Protein function prediction is one of the most challenging problems in the post-genomic era. Previous prediction methods using protein-protein interaction networks relied on the neighborhoods or the connected paths to known proteins. Still new algorithm is required to increase the accuracy. In this paper, we propose a novel protein function prediction approach on the basis of frequent pattern mining in graph data. A protein-protein interaction network is represented as an unweighted, undirected graph with nodes denoting proteins and edges denoting interactions between proteins. Each node is labeled with a set of corresponding protein functions. The function prediction method is processed in three steps, neighbor finding, pattern finding and function annotation. Using our approach we predict protein functions on a core set of protein-protein interaction data from DIP (Database of Interacting Proteins) and function annotation data from FunCat of MIPS (the Munich Information Center for Protein Sequences). The experimental results show better performance in prediction accuracy than existing neighbor counting methods.