Transformer-based Deep Learning for Glycan Structure Inference from Tandem Mass Spectrometry.

Area of Science:

• Glycomics and Computational Biology
• Biochemistry and Structural Biology
• Artificial Intelligence in Life Sciences

Background:

• Glycan structural analysis via tandem mass spectrometry (MS/MS) is crucial but challenging due to complex structures.
• Existing computational methods, including database searching and deep learning, face limitations in accuracy and scope.
• Accurate glycan inference requires capturing complex dependencies within MS/MS spectra.

Purpose of the Study:

• To develop advanced computational models for improved glycan structure prediction from MS/MS data.
• To overcome limitations of existing methods in handling glycan complexity and enabling novel structure discovery.
• To establish a new benchmark for glycan analysis using transformer-based deep learning.

Main Methods:

• Development of GlycoBERT, a transformer-based sequence classifier for glycan structure prediction.
• Development of GlycoBART, a transformer-based sequence-to-sequence model for de novo glycan inference.
• Validation of models on independent datasets and application to real-world MS/MS data.

Main Results:

• GlycoBERT achieved 95.1% structural accuracy, outperforming the state-of-the-art CandyCrunch model.
• GlycoBART successfully generated de novo glycan structures, including a novel structure not found in major databases.
• Both models demonstrated robust performance on independent validation datasets.

Conclusions:

• GlycoBERT and GlycoBART represent a significant advancement in computational glycan analysis.
• These models provide a powerful framework for accurate glycan structure prediction and the discovery of novel glycan diversity.
• The developed models set a new benchmark, enabling more comprehensive exploration of the glycome.