Several studies have shown that the prediction of protein function using PPI data is promising. However, the PPI data generated from experiments are noisy, incomplete and inaccurate, which promotes to represent PPI dataset as an uncertain graph. In this paper, we proposed a novel algorithm to mine maximal dense subgraphs efficiently in uncertain PPI network. It adopted several techniques to achieve efficient mining. An extensive experimental evaluation on yeast PPI network demonstrated that our approach had good performance in terms of precision and efficiency.