The tumors in both studies could be partitioned into two majors subgroups based on their estrogen receptor (ER) status as suggested by others (4, 5). Additionally, these expression cassettes could provide a refined estimate of prognosis, perhaps beyond those clinical indicators currently available to us. Sorlie et al. (3) identified a luminal subgroup (subgroup A) of ER-positive tumors associated with the best outcome. van't Veer et al. (3) addressed this problem, by investigating a narrow subset of node-negative breast cancer patients. They found 231 genes significantly associated with disease outcome as defined by the presence of distant metastasis at the 5-year mark. Van de Vijver et al. (6) provided a validation of the van't Veer predictor applied to 234 new patients from the same institution and using the same array platform.

Relapse-free survival (RFS) was defined as the interval elapsed between the date of breast surgery and the date of diagnosed further episode of breast cancer, whether the breast cancer was classified as a recurrence or second primary, and whatever the histology. Breast cancer survival (BCS) was defined as the interval elapsed between the date of breast surgery and the date of breast cancer-related death (documented from hospital records). ER status was determined by using ligand-binding assays and immunohistochemistry. Grade was determined by using the Elston-Ellis grade system (7).

RNA Extraction and Probe Preparation. Isolation of RNA was performed by using the TRIzol method (Invitrogen) according to the manufacturer's instructions. RNA quality from each tumor biopsy was assessed by visualization of the 28S/18S ribosomal RNA ratio on 1% agarose gel. Total RNA was linearly amplified by using a modification of the Eberwine method (8, 9). Total RNA from the Universal Human Reference (Stratagene) was amplified and used as reference for cDNA microarray analysis. The cDNA microarray chips consisted of 7,650 total features and were manufactured at the National Cancer Institute microarray facility.

A detailed protocol for RNA amplification and cDNA probe labeling and hybridization is available at http://nciarray.nci.nih.gov/reference/index.shtml. genepix software (Axon Instruments, Union City, CA) was used to analyze the raw data, which were then uploaded to a relational database maintained by the Center for Information Technology at the National Institutes of Health.

Data Analysis. Images of all of the scanned slides were meticulously inspected for artifacts, and aberrant spots and slide regions were flagged for exclusion from analyses. Log (base 2) ratios for each spot were calculated as follows. In each channel, signal was calculated as foreground median minus background median. If the signal was <100 in any single channel, the signal value in that channel was set to 100. If the signal was <100 in both channels, the spot was flagged as unreliable and not used in any further analyses. Also, if >50% of the pixels in the foreground in either channel reached the saturation threshold, the spot was flagged and not used in analyses. For all remaining (nonflagged) spots, a log ratio was calculated as log₂[(red signal)/(green signal)]. The log ratios were then normalized within each array by subtracting from each the median log ratio value across the spots on the array. The channel-specific intensity data and normalized log ratios of all 99 experiments are available in Tables 3-5, which are published as supporting information on the PNAS web site.

The first phase of the analysis was to compare expression profiles between specimens segregated according to values of standard prognostic variables. In particular, we considered the following comparisons: tumor grade 1 or 2 vs. 3; tumor size ≤2 cm vs. >2 cm; age <50 years vs. age ≥50 years (menopausal status); node negative vs. node positive; and ER-vs. ER+. These comparisons were made by parametric t tests using the statistical software splus (SPLUS 6.0 Professional, Insightful, Seattle). To control for multiple comparisons, we reported as significant genes only those that reached significance at level P = 0.001. Testing 7,650 probes at this significance level, we expect that the average number of spuriously significant (false positive) results will be eight or less.

Cluster analyses were conducted to search for natural groupings in the profiles. Before clustering, a screening procedure was applied to eliminate genes showing minimal variation across the set of 99 specimens. Specifically, for each gene, the 5th and 95th percentiles of the ratios were calculated. If the ratio of the 95th to 5th percentile was <3 that gene was not included in the cluster analysis. This process left 706 probe elements for the cluster analyses. Hierarchical agglomerative clustering using the statistical package brb-arraytools software (available at http://linus.nci.nih.gov/BRB-ArrayTools.html) was applied to these normalized log ratios by using both compact linkage and average linkage and both Euclidean and one minus Pearson correlation distance metrics. Normalized log ratios were median-centered within each gene for all of the cluster analyses. The clustering results obtained by using compact linkage with one minus Pearson correlation distance applied to the 706 probe elements appeared by visual inspection to yield the most distinctive clusters (remaining blinded to any clinical or outcome variables), and hence this was the clustering algorithm used for the unsupervised cluster analyses based on these probe elements (Fig. 1). The presence of significant clustering was assessed by applying the global test of clustering proposed by McShane et al. (10). The same techniques were applied for the clustering analyses using gene subsets sets derived from the van't Veer and Sorlie studies (Figs. 3, 4, 5, 6 which are published as supporting information on the PNAS web site).

Unsupervised Clustering Identifies Natural Groups Corresponding to Breast Lineage Markers. Using an unsupervised hierarchical clustering approach, we sought to define natural subclasses of breast tumors as determined by gene expression profiles. We performed the unsupervised clustering on the 706 probe elements selected as exhibiting high variability across all tumors (see Materials and Methods). Included in this list are genes corresponding to the standard prognostic parameters, such as ER and HER-2/neu status. Application of the global test of clustering and reproducibility measures of McShane et al. (10) showed borderline statistically significant evidence for clustering (P = 0.057), and the robustness indices suggested (robustness >75%) that the most reproducible clustering structure was evident when the dendrogram was cut at two clusters. Our results show that the tumor samples could be confidently separated into two main groups primarily associated with ER status as determined by the ligand-binding assay and confirmed by immunohistochemistry (Fig. 1). The entire cluster diagram including gene expressions is included in Fig. 7, which is published as supporting information on the PNAS web site. Eighty-two percent of the tumors in the left main branch are ER-, and 88% of the tumors in the right main branch are ER+ by immunohistochemistry. This finding corroborated the earlier analysis (Table 1) that the major factor discriminating the expression phenotype is ER status. As expected, the ER-cluster had a higher percentage of high-grade tumors than the ER+ cluster (Fig. 1).

The dendrogram further branched into smaller subgroups within the ER+ and ER-classes. Although our sample size was not large enough to establish high reproducibility of smaller subgroups, we were able to locate in our dendrogram several subgroupings previously identified in other studies. We describe here what we observed in our data relative to those clusters that have been reported by others. Within the ER-cluster are tumors with “basal”-like expression characteristics as defined by higher gene expression of keratin 5, keratin 6, metallothionein 1X, and fatty acid binding protein 7 as reported (Table 8, which is published as supporting information on the PNAS site) (2, 18). Furthermore, they exhibited higher expression of the secreted frizzled-related protein 1 (SFRP1) and the oncogene c-kit and lower expression of fibronectin 1 and mucin 1. Basal 1 subgroup was differentiated by higher expression of matrix metalloproteinase 7 and cell growth-related genes such as topoisomerase II α, mitotic feedback control protein Madp2 homolog (MAD2L1), cell division control protein 2 homolog (CDC2), and PCNA, suggesting a signature for a high proliferation rate (Table 8). In contrast, the basal 2 subgroup was distinguished by higher expression of many components of the transcriptional factor AP-1 such as c-fos, c-jun, and fos B, as well as by overexpression of activating transcription factor 3, caveolin 1 and 2, hepatocyte growth factor, and transforming growth factor β receptor II. A subgroup distinct from the basal-like groups in the ER-subset is defined by a high rate of HER-2/neu overexpression (HER-2/neu: 5/7 tumors). This HER-2/neu subgroup was further distinguished from the basal-like subgroups by the higher expression of MDR1, S100 calcium-binding protein P, fatty acid synthase, RAL-B, RAB6A, fibronectin 1, and syndecan 1 and lower expression of c-kit and c-myc.

The ER+ subgroup showed differential expression of genes associated with ER activation such as LIV-1, trefoil factor 3, neuropeptide Y receptor Y1, keratin 8, GATA3, and X-box binding protein 1. These are also genes that define the ER+ cluster as having “luminal” characteristics as defined (2, 18). Moreover, the ER+ cluster could be further segregated into three smaller subclasses: luminals 1, 2, and 3 (Fig. 1).

We then asked whether the array-derived tumor groups demonstrated any differences with respect to survival. When the basal/neu (predominantly ER-) and the luminal-like (predominantly ER+) cluster were compared by Kaplan-Meier and Cox regression analysis, the luminal-like subgroup had a significant advantage in both RFS and BCS (Fig. 2 A and B). The three subclasses within the basal/neu (ER-) cluster appeared to have similarly poor survival characteristics (Fig. 2 C and D). By contrast, the three subgroups within the luminal-like (predominantly ER+) cluster showed distinct differences in survival (Fig. 2 E and F). Luminal 1 had the best outcome with an 80% 10-year RFS. This luminal 1 subgroup was also correlated with lower-grade tumors and was further characterized by differential higher expression of c-kit, hepatocyte growth factor, insulin-like growth factor-binding protein-3, ATF-3, and components of the AP-1 transcriptional factor such as c-fos, c-jun, fosB, and jun-D, and as expected by lower expression of cell growth related genes such as topoisomerase II α, mitotic kinesin-like protein-1, PCNA, CDC2, BUB1, and MAD2L1 (Table 8). This was significantly different from the luminal 2 subgroup that had the worst outcome with a 10-year RFS of 40% (P = 0.022 luminal-like 1 vs. luminal-like 2). This subgroup showed higher expression of a protein tyrosine phosphatase type IVA member, tumor necrosis factor receptor-associated factor 3, RAD21, and BRCA1-associated protein 1 (BAP1) and lower expression of FGFR1, CXCR4, ATF-3, and vascular cell adhesion molecule 1. The luminal 3 subgroup had an intermediate survival outcome of 60% at 10 years. Intriguingly, although some of the ER+ tumors were HER-2 overexpressors (6 of 66), they were dispersed among the luminal subgroups.

In the Sorlie et al. study (3), tumor classification was performed based on the 456 cDNA clones (427 unique genes) in their “intrinsic” gene list. To assess how these genes could perform in classifying our tumor set, we sought the overlap between either the intrinsic gene list and our cDNA array or the overlap between the intrinsic gene list and the 706 variably expressed probe elements in this study. We found 332 (285 unique genes) that overlapped with the full set of probe elements on our cDNA array and 105 (96 unique genes) probe elements that overlapped with our 706-gene list. Interestingly, in both situations hierarchical cluster analysis segregated our tumors into two distinct subgroups mainly based on their basal (predominantly ER-) and luminal (predominantly ER+) characteristics (Figs. 5 and 6).

The luminal-like tumors were further segregated into at least two (possibly three) smaller subgroups, which may correspond to luminal A, B, and C subtypes. Additional analysis revealed that 100% of the specimens in the luminal A subtype in the 332-gene cluster and 77% in the 105-gene cluster were assigned as luminal-like 1 tumors in our unsupervised cluster analysis (based on the 706-element cluster, Fig. 1). In contrast, the specimens in the luminal B and C subtypes were assigned either as luminal-like 2 or luminal-like 3 tumors.

Although the luminal A subtype showed a favorable clinical outcome when compared with other luminal subtypes, this difference was not statistically significant.

Identification of Gene Clusters Associated with Survival. Our database includes the survival data of all patients studied with a median follow-up of 6.1 years. To determine the genes associated with improved RFS, we performed Cox proportional hazards regressions on each of the 7,650 probe elements on our array. A group of 485 different probe elements were identified that could separate RFS in the 99 patients with a P value <0.05, but only 16 probe elements were significant at the more stringent P < 0.001 level used to control for multiple comparisons (Table 9, which is published as supporting information on the PNAS web site). We discuss results for the larger list of 485 so that our results may be compared with results of previous studies. Of the genes reported to be associated with prognosis in breast cancer, only PCNA and topoisomerase II α were found to be elevated in the poor prognostic group (HER2-neu/basal-like). Interestingly, this gene list did not include the ER, Her-2/neu, or p53. This is consistent with the findings of van't Veer et al. (3), who found no association between HER2 and ER expression levels and survival in the node-negative cases included in their study.

Here, we present an analysis on 99 tumors from node-positive and node-negative patients, the majority receiving adjuvant treatment according to accepted practice guidelines at the time of the diagnosis. Our results were significant in their concordance with those of the earlier studies despite the differences in patient populations, treatments used, and technology platforms used. Thus, our results provide supporting evidence for the prognostic importance of genes identified in previous reports on a completely independent patient cohort with an independent microarray platform. We found that the ER status of the tumor was, indeed, the most important discriminator of expression subtypes and that tumor grade was a distant second. Other clinical features, namely lymph node positivity, menopausal status, and tumor size were not strongly reflected in the expression patterns obtained with the 7,650-feature microarray in this investigation. This finding confirms that ER biology plays a central role in breast carcinogenesis defining the configuration of the final tumor. Furthermore, investigation of gene expression in primary tumors may be unlikely to identify a set of genes whose expression reliably correlates with lymph node metastasis. This finding is consistent with data showing that only a small fraction of cells in a tumor mass have metastatic potential (19, 20). The genetic signature from this metastatic fraction would be “diluted” by the signals from nonmetastasizing cells.

Similar to the findings of Sorlie et al. (2), unsupervised hierarchical clustering analysis segregated the tumors into two main clusters based on their basal (predominantly ER-) and luminal (predominantly ER+) characteristics. Furthermore, within each of these clusters we could identify smaller subgroups that were characterized by distinct gene expression signatures involving potential different oncogene-specific pathways. A HER-2/neu subgroup was characterized by higher expression of the oncogene her-2/neu and higher expression of genes involved in the ras pathway such a Ras-related GTPases, RALB, and RAB6A. Convergence of neu and ras pathways in breast cancer tumorigenesis has already been documented (21). In contrast, basal 1 and 2 subgroups were characterized by higher expression of the oncogenes c-kit, c-myc, and SFRP1. SFRP1 is a modulator of Wnt signaling. Recently, aberrations of Wnt signaling were reported to be involved in the pathology of various human neoplasms (22). Activation of the Wnt signaling pathway appears to lead to the cytosolic stabilization of a transcriptional cofactor, β-catenin, that can regulate the transcription from a number of target genes including the cellular oncogene c-myc. In breast carcinoma, SFRP1 expression has been associated with loss of ER and the presence of lymphoplasmocytic reaction around the tumor associated with a more aggressive disease (23, 24). Furthermore, basal 1 type exhibited higher expression of genes involved in cell cycle and growth such as PCNA, CDC2, and BUB1 whereas the basal 2 type showed higher expression of transcription factors such as c-fos and ATF3, expression signatures that both could be modulated by c-myc (21). These data raise the possibility that Myc, which is amplified in 15% of breast cancers, may have a more important role in determining the expression profile of a breast cancer than previously thought.

Moreover, as noted in earlier studies, our basal and luminal subgroups also showed the expected differences in survival with a better outcome in the luminal group. More interestingly, we found that the 231 genes described by van't Veer et al. (3) as separating survival groups in node-negative untreated patients may have distinct prognostic capabilities in a more heterogeneous population of node-positive/negative patients treated with adjuvant therapy. Using the 93 probe elements, 56 unique genes, from the van't Veer prognostic set represented in our microarrays, we could easily separate our 99 patients into two prognostic subsets. This finding appears to confirm the importance of some subset of these 56 genes as bona fide prognostic markers. In particular, the overlap between the van't Veer 231 genes and our 485 probes associated with survival (at P < 0.05 level) was 14 probe elements representing 11 unique genes. Five of the 11 genes in this set are involved in cell replication and chromosomal stability and were up-regulated in the poor prognostic setting, suggesting a molecular mechanism for this clinical outcome. An intriguing question is what might be a “minimal” set of genes necessary to establish a reproducible prognostic classification. In leukemia, with well-defined genetic changes, gene profiles segregate with particular translocations (25). In our breast cancer series the major groupings follow previously defined signaling pathways such as ER+, ER-, and c-erbB2/ras. Our study design did not permit us to relate BRCA1 and BRCA2 status to these expression patterns.