The accuracy of the histological classification of cells plays a determining role in disease diagnosis and treatment. Recent studies have shown that the distribution of chromatin-associated proteins reflects alterations in cell phenotype. Using 3D fluorescence images of cultured human breast epithelial cells with multiple known phenotypes, we have developed an automated method to classify the phenotype of epithelial cells based on their nuclear protein distribution. Features which describe the distribution of specific nuclear proteins are first measured, on a per nucleus basis, by our local bright feature (LBF) analysis technique. Features from thousands of nuclei with multiple, known phenotypes were then grouped by a novel voting-based clustering method into a number of clusters of similar pattern. This allows us to establish the statistical link between clusters and the phenotypes of the cells. Finally, we used this statistical link to predict the probable phenotype of individual or groups of nuclei. The results show that the combined use of 3D confocal imaging, image feature analysis, and clustering analysis provides an efficient way to predict the phenotype of epithelial cells based on the nuclear distribution of chromatin-associated proteins.