Prediction of Protein-Protein Binding Interfaces - Identification of Ligand Binding Site and Protein-Protein Interaction Area - page 104

In spite of the ever-increasing quality and cost-effectiveness of experimental
techniques, in silico protein interaction models remain an important tool for the
study of biological processes which involve molecular complexation - such as sig-
naling pathways and biological clock systems (Johnson et al. 2008 ; Bass and
Takahashi 2010 ; Duong et al. 2011 ; van Gelder et al. 2003 ) . Analysis of protein-
protein interactions may be performed on multiple levels of accuracy. On the most
basic level it is usually sufficient to determine whether interaction occurs at all
under certain conditions. This is done in order to infer the so-called protein-protein
interaction (PPI) networks. Such networks can then be refined by predicting protein
binding interfaces (to determine their approximate steric relationships). Accurately
modeling the 3D structure of the entire complex is the most challenging task, requir-
ing knowledge of molecular interactions on the atomic level.
Over the past two decades many methodologies and algorithms have been devel-
oped and applied (with varying success) on each of these three distinct levels.
An objective measure of the accuracy of protein complexation models is pro-
vided by the CAPRI (Critical Assessment of Predicted Interactions) challenge (see
preceeding chapter). Similarly to CASP contestants are provided with the necessary
input data (structures of individual monomers or amino acid sequences - if the
structure is easy to predict). Since the goal is to determine subtle details of protein-
protein interfaces, prediction quality is dependent not only on the number of cor-
rectly modeled contact points, but also on the accuracy of atom positions within the
interface zone. Solutions are penalized by the number of steric clashes between
interacting chains. Successive editions of the CAPRI challenge are being organized
on a regular basis since 2001. Numerous publications are available regarding the
challenge itself and its most accurate prediction pipelines (Janin et al. 2003 ; Janin
and Wodak 2007 ; Janin 2007 , 2010a, b ; Kastritis et al. 2011 ).
Docking analysis is a complex process, typically composed of three phases:
(1) selection of interface candidates based on experimental data (or predictions) to
focus the conformational search; (2) generation of the protein complex via rigid-
body docking; (3) ranking and scoring of results. Thus, we limit our description to
tools most often used in these sophisticated algorithms.
Our study focuses on prediction of protein interfaces - a task which should enable
us to detect protein complexation events and may also serve higher-order structure
prediction workflows. Even on this basic level existing tools are incapable of identi-
fying interface residues with consistent accuracy. The varying difficulty of modeling
complexation events suggests that many different binding mechanisms come into
play and thus many different kinds of interacting residues can be observed (differing
with respect to their specificity, sequence conservation, hydrophobicity, etc.)
This chapter presents four differing approaches to prediction of protein-protein
binding interfaces by means of molecular docking. We will compare FOD with three
state-of-the-art models: HADDOCK, ZDOCK and RosettaDock, each of which is an
implementation of rigid-body grid-based docking algorithms. Besides the specifics
of formulation and parameterization of force fields, the main difference between
these three distinct programs lies in their approach to focusing the conformational
space search. ZDOCK allows the user to select specific amino acids while HADDOCK
calls for experimental preselection of interface residues. Users of RosettaDock may