Aiming at the situation that multi-label learning and ensemble learning are not applied maturely in solving the problem of protein subcellular multi-localization prediction, an ensemble multi-label learning based method was studied to address this issue. Firstly, from the view of combination of multi-label learning and ensemble learning, a three-layer ensemble multi-label learning framework was proposed. Learning algorithms and classifiers were both categorized into three groups according to the corresponding three layers of the proposed framework. In this framework, binary classification learning, multi-class classification learning, multi-label learning and ensemble learning were all integrated together effectively, and thus a general-purpose ensemble multi-label learning model was constructed. Secondly, a learning system with good expansibility was designed using the object-oriented technology and Unified Modeling Language (UML), which can enhance the function and improve the performance of the system. Finally, by extending the model, a Java-based learning system was developed and applied successfully to predict protein's multiple subcellular localizations. The test results on a gram positive bacteria dataset indicate the operability of the system function as well as better prediction performance, the proposed system may become a useful tool to predict protein multiple subcellular localizations.