Co-folding of Carbohydrate-Protein Complexes

Abstract

Deep learning–based methods have transformed in silico protein structure prediction, and co-folding approaches now enable the generation of three-dimensional protein models from sequence in the presence of cofactors and custom ligands, unlocking major opportunities in drug discovery and design. Nonetheless, co-folded models remain limited by factors such as a difficult-to-quantify reliance on pose memorization from training data, and violations of geometric and stereochemical constraints. In this context, protein–carbohydrate interactions are particularly challenging to model accurately owing to the high density of chiral centers of the ligands, the flexibility of glycosidic linkages, and the typically exposed nature of binding sites. In this contribution, we assess the performance of two widely used deep learning models, AlphaFold3 and Boltz-2, in predicting carbohydrate–protein complex structures. Using three carbohydrate-binding protein families, we show that accurate geometries of bound states and insights into protein–carbohydrate recognition specificity can be obtained directly from co-folded models generated using only protein sequence and a simple ligand string notation, paving the way for the high-throughput screening of glycans and glycomimetics. Supplementary data available at https://zenodo.org/records/17618082