Protein function prediction is a challenging but important task in bioinformatics. Many prediction methods have been developed, but are still limited by the bottleneck on training sample quantity.