A candidate molecular signature associated with tamoxifen failure in primary breast cancer

A first cohort of 35 postmenopausal Caucasian women (age range 41 to 92 years; median age 74 years) with ER-α-positive primary breast cancer diagnosed at the Ninewells Hospital (Dundee, UK), for whom complete clinical and pathological data were available, were studied (Table 1). Eighteen women were considered initially unfit for surgery or declined surgical intervention and received neoadjuvant tamoxifen therapy only (20 mg/day). Tumors were monitored by clinical caliper measurement and mammography. After an initial response, as indicated by a clinical and mammographic reduction in tumor size, the tumors increased in size despite tamoxifen therapy and were removed by surgical excision (tamoxifen failure [TF] group; n = 18) and used in the study. No tissue from these patients prior to endocrine therapy was available. Breast tumor samples from the control group (age-matched women selected from the same geographic and ethnic population who underwent surgical resection before any endocrine therapy, and who did not exhibit any cancer recurrence for 5 years after primary surgical resection and subsequent adjuvant treatment with tamoxifen alone) were excised. Using the χ² test, there was no significant difference in age, histologic grade, lymph node status, or tumor size between the two groups (Table 1). Informed consent was obtained from all patients, and the study was approved by the ethics committee of the institution.

From this cohort, 18 tumor samples (from eight control patients and 10 TF patients), for which sufficient total RNA could be obtained, were used in the initial cDNA array investigation; the whole set (17 control samples and 18 TF samples) was used for RTQ-PCR measurements; and 33 samples were used to conduct the TMA experiments (16 control samples and 17 TF samples; formalin-fixed paraffin-embedded blocks were not available for two patients).

Surgical resection specimens were transported fresh to the adjacent pathology laboratory, and small pieces of tumor tissue were dissected out by a pathologist and snap frozen in liquid nitrogen before storage at -80°C. Approximately 10 mg tissue was homogenized in 750 μl QIAzol lysis reagent (Qiagen Ltd, Crawley, West Sussex, UK). RNA quality was assured using the BioAnalyzer 2100™ (Agilent Technologies, Palo Alto, CA, USA).

Eighteen tumor samples (from eight control patients and 10 TF patients) were used in cDNA array analysis (screening set). Gene expression levels were determined using large-scale measurement experiments using customized nylon cDNA arrays (7.5 × 11.5 cm; 1,034 human genes; 12 genes/cm²) produced in our facility (TAGC Laboratory, University of Aix-Marseille II, France), as previously described [14‐16]. Following hybridization to each array with a ³³P-labeled probe synthesized by reverse transcribing 5 μg total RNA [14], hybridization signals were scanned with a FUJI BAS 5000 beta imager (Raytest, Asnieres, France) and then quantified with the BZScan software, in accordance with the developers' recommendations [17] (TAGC Laboratory, Marseille, France). Intensity values were then adjusted using a normalization step based on the DNA quantification of each spot and the sum of intensities detected in each experiment.

We conducted supervised analyses of genes that could allow discrimination between the two classes of tumor samples (control and TF) by cross-analyzing the results given by three independent methods: supervised analysis using a signal-to-noise metric [18]; significance analysis of microarrays [19]; and Mann-Whitney test (P < 0.05). For each method, we considered the ranks assigned for each gene, and we selected the genes with the best sum of the ranks obtained using the three methods. Expression profiles were then analyzed by hierarchical clustering using the Cluster program developed by Eisen and colleagues [20], and the results visualized using Treeview software (Eisen Laboratory, Berkeley, CA, USA).

Using RTQ-PCR, tumor samples were examined as previously described [15], using a LightCycler^® 1.5 (Roche, Meylan, France) in combination with the LightCycler Faststart DNA Master Sybr Green I (Roche), in accordance to the manufacturer's recommendations. For each gene, the amount of target was calculated as follows by normalization to the expression of the 28S gene and relative to the calibrator: E^{-(ΔCTsample-ΔCTcalibrator)}, where E is the efficiency of the RTQ-PCR reaction calculated with the slope of the corresponding standard curve, C_T is the threshold cycle, and ΔC_T is (C_T target gene – C_T 28S). Statistical analysis of RTQ-PCR measurements was performed using the Mann-Whitney test and the Spearman's rank test, by Statgraphics^® 3 plus software (Statgraphics Centurion, Herndon, VA, USA). The results were judged to be statistically significant at a confidence level greater than 95% (P <0.05).

Thirty-three tumor samples (16 control and 17 TF) were used to construct a TMA, containing up to six 0.6 mm diameter cores from each invasive breast tumor using a manual tissue arrayer (Beecher Instruments Inc., Sun Prairie, WI, USA). Briefly, hemotoxylin and eosin stained tumor sections were reviewed by a single pathologist (KER), and areas suitable for inclusion in the TMA marked. Sections were matched to their corresponding wax blocks (the donor blocks), and 0.6 mm diameter cores of tumor were removed from these donor blocks and inserted into the recipient paraffin TMA block in a grid arrangement.

Four micrometer sections were cut from the TMA block and placed onto poly-L-lysine coated glass slides (VWR International Ltd., Lutterworth, UK) and dried for 1 hour at 60°C, before being de-paraffinized in Histoclear (National Diagnostics, Hessle, UK) and rehydrated through a graded alcohol series. Citric acid buffer (10 mmol/l, pH 6.0) was used as a standard microwave-based antigen retrieval method. Sections were microwaved in a microwave compatible pressure vessel for 15 minutes before being immunostained on a Dako Autostainer Universal Staining System (Dako, Ely, UK) using Vectastain^® ABC kits (Vector Labs, Peterborough, UK), in accordance with the manufacturer's protocol. Briefly, sections were blocked by either normal goat or horse serum containing 10% (vol/vol) from stock avidin solution (Vector Labs) for 20 minutes followed by incubation with primary antibody, including 10% (vol/vol) from stock biotin solution (Vector Labs) for 1 hour to reduce nonspecific background staining. The following anti-human antibodies were used as primary antibodies: anti-insulin-like growth factor binding protein 4 (anti-IGFBP4; ab4252); anti-c-Fos (ab7963); anti-γ-synuclein (anti-SNCG; ab6169; Abcam Ltd, Cambridge, UK); anti-Bcl2 (clone 124; Dako); anti-ER-α clone 6F11 (Vision BioSystems, Newcastle-upon-Tyne, UK); and anti-c-Met (CVD13; Zymed^® Laboratories Inc., Paisley, UK). Sections were then incubated with either biotinylated anti-rabbit or anti-mouse antibody for 30 minutes followed by Vectastain^® Elite ABC reagent for another 30 minutes. Liquid diaminobenzidine (Dako) was used as a chromogenic agent for 5 minutes and sections were counterstained with Mayer's hematoxylin. In between each immunostaining step, slides were washed briefly in Tris-buffered saline (pH 7.6). Sections known to stain positively were included in each batch, and negative controls were prepared by replacing the primary antibody with Tris-buffered saline.

TMA scoring was carried out independently by one of the authors (KER, SMH), and concordance was confirmed by a specialist breast pathologist (CAP) using a Nikon Eclipse E600 light microscope. Antibody staining of cores containing tumor were assessed using a scoring system based on the quickscore method [21]. Briefly, the proportion of positive cells was estimated and given a score on a scale from 1 to 6 (1 = 0% to 4%; 2 = 5% to 19%; 3 = 20% to 39%; 4 = 40% to 59%; 5 = 60% to 79%; and 6 = 80% to 100%). The average intensity of the positively staining cells was estimated and given a score from 0 to 3 (0 = no staining; 1 = weak staining; 2 = intermediate staining; and 3 = strong staining). A quickscore was then calculated by multiplying the percentage of cells staining score by the intensity score, to yield a minimum value of 0 and a maximum value of 18.

A separate cohort of 33 Caucasian women (age range 31 to 77 years; median age 55.5 years) with ER-positive primary breast cancer diagnosed at Centre Léon Bérard (Lyon, France) were selected (Table 2) and provided by the Centre de Ressources Biologiques of the Centre Léon Bérard (Lyon, France). The breast tumor samples were excised from women who did not receive endocrine therapy, chemotherapy, or radiotherapy before surgery. Complete clinical, histologic, and biologic information was available. All patients received postoperative adjuvant endocrine therapy alone for 5 years (tamoxifen 20 mg/day) and no chemotherapy. Fourteen patients relapsed under tamoxifen treatment (relapsing group) and 19 patients did not have a recurrence after 5 years of tamoxifen treatment (nonrelapsing group). Informed consent was obtained from all patients. Using the χ² test, there were no significant differences between groups in age, histologic grade, lymph node status, or tumor size (Table 2).

RNA extraction was performed as described above. RTQ-PCR experiments were conducted using a LightCycler 480^® (Roche) in combination with the LC480 SybrGreen I Master Mix (Roche), in accordance with the manufacturer's recommendations. The expression of the six genes was investigated using the same pair of primers, and the same calculation and normalization methods as described above. Statistical analysis of RTQ-PCR measurements was performed using the Mann-Whitney test using the Statgraphics^® 3 plus software (Statgraphics Centurion). The results were judged statistically significant at a confidence level greater than 95% (P <0.05).

For each gene, the 33 ER-positive breast tumors were then divided into two groups: one of 16 tumors with 'low' mRNA level (lower than the median of the mRNA levels of the 33 breast tumor samples) and another of 17 tumors with 'high' mRNA level (higher than the median of the mRNA levels of the 33 breast tumor samples). Outcomes of interest were overall survival (OS) and relapse-free survival (RFS). OS was measured from the date of diagnosis to death or censored at the last follow up. RFS was measured from the date of diagnosis to relapse or censored at the last follow-up. Survival distributions were estimated using the Kaplan-Meier method and the significance of differences between survival rates was ascertained by the log-rank test, using the SPSS^® Software (SPSS Inc., Chicago, IL, USA). Candidate prognostic factors for RFS with a 0.05 significance level in univariate analysis were entered in a multivariate Cox model, and a backward selection procedure was used to build the final model [22]. Likelihood ratio test was used to select the best fit between models [23].

Because this study is an exploratory analysis, all of the statistical analyses performed in this work were done at the 0.05 significance level, and no correction was applied for multiple testing.

cDNA arrays were used to identify candidate genes associated with tamoxifen failure. From a training set of 18 tumor samples (eight control and 10 TF), total RNA was extracted from each sample and used to synthesize the corresponding complex probe to be hybridized on the cDNA arrays. With the aim being to identify a molecular signature that might allow discrimination between the two classes of tumor samples (control and TF), a cross-analysis based on three different statistical methods (significance analysis of microarrays, signal-to-noise statistic method, and Mann-Whitney test) was applied to the normalized cDNA array data. For each method, the ranks assigned to each gene were considered, and the genes selected with the best sum of the ranks obtained using the three methods. Forty-nine discriminant genes arose, and hierarchical clustering of the gene expression profiles discriminated between the two groups of patients (control and TF; Figure 1). Two out of the 49 discriminating genes (HLA-DRA and STAT1) were selected by two individual spots located at different places on the array, emphasizing the reproducibility of the gene signature identified.

Functional annotation of the 47 genes in the signature showed that 34% of the genes were involved in immune response (B2M, CCL2, CCL3, CD22, CD33, CXCL12, HLA-A, HLA-DPB1, HLA-DRA, IL10RA, IL16, IL2RG, IL6, ILF2, TCRA, and TNFRSF7), 23% in transcription regulation (ELF1, ESR1/ERα, FOS, JUN, JUNB, MAZ, NFYA, PBX1, RXRA, STAT1, and ZNF607), 23% in cellular proliferation regulation and mitosis control (DLEU2, IGFBP4, MAP2K2, MET, NF1, NOTCH4, PAK1, SNCG, SSSCA1, TEK, and TGFBR2), 9% in tumor invasion (AMFR, FLT1, MMP13, and MMP16), 4% in apoptosis (BCL2 and FAS), 4% in cell adhesion (CD34 and RDX), and 3% in other functions. Sixteen genes encoded proteins of the immune system: major histocompatibility complex proteins (B2M, HLA-A, HLA-DPB1, HLA-DRA, and TCRA), cell antigen (CD22 and CD33), receptors (IL2RG, IL10RA, and TNFRSF7), cytokines and chemokines (IL6, IL16, CCL2, CCL3, and CXCL12), and a transcription factor (ILF2). In addition, three of the 23% of genes involved in transcription regulation were genes that take part in the activator protein-1 transcription complex (FOS, JUN, and JUNB). Interestingly, expression of 17 genes from the 47-gene signature selected in our study (Figure 1 [red asterisks]) was associated with estrogen action, because their expression was modulated by estradiol treatment in breast cancer cells in vitro [15, 24‐31]. This suggests that identification of estradiol-regulated genes [32] or tamoxifen-regulated genes [33]in vitro might be a good approach to selecting prognostic or predictive molecular markers of ER-positive breast cancer [32‐34].

Considering the rank assigned to each gene following the statistical cross-analysis performed on the cDNA array data (as described in the Materials and methods section [above]), ESR1/ERα emerged as the most discriminating gene, as expected [35‐38]. We then investigated the expression of ESR1/ERα and that of the five following top-ranked genes (MET, FOS, SNCG, IGFBP4, and BCL2) by RTQ-PCR on the initial set of 18 tumor samples (eight control and 10 TF). Using Spearman rank correlation, for five genes (ESR1/ERα, FOS, IGFBP4, MET, and SNCG) the data demonstrated a positive and significant (P < 0.05) correlation between mRNA levels measured by cDNA array and RTQ-PCR, indicating consistency between the cDNA array and the RTQ-PCR measurements (Table 3).

To explore the reliability of the expression signature previously identified by cDNA array, RTQ-PCR of ESR1/ERα, FOS, IGFBP4, MET, BCL2, and SNCG gene expression was examined in a larger set of 35 tumor samples (17 control and 18 TF), including the 18 samples used in the initial screen. The data presented in Table 4 demonstrate a significant difference of expression between the control and TF groups for the six genes (Mann-Whitney test). As expected, ESR1/ERα emerged as the most significant gene (P < 10^-6). mRNA levels of four genes (IGFBP4, SNCG, BCL2, and FOS) were significantly lower in the TF groups, whereas MET mRNA levels were significantly higher in the TF group.

Using the Mann-Whitney test, the expression of the six genes was compared with clinical and pathological parameters Table 1). ESR1/ERα (P = 0.002) and BCL2 (P = 0.003) genes had significantly higher expression in axillary node metastasis-negative patients (n = 14) than in node-positive patients (n = 21). Moreover, gene expression levels of ESR1/ERα (P = 0.004), BCL2 (P = 0.003), and FOS (P = 0.01) were significantly lower in grade III (n = 11) than in grade I + II (n = 22) tumor samples. Finally, MET mRNA levels were significantly increased (P = 0.016) in tumor grade III samples compared with grade I + II samples.

In order to explore, at the protein level, the expression variations of the six top-ranked genes and to ascertain whether immunohistochemical detection of these proteins on tissue sections could differentiate between the control and the TF groups, immunohistochemistry was performed on TMAs containing cores of tissue from 33 out of the 35 patients (16 control and 17 TF; Figure 2). For two samples (one control and one TF), tumor sections were not available. A Spearman rank correlation test between the TMA data and the RTQ-PCR measurements revealed a significant and positive correlation for ESR1/ER-α (P < 10^-4), Bcl2 (P = 0.003), and IGFBP4 (P = 0.05; data not shown). The association of protein immunohistochemical detection with the control or TF group was studied using the Fisher exact test (Table 5). There was a significant statistical association between ER-α (P = 0.0004), SNCG (P = 0.02), and IGFBP4 (P = 0.03) immunohistochemical staining with patient group (control or TF). No significant association was observed for c-Fos, c-Met, and Bcl2 protein detection (P > 0.05).

We examined an independent cohort of ER-positive breast cancer patients from a different geographic location (Centre Léon Bérard, Lyon, France) to assess the pertinence of the biomarkers identified in the study. The separate cohort (n = 33) included 14 patients who relapsed under tamoxifen treatment (relapsing group) and 19 patients who did not relapse after 5 years of tamoxifen treatment (nonrelapsing group). In this cohort, low mRNA levels of FOS, BCL2, SNCG, and IGFBP4 were significantly associated with tamoxifen failure (P < 0.05, Mann-Whitney test; Table 6). MET was the only biomarker that could not be confirmed.

We then used univariate analysis (log-rank test) to study further the prognostic value of the biomarkers identified. Apart from (as expected) ESR1/ERα (P < 10^-4), univariate analysis revealed that low expression mRNA levels of FOS (P < 10^-4), BCL2 (P < 10^-4), SNCG (P = 0.008), and IGFBP4 (P = 0.038) were significantly associated with shorter OS (Figure 3 and Table 7). Low expression mRNA levels of ESR1/ERα, BCL2, and FOS were also significantly associated with shorter RFS (P < 10^-4; Figure 4 and Table 8), whereas low expression of SNCG and IGFBP4 was close to statistical significance (P = 0.078 and P = 0.083, respectively; Figure 4). No significant association between MET mRNA level and OS (P = 0.879) or RFS (P = 0.449) could be identified. In Cox multivariate regression analysis of OS, only the prognostic significance of ESR1/ERα (P = 0.032; hazard ratio [HR] = 0.17, 95% confidence interval [CI] = 0.03 to 0.86) and FOS (P = 0.049; HR = 0.22, 95% CI = 0.06 to 1.00) persisted (Table 7). Cox multivariate regression analysis of RFS revealed that BCL2 (P < 10^-4; HR = 0.10, 95% CI = 0.03 to 0.34) and FOS (P = 0.038; HR = 0.15, 95% CI = 0.03 to 0.89) were independent prognostic markers and that patients with low mRNA levels of BCL2 or FOS were at greater risk for relapse (Table 8).

We then defined a two-gene signature based on BCL2 and FOS mRNA expression levels. This signature was constructed by selecting patients who expressed low mRNA levels of both BCL2 and FOS (group A; n = 14) versus patients exhibiting high expression levels of at least one of these two genes (group B; n = 19). Among these 19 patients, concomitant high levels of both BCL2 and FOS were detected in 15 samples. Concerning the four patients with high expression levels of either BCL2 or FOS, the Kaplan-Meier curve for RFS could be superimposed over those of the patients having high levels of both BCL2 and FOS. The resulting Kaplan-Meier curves for RFS according to the BCL2/FOS signature are illustrated in Figure 5 (univariate analysis, P < 10^-4; HR = 0.014, 95% CI = 0.002 to 0.117).

Finally, we tested the BCL2/FOS signature with respect to RFS to determine which final model is better. The model with the BCL2/FOS signature was better fitting (likelihood = 90.35) than the model with ESR1/ERα alone (likelihood = 109.28; P < 10^-4) or with only BCL2 (likelihood = 102.64; P = 0.0005) or FOS (likelihood = 114.47; P < 10^-4), demonstrating that our two-gene signature has a better prognostic value than ESR1/ERα alone for RFS. We then tested the model with both BCL2/FOS and ESR1/ERα and found that the combination was better fitting (likelihood = 89.82) than the model with ESR1/ERα alone (likelihood = 109.28; P < 10^-4), demonstrating that the BCL2/FOS signature could improve the prognostic value of ESR1/ERα for RFS.