The coverage and reliability of protein-protein interactions determined by high-throughput experiments still needs to be improved, especially for higher organisms, therefore the question persists, how interactions can be verified and predicted by computational approaches using available data on protein structural complexes. Recently we developed an approach called IBIS (Inferred Biomolecular Interaction Server) to predict and annotate protein-protein binding sites and interaction partners, which is based on the assumption that the structural location and sequence patterns of protein-protein binding sites are conserved between close homologs. In this study first we confirmed high accuracy of our method and found that its accuracy depends critically on the usage of all available data on structures of homologous complexes, compared to the approaches where only a non-redundant set of complexes is employed. Second we showed that there exists a trade-off between specificity and sensitivity if we employ in the prediction only evolutionarily conserved binding site clusters or clusters supported by only one observation (singletons). Finally we addressed the question of identifying the biologically relevant interactions using the homology inference approach and demonstrated that a large majority of crystal packing interactions can be correctly identified and filtered by our algorithm. At the same time, about half of biological interfaces that are not present in the protein crystallographic asymmetric unit can be reconstructed by IBIS from homologous complexes without the prior knowledge of crystal parameters of the query protein.