Abstract: Predicting protein function is one of the key challenges in the post genome era. Functional annotation databases of proteins mainly provide the knowledge of positive examples that proteins carrying out a given function, and rarely record the knowledge of negative examples that proteins not carrying out a given function. Current computational models almost only focus on utilizing the positive examples for function prediction and seldom pay attention to these scarce but informative negative examples. It is well recognized that both positive and negative examples should be used to achieve a discriminative predictor. Motivated by this recognition, in this paper, we propose a protein function prediction approach using positive and negative examples (ProPN) to bridge this gap. ProPN first utilizes a direct signed hybrid graph to describe the positive examples, negative examples, interactions between proteins and correlations between functions; and then it employs label propagation on the graph to predict protein function. The experimental results on several public available proteomic datasets demonstrate that ProPN not only makes better performance in predicting negative examples of proteins whose functional annotations are partially known than state-of-the-art algorithms, but also performs better than other related approaches in predicting functions of proteins whose functional annotations are completely unknown.