A Glycan-Aware Diffusion Model for Carbohydrate and Glycoprotein Structure Prediction

Area of Science:

• Computational Biology
• Structural Biology
• Carbohydrate Chemistry

Background:

• Biomolecular diffusion models excel at protein structure prediction.
• Predicting glycans is challenging due to their complex topology and flexibility.
• Existing models struggle to simultaneously capture glycan stereochemistry and connectivity.

Purpose of the Study:

• To develop a novel method for accurate glycan structure prediction.
• To adapt diffusion models for carbohydrate chemistry.
• To improve the prediction of glycoproteins and protein-glycan complexes.

Main Methods:

• Developed SweetFold, a glycan-aware adaptation of the Boltz-1x diffusion model.
• Represented glycans as pseudo-polymers to preserve chemical details.
• Incorporated glycan-specific architecture, stereochemical supervision, and a sugar-centric training curriculum.

Main Results:

• SweetFold demonstrated improved structural metrics on monosaccharide, oligosaccharide, lectin, and glycoprotein benchmarks.
• The model successfully preserved monosaccharide identity, anomeric state, and stereochemistry.
• Protein-only prediction performance was maintained compared to baseline models.

Conclusions:

• Chemically localized representation and supervision are key to extending diffusion models to carbohydrate chemistry.
• SweetFold represents a significant advancement in predicting the structure of glycans and their complexes.
• This work opens new possibilities for understanding glycan function through accurate structural modeling.