Graph-Based Clustering of Predicted Ligand-Binding Pockets on Protein Surfaces.

Detecting appropriate ligand binding pockets on protein surfaces has several important applications in the drug discovery process. In pocket sets identified by two software packages, PASS and Fpocket, we found a sizable number of protein-ligand complexes where more than one pocket overlaps with the ligand. In such cases, it would be desirable if a merged set of contacting pockets would represent the small molecule. Thus, we tested three clustering approaches to merge the given pockets, a classical clustering method and two methods based on algorithms from graph theory. We found that hierarchical clustering, as well as an approach based on the concept of maximum flow, could be favorably used for clustering pockets predicted either by PASS or by Fpocket.