Despite the improvement of predicting protein-protein interactions, little progress has been achieved in the network based prediction of protein complexes. We assume that the problem of network based predictions is caused by the coexistence of incompatible interactions which cannot occur simultaneously. Therefore, we propose the method to predict protein complexes eliminating incompatible interactions which cannot occur to single interface at the same time. To obtain domain interface data, we integrate several biological databases: PSIbase, SwissProt, SCOP, and ASTRAL. This method integrate the existing prediction methods, MCODE and LCMA, based on the connectivity in the protein interaction graph. The prediction accuracy are compared with the existing methods and known protein complexes of MIPS CYGD. As a result, the prediction accuracy shows improvement compared to the existing methods, MCODE and LCMA.