With the explosively increased amount of newly discovered proteins, predicting the function of these proteins from amino acid sequencesis becoming one of the main challenges in functional annotation of genomes. Nowadays a number of computational approaches have been developed to predict DNA-binding proteins effectively and accurately from amino acid sequences, such as SVM, DNABP and CNN-RNN. However, these methods do not consider the context in amino acid sequences, which makes it difficult for them to capture sequence features adequately. In this paper, we propose CNN-BiLSTM, a new method for predicting DNA-binding proteins, elaborately reconciling convolution neural network and bi-directional long short-term memory recurrent neural network. CNN-BiLSTM can explore the potential contextual relationships of amino acid sequences to obtain more features than traditional models. The experimental results show that the predication accuracy of the proposed CNN-BiLSTM method on the test set is 96.5%, which is 7.8% higher than that of SVM, 9.6% higher than that of DNABP and 3.7% higher than that of CNN-RNN respectively. Being tested on 20, 000 independent samples provided by UniProt that weren't involved in model training, the accuracy of CNN-BiLSTM is 94.5%, which is 12% higher than that of SVM, 4.9% higher than that of DNABP and 4% higher than that of CNN-RNN respectively.The model training process is visualized and compared with that of CNN-RNN, and it is found that the training process of CNN-BiLSTMsupport better generalization from the training data set, which shows that CNN-BiLSTM has a wider range of adaptations to protein sequences. On the independent samples set, CNN-BiLSTM presents better credibility, for its predicted scores are closer to the labels of the samples than those of CNN-RNN. Therefore, the proposed CNN-BiLSTM is a more powerful method for identifying DNA-binding proteins.