Membrane proteome is the “doorbell” and “portal” of the cell. The subcellular localization plays a key role to elucidate these membrane proteins functions. This study introduced a novel model that not only learns the features from protein sequences, but also learns features from cross-species protein-protein interaction network. The protein GO features were obtained by deep learning method. The major low dimensional structure was built by linear discriminant analysis (LDA). Then the support vector machines (SVM) was applied to predict the membrane proteins subcellular location. The final evaluation results by k-fold cross-validation test showed that the accuracy of the proposed method has been improved a lot compared with the BLAST method. In this study, the overall accuracy could be over 95% in the situation that the training sets contains only 10% samples and the rest samples as testing sets, which means that the proposed method has robust performance for small sample train dataset.