Drug-target interaction (DTI) provides novel insights about the genomic drug discovery. The wet experiments of identifying DTIs are time-consuming and costly. Recently, the increase of available data provides the opportunity to the development of computational methods. Although many computational methods have been proposed (such as classification-based methods, graph-based methods and network-based methods), there is still room for improvements. On one hand, there are much more non-interaction drug-target pairs than interaction pairs, and the classification-based methods are undermined by the imbalanced data and heavy computational burden. On the other hand, the graph-based methods and the network-based methods are incapable of predicting the interactions between new drugs and new targets. In the paper, we investigate the correlation of drugs and targets that interact, based on four classes of drug–target interaction data involving enzymes, ion channels, G-protein-coupled receptors and nuclear receptors. By exploiting the global information from interaction data, we compile the drug-target interaction networks as the binary classification datasets with positive and negative drug-target pairs. Then, we develop a representation of drug-target pairs based on drug chemical similarity and target sequence similarity, and adopt the random forest as classification engine to build the prediction models. Compared with the state-of-the-art methods, our method produces satisfying performance on the benchmark datasets. In general, our method can predict the interactions between know drugs and targets as well as the interactions between new drugs and new targets. In conclusion, our method is a promising tool for the drug–target interaction prediction.