Breast cancer metastasis can be predicted by morphometric quantification of mitotic activity in the tumor perimeter according to the Mitotic Activity Index (MAI). We used Affymetrix HuFL GeneChips to compare approximately 6,000 RNAs from 12 high-MAI (10 mitoses per 10 h.p.f., high metastatic risk) and 11 low-MAI, premenopausal, lymph-node-negative, primary breast cancers. We identified 51 discriminating genes as having t-statistics above the randomly permuted data set background (Permax 0.96) and means differing by 100 expression units, and a twofold ratio. The largest category consisted of cell cycle and division genes, usually (12/13) those with the highest levels of expression in high-MAI tumors. Internal validation of the list was provided by the following: one gene independently selected by two probe sets during analysis; five pairs of genes known to be co-expressed and one instance of reciprocal regulation of ligand and receptor. Genes that may contribute to divergent aggressiveness between groups include seven membrane-associated genes participating in cell–cell interactions, five extracellular matrix genes, and five genes encoding matrix-modifying proteolytic enzymes. A graphical classification model was tested using a 10% jackknife and 50 genes selected in each cycle. The mean level of expression in the low-MAI group and the difference of means between the groups were rank-ordered. These two variables were plotted on a triangular coordinate system against actual expression values for each tumor. Distinctive bimodally shifting cloud patterns of data points were used to assign 78% (18/23) of jackknifed tumors to their correct MAI groups. Expression profiling can be used to discover genes associated with breast cancer metastasis, and these may be of practical use in prospective risk classification.