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

Abstract

Biomolecular diffusion models can now predict proteins and heterogeneous complexes, but glycans remain difficult because their branched topology, conformational flexibility, and strict stereochemical rules must be captured simultaneously. We developed SweetFold, a glycan-aware adaptation of Boltz-1x for the structure prediction of free glycans, glycoproteins, and protein-glycan complexes. SweetFold represents glycans as pseudo-polymers rather than generic ligands, preserving monosaccharide identity, anomeric state, glycosidic connectivity, and atom-level stereochemistry. We pair this representation with glycan-specific architecture, stereochemical supervision, and a sugar-centric training curriculum. Across monosaccharide, oligosaccharide, lectin, and glycoprotein benchmarks, SweetFold improves structural metrics relative to baseline all-atom diffusion models while retaining protein-only benchmark performance. These results show that chemically localized representation and supervision can extend biomolecular diffusion models to carbohydrate chemistry.