In this paper, we propose a pattern-based protein function annotation framework, employing protein interaction networks, to predict annotation functions of proteins. More specifically, we first detect patterns that appear in the neighborhood of proteins with a particular functionality, and then transfer annotations between two proteins only if they have similar annotation patterns. We show that, in comparison with other techniques, our approach predicts protein annotations more effectively. Our technique (a) produces the highest prediction accuracy of 70-80% precision and recall for different organism specific datasets, and (b) is robust to false positives in protein interaction networks.