The cellular processes in cells are controlled by protein-protein interactions (PPI), and comprehensive PPI maps are important to understand the complicated regulatory, metabolic and signaling pathways. Recently, new parameters for PPI prediction are under discovering. Here, a Na?ve Bayesian algorithm with the correlated Gene Ontology (GO) was used for PPI prediction. The characteristic pairs of GO terms was demonstrated by training a non-redundant PPI data set from two online budding yeast databases, and the probability about this two correlated GO terms was also obtained. The accuracy of the prediction was tested by both positive and negative control data. The approach can distinguish them properly, with a satisfied sensitivity and low false positive rate. After comparing the prediction result to the data derived from high-throughput experiments, it is proved that the method is more sensitive and efficient than other means. Furthermore, some new insightful knowledge about interactions of the proteins will be found using this prediction approach, and the prediction is very helpful to the laboratory experiments.