In this work, a hybrid method is proposed to eliminate the limitations of traditional protein-pro- tein interactions (PPIs) extraction methods, such as pattern learning and machine learning. Each sentence from the biomedical literature containing a protein pair describes a PPI which is predicted by first learning syntax patterns typical of PPIs from training corpus and then using their presence as features, along with bag-of-word features in a maximum entropy model. Tested on the BioCreAtIve corpus, the PPIs extraction method, which achieved a precision rate of 64%, recall rate of 60%, improved the performance in terms of F1 value by 11% compared with the component pure pattern- based and bag-of-word methods. The results on this test set were also compared with other three ex- traction methods and found to improve the performance remarkably.