The prediction of new protein-protein interactions is important due to many unknown functions of biological pathways. In addition, many protein-protein interaction databases contain different types of protein interactions, i.e., protein associations, physical protein associations and direct protein interactions. Moreover, discovering new crucial protein-protein interactions through biological experiments is still difficult. Therefore, there is increasing demand to discover not only protein associations but also direct protein interactions. Many studies have predicted protein-protein interactions by directly using biological features, such as Gene Ontology (GO) functions and domains of protein structure between two interacting proteins. In this article, we propose TransDomain, a new method of predicting potential protein-protein interactions by using a new strong transitive relationship between interacting protein domains. Our results demonstrate that TransDomain can effectively predict potential protein-protein interactions from existing identified protein interaction relationships. TransDomain achieved 90% precision rate and 91% accuracy in the prediction of all types of protein-protein interactions and outperformed the existing PPI prediction systems and simulated GO-based prediction methods.