Although distinct pathological stages of breast cancer have been described, the molecular differences among these stages are largely unknown. Here, through the combined use of laser capture microdissection and DNA microarrays, we have generated in situ gene expression profiles of the premalignant, preinvasive, and invasive stages of human breast cancer. Our data reveal extensive similarities at the transcriptome level among the distinct stages of progression and suggest that gene expression alterations conferring the potential for invasive growth are already present in the preinvasive stages. In contrast to tumor stage, different tumor grades are associated with distinct gene expression signatures. Furthermore, a subset of genes associated with high tumor grade is quantitatively correlated with the transition from preinvasive to invasive growth.

Tamoxifen significantly reduces tumor recurrence in certain patients with early-stage estrogen receptor-positive breast cancer, but markers predictive of treatment failure have not been identified. Here, we generated gene expression profiles of hormone receptor-positive primary breast cancers in a set of 60 patients treated with adjuvant tamoxifen monotherapy. An expression signature predictive of disease-free survival was reduced to a two-gene ratio, HOXB13 versus IL17BR, which outperformed existing biomarkers. Ectopic expression of HOXB13 in MCF10A breast epithelial cells enhances motility and invasion in vitro, and its expression is increased in both preinvasive and invasive primary breast cancer. The HOXB13:IL17BR expression ratio may be useful for identifying patients appropriate for alternative therapeutic regimens in early-stage breast cancer.

Gene expression profiles of thousands of genes can now be examined en masse through cDNA and oligonucleotide microarrays 1-3 . Recently, studies have been reported that examined gene expression changes in yeast 4,5 , as well as in mammalian cell lines 6 , primary cells 7 and tissues 8 . However, present applications of microarray technology do not include the study of gene expression from individual cell types residing in a given tissue/organ (that is, in situ). Such studies would greatly facilitate our understanding of the complex interactions that exist in vivo between neighboring cell types in normal and disease states. We demonstrate here that gene expression profiles from adjacent cell types can be successfully obtained by integrating the technologies of laser capture microdissection 9 (LCM) and T7-based RNA amplification 10 with cDNA microarrays 11 . Neighboring small and large neurons are individually capturedTo demonstrate this integration of technologies, we examined the differential gene expression between large-and small-sized neurons in the dorsal root ganglia (DRG). In general, large DRG neurons are myelinated, fast-conducting and transmit mechanosensory information, whereas small neurons are unmyelinated, slow-conducting and transmit nociceptive information 12 . We chose this system because numerous differentially expressed genes (small versus large) have been reported, thus the success of this experiment could be assessed; and because many small and large neurons are adjacent to each other, thus we could test whether individual neurons can be cleanly captured. Large (diameter of >40 µm) and small (diameter <25 µm and with identified nuclei) neurons were cleanly and individually captured by LCM from sections (10 µm in thickness) of Nissl-stained rat DRG (Fig. 1). For this study, two sets of 1,000 large neurons and three sets of 1,000 small neurons were captured for cDNA microarray analysis. RNA amplification is reproducible between individual capturesRNA was extracted from each set of neurons and linearly amplified (independently) an estimated 10 6 -fold using T7 RNA polymerase. After being amplified, one fluorescently labeled probe was synthesized from an individually amplified RNA (aRNA), divided equally into three parts and hybridized in triplicate to a microarray ('chip') containing 477 cDNAs (see Methods for chip design) plus 30 cDNAs encoding plant genes (for the determination of non-specific nucleic acid hybridization). Expression in each neuronal set (called S1, S2 and S3 for small and L1 and L2 for large neurons) was thus monitored in triplicate, requiring a total of 15 microarrays. The quality of the microarray data is demonstrated by pseudocolor arrays, one resulting from hybridization to probes derived from neuronal set S1 and the other from neuronal set L1 (Fig. 2a). In Fig. 2a, the enlarged part of the chip shows some differences in fluorescence intensity (that is, expression levels) for particular cDNAs and demonstrates that spots containing the different cDNAs are relatively uniform ...