Modulating protein-protein interactions (PPIs) by small ligands is a challenging but attractive strategy for therapeutic intervention of various human diseases. To date, over 30 PPIs such as MDM2/TP53, BCL-X
L(BCL-2)/BAK, and IL2/IL2 receptor alpha have been intensively studied as target for small ligands. In the era of rapid accumulation of a huge number of PPI data, lack of the methodology that aims to efficiently select drug target PPIs by holistically assessing druggability of each PPI is an issue that should be immediately resolved. To address the issue, we have recently proposed an integrative in silico approach for discovering druggable PPIs by detecting interacting domains, evaluating similarity in biological function between interacting proteins with Gene Ontology terms, and finding ligand-binding pockets on protein surface. Application of our approach to a large number of human PPI data showed its effectiveness for assessing druggabilities of PPIs and selecting some promising candidates of target PPIs.
As the next step, we introduce a support vector machine (SVM)-based machine-learning method to our approach and apply it to human PPIs. In this study, we focus on human PPIs with tertiary structures of their protein complex already resolved. Known target PPIs, carefully selected from previous studies focusing on the development of PPI-inhibiting ligands, were used as training data in our SVM-based method. Twenty-six physicochemical, 16 drug/chemical, and 26 functional attributes of interacting proteins were utilized as feature vector of PPIs. The best SVM model constructed can distinguish known target PPIs from others at a high accuracy of 85% (sensitivity, 89%; specificity, 81%). We will discuss the effectiveness of our SVM-based approach and novel target PPIs predicted as potentially druggable by our method. To our knowledge, this study is the first application of a machine-learning method to the prediction of druggable PPIs.
抄録全体を表示