CLIMBS: assessing Carbohydrate-Protein interactions through a graph neural network classifier using synthetic negative data

Carbohydrate-protein interactions are essential for biological processes such as cellular signaling and metabolism. However, current prediction methods cannot accurately evaluate the affinity and specificity of proteins for carbohydrates such as glucose and galactose. Here, we develop a machine learning classifier, named CLIMBS, and train it on crystal structures and synthetic data from unsuccessfully designed binders, to effectively assess if carbohydrate-protein complexes represent realistic, native like structures. Compared to other methods, CLIMBS has outstanding accuracy, excellent carbohydrate specificity, sub-second runtime per sample, and no prediction bias towards either negative or positive samples. CLIMBS can work as an additional criterion in carbohydrate-protein interaction modeling for areas such as docking, protein design and enzyme design.