AbstractComputational methods that allow predicting the effects of nonsynonymous substitutions are an integral part of exome studies. Here, we validated and improved their specificity by performing a comprehensive bioinformatics analysis combined with experimental and clinical data on a model of glucokinase (GCK): 8835 putative variations, including 515 disease-associated variations from 1596 families with diagnoses of monogenic diabetes (GCK-MODY) or persistent hyperinsulinemic hypoglycemia of infancy (PHHI), and 126 variations with available or newly reported (19 variations) data on enzyme kinetics. We also proved that high frequency of disease-associated variations found in patients is closely related to their evolutionary conservation. The default set prediction methods predicted correctly the effects of only a part of the GCK-MODY-associated variations and completely failed to predict the normoglycemic or PHHI-associated variations. Therefore, we calculated evidence-based thresholds that improved significantly the specificity of predictions (≤75%). The combined prediction analysis even allowed to distinguish activating from inactivating variations and identified a group of putatively highly pathogenic variations (EVmutation score <−7.5 and SNAP2 score >70), which were surprisingly underrepresented among MODY patients and thus under negative selection during molecular evolution. We suggested and validated the first robust evidence-based thresholds, which allow improved, highly specific predictions of disease-associated GCK variations.