A detailed and complete structural knowledge of the interactome is one of the grand challenges in Biology, and a variety of computational docking approaches have been developed to complement experimental efforts and help in the characterization of protein-protein interactions. Among the different docking scoring methods, those based on physicochemical considerations can give the maximum accuracy at the atomic level, but they are usually computationally demanding and necessarily noisy when implemented in rigid-body approaches. Coarser-grained knowledge-based potentials are less sensitive to details of atomic arrangements, thus providing an efficient alternative for scoring of rigid-body docking poses. In this study, we have extracted new statistical potentials from intermolecular pairs of exposed residues in known complex structures, which were then used to score protein-protein docking poses. The new method, called SIPPER (scoring by intermolecular pairwise propensities of exposed residues), combines the value of residue desolvation based on solvent-exposed area with the propensity-based contribution of intermolecular residue pairs. This new scoring function found a near-native orientation within the top 10 predictions in nearly one-third of the cases of a standard docking benchmark and proved to be also useful as a filtering step, drastically reducing the number of docking candidates needed by energy-based methods like pyDock.