BACKGROUND: Similarity based computational methods are a useful tool for predicting protein functions from protein-protein interaction (PPI) datasets. Although various similarity-based prediction algorithms have been proposed, unsatisfactory prediction results have occurred on many occasions. The purpose of this type of algorithm is to predict functions of an unannotated protein from the functions of those proteins that are similar to the unannotated protein. Therefore, the prediction quality largely depends on how to select a set of proper proteins (i.e., a prediction domain) from which the functions of an unannotated protein are predicted, and how to measure the similarity between proteins. Another issue with existing algorithms is they only believe the function prediction is a one-off procedure, ignoring the fact that interactions amongst proteins are mutual and dynamic in terms of similarity when predicting functions. How to resolve these major issues to increase prediction quality remains a challenge in computational biology. RESULTS: In this paper, we propose an innovative approach to predict protein functions of unannotated proteins iteratively from a PPI dataset. The iterative approach takes into account the mutual and dynamic features of protein interactions when predicting functions, and addresses the issues of protein similarity measurement and prediction domain selection by introducing into the prediction algorithm a new semantic protein similarity and a method of selecting the multi-layer prediction domain. The new protein similarity is based on the multi-layered information carried by protein functions. The evaluations conducted on real protein interaction datasets demonstrated that the proposed iterative function prediction method outperformed other similar or non-iterative methods, and provided better prediction results. CONCLUSIONS: The new protein similarity derived from multi-layered information of protein functions more reasonably reflects the intrinsic relationships among proteins, and significant improvement to the prediction quality can occur through incorporation of mutual and dynamic features of protein interactions into the prediction algorithm.