Correlation of microarray-based breast cancer molecular subtypes and clinical outcomes: implications for treatment optimization

Fresh frozen breast cancer tissue from every third patient diagnosed and treated between 1991 and 2004 at the Koo Foundation Sun-Yat-Sen Cancer Center (KFSYSCC) were randomly selected for the study. Patients with follow-up periods shorter than three years were excluded, with the exception of those who died of the disease within three years of the initial treatment. In cases of ineligibility, the following sample was selected. The selected tissue samples spanned the major transition periods of adjuvant chemotherapy from CMF (cyclophosphamide, methotrexate and fluorouracil) to CAF (cyclophosphamide, doxorubicin, fluorouracil) and to taxane-based regimens. Four hundred forty seven samples were obtained, but 135 samples were excluded due to insufficient RNA (n = 1), poor RNA quality (n = 116), or unacceptable microarray quality (n = 18). A total of 312 samples were eligible for the study (Cohort 1). Gene expression profiles of an additional 15 lobular breast carcinoma samples, collected between 1999 and 2004 and previously studied, were also included (Cohort 2). All patients were treated by a multidisciplinary team according to the guidelines consistent with the National Comprehensive Cancer Network [18]. Following modified radical mastectomy or breast-conserving surgery plus dissection of axillary nodes, patients received radiotherapy, adjuvant chemotherapy, and/or hormonal therapy, if indicated. Neoadjuvant chemotherapy was administered to patients with locally advanced disease. The study was approved by the institutional review board (ID number 20020128A) and ethical approval was obtained from the same board for samples without obtainable informed consent.

Total RNA was isolated using Trizol (Invitrogen, Carlsbad, CA) and purified with the RNeasy Mini Kit (Qiagen, Valencia, CA). RNA quality was assessed using an RNA 6000 Nano Kit and an Agilent 2100 Bioanalyzer (Agilent Technologies, Waldbronn, Germany). The RNA samples used for the study had an average RNA Integrity Number of 7.85 ± 0.99 (mean ± SD). Hybridization targets were prepared from total RNA according to the Affymetrix protocol and hybridized to U133 plus 2.0 arrays. The expression intensity of each gene was scaled to a trimmed-mean of 500, logarithmically transformed to base 2 and normalized using quantile normalization. The dataset and MIAME compliant information had been deposited in the GEO database (GSE20685).

Although the classifier genes of Perou-Sørlie intrinsic types [2] could be applied to our datasets, such direct application, crossing to a different microarray platform for molecular subtyping could compromise the robustness and accuracy of the classification. To establish a reliable classification method specific to the Affymterix microarray platform, we decided to develop and validate a platform-specific methodology for the molecular subtyping of breast cancer. From the literature, we selected 23 pivotal genes known to play important roles in the biology of breast cancer (Additional file 1, Table S1), and subsequently conducted linear and quadratic correlations with each of the 23 pivotal genes for all probe-sets. The probe-sets showing significant degree of correlation with any of the pivotal genes were further selected according to their expression intensities, range of expression, and density plot kurtosis. Finally, 783 probe-sets were selected and used for molecular subtyping (Additional file 1, Table S2). The procedures associated with probe-set selection and two-step k means clustering for classification are detailed in the methodology in the supplemental files (Additional File 2).

The genes used for our molecular subtyping were applied to three independent datasets for validation [10, 19, 20]. Genes corresponding to our classification probe-sets were identified in the published datasets. If one probe-set was mapped to multiple genes in the independent datasets, the average intensity was calculated and applied. Centroid analysis was used to determine subtypes of breast cancer [5]. Hierarchical clustering analysis was conducted to examine whether the same subtypes identified in ours and three other independent datasets shared the same differential expression patterns for genes of wound-response [9], tumor stromal reaction [21], tumor vascular endothelial normalization [22, 23], and cell cycle proliferation (Additional file 1, Table S3).

In addition to examining the relationship between the molecular subtypes of breast cancer identified in this study and various clinical parameters, our classifier genes were also applied to the other two published independent breast cancer datasets for confirmation [10, 24]. In addition, we used the reported genes of the OncotypeDX [8] and MammaPrint [3] predictors to assess the risk of distant recurrence for cases in all three datasets. For prediction of recurrence risk by the genes of OncotypeDx predictor, we adopted the same statistical predictive model used by Paik et al. [8]. The molecular subtypes were correlated with the predicted risk of recurrence. The procedures of these studies are detailed in the methodology section of Additional file 2.

To quantitatively determine the status of ER, PR, and HER2, we used the intensity of gene expression measured by a microarray, because not all of the patients had results for ER, PR, and HER2 by immunohistochemistry (IHC). The values of gene expression used to determine the positive or negative status of ER, PR, and HER2 were based on density plots of 312 breast cancer samples in Cohort 1 (Additional file 3, Figure S1). Bimodal distribution was observed for all three genes, and the cut-points were statistically determined, according to the method described in the methodology section of Additional file 2. Studies into the correlation between the results of IHC and gene expression for the status of ER and HER2 showed significant positive correlations (Additional file 3, Figure S2). This finding supports the approach of using the intensity of gene expression to determine the status of ER, PR, and HER2.

All statistical analysis was conducted using the SAS/STAT software (ver. 9.1.3) (SAS Institute, Inc.) and the R software package (v2.6) from Bioconductor (http://​www.​bioconductor.​org). Heat-maps were generated using the R software (v2.9.1). All comparisons of survival were performed using the log-rank test and all Kaplan-Meier survival curves were plotted using S-Plus software (ver. 6.0.2).

Table 1 summarizes the clinical characteristics of the 327 patients in our cohorts. Fifteen samples in cohort 2 were lobular carcinomas. Consequently, most breast cancer samples in cohort 2 were positive for ER and PR with a lower nuclear grade and fewer HER2 positive cases.

As shown in Figure 1, we classified breast cancer into six different molecular subtypes. The 783 probe-sets used for classification were grouped into 13 clusters enriched with genes associated with cell cycle/proliferation, cell movement, metabolism, and reproductive system development (Additional file 3, Figure S3). We then conducted statistical analysis between the molecular subtypes and various clinical parameters (Table 2). The results summarized in Table 2 show that by the T stage, smaller tumors dominated in subtypes V and VI, while larger tumors dominated in subtypes II, III and IV (p = 2 × 10^-5). The majority of patients in subtypes IV, V and VI were positive for ER and PR (p = 6.3 × 10^-51 and 2.3 × 10^-18, respectively). Interestingly, all subtype V breast cancers were positive for ER and PR and negative for HER2. In contrast, all subtype I breast cancers were negative for ER. Nearly all subtype II breast cancers were negative for ER (97%), and the majority had over-expression of HER2 (76.5%) (p = 9.1 × 10^-20). Subtype III comprised breast cancers that had weaker ER and variable PR and HER2 expression (data not shown). Subtype IV had full range expression of HER2. Subtype II had the greatest propensity to develop distant metastases (47%) followed by subtypes IV (36%) and VI (24%), while subtype V was least likely to metastasize (5%) (p = 2.5 × 10^-5). Figure 2 shows the survival curves of all six molecular subtypes. The statistical results of comparing survival outcomes between any two molecular subtypes are summarized in Table 3.

To demonstrate the biologically distinctive nature of six different subtypes of breast cancer, we studied the differential expressions of genes associated with cell cycle/proliferation, wound-response [9], stromal reaction [21] and vascular endothelial normalization [22, 23] using one-way clustering analysis. Genes used in this study were not used for molecular subtyping. As shown in Figure 3, all six molecular subtypes demonstrated distinct gene expression characteristics. The dendrograms of the probe-sets and the probe-set IDs are summarized in Figure S4 of Additional file 3. For validation, we used our classifier genes with centroid analysis to determine molecular subtypes of breast cancer samples in three independent datasets [10, 19, 20]. We then compared differential gene expression patterns associated with cell cycle/proliferation, wound-response, stromal reaction and vascular endothelial normalization for the same molecular subtypes between our dataset and the other three independent datasets. The same molecular subtypes in all four datasets were shown to share the same differential gene expression patterns (Figure 3). For further validation, we employed a different approach. We selected five genes (CAV1, DHFR, TYMS, VIM, ZEB1) known to be associated with drug sensitivity and the epithelial-mesenchymal transition of breast cancer [25‐29]. The intensity of expression of these genes was plotted according to molecular subtypes. Again, each molecular subtype shared the same unique molecular characteristics across all four datasets (Additional file 3, Figure S5).

To study how the molecular subtypes of breast cancer used in this study are correlated with the Perou-Sørlie intrinsic types [1, 2], we applied the classifier genes used by Perou-Sørlie [2] to our samples. As shown in Figure 4, there were both similarities and considerable differences between the two classification methods. A high degree of concordance was noted between our subtype I and the basal-like intrinsic type. When we applied our classification genes to the NKI dataset [24], wherein we also noticed an 89% concordance between our subtype I and the basal-like intrinsic type. The high degree of concordance was likely the result of the very distinctive features of this subtype of breast cancer.

Nevertheless, most of the Perou-Sørlie luminal A intrinsic type was divided into subtypes V and VI according to our classification genes (Figure 4). When we compared metastasis free survival between subtypes V and VI of the luminal A intrinsic type breast cancer patients in our cohorts, significantly better metastasis-free survival was observed for the subtype V patients comparing to the subtype VI (p = 0.025) (Additional file 3, Figure S6). There were no significant differences in disease severity (T stage p = 0.33, N stage p = 0.50, M stage p = 1, positive axillary lymph node number p = 0.50, and nuclear grade p = 1), however. The differentiation of subtypes V and VI breast cancer within the luminal A intrinsic type is therefore clinically significant. The distinction was further supported by the differential gene expression patterns for wound response and vascular endothelial normalization between these two subtypes of breast cancer (Figure 3).

The cases of HER2 over-expressing intrinsic type, they were divided into molecular subtypes II and III (Figure 4). We noted that cases of molecular subtype III cases classified as HER2 over-expressing intrinsic type expressed ER at a level higher than the subtype II and HER2 over-expressing intrinsic type (the average intensity of gene expression in logarithm to base 2 were 9.9 ± 0.96 vs. 8.6 ± 1.0, p < 0.0001). It appears that our molecular subtyping have discerned different subsets within the HER2 over-expressing intrinsic type. In this study, we did not find normal-breast like intrinsic type breast cancer in our cohorts.

The breast cancer samples included in this study covered the period of transition of adjuvant chemotherapy regimen from CMF to CAF and to taxane-based regimens. These archival samples provided an opportunity to examine how breast cancer subtypes might have responded differentially to CMF and CAF regimens of adjuvant chemotherapy. The results of this study show that the change from methotrexate to doxorubicin had a major impact on the survival of patients with subtype IV breast cancer (Figure 5). None of the pertinent clinical factors between these two groups of patients showed a significant difference except for the N stage. The N stage was higher in the CAF group (Table 4). In spite of the higher N stage, significantly better metastasis-free and overall survival was observed for subtype IV patients treated with CAF than with CMF (Figure 5). For other molecular subtypes, we did not find significant differences in survival between groups receiving treatment with CAF or CMF (Additional file 1, Table S5). The small sample size did not allow us to draw firm conclusions regarding survival in molecular subtypes other than subtype IV.

A number of patients in our cohorts opted not to receive adjuvant chemotherapy even though it was indicated. This allowed us to study how the omission of adjuvant chemotherapy might have influenced patient survival among various subtypes of breast cancer. In a comparison of disease severity between those with and without adjuvant chemotherapy in each subtype, only patients with subtype V showed no significant difference (Table 5), thereby offering an interpretable comparison. As shown in Figure 6, the metastasis-free and overall survivals of subtype V were essentially the same between those who received adjuvant chemotherapy and those who did not. This suggests that adjuvant chemotherapy did not provide survival benefits to subtype V patients in the early stages; however, this would require further confirmation due to small sample size.

To seek further support of our finding, we studied patients from the NKI dataset with subtype V breast cancer of N1 stage [24]. This dataset includes treatment and survival outcome information, and many patients in this dataset did not receive adjuvant chemotherapy. The distribution of tumor size and fraction of patients treated with hormone therapy were not significantly different between those who received adjuvant chemotherapy and those who did not. The p values determined by Fisher's exact test were 0.32 and 1.0, respectively. Stage N0 patients were excluded because an overwhelming number were not treated with adjuvant chemotherapy. The results showed that there was no difference in survival between stage N1 subtype V patients treated with adjuvant chemotherapy and those who were not (Figure 6).

As mentioned earlier, subtype I breast cancer was essentially the same as the basal-like intrinsic type (Figure 4), and this subtype of breast cancer is known to be chemosensitive [30]. The five and ten year survival rates of patients with basal-like breast cancer who did not receive adjuvant chemotherapy were 64% and 44%, respectively [24]. This is consistent with the fact that basal-like breast cancer has aggressive clinical course and poor survival without adjuvant chemotherapy [31]. When we studied the survival of patients with subtype I breast cancer following CMF or CAF adjuvant chemotherapy, it was noticed that both groups had good long-term survival outcome based on the results from a limited number of patients (Figure 7). This suggests that subtype I breast cancer responds well to CMF adjuvant chemotherapy, and this finding is supported by a recent study of two large randomized clinical trials in which patients with node negative basal-like breast cancer were sensitive and responsive to CMF adjuvant chemotherapy and had good long-term survival following treatment [32]. Adjuvant chemotherapy is therefore critical for the long-term survival of patients with early stage subtype I breast cancer. The use of less toxic CMF could be as effective as CAF and deserves further study.

Oncotype and MammaPrint predictors are used to predict the risk of distant recurrence in breast cancer patients for the optimization of treatment [12, 13, 33]. To learn how the groups with varying levels of the predicted risk are correlated with molecular subtypes of breast cancer, we conducted a study on patients in our dataset and the other two independent datasets [10, 24]. We determined molecular subtypes and the scores of relative risk of distant recurrence. The results, summarized in Figure 8, reveal that patients with a high risk of distant recurrence according to the genes of Oncotype predictor included both subtypes I (basal-like) and II (HER2 over-expressing). Low risk cases were mostly subtypes V and VI, while most intermediate risk cases were subtypes III and IV. The high-risk cases predicted by the genes of MammaPrint included most of subtype I and many of subtypes II, III and IV, with low risk cases limited to subtypes V and VI (Figure 8). The results were consistent across all three datasets; thus, breast cancer within the same predicted risk group is heterogeneous according to molecular subtype. Patients within the same risk group may require different therapeutic approaches for better survival outcome.