GlycanGT: a pretrained graph transformer framework for glycan graph representation and generative learning.

Area of Science:

• Computational biology
• Glycomics
• Bioinformatics

Background:

• Glycan functional understanding lags behind proteins and nucleic acids.
• Ambiguous glycan annotations limit computational analysis.
• Existing graph-based methods struggle with global patterns and incomplete sequences.

Purpose of the Study:

• Introduce GlycanGT, a novel graph-transformer model for glycan analysis.
• Improve computational modeling of glycan structures and functions.
• Address limitations in current glycan annotation and analysis methods.

Main Methods:

• Represented glycans as graphs of monosaccharides and glycosidic bonds.
• Pretrained GlycanGT using a masked language modeling objective.
• Evaluated GlycanGT on 8 benchmark tasks, including immunogenicity classification.

Main Results:

• GlycanGT outperformed existing methods across 8 benchmark tasks.
• Achieved 0.844 AUPRC for immunogenicity classification.
• Generated biologically relevant embeddings clustering known N- and O-glycan categories.
• Accurately predicted candidates for ambiguous glycan sequences with >80% top-5 accuracy.
• Demonstrated high accuracy in monosaccharide and glycosidic bond predictions under high masking levels.

Conclusions:

• GlycanGT offers a powerful new tool for glycan analysis.
• The model improves the interpretation of complex and ambiguous glycan structures.
• This work facilitates deeper understanding of glycan functions.