To evaluate the current status of the protein–protein docking field, the CAPRI experiment came to life. Researchers are given the receptor and ligand 3-dimensional (3D) coordinates before the cocrystallized complex is published. Human predictions of the complex structure are supposed to be submitted within 3 weeks, whereas the server ClusPro has only 24 h and does not make use of any biochemical information. From the 10 targets analyzed in the second evaluation meeting of CAPRI, ClusPro was able to predict meaningful models for 5 targets using only empirical free energy estimates. For two of the targets, the server predictions were assessed to be among the best in the field. Namely, for Targets 8 and 12, ClusPro predicted the model with the most accurate binding-site interface and the model with the highest percentage of nativelike contacts, among 180 and 230 submissions, respectively. After CAPRI, the server has been further developed to predict oligomeric assemblies, and new tools now allow the user to restrict the search for the complex to specific regions on the protein surface, significantly enhancing the predictive capabilities of the server. The performance of ClusPro in CAPRI Rounds 3–5 suggests that clustering the low free energy (i.e., desolvation and electrostatic energy) conformations of a homogeneous conformational sampling of the binding interface is a fast and reliable procedure to detect protein–protein interactions and eliminate false positives. Not including targets that had a significant structural rearrangement upon binding, the success rate of ClusPro was found to be around 71%. Proteins 2005;60:239–244. © 2005 Wiley-Liss, Inc.