Abstract：Prediction of protein function is the most challenging problem of the post-genomic era. The methods on the basis of protein-protein interaction networks are widely used at present. But when the interacting partners is small the success rate is decreased. Here we propose a "Small-world Networks"-based algorithm (SWN-BA) to infer protein function from Saccharomyces cerevisiae protein-protein interaction network, and get higher success rate with　SWN-BA comparing with Global Optimization (GO) method when . The thought of the "Small-world Networks" can well be used in the study in protein function from protein-protein interaction networks especially when k is small.