Virtual screening (VS) is widely used in the process of a computational drug discovery for reducing a large amount of cost. Chemical genomics-based virtual screening (CGBVS) which is a kind of VS predicts new protein-compound interactions (PCIs) from known PCIs data using several methods of machine learning or data mining. Although CGBVS provides highly efficient and accurate PCIs prediction, CGBVS has poor performance on prediction for new compound for which PCIs are unknown. Pairwise kernel method (PKM) is one of the state-of-the-art methods of CGBVS, that showed highest accuracy on prediction for new compounds. In this study, from the viewpoint of link mining we improved PKM by combining link indicator and chemical similarity, and evaluated their accuracy. The proposed method obtained AUPR value of 0.562 which is higher than that achieved by using normal PKM (0.468).