Atom-level machine learning of protein-glycan interactions and cross-chiral recognition in glycobiology.

Abstract

We describe a machine-learned (ML) model, MCNet, which predicts interactions between proteins and glycans. MCNet predicted quantitative interactions between glycan-binding proteins (GBPs) and enantiomers of common glycans, which were not part of the original training datasets. l-glycans are rare in nature but are important in consideration of safety of putative mirror-image life-forms. Current ML models that predict properties of glycans from their monosaccharide composition cannot extrapolate properties of mirror glycans. Instead, MCNet uses an atom-level description of the glycan to output an estimate of binding to GBPs. MCNet is trained using data from glycan microarrays and affinity measurements unified using a "fraction bound" parameter. Trained MCNet predicted unexpected binding of l-glucose to some fucose-binding GBPs. Both glycan and lectin arrays conformed these predictions. ML models akin to MCNet reach beyond traditional glycobiology and make it possible to anticipate interaction between biomolecules in mirror-life forms and present-day life-forms.