Functional Prediction of Binding Pockets

Abstract

Putative function for targets with no known ligands has typically been determined from liganded homologous proteins using sequence and structure comparisons. However, it is debatable what percentage of sequence identity implies similar function, whereas structural similarity is focused on global folds and could miss divergent structures and novel global folds. The present study describes an approach to classify a diverse set of proteins and predict their function. Descriptors corresponding to structural, physicochemical, and geometric properties of the ligand-binding cavities of a collection of 434 complexes (17 protein families) were calculated and analyzed by statistical methods. The best model using discriminant function analysis (DFA) consisted of 371 proteins (15 families) and had correct classification rates of 90% and cross-validation 86%. DFA with one protein and a random sample of the remaining proteins led to 100% correct prediction of putative protein function for 10 of the 15 protein families.