MIEN: a deep-learning model for pocket-conditioned glycan pose and conformation discrimination

Abstract

In protein-carbohydrate interaction, Molecular docking can generate many glycan poses that appear plausible, but the top-ranked pose may differ from the experimentally observed conformation. We introduce Minimal Interface Equivariant Network (MIEN), a structure-based equivariant graph network that scores the compatibility of a glycan pose with a fixed protein pocket. The model operates on a minimal-interface graph that contains only protein atoms forming polar or CH-pai contacts with the glycan; protein and sugar residue names are excluded. MIEN was trained on 5,241 experimental complexes and one unified set of contact-preserving rigid-body negatives grouped into three semantic-contact tiers. On a protein-disjoint benchmark of 260 held-out complexes, MIEN separated experimental poses from non-native poses and ranked experimental poses more highly within matched pockets. We then evaluated MIEN on AutoDock-GPU and GlycanDock candidate pools generated independently for the same complexes. Across three training seeds, MIEN selected lower-RMSD top-1 poses more often than either docking engine and increased the number of best poses below 1, 2 and 4 Angstrom. These results provide a practical route to more realistic protein-carbohydrate complex poses and a stronger basis for structural and mechanistic studies of protein-carbohydrate interactions.