A series of 212 primary breast cancer cases were studied; 80 of these tumors were analyzed using cDNA microarrays, along with one normal breast tissue sample collected from breast reduction surgery. Patient samples were sequentially collected at Ullevål University Hospital from 1990 to 1994 (IRB approval 350, protocol 75026). The last update of patient information was done in 2006, providing an observation time of 12 to 16 years. Patients were followed until death or emigration, and only 12 patients were lost to follow-up. The average age of the 80 cases analyzed by cDNA microarrays was 65.0 years at time of primary surgery (range 28.2 to 87.7 years), similar to the average age of 64.4 years (range 28.2 to 91.5 years) for the total series. The 80 cases were selected from the total series based only on sufficient amount of fresh frozen tissue for microarray analysis. Consequently, a slightly higher fraction of patients with larger tumor size was observed in this subcohort. A summary of the clinical and histopathological data of the patients is shown in Table 1 (see Additional file 1 for more detailed information). All patients were treated according to Norwegian national guidelines at the time of diagnosis 19. Patients receiving adjuvant systemic therapy were given nine courses of CMF (cyclophosphamide, methotrexate, 5-fluorouracil) and/or Tamoxifen for two years. Dosage of radiation given as adjuvant treatment was dependent on indication; after breast conserving therapy the mammary gland was given 50 Gy (2 Gy × 25). The number of samples entered into the survival analyses is smaller than 212 (full dataset) and 80 (subset with gene expression data); excluding patients with missing information or distant metastases at the time of diagnosis and primary surgery, leaves us with a maximum number of 200 (full set) and 77 (subset) patients.

Primary breast carcinoma tissue was snap frozen and stored at -80°C. Frozen sections stained with hematoxylin/eosin were reviewed to confirm tumor content, and specimens in which at least 5% of the cells were carcinoma cells were included in this study. The majority of samples (80%) analyzed using microarrays had at least 40% tumor cell content. Sections from paraffin embedded tissue were re-evaluated by an experienced breast pathologist (JMN) to classify and grade the carcinomas according to the modified Scarff-Bloom-Richardson method 20 (Table 1).

DNA was isolated from both peripheral blood cells (leukocytes) and tumor tissue using chloroform/phenol extraction followed by ethanol precipitation (Nuclear Acid Extractor 340A; Applied Biosystems, Foster City, California, USA) according to standard procedures. TP53 mutation detection in tumor DNA was performed by prescreening exon 2–11 using temporal temperature gradient gel electrophoresis (TTGE) 21. Samples with aberrant migrating bands from TTGE were sequenced (ABI PRISM™ 377 DNA Sequencer, Applied Biosystems) to determine the nature of the mutation. AI analysis was performed using the ABI 310 capillary electrophoresis on two different highly polymorphic markers in the TP53 locus, one located in intron 1 22 and the other downstream of exon 11 23. AI was scored according to a threshold of 0.84 (ratio between two allele variants in tumor divided by ratio between two alleles in blood) 24. At least one of the polymorphic markers had to show AI to score positive. The ER and PR were analyzed using both immunohistochemistry (IHC) and biochemical/ligand-binding assay (Abbott Diagnostics, Abbott Park, Illinois, USA). The results from IHC were used in our data analysis since that is the current recommended method; however, in a few cases where IHC had not been performed, results from a biochemical/ligand binding assay were used.

Total RNA was isolated from snap frozen tumor tissue using TRIzol^® solution (Invitrogen™, Carlsbad, California, USA). The concentration of total RNA was determined using an HP 8453 spectrophotometer (Hewlett Packard) and the integrity of the RNA was assessed using a 2100 Bioanalyzer (Agilent, Santa Clara, California, USA). Linear amplification of total RNA was performed using an optimized protocol previously described 25. Amplified tumor RNA was labeled by Cy5 and amplified RNA from Universal Human Reference total RNA (Stratagene^®, La Jolla, California, USA) was labeled by Cy3. The labeling and hybridization of amplified RNA to cDNA microarrays, containing more than 42,000 elements, was performed as previously described 25. Experimental protocols can be found at the referred web site 26. Images of the arrays was acquired using a Gene Pix 4000B scanner (Axon Instruments, Sunnyvale, California, USA), and analyzed using Gene Pix Pro 3.0/4.0/4.1 software and by visual inspection. Data were entered in the Stanford Microarray Database 27, and intensity levels were normalized. Data were filtered for spot quality and included in the analysis if the pixels within a spot showed a regression correlation of at least 0.6 or if the signal intensity of both sample and reference were at least 1.5 over background. A hierarchical clustering algorithm integrated into the Stanford Microarray Database was applied to group genes and samples on the basis of their similarities in expression, and the results were visualized using TreeView 28 and Java TreeView 29 software (for analysis software links, see 30). Prior to clustering analysis, the data were centered to median expression of each gene across the dataset. The hierarchical clustering shown in Figure 1 was performed using previously described 'intrinsic' genes 3 consisting of 540 clones (representing 496 genes corresponding to a single unique UniGene cluster) whose expression was measurable in at least 70% of the samples. The 'intrinsic' set of genes consisted of genes with significantly greater variation in expression between different tumors than between paired samples from the same tumor, the name reflecting that genes were selected to optimally identify intrinsic characteristics of breast tumors.

In the analysis of breast cancer death for all patients and for patients with gene expression data, we used the Kaplan-Meier estimator and univariate Cox regression to assess the marginal effect of each factor (that is, when not correcting for the effect of other factors). The joint effect of two or more factors was assessed using multivariate Cox regression. A parsimonious Cox regression model with only significant factors was obtained by backwards elimination starting with all factors, and the final model was checked for all possible two-factor interactions. The P value for the total effect of a factor was calculated from the (partial) likelihood ratio statistic, while the Wald test statistic was used to compute the P value for each level of a multilevel factor. The proportional hazards assumption of Cox regression was checked using the test of Grambsch and Therneau 31 with Kaplan-Meier weights as implemented in S-PLUS (version 6.1 for Windows. Insightful Corporation, Seattle, Washington, USA).

Pearson correlation analysis was used (Microsoft Excel 2000) to find the correlation between the gene expression profile of a single sample and five previously defined centroids 3. The five centroids were based on a set of tumor samples from advanced breast cancer cases and represent the average expression profiles of sample subgroups defined by hierarchical clustering using genes that showed more similar expression between two samples from the same tumor than between any other tumor ('intrinsic' genes). Cross-tabulation and Pearson X²-test or Fishers Exact test (when appropriate) were performed using SPSS (version 13.0. SPSS Inc, Chicago, Illinois, USA) when studying distribution of clinical, histopathological or molecular genetic parameters. Genes with potential significant changes in expression between tumors having TP53 mutations and tumors with wild-type TP53 were identified using the significance analysis of microarray (SAM) procedure 3233. Data from 27,393 clones whose expression was measurable in at least 80% of the samples were included in the analysis.

The 80 breast tumor samples were assigned to five different subgroups according to their gene expression pattern; luminal A, highly proliferating luminals, normal-like, basal-like and ERBB2⁺. The assignment of tumors into subgroups was based on hierarchical clustering using the 540 previously identified 'intrinsic' genes. The resulting dendrogram showed two main groups of breast tumors (Figure 1); those with high expression of the ER-related gene cluster and luminal type characteristics (left branch), and those showing a lower relative expression of the ER-related cluster (right branch). Further subdivision of the samples identified five groups similar to those previously described 2 also in this set of breast tumors. The correlation between the gene expression profile of each sample and five previously described centroids 2 were calculated and a scatter chart was made (Figure 2) according to the order of samples in the dendrogram from hierarchical clustering (Figure 1). The correlation of each sample to each of the centroids showed a continuous wave pattern over the sample set, and visualized how each sample carries elements from different profiles. The luminal A and basal-like breast tumors showed a pronounced opposite proportional pattern.

In the subsequent Cox regression analysis the breast tumor samples were assigned to five subgroups based on hierarchical clustering, combined with the pattern of centroid correlation and fine-tuned by gene expression pattern of characteristic subgroup markers. It should be emphasized that the classification method is unsupervised, meaning that the samples are grouped solely based on the gene expression data.

In Table 2 (Univariate analysis) relative risks (hazard ratios) with 95% confidence intervals (CI) of univariate Cox regressions for all factors considered are shown. P values for testing the hypothesis of no marginal effect of a factor and its levels are also given. Tumor size and TP53 status are the two most significant factors, but lymph node status also has a significant effect. Patients with TP53 mutations have a breast cancer death rate that is four to five times higher than for those without mutations. Patients with tumor size T3–T4 have a breast cancer death rate that is about four times higher than for those with tumor size T1, while patients with tumor size T2 have a breast cancer death rate that is about double that of those with size T1. Patients with four or more positive lymph nodes have a breast cancer death rate that is about three times higher than for those with negative lymph nodes. The effect of TP53 status and tumor size on breast cancer death is also shown in the Kaplan-Meier plots of Figure 2a. The high number of patients with the heterogeneous grade 2 probably explains why grade is not as strong a prognostic marker as expected in this series of samples. In a multivariate Cox regression model obtained after elimination of all non-significant factors, the TP53 status, tumor size and lymph node status are the only significant remaining factors (Table 2, Multivariate analysis), and their effects are about the same as in the univariate analysis.

By performing univariate Cox regressions on the samples included in the array experiment for all factors considered, TP53 mutation status, gene expression group, tumor size and lymph node status were all significant factors for survival (Table 3, Univariate analysis). The effect of tumor size is somewhat larger than for the analysis of all patients, while the effect of TP53 status is somewhat smaller. The gene expression groups have a large effect on survival. The breast cancer death rates for the highly proliferating luminal, the basal-like and the ERBB2⁺ groups are about six times higher than for the luminal A group, while the breast cancer death rate for the normal-like group is almost three times that of the luminal A group. As the assumption of proportional hazards is violated for the highly proliferating luminals, the relative risk estimate for this group should be interpreted as an average over the time period considered. The effect of TP53 mutation status and tumor size on breast cancer death is also shown in the Kaplan-Meier plots in Figure 3b, while Figure 3c gives the Kaplan-Meier plot for the five gene expression groups. The non-proportionality of the hazard of highly proliferating luminals is clearly seen in the latter.

In a multivariate Cox regression model obtained after elimination of all non-significant factors, TP53 mutation status and lymph node status are the significant remaining factors (Table 3, Multivariate analysis). Gene expression groups and tumor size were the last factors to be eliminated. TP53 mutation status and gene expression group are closely related; 17 of the 20 patients with TP53 mutation are in the basal-like or ERBB2⁺ group. This makes it difficult to separate the effect of TP53 from the effect of the gene expression groups. TP53 mutation and lymph node status are the strongest predictors of survival in the multivariate analysis, with effects of about the same size as in the univariate analyses.

Our set of samples was also classified according to three previously described gene lists 489. By clustering the samples using the genes overlapping with our arrays, the samples were separated into two main branches in each dendrogram; predicted to be a good prognosis group and a poor prognosis group. Although the genes on our arrays did not overlap with all genes from the respective lists and the respective score procedure in each paper was not followed, it was interesting to see whether the gene lists had significance using this simple approach. In univariate Cox regression analysis two of the gene lists were significant in predicting breast cancer survival (15/21 genes: relative risk (RR) 3.70, 95%CI 1.64 to 8.34, P = 0.0007; 63/76 genes: RR 2.24, 95%CI 1.09–4.61, P = 0.028), while the third was close to being significant despite the limited number of overlapping genes (26/70 genes: RR 1.99, 95%CI 0.93 to 4.26, P = 0.067). None of the classifiers were significant when included in multivariate analysis together with the variables listed in Table 3, a result where the interpretation should be made according to the previous stated limitations of our approach.

The most striking characteristic of the basal-like and ERBB2⁺ subclusters (Figure 4) was that most cases carried a TP53 mutation in their tumor. In the basal-like dendrogram branch 83% (10/12) of the carcinomas harbored a TP53 mutation, and in the ERBB2⁺ subcluster the fraction was 70% (7/10). In the two basal-like tumors with wild-type TP53 (Figure 4) the mutations may have been missed, or other components of the pathway may have been affected. The basal-like subcluster showed a higher correlation between the samples in the dendrogram compared to other subclusters, as well as higher correlation to the 'basal-like centroid' (Figure 2), which may reflect the strong impact of a TP53 mutation on the global transcription program.

Most of the TP53 mutations detected among the 80 samples analyzed on the microarray (17/20, 85%) and among the total series of samples (39/48, 81%) were missense mutations, which is the type of mutation most frequently found in TP53 34. Only three samples outside the basal-like and ERBB2⁺ clusters were affected with TP53 mutations; two missense mutations outside the DNA binding domain (codon 113, codon 138) affecting the highly proliferating luminals and one frequent missense mutation in the DNA major groove interacting domain (codon 273) affecting the luminal A group. Figure 4 further shows that IHC detected only 50% (10/20) of the mutations detected by TTGE. Almost half of the samples analyzed had AI (LOH) of TP53 in their tumor tissue (array: 19/41, 46%; total: 47/98, 48%), and LOH was strongly associated with the presence of a TP53 mutation (p < 0.001). Among the samples analyzed using microarray analysis, the samples with the highest frequency of AI seemed to cluster in the outermost three subgroups with the poorest outcome (basal-like, ERBB2⁺ and highly proliferating luminal groups). Interestingly, the highly proliferating luminals showed a high frequency of AI (7/8, 88%) despite a low frequency of mutations in the TP53 gene.

Although the relative expression level of TP53 mRNA measured using microarrays showed variation between individual samples, the basal-like and ERBB2⁺ groups, which had the most TP53 mutants, clearly showed the highest mRNA expression, while the luminal group had intermediate expression and the normal-like group had a lower expression (the data were centered to median expression across the dataset; Figure 4). Two of the three mutated samples falling outside the basal-like and ERBB2⁺ groups had mutations that gave a lower relative expression of TP53 mRNA.

SAM analysis was performed to find the gene expression pattern specific for tumors containing a TP53 mutation. With an estimated median number of false negatives being zero (delta slider 1.56 and fold change 2.0), 377 significant clones were selected when analyzing a set of 27,393 cDNA clones (Additional file 3). The highly specific gene expression pattern associated with TP53 mutation status is illustrated in Figure 5, where hierarchical clustering of the 80 samples and 80 selected genes (the 40 most significantly upregulated and 40 most downregulated genes of the 377 genes) are shown. Many genes that showed higher relative expression in carcinomas with a TP53 mutation were involved in the cell cycle and cell proliferation (for example, CCNB2, CDCA5, CENPA, and UBE2C), while the genes with lower relative expression showed more diverse functions and were highly associated with ER status (for example, IRS1, ESR1, DNAL1 and NAT1). Some of the genes (for example, MYBL2, CDCA8, DNALI1 and DACH1) were also identified recently by Miller and colleagues 35 as part of a 32-gene expression signature that distinguishes TP53 mutant and wild-type tumors. Further investigation of the gene expression pattern of different TP53 mutations is needed to understand more about the different effects they have in breast cancer.

The distribution of clinicopathological markers between the gene expression subgroups is shown in Figure 4. The subgroups that showed the poorest survival had the largest difference in median age, members of the basal-like and ERBB2⁺ groups being the youngest (median age 56 years (28 to 75) and 60 years (47 to 87), respectively) and those of the highly proliferating luminal group the oldest (median age 75 years (59 to 82)). Dividing the patients into two equally large groups (<65 years and ≥65 years) gave a significant skewed distribution between the five gene expression groups (p = 0.019), but there was no significant difference when dividing them into pre- and post-menopausal women (<55 years and ≥55 years) (Figure 4). Tumor size did not show a statistically significant skewed distribution between the gene expression groups, suggesting that the gene expression patterns provide new and different information compared to this established marker. Neither was node status associated with any particular gene expression group.

The histological type and grade, as well as ER and TP53 mutation status all showed a highly significant skewed distribution between the gene expression groups (p < 0.001; Figure 4). The ERBB2⁺ and basal-like groups showed the highest fraction of grade 3 tumors (80% and 58%, respectively). Although carcinomas of the luminal A and highly proliferating luminal subgroups were mainly grade 2, they differed by the fact that the luminal A group included 22% of cases with grade 1 and no grade 3 cases, while the highly proliferating luminal group included 27% of cases with grade 3 and no grade 1 cases. Among the luminal samples, 89% were ER positive, while none of the basal-like and only 30% of the ERBB2⁺-group were ER positive. The relative mRNA expression of ER (centered over the dataset) was compared to protein expression (IHC) and showed a strong correlation (Figure 4).

Tissue samples with a low fraction of malignant epithelial cells were included in this study to increase our understanding of how this may affect a tumor's characteristic gene expression pattern. Samples with different percentages of malignant epithelial cells were distributed among all subclasses, although an over-representation (not statistically significant) of tumors with low tumor content were seen in the normal-like subgroups (Figure 4). The invasive lobular carcinomas tended to have lower tumor content than invasive ductal carcinomas. Hierarchical clustering of samples with ≥40% tumor cells results in the same subgroups as when all samples were included. These results suggest that the gene expression profile of a grossly dissected tumor sample in many cases is not heavily influenced by a relatively low quantity of malignant cells, and that tumor stromal cells also may affect the gene expression profile.

Invasive ductal carcinomas were distributed among all subclusters, but entirely dominated the highly proliferating luminal, basal-like and ERBB2⁺ groups (Figure 4). Invasive lobular carcinomas were divided into two groups, those clustering with the normal-like subgroup, previously referred to as 'typical-lobular' 36, and the 'ductal-like lobular' clustering with the two luminal groups (14/22 and 7/22, respectively). The only lobular samples clustering on the edge of the basal-like group were alveolar lobular with pleomorphic areas. Two tubulolobular carcinomas clustered together in the highly proliferating luminal subgroup, and three mucinous carcinomas clustered together in the luminal A subgroup, showing that the phenotypes distinguished by pathologists are also distinct on a molecular level. A fourth sample, an invasive ductal carcinoma, was part of the 'mucinous cluster,' showing a gene expression pattern similar to these tumors despite the fact that no typical mucinous elements in paraffin-embedded or frozen tissue sections from this patient were seen. The mucinous samples were all ER positive by IHC, TP53 wild type, ERBB2 negative and E-cadherin positive. A ductal carcinoma in situ with microinvasion and a metaplastic carcinoma clustered in the normal-like subgroup. The invasive ductal carcinoma sample from a man (ULL-020) clustered with the highly proliferating luminal group.