Most biological actions of proteins, including their ability to interact with one another, involve some specific parts of their three-dimensional structure, called binding sites. These have evolved for their ability to bind other molecules effectively and are often conserved in different proteins. Identifying protein-protein binding sites in a protein that is known to interact with other proteins can provide important clues to the function of the protein and can also be used in protein-protein docking studies to reduce the search space explored by docking algorithms. We have developed an algorithm for structural similarity search in a database of non-redundant protein structures to find conserved binding regions on proteins involved in protein-protein interactions. We have used this algorithm to find conserved regions on a protein surface. The structurally conserved residues found were labeled as a protein-protein binding site, which allowed us to tune the AutoDock docking algorithm to predict the native protein complex structure from unbound protein structures. The conservation of protein structures that correctly predicted protein-protein binding site was used in AutoDock program to improve protein-protein docking. A web application based on our method is available at http://probis.cmm.ki.si.