New method of glycoproteomics, Fudan developed a hybrid end-to-end framework based on Transformer and GNN, published in Nature sub-journal

Editor | Radish peel

Protein glycosylation is a post-translational modification of proteins by sugar groups, which plays an important role in various physiological and pathological functions of cells.

Glycoproteomics is the study of protein glycosylation within the proteome, using liquid chromatography coupled with tandem mass spectrometry (MS/MS) technology to obtain combined information on glycosylation sites, glycosylation levels and sugar structures. .

However, current database search methods for glycoproteomics often have difficulty determining glycan structures due to the limited occurrence of structure-determining ions. Although spectral search methods can exploit fragmentation intensity to facilitate structural identification of glycopeptides, difficulties in spectral library construction hinder their application.

In the latest study, researchers from Fudan University proposed DeepGP, a hybrid deep learning framework based on Transformer and graph neural networks, for predicting MS/MS spectra and retention times (RT) of glycopeptides.

Two graph neural network modules are used to capture branched sugar structures and predict sugar ion strengths respectively. Additionally, a pre-training strategy was implemented to alleviate the shortage of glycoproteomic data.

The research was titled "Deep learning prediction of glycopeptide tandem mass spectra powers glycoproteomics" and was published in "Nature Machine Intelligence" on July 30, 2024.

Protein post-translational modifications (PTMs) significantly increase the complexity of the proteome. As one of the most important PTMs, glycosylation affects more than 50% of mammalian proteins and plays a key role in many physiological and pathological processes.

During the glycosylation process, sugar molecules are attached to the side chains of specific amino acid residues, resulting in structural heterogeneity, resulting in the diversity of glycopeptide isomers and increasing the difficulty of identification.

Liquid chromatography tandem mass spectrometry (LC-MS/MS) is the primary technique to identify glycopeptides by fragment ions and molecular weight combined with RT. Mass-to-charge ratio (m/z) alone is not enough to determine the sugar structure, so scientists use spectral matching methods to improve identification sensitivity. However, constructing glycopeptide MS/MS spectral libraries is costly and complex.

In recent years, deep learning has made progress in peptide MS/MS spectrum prediction. However, the relatively small number of current glycopeptidomics data sets and the lack of standardized protocols for generating glycopeptide mass spectrometry data limits the availability of suitable data for deep learning model training.

To this end, researchers from Fudan University propose DeepGP, a deep learning-based hybrid end-to-end framework for complete N-glycopeptide MS/MS spectra and RT prediction. The deep learning framework consists of a pre-trained Transformer module and two graph neural network (GNN) modules.

Illustration: model architecture and glycopeptide MS/MS spectral prediction.

DeepGP model

• accepts glycopeptides as input

• Encoded glycopeptide features:

  1. Glycostructure
  2. Amino acid sequence
  3. PTM type
  4. PTM position
  5. Precursor charge state

• The sugar structure is embedded through GNN, converting the glycopeptide into a graph:

  • Node: Monosaccharide

    

    Illustration: Distinguishing similar glycan compositions on synthetic data sets based on DeepGP. (Source: Paper)

  Two GNN modules capture sugar structure and predict sugar ion intensity
  Researchers evaluated three GNN architectures, including graph convolutional network (GCN), graph isomorphism network (GIN) and graph attention network (GAT) for sugar intercalation and B/Y ion intensity prediction.
  GCN utilizes convolution operations to obtain node representations and implements a message passing protocol to aggregate the representations of adjacent nodes; GIN performs well in graph isomorphism tests; GAT incorporates an attention mechanism to enable the model to focus on the most relevant parts of the input.
  Experimental results show that GCN performs best in the sugar embedding task, while GIN performs well in the B/Y ion intensity prediction task, so GCN and GIN were selected for corresponding analysis.

  

  Illustration: DeepGP performance in MS/MS prediction. (Source: paper)

Pre-training strategy to alleviate the shortage of glycoproteomics data

DeepGP uses a large amount of unlabeled natural language data for pre-training, similar to models such as BERT. Pre-training enables the model to have a knowledge base before formal training, thereby enhancing its performance in dealing with small-scale annotated data.

Testing on multiple biological data sets

연구원들은 마우스와 인간 샘플 데이터 세트를 사용하여 MS/MS 및 RT 예측에서 DeepGP의 높은 정확도를 입증했습니다.

그림: 글리코펩타이드 식별을 위해 pGlyco3(글리코펩타이드 검색 방법)과 결합된 DeepGP. (출처: 논문)

합성 및 생물학적 데이터 세트에 대한 DeepGP의 포괄적인 벤치마킹은 유사한 글리칸을 구별하는 데 있어 DeepGP의 효율성을 검증합니다. 데이터베이스 검색과 결합된 DeepGP는 글리코펩타이드 검출 감도를 향상시킵니다.

논문 링크:
https://www.nature.com/articles/s42256-024-00875-x

The above is the detailed content of New method of glycoproteomics, Fudan developed a hybrid end-to-end framework based on Transformer and GNN, published in Nature sub-journal. For more information, please follow other related articles on the PHP Chinese website!