PocketGraph: graph representation of binding site volumes

Abstract

The representation of small molecules as moleculargraphs [1] is a common technique in various fields ofcheminformatics. This approach employs abstractdescriptions of topology and properties for rapid analysesand comparison. Receptor-based methods in contrastmostly depend on more complex representations imped-ing simplified analysis and limiting the possibilities ofproperty assignment. In this study we demonstrate thatligand-based methods can be applied to receptor-derivedbinding site analysis.We introduce the new method PocketGraph that trans-lates representations of binding site volumes into lineargraphs and enables the application of graph-based meth-ods to the world of protein pockets. The method uses thePocketPicker [2] algorithm for characterization of bindingsite volumes and employs a Growing Neural Gas [3] pro-cedure to derive graph representations of pocket topolo-gies.Self-organizing map (SOM) projections revealed a limitednumber of pocket topologies. We argue that there is onlya small set of pocket shapes realized in the known ligand-receptor complexes.