| Biological function is the consequence of interactions between molecules. While some binding sites are well characterized and easily detected (e.g. the catalytic triad of serine proteases), many protein-protein interfaces are difficult to characterize. Therefore, predicting potential binding surfaces is an important challenge in structural bioinformatics. The aim of this project is to develop a method for predicting potential protein binding surfaces using profiles constructed from observed patches at protein-protein interfaces. Our method uses small surface patches consisting of triplets of adjacent surface atomic groups that can be touched simultaneously by a probe sphere representing a solvent molecule. Protein atomic groups are classified here into 13 types based on heavy-atom types, the number of covalently attached hydrogen atoms and the number of all covalently attached atoms (as proposed by Tsai et al., J. Mol. Biol., 1999, 290:253-266), giving a total of 455 distinct triplets. We have compiled a set of profile scores by comparing the relative abundance of the different triplets in a large set of protein-protein interfaces with their abundance on exterior protein surfaces. The proposed method has been tested to assess its ability to predict protein binding surfaces and to filter docking orientations generated by docking programs. |