Protein functions are rather difficult to predict, i.e., proteins are often multifunctional and promiscuous, and the functional annotations of proteins are incomplete and error prone [2,3,21]. There are many avenues for future improvements for protein function prediction. For example, incorporating pathway information, evolutional knowledge, and reducing the noisy in the individual networks before the integration. In addition, an input network can have some high quality subnetworks and low quality subnetworks; it is promising to design algorithms to discover and differentiate these subnetworks, and to integrate the high quality subnetworks to enhance the prediction accuracy, and to discard the low quality ones (or assign very small weights to these subnetworks) to reduce their destructive effects.