In the post-genomics era in which gene sequences have been decoded, large-scale protein-protein interaction data are generated with the rapid development of system biology experiments. It is important in functional genomics to search for function modules and predict protein functions from the data. A new method called modularized clustering method(MCM), which are based on the direct and second-order interactions of modules, is applied to the latest high-throughput protein-protein network of yeast to predict the function of unknown proteins in the modules. P value of hypergeometric cumulative distribution of modules and the disturbance analysis on the data, including adding, removing and rewiring interactions, are employed to evaluate the prediction quality and robustness of the method. The results show that MCM has high prediction precise rate and coverage, and it is robust to high false-positive data and missing data. The predicted results of unknown proteins with high prediction precise rate can be instructive in biological analysis and the algorithm can be generalized to other networks with the similar structures.