -p prosit_att.pt

• -p: Pre-trained model from Phase 1.
• A for-loop is needed to convert all tsv files to pkl files.

Pre-trained model: marine_att.pt, https://figshare.com/articles/dataset/Models/25513531

---

CNN-Based WinnowNet

Phase 1: Training on Easy Tasks (Synthetic Data)

python SpectraFeatures_training.py -i filename.tsv -s filename.ms2 -o spectra_feature.pkl -t 20 -f cnn
python WinnowNet_CNN.py -i spectra_feature_directory -m prosit_cnn.pt

Phase 2: Training on Difficult Tasks (Real Data)

python SpectraFeatures_training.py -i filename.tsv -s filename.ms2 -o spectra_feature.pkl -t 20 -f cnn
python WinnowNet_CNN.py -i spectra_feature_directory -m cnn_pytorch.pt -p prosit_cnn.pt

Pre-trained model: cnn_pytorch.pt, https://figshare.com/articles/dataset/Models/25513531

---

Notes

• All input MS2/TSV files must be preprocessed properly.
• Models trained in Phase 1 are reused to initialize weights in Phase 2.
• Training with GPU is recommended for performance.

Inference

PSM Rescoring

Self-Attention-Based WinnowNet

To generate input representations for PSM candidates and perform re-scoring using the self-attention model, run:

python SpectraFeatures.py -i tsv_file -s ms2_file -o spectra.pkl -t 48 -f att 
python Prediction.py -i spectra.pkl -o rescore.out.txt -m att_pytorch.pt  

CNN-Based WinnowNet

To generate input representations for PSM candidates and perform re-scoring using the CNN model, run:

python SpectraFeatures.py -i filename.tsv -s filename.ms2 -o spectra.pkl -t 48 -f cnn
python Prediction_CNN.py -i spectra.pkl -o rescore.out.txt -m cnn_pytorch.pt 

Explanation of options:

• -i: Input tab-delimited file with PSMs
• -s: Corresponding MS2 file (filename should match TSV).
• -o: Output file to store extracted features as a pkl file.
• -t: Number of threads for parallel processing.
• -f: Feature type (att for self-attention model, cnnfor CNN model).
• -m: Filename to save the trained model.
• A for-loop is needed to convert all tsv files to pkl files.

Evaluation

FDR Control at the PSM/Peptide Levels

Filter the re-scored PSM candidates to control the false discovery rate (FDR) at both the PSM and peptide levels (targeted at 1% FDR). You will need both the original PSM file and the re-scoring results.

python filtering.py -i rescore.out.txt -p tsv_file -o filtered -d Rev_ -f 0.01

Explanation of options:

• -i: Rescoring file from WinnowNet
• -p: Input tab-delimited file with PSMs
• -o: filtered results' prefix
• -d: Decoy prefix used for target-decoy strategy. Default: Rev_
• -f: False Discovery Rate. Default: 0.01
• A for-loop is needed to convert all tsv files to pkl files.

• The filtered output files include updated PSM information (new predicted scores, spectrum IDs, identified peptides, and corresponding proteins).
• Assembling filtered identified peptides into proteins
• This script is needed to run at the working directory inlucding filtered results at PSM and Peptide levels.

python sipros_peptide_assembling.py

When assembling filtered, identified peptides into proteins, the overall protein-level FDR depends on the quality of the filtered peptide list. An initial peptide-level FDR (for example, 1%) may lead to a protein-level FDR that is higher than desired. In such cases, you need to re-filter the peptides using a stricter (i.e., lower) FDR threshold until you achieve a 1% protein-level FDR.

Contact and Support

For further assistance, please consult the GitHub repository or reach out to the project maintainers.