Clustering methods have been often used to find biologically relevant groups of genes or conditions based on their expression levels. Since many functionally related genes tend to be coexpressed, by identifying groups of genes with similar expression profiles, the functionalities of unknown genes can be inferred from those of known genes in the same group. In this paper we address a novel clustering approach, called seed-based clustering, where seed genes are first systematically chosen by computational analysis of their expression profiles, and then the clusters are generated by using the seed genes as initial values for k-means clustering. The seed-based clustering method has strong mathematical foundations and requires only a few matrix computations for seed extraction. As a result, it provides stability of clustering results by eliminating randomness in the selection of initial values for cluster generation. Our empirical results reported here indicate that the entire clustering process can be systematically pursued using seedbased clustering, and that its performance is favorable compared to current approaches.