Abstract
Protein structure comparison algorithms are useful for predicting aspects of protein function. Some algorithms identify remote homologs, while others distinguish closely related proteins that prefer different substrates. Most of these methods assume that proteins are rigid in order to perform comparisons more rapidly, while others compensate for flexibility by representing proteins as a connected group of rigid components. To consider the motion of individual atoms, this paper presents a method for generating a map of binding cavity conformations based on conformational snapshots. We use clusters of protein conformations to distinguish proteins that have different binding preferences. Our results, on the serine proteases and enolase superfamilies show that, despite structural flexibility in binding sites, our methods correctly classify proteins with different binding specificities both qualitatively and quantitatively.