MS2Glycan: a model for generating glycan structures by mass spectrometry

Authors: Defeng Li, Zizheng Nie, Yanmin Liu, Xiaojun Cai, Yong Zhang*,  Xuefeng Cui*
    - *: To whom correspondence should be addressed.
Contact: xfcui@email.sdu.edu.cn

Introduction

We introduce MS2Glycan, a deep learning method for iteratively generating novel glycan structures from mass spectrometry data. It models component generation similarly to an object detection task, incorporating both localization and classification sub-tasks, and constructs the glycan structure component by component. MS2Glycan offers a key improvement over GlycanFinder by using a data-driven approach to predict component locations, rather than relying on predefined rules.

Usage

MS2Glycan is based on the framework and training principles of GlycanFinder¹, utilizing data from five types of mouse tissues proposed by pGlyco 2.0². It employs a five-fold cross-validation approach, where four tissues are used for training and one tissue is used for prediction.
The five mouse datasets were downloaded from the PRIDE repository with accession numbers: PXD005411 (mouse brain), PXD005412 (mouse kidney), PXD005413 (mouse heart), PXD005553 (mouse liver), PXD005555 (mouse lung).
Note that our model uses mass spectrometry data processed and identified by the PEAKS Studio³ software, rather than the raw mass spectrometry data from pGlyco 2.0. The experimental data was shared with me by the authors of Glycanfinder, so it cannot be publicly released. If you wish to replicate the results of this experiment, please contact the authors of Glycanfinder directly or use Peaks Studio for data processing.

Run

This repository provides example data ,model parameters (lung for testing) and one-click code execution.

Clone this repository by:

git clone https://github.com/xfcui/MS2Glycan.git

Make sure the python version you use is >= 3.9, and install the packages by:

bash install.sh

Notes

• fairseq requires a pip version lower than 24.1
• Make sure the LD_LIBRARY_PATH includes the CUDA path
  export LD_LIBRARY_PATH=/~/envs/~/lib/python3.9/site-packages/nvidia/cuda_nvrtc/lib:$LD_LIBRARY_PATH

Train the model: If you are training with your own dataset, you need to provide a glycan library in GlycoCT format and use create_combination function to generate the combination.pkl file.

# Specify the MGF mass spectrometry file, glycan database, and CSV mass spectrometry identification results and the model parameter path.
python train.py --mgf_file=sample_data/mouse-tissue.mgf --glycan_db=sample_data/all.pkl --csv_file_train=sample_data/test.csv --graph_model=ckpts/graphormer_nolung1.pt --cnn_model=ckpts/allmodel_nolung1

Evaluate the framework:

# Specify the MGF mass spectrometry file, glycan database, and CSV mass spectrometry identification results and the model parameter path.
python inference.py --mgf_file=sample_data/mouse-tissue.mgf --glycan_db=sample_data/all.pkl --csv_file_predict=sample_data/test.csv --graph_model=ckpts/graphormer_nolung.pt --cnn_model=ckpts/allmodel_nolung.pt

Peak matching on the original mass spectrum:

match_peaks.ipynb

Reference

- [1] Sun W, Zhang Q, Zhang X, et al. Glycopeptide database search and de novo sequencing with PEAKS GlycanFinder enable highly sensitive glycoproteomics[J]. Nature Communications, 2023, 14(1): 4046.

- [2] Liu M Q, Zeng W F, Fang P, et al. pGlyco 2.0 enables precision N-glycoproteomics with comprehensive quality control and one-step mass spectrometry for intact glycopeptide identification[J]. Nature communications, 2017, 8(1): 438.

- [3] https://www.bioinfor.com/