We studied 240 tumors from a cohort of 445 patients surgically treated for stage II breast cancer (age 31 to 81 years), diagnosed in the South Swedish Health Care Region between 1985 and 1994 and originally participating in two randomized clinical trials 1617. All patients received 2 years of adjuvant tamoxifen treatment, without stratification according to ER status. The median follow-up time for patients alive and free from metastasis at the last follow-up visit was 5.3 years. The current study was approved by the Lund University Medical Ethics Committee.

Fresh-frozen tumor tissue was used for routine determination of cytosolic ER and PgR, as well as the S-phase fraction, using an enzyme immunoassay and DNA flow cytometry, respectively, as described earlier 1819. Cores of 0.6 mm diameter formalin-fixed, paraffin-embedded tumor tissue were used to generate tissue microarrays (TMAs) for the 445 cases. Three cores from every individual tumor were arrayed. These TMAs have been used for immunohistochemical staining of CK5, CK14, EGFR and cytokeratin clone AE1/AE3 as described previously 1220. A pathologist re-evaluated the histological type on whole formalin-fixed paraffin-embedded sections 21.

An additional TMA consisting of tumors from 23 BRCA1 germline mutation carriers, diagnosed in Sweden between 1980 and 2001, was used for evaluation purpose. This TMA was generated as described above, including two or three cores from each tumor.

Sections (4 μm) of the TMA blocks were mounted on Dako REAL™ Capillary Gap Microscope Slides (DAKO, Glostrup, Denmark), were deparaffinized in xylene and were rehydrated in ethanol. Antigen retrieval was achieved either by placing slides in Tris-ethylenediamine tetraacetic acid buffer (pH 9.0) at 125°C in a 2100 Retriever (PickCell Laboratories, Amsterdam, the Netherlands) for 5 minutes (CD44/CD24), or by microwaving the slides in Tris-ethylenediamine tetraacetic acid buffer (pH 9.0) for 7 minutes at 800 W followed by 15 minutes at 350 W (HER2).

Double-immunostaining with antibodies for detection of CD44 and CD24 was performed by an Autostainer (DAKO) using EnVision G|2 Doublestain System Rabbit/Mouse (DAB⁺/Permanent Red) (DAKO) according to the manufacturer's instructions. Antibodies for the detection of CD44 (Clone 156-3C11, 1:800) and CD24 (Clone SN3b, 1:400) were purchased from Neomarkers (Fremont, CA, USA). CD44 was detected with Permanent Red and CD24 was detected using diaminobenzidene (DAB).

HER2 was detected with a rabbit polyclonal primary antibody (A0485, 1:1,000; DAKO) followed by EnVision™ on a TechMate™ (DAKO). All slides were counterstained with hematoxylin for the identification of nuclei.

The scoring was performed twice by one person in a blinded fashion. All unclear cases were discussed with a pathologist. In case of discrepant staining between the three cores from the same patient, an average was used.

HER2 scoring was carried out according to the standard procedure (DAKO): 0, < 10% of the tumor cells showed membranous staining; 1, > 10% of the tumor cells were positive but not circumferential; 2, weak staining around the whole membrane in > 10% of the tumor cells; and 3, strong staining around the membrane in > 10% of the tumor cells. CD44 staining was detected mainly in the membrane and the scoring was as follows: 0, 0% positive tumor cells; 1, 1% to 10% positive cells; 2, 11% to 50% positive cells; 3, 51% to 75% positive cells; and 4, 76% to 100% positive cells. CD24 staining was detected mainly in the cytoplasm and the scoring was performed as described for CD44.

The proportion of CD44⁺/CD24^- tumor cells was determined as the percentage of cells positive for Permanent Red staining but negative for DAB staining. The frequencies of CD44^-/CD24⁺ cells and of CD44⁺/CD24⁺ cells were determined in a similar fashion.

Associations between the presence of CD44, CD24 or different CD44/CD24 phenotypes and clinical variables as well as breast cancer subgroups were assessed by Fisher's exact test, except for age where the Mann-Whitney U test was used. The Kaplan-Meier method was used to estimate distant disease-free survival, and the log-rank test to compare survival between two strata. All tests were two-sided and P < 0.05 was considered significant. Statistical analyses were carried out using Stata 10.0 software (Stata Corporation, College Station, TX, USA).

For 168 of the 445 tumors in our cohort, mRNA expression analysis has previously been performed using cDNA microarrays with 27,648 reporters 2223. The microarray data for these 168 tumors are available through the Gene Expression Omnibus database (accession numbers GSE6577 and GSE5325). Data pre-processing and filtering for the selected 168 tumors were performed using the BioArray Software Environment 24 as previously described 23, leaving 15,040 reporters that were used for all subsequent analyses.

Three independent sets of gene signatures were used to further characterize the tumors. Genes were matched across datasets based on gene symbols, and matched genes were centralized in our dataset across the 168 tumors. Nearest centroid classifiers were used for the Sørlie and colleagues' 9 and the Hu and colleagues' 10 subtype classifications, with each tumor classified based on to which centroid it was most correlated using Pearson correlation. The average expression level for the genes in a signature was used to characterize tumors for the Shipitsin and colleagues' gene signatures 5.

Hierarchical clustering was performed using MeV in the TM4 system 25 with complete linkage, Pearson correlation distance and gene centralization. Genes with different expression levels in tumors containing CD44⁺/CD24^- cells and tumors lacking cells with this phenotype were identified using a two-sided t test. To account for multiple comparisons, the false discovery rate was calculated for gene lists. Sixty-nine of the 168 tumors with gene expression profiles were among the 240 tumors for which CD24 and CD44 immunohistochemical stainings were obtained.

We analyzed the presence of CD44 and CD24 antigens on human breast cancer tissues using double-staining immunohistochemistry. The CD44 and CD24 expression was successfully determined in 240 cases after excluding tumors with scarce tissue on the TMA. These 240 tumors did not significantly differ from the excluded 205 tumors in regard to tumor size, lymph node status, S-phase fraction, ER status or disease-free survival. The median age was slightly higher in the included patients (64 years versus 62 years for excluded patients), although not reaching significance.

Figure 1 displays representative staining patterns of various breast tumors. Tumor cells were distinguishable from stromal cells and inflammatory cells (generally CD44⁺ and CD24^-) by expression of cytokeratin clone AE1/AE3 in adjacent sections. The frequency of tumors with varying degrees of CD44⁺ and CD24⁺ tumor cells is presented in Table 1. CD44 showed mainly membranous staining (Permanent Red) with only six tumors displaying cytoplasmic staining, four of which had also membrane staining. Thirty-two percent (77/240) of the tumors had ≥ 1% cells with membranous and/or cytoplasmic CD44 expression and were considered CD44⁺. CD24 was almost exclusively detected in the cytoplasm, with only four tumors displaying membrane DAB staining (all four with positive cytoplasmic staining as well). Forty-six percent (110/240) of the tumors had ≥ 1% cells with CD24 staining and were considered CD24⁺.

To determine the proportion of putative tumorigenic CD44⁺/CD24^- cells within each tumor, we scanned for the presence of Permanent Red staining without any DAB interference. CD44⁺/CD24^- tumor cells were detected in 31% (75/240) of the tumors, with the proportion of this phenotype ranging from only a few cells to almost all tumor cells (Figure 1a to 1c). The frequency of tumors with different proportions of CD44⁺/CD24^- tumor cells is presented in Table 1. We also determined the proportion of CD44^-/CD24⁺ and double-positive (CD44⁺/CD24⁺) cells in each tumor (Table 1). CD44⁺/CD24⁺ cells (Figure 1c) were detected in 6% (15/240) of the tumors, while CD44^-/CD24⁺ cells (Figures 1a, d) were detected in 45% (108/240) of tumors. The majority of tumors included at least a few double-negative (CD44^-/CD24^-) tumor cells, while a complete lack of both proteins was evident in 35% (83/240). Thirteen percent (30/240) of the tumors were positive for more than one of the CD44⁺/CD24^-, CD44^-/CD24⁺ and CD44⁺/CD24⁺ phenotypes, with 5% (11/240) displaying tumor cells of all three phenotypes. In 12% (28/240) of the tumors, the majority (> 50%) of tumor cells was CD44⁺/CD24^-.

HER2 expression was successfully determined in all 240 cases with CD44/CD24 data, and 16% (38/240) were considered strongly positive (score = 3). Information on other clinical variables and patient survival was available from previous studies. The CD44⁺/CD24^- status was significantly correlated to biological characteristics such as low/negative HER2 expression (P = 0.002), elevated expression of CK5/14 (P = 0.012) and EGFR (P = 0.007) (Table 2). Interestingly, all medullary tumors (n = 8) were positive for the CD44⁺/CD24^- phenotype. The presence of CD44⁺/CD24^- tumor cells was not associated with ER and PgR status, with S-phase fraction or with tumor size, nor with lymph node status, menopausal status or age, although showing a tendency of being more common in premenopausal women. Using higher cutoff criteria for the CD44⁺/CD24^- status (10% or 50%) and comparing these 50 tumors or 28 tumors, respectively, with all other tumors gave more or less comparable results, although overall slightly higher P values (data not shown). The only additional significant association was with ER-negative status using a 10% cutoff value for the CD44⁺/CD24^- status (P = 0.041).

The presence of CD44^-/CD24⁺ tumor cells was solely associated with strong HER2 staining (P = 0.002) and not with any other tumor characteristics. The presence of double-positive (CD44⁺/CD24⁺, n = 15) tumor cells was not associated with any tumor features, although an increase of tumors of medullary type was indicated compared with tumors lacking cells with this phenotype (23% versus 2%).

We did not see any correlation between CD44⁺/CD24^- status and distant disease-free survival, nor between CD44⁺/CD24^- status and site of distant recurrence. Factors significantly correlated to favorable distant disease-free survival at 5-year follow-up (in this cohort of patients receiving adjuvant tamoxifen therapy) included positive ER and PgR status (P = 0.005 and P = 0.037, respectively) and negative CK5/14 and EGFR status (P = 0.035 and P = 0.005, respectively), but not HER2 and lymph node status, while tumor size and a high S-phase fraction reached marginal significance. Lymph node status, however, was correlated to distant disease-free survival in the entire cohort of 445 tumors (P = 0.005).

The CD44⁺/CD24^- phenotype was clearly related to certain tumor biological characteristics. To study this relation in more depth, we used five protein markers available for 232 out of the 240 cases with CD44/CD24 data to define five tumor subgroups. Tumors positive for ER and/or PgR were designated steroid receptor positive (SR+). The SR and HER2 status was used to broadly divide tumors into four subgroups; SR+HER2- (n = 150), SR+HER2+ (n = 14), SR-HER2+ (n = 24), and SR-HER2- (triple negative, n = 44). The latter group was further subdivided into a subgroup expressing basal CK5/14 and/or EGFR (SR-HER2- basal-like, n = 30), and a subgroup negative for all five markers (SR-HER2- nonbasal, n = 14).

Gene expression data were available for 69 of the 232 tumors, which allowed us to correlate our five subgroups, as defined by the five tumor markers, with breast cancer subtypes defined by gene expression profiling and intrinsic gene lists 910. As seen in Figure 2, SR+HER2- tumors were clearly enriched for luminal A tumors, while SR-HER2- basal-like tumors corresponded prominently to the basal-like subtype defined by gene expression profiling. As expected, SR-HER2+ tumors correlated well with the HER2+ subtype. The SR-HER2- nonbasal subgroup showed no clear association with any tumor subtype – and only two tumors were SR+HER2+, making it difficult to draw any conclusions for this subgroup. We therefore conclude that tumor classification based on a combination of five commonly used tumor protein markers is biologically relevant for subgroup analysis in our tumor material.

The expression of CD44 and CD24 differed significantly between the subgroups (P = 0.001 and P = 0.035, respectively) (Table 3). CD44 was highly expressed in the SR-HER2- basal-like subgroup, with 63% of the tumors being positive compared with 32% for the entire cohort, and was very lowly expressed in the HER2+ groups, with only 14% and 17% of the tumors being positive in the SR+HER2+ and SR-HER2+ groups, respectively. CD24 was highly expressed in the SR-HER2+ group (75% compared with 47% for the entire cohort).

The frequency of tumors positive for the CD44⁺/CD24^- phenotype varied significantly between the different subgroups (P < 0.001, Table 3). While close to two-thirds (63%) of tumors resembling the basal-like subtype (SR-HER2- basal-like) expressed CD44⁺/CD24^- cells, this phenotype was very low in HER2+ tumors, regardless of SR status: 8% in SR-HER2+ tumors and 14% in SR+HER2+ tumors (Table 3). The divergence between subgroups remained when the cutoff level for CD44⁺/CD24^- was raised to 10% or 50% (P = 0.003 and P = 0.033, respectively). The frequency of CD44⁺/CD24^- cells within tumors positive for this phenotype was also higher in basal-like tumors, as indicated in Figure 2 for the 69 tumors with gene expression data.

The proportion of tumors with CD44^-/CD24⁺ cells also differed significantly between subgroups (P = 0.027), with the highest proportion in the two HER2+ groups (SR-HER2+ and SR+HER2+) (Table 3). The presence of double-positive (CD44⁺/CD24⁺) and double-negative (CD44^-/CD24^-) tumor cells did not differ between subgroups (Table 3).

Since there was a clear correlation between the CD44⁺/CD24^- status and the basal-like tumor subtype, we extended the immunohistochemical analysis to an additional TMA including tumors from BRCA1 hereditary breast cancer patients, known to be of predominantly basal-like phenotype 9. A basal-like status was verified for nine of our BRCA1 hereditary tumors for which gene expression data were available.

Seventeen of the 23 BRCA1-defective tumors were successfully stained for CD44 and CD24 expression. The frequency of tumors with different proportions of CD44⁺/CD24^- cells is presented in Table 4. Ninety-four percent (16/17) of the tumors were positive for this phenotype, thus corroborating the finding of a high proportion of CD44⁺/CD24^- cells in the SR-HER2- basal-like subgroup (Table 4). BRCA1 germline mutant tumors, however, also had a high proportion (70%) of cells positive for the CD44^-/CD24⁺ phenotype, considerably higher than the 40% seen among SR-HER2- basal-like tumors.

We correlated the presence of CD44⁺/CD24^- cells to two gene signatures with prognostic value specific for either CD44⁺ or CD24⁺ breast cancer cells published by Shipitsin and colleagues 5. Signatures A and B are associated with shorter and longer distant recurrence-free survival time, respectively. A positive correlation of CD44⁺/CD24^- cells to signature A was seen if the cutoff level for CD44⁺/CD24^- was 50% or 75% positive cells (P = 0.05 and P = 0.008, respectively) but not when using lower cutoff levels. A negative correlation to signature B was seen when using cutoff levels of 0%, 10% or 75% positive cells (P < 0.001, P < 0.001 and P = 0.01, respectively).

To identify a gene expression signature for the CD44⁺/CD24^- phenotype, we used the 69 tumors with gene expression data and looked for genes differentially expressed between tumors containing CD44⁺/CD24^- cells and tumors lacking such cells. The top 20 genes are displayed in Table 5. Interestingly, CD44 emerged as the second ranked gene, demonstrating a good correspondence between protein and mRNA levels for this gene – although the top 20 genes collectively had a relatively high false discovery rate of 28%.