Carbohydrates play a key role in a variety of physiological and pathological processes and, hence, represent a rich source for the development of novel therapeutic agents. Being able to predict binding mode and binding affinity is an essential, yet lacking, aspect of the structure-based design of carbohydrate-based ligands. We assembled a diverse data set comprising 273 carbohydrate-protein crystal structures with known binding affinity and evaluated the prediction accuracy of a large collection of well-established scoring and free-energy functions, as well as combinations thereof. Unfortunately, the tested functions were not capable of reproducing binding affinities in the studied complexes. To simplify the complex free-energy surface of carbohydrate-protein systems, we classified the studied proteins according to the topology and solvent exposure of the carbohydrate-binding site into five distinct categories. A free-energy model based on the proposed classification scheme reproduced binding affinities in the carbohydrate data set with an r(2) of 0.71 and root-mean-squared-error of 1.25 kcal/mol (N = 236). The improvement in model performance underlines the significance of the differences in the local micro-environments of carbohydrate-binding sites and demonstrates the usefulness of calibrating free-energy functions individually according to binding-site topology and solvent exposure.