## INTRODUCTION

Protein–protein interactions play an important role in various biochemical processes. With the development of genomics and proteomics, more and more protein interaction data are obtained by experimental methods1, 2 but their structural annotations are lacking. The complete structural annotations for these interaction data will take years with the current structure determination methods.3 Therefore, to solve the problem, computational methods that predict structural details of protein interaction have been designed and continuously improved in recent years.

Docking methods are computational methods, which predict interactions between two proteins. In the last 2 decades, various docking methods have been designed.4–23 Jiang and Kim designed a soft docking method of matching surface cubes and surface normals.6, 7 Grid method was also used by Cherfils *et al.*8 Katchalski-Katzir *et al.*9 first used FFT to sample the conformational space. Vakser10 used a sparse cube size with the FFT procedure to dock various protein complexes. Sternberg and coworkers11, 12 designed FTDOCK, which performed search with FFT and used a soft electrostatic function as a filter for docking unbound structures. Abagyan and coworkers13–18 used pseudo-Brownian search and biased probability Monte Carlo search to deal with rigid-body search and side-chain conformation search, respectively. Weng and coworkers19–21 developed ZDOCK with a geometry function counting surface atom pairs and the ACE solvation function in addition to a novel implementation of the original FFT method. Baker and coworkers22 devised RossettaDock using a two-stage searching method sampling conformational space.

One of the most important tasks in docking is calculating scores that can discriminate true complexes from false ones. Of the various scoring functions that have been applied to different docking procedures, the most useful and efficient ones are based on the concept of geometry or shape complementarity. For example, they are used in SOFTDOCK,6, 7 FTDOCK,11 and ZDOCK.19, 20 A more physical version, namely, Lennard–Jones potential is used in ICM-DISCO13 and RossettaDock.22 Besides geometry scoring functions, other types of functions have been developed to represent other aspects of complementarity in the process of complex formation, including electrostatics, solvation, hydrogen bonding, and so forth. Electrostatics energy is generally calculated by a modified Coulomb equation.11, 15, 22 As for solvation energy, several methods15, 22, 24 have been used based on various solvation models. Most of them are knowledge-based statistical potentials using either contact surface areas or contact pairs of atoms with predefined atomic types. Hydrogen bonding energy is estimated in some docking methods.15, 22

A community-wide assessment, Critical Assessment of PRedicted Interactions (CAPRI), has been set up to help developers improve their methods. Reviews for rounds 1–525, 26 as well as more docking methods participating in CAPRI can be found.26 We have attended CAPRI from rounds 6–11 and predicted eight targets.27–33

In this article, we will present an improved rigid-body docking procedure that uses a large cube size in sampling the conformation space. We systematically test the method with protein–protein interaction benchmark V2.0.34 We will show that our coarse grained rigid-body docking procedure is able to find correct solutions and rank them within top 2000 for 66 out of 83 complexes. We will also present our results in CAPRI rounds 6–11.