Chemical foundation models reveal molecular principles underlying glycobiology

Abstract

Glycans encode biological specificity through branching, linkage and stereochemical diversity, and have traditionally been modelled using specialised computational representations distinct from those developed for general chemistry. It remains unclear to what extent representations learned from general molecular chemistry capture the molecular principles of glycobiology. Here, we systematically evaluate the transfer of chemical foundation models to glycobiology using standardized benchmarks spanning taxonomy, immunogenicity, glycosylation type and protein-glycan interactions. We show that atom-level representations learned from general chemistry encode biologically meaningful molecular organization, providing particularly strong performance for molecular recognition, including immunogenicity and protein-glycan interaction prediction. Fine-tuning reorganizes these representations into latent spaces aligned with glycan architecture, glycosylation programmes and molecular recognition. Mechanistic analyses identify stereochemistry as an important determinant of successful transfer, while tokenization, context length and conformer generation influence the representation of large, branched glycans. These representations generalize from natural glycans to synthetic glycomimetics and recover experimentally validated stereochemistry-dependent glycan-lectin recognition. Together, our findings demonstrate that molecular principles learned from general chemistry are sufficient to capture fundamental features of glycobiology. They position glycans as a specialised region of chemical space and chemical foundation models as a broadly applicable framework for glycan function prediction, molecular recognition and glycomimetic discovery.