Thermodynamic information can be inferred from static atom configurations. To accurately model the thermodynamics of carbohydrate–protein binding, a large binding data set has been assembled from the literature. This data set contains information from 262 unique protein–carbohydrate crystal structures for which experimental binding information is known. Hydrogen atoms were added to the structures and training conformations were generated with the automated docking program AutoDock (AD) 3.06, resulting in a training set of 225,920 all-atom conformations. Sixty-four formulations of the AD 3.0 free energy model were trained against the data set, testing each of four alternate methods of computing the van der Waals, solvation, and hydrogen bonding energetic components. The van der Waals parameters from AD1 produced the lowest errors, while punishing buried potential hydrogen-bond-forming atoms produced larger errors. Ten models were found that predict the Gibbs free energy of binding with less than half the error of any similar models previously published. We have used the best of these to determine free energies generated by the automated docking of disaccharides with mannosyl, allosyl, glucosyl, galactosyl, 6-phosphogalactosyl, and 6-phosphoglucosyl glycons, all linked by beta-(1,2), beta-(1,3), beta-(1,4), and beta-(1,6)-glycosidic bonds, into the fourteen available tertiary structures of glycoside hydrolase family 1 members (eight beta-glucosidases, one beta-galactosidase, one 6-phospho-beta-galactosidase, one myrosinase, and three beta-glycosidases) to investigate the structural determinants of their enzyme specificities. Analysis of variance (ANOVA) has been used to find significant interactions between active-site mutations and glycon specificity. Five active-site residues (T194, F205, N285, R336, and N376) (Zea mays beta-glucosidase numbering)) control a significant amount of glycon specificity.

Extended Abstract Status: Not Uploaded