Predicting protein-ligand and protein-peptide interfaces
Istituto di Analisi dei Sistemi ed Informatica “A.Ruberti” (IASI), CNR, Rome, Italy
2 College of Computing, Georgia Institute of Technology, Atlanta, GA, USA
* e-mail: firstname.lastname@example.org
Accepted: 15 May 2014
Published online: 24 June 2014
The paper deals with the identification of binding sites and concentrates on interactions involving small interfaces. In particular we focus our attention on two major interface types, namely protein-ligand and protein-peptide interfaces. As concerns protein-ligand binding site prediction, we classify the most interesting methods and approaches into four main categories: (a) shape-based methods, (b) alignment-based methods, (c) graph-theoretic approaches and (d) machine learning methods. Class (a) encompasses those methods which employ, in some way, geometric information about the protein surface. Methods falling into class (b) address the prediction problem as an alignment problem, i.e. finding protein-ligand atom pairs that occupy spatially equivalent positions. Graph theoretic approaches, class (c), are mainly based on the definition of a particular graph, known as the protein contact graph, and then apply some sophisticated methods from graph theory to discover subgraphs or score similarities for uncovering functional sites. The last class (d) contains those methods that are based on the learn-from-examples paradigm and that are able to take advantage of the large amount of data available on known protein-ligand pairs. As for protein-peptide interfaces, due to the often disordered nature of the regions involved in binding, shape similarity is no longer a determining factor. Then, in geometry-based methods, geometry is accounted for by providing the relative position of the atoms surrounding the peptide residues in known structures. Finally, also for protein-peptide interfaces, we present a classification of some successful machine learning methods. Indeed, they can be categorized in the way adopted to construct the learning examples. In particular, we envisage three main methods: distance functions, structure and potentials and structure alignment.
© Società Italiana di Fisica and Springer-Verlag Berlin Heidelberg, 2014