A clinically relevant gene signature in triple negative and basal-like breast cancer

In order to facilitate pooling of data sets from different laboratories we only used data from a single platform (Affymetrix U133A and U133 Plus 2.0 chips) and included only samples that were defined as triple negative based on the mRNA expression of ER, PgR, and HER2 as previously described [13‐15]. To obtain a large enough sample size for discovery it was necessary to pool several datasets. A major concern during this exercise is the possible confounding effect of systematic technical differences that exist between individual datasets. These could lead to false discovery during metagene definition and could also weaken the power of validation. We applied two different strategies to minimize this problem. First, we selected only highly comparable datasets for discovery. We initially identified 579 TNBC from a total of 3,488 publicly available primary breast cancer gene expression profiles representing 28 individual datasets (Additional file 2, Supplementary Table S1). We excluded 13 datasets contributing 185 TNBC cases from the discovery cohort because they did not fulfill our criteria of comparability of the microarray data (for details see Additional file 4, Supplementary Methods Section 1 and Additional file 1, Supplementary Figure S2). The final discovery cohort to identify metagenes included 394 TNBC from 15 datasets (cohort-A). The 185 samples excluded from discovery were retained as a validation set (cohort-B) to assess correlations between various metagenes and between metagenes and clinical outcome (Additional file 1, Supplementary Figure S1). This strategy maximized the integrity of metagene discovery at the cost of possibly reducing the power of the validation study. The two cohorts did not significantly differ with respect to age, tumor size and histological grade. However, the validation cohort-B contained a larger number of lymph node positive patients and a higher proportion of fine needle aspiration (FNA) samples. Follow-up data were available for 2,348 of the total 3,488 samples and 327 of the 579 TNBC samples. Since the number of patients with follow-up in validation cohort B was too small (n = 30 of 185) an additional independent validation cohort-C [16] (n = 76) was included to assess the prognostic value of the metagenes (Additional file 1, Supplementary Figure S1). The patient characteristics of the discovery and validation cohorts are given in Table 1. For analysis of normal tissue a dataset from a benign breast was used (Additional file 2, Supplementary Table S1).

Unsupervised analysis, without input of clinical variables, was performed to identify metagenes that were defined as the arithmetical average expression of highly correlated genes. Gene clusters were selected with either a minimal membership of 10 genes and a minimal correlation threshold of 0.7, or a minimum of 25 genes and a correlation of 0.6, respectively (for details see Additional file 4, Supplementary Methods Section 2). We also employed a screen to remove genes that showed data-set bias. The dependence of the expression levels of the metagene probesets on the dataset vector was analyzed using the Kruskal-Wallis statistic (Additional file 4, Supplementary Methods Section 3). Only Stroma and Hemoglobin metagenes displayed a bias for FNA samples that reflect frequent contamination of these types of samples with blood and the lack of stromal elements compared to core needle or surgical biopsies (Additional file 1, Supplementary Figure S3 and Additional file 4, Supplementary Methods). Therefore, these two metagenes were analyzed only in surgical biopsies.

No systematic bias was observed between the U133A and U133 Plus2.0 arrays, which differ only in the spatial feature size of the probesets (for details see Additional file 4, Supplementary Methods Section 4). Both metagene distributions and "Centroid methods" were used to classify subtypes of TNBC as given in Additional file 4, Supplementary Methods Sections 8 and 9).

Relapse free survival (RFS) was preferentially used as a clinical endpoint for event free survival (EFS). Only if RFS was not available in some datasets was it replaced by distant metastasis free survival (DMFS). Details on used endpoints, Kaplan-Meier and Cox regression analysis are given in Additional file 4, Supplementary Methods Section 5. Optimized cutoffs for dichotomizing of metagene scores to plot survival curves were derived from the discovery cohort and were applied without modification to the validation cohorts (Additional file 4, Supplementary Methods Section 6). All P-values are two-sided and 0.05 was considered as a significant result. Analyses were performed using the R software [17] and SPSS version 17.0 (SPSS Inc. Chicago, IL).

In our discovery cohort we identified 16 clusters of correlated genes by unsupervised methods whose expression values were averaged as metagenes (Figure 1). As expected, no cluster of genes correlated with ER, PgR, and HER2 status [4] were identified. In contrast the identified metagenes presented in Table 2 included the basal-like phenotype [4], an apocrine/androgen receptor signaling signature [18, 19], five signatures related to different types of immune cells [4, 20‐25], a stromal signature [26, 27], the claudin-CD24 signature [28, 29], markers of blood [30] and adipocytes [4], as well as an inflammatory signature [31‐33] and an angiogenesis signature [23, 34]. These phenotypes corresponded to previously described gene signatures that have also been used to define subsets of TNBC in a recent smaller study [9]. The angiogenesis signature (VEGF metagene) has been described very recently as a "hypoxia signature" associated with poor outcome and expressed in distant metastases [34]. As shown in Figure 1, we observed the highest correlation between different types of immune cell metagenes. Similar relationships between the metagenes were detected in the validation cohort-B (Figure 1) and -C (Additional file 1, Supplementary Figure S4). The presence of B-lymphocytes in the tumor is the primary source of the expression of the B-Cell metagene that is largely composed of immunoglobulin genes [20, 22]. In contrast, immunohistochemical analyses of IL-8 expression and analysis of gene expression data of breast cancer cell lines indicate that carcinoma cells are the main source of the IL-8 metagene (Figure 2).

We observed a clear bimodal distribution of the basal-like metagene score among TNBC (Figure 3). This bimodal distribution allows us to derive a cutoff to separate cases into high and low expression groups by fitting two normal distributions to the data (Figure 3). According to this cutoff, 72.8%, 73.0% and 69.7% of TNBC were defined as BLBC in the discovery cohort-A, validation cohort-B, and validation cohort-C, respectively. Table 3 compares the clinical characteristics of BLBC or non-BLBC triple negative cancers the discovery cohort-A. The positive association between high histological grade (G3, P < 0.001), younger age (P = 0.004) and BLBC were also observed in the validation cohort-C and validation cohort-B, respectively (Additional file 2, Supplementary Table S2).

In unsupervised clustering of the metagenes the basal-like metagene clustered next to the apocrine metagene but showed a strong inverse correlation (Figure 1). To quantify the correlation between the basal-like metagene and all other metagenes from Table 2 we used quartiles of the respective metagenes. Additional file 2, Supplementary Table S3 presents the six metagenes that displayed significant correlations with the BLBC phenotype in both the discovery and validation cohorts. A positive correlation was found between the BLBC phenotype and the proliferation and angiogenesis (VEGF) metagenes. A negative correlation was observed for the apocrine/androgen receptor signaling and two immune system related metagenes (MHC-2 and T-Cell metagenes), as well as an adipocyte related signature.

Since we observed a negative correlation between the basal-like metagene and potential markers of normal breast tissue, such as the adipocyte metagene, we had to exclude the possibility that we are only distinguishing stroma-rich and stroma-poor samples. As shown in Additional file 1, Supplementary Figure S5, when metagenes for proliferation, adipocytes and histones were compared between BLBC, non-BLBC, and normal breast samples it is clearly demonstrated that the non-BLBC subtype is distinct from normal breast tissues in the expression of several metagenes. Proliferation genes have been previously shown to be the most important determinant of cancer vs normal signatures [35]. Furthermore, the strong bimodal distribution of the basal-like metagene argues against the possibility that this metagene is inversely describing the degree of contamination with normal tissue which should rather result in a continuous distribution. The non-BLBC tumors in our TNBC dataset mainly represent samples of the "molecular apocrine" type (16.5%), which demonstrates the inverse bimodal distribution as the basal-like metagene, and a relatively small group of "claudin-low" tumors (6.3%). The mutual relationship of these three metagenes is shown in Additional file 1, Supplementary Figure S6.

To assess the prognostic value of the metagenes, we analyzed the event free survival of patients as a function of metagene expression. The basal-like metagene had no significant effect on survival (Additional file 1, Supplementary Figure S7). In contrast, five other metagenes including the IL-8, Histone, VEGF, B-Cell, and T-Cell metagenes showed significant prognostic values when considered as continuous variables in univariate analysis (Additional file 2, Supplementary Table S4). In a stepwise multivariate Cox regression analysis only three of these, the IL-8, Histone, and the B-Cell metagenes, remained significant (Additional file 2, Supplementary Table S5). The IL-8 and Histone metagenes were positively correlated with one another in all data sets (see Figure 1). The B-cell and IL-8 metagenes were associated with prognosis but with an opposing direction. Based on these observations, we derived a B-Cell /IL-8 metagene ratio as a prognostic index for TNBC. Figure 4A demonstrates that patients with a high expression of the B-Cell and low expression of the IL8 metagene have significantly better prognosis than other TNBC patients (HR 0.37, 95% CI 0.22 to 0.61; P < 0.001). The five-year event-free survival was 84 ± 4% for the good prognosis group (n = 95) compared to 59 ± 4% for the rest of the patients. In validation cohort B (n = 30), there was a non-significant trend for better survival for patients with high B-cell low IL8 metagene expression (P = 0.3, Figure 4B). Since this cohort has limited power due to the small sample size, we also tested the prognostic value on a separate and larger (n = 75) validation cohort of TNBC samples [16]. The B-cell/IL8 metagene ratio had significant prognostic value in this second validation cohort C, the hazard ratio (HR) was 0.26, (95% CI 0.10 to 0.68) and the five-year DFS was 78 ± 9% vs. 45 ± 8%, (P = 0.003) (Figure 4C). The prognostic value was independent of histological grade; Figure 4D, E shows pooled data from all three cohorts to increase sample size, (see also Additional file 1, Supplementary Figure S8 for the individual cohorts). Moreover, the prognostic value of the B-cell/IL8 metagene ratio was observed both in BLBC and non-BLBC TNBCs (P = 0.001 and P = 0.006, respectively; Additional file 1, Supplementary Figure S9). The proportion of BLBC cases was similar in the Good and Poor prognosis groups defined by the B-cell/IL8 metagene ratio (75.2% and 71.8%, respectively; P = 0.54).

To assess a potential predictive value for sensitivity to systemic adjuvant chemotherapy, the patients were stratified by adjuvant treatment. In the discovery cohort, 186 patients received no adjuvant systemic treatment and 81 patients received chemotherapy (mostly Cyclophosphamide Methotrexate Fluorouracil; CMF)). Better prognosis was observed for the high B-cell/low IL8 group in both untreated (P = 0.001) as well as chemotherapy treated patients (P = 0.05; not shown). A potential predictive value of the B-cell and IL8 metagenes was also analyzed in 191 patients with TNBC who received neoadjuvant chemotherapy. We assembled this cohort of samples with information on pathologically complete response (pCR) from seven datasets. As shown in Additional file 1, Supplementary Figure S10 the B-cell metagene had a modest predictive value with an area under the curve (AUC) of 0.606 consistent with our previous results [22]. The predictive value for the IL8 metagene was smaller (AUC -0.552). Combining both metagenes increased the AUC to 0.612 (95% CI 0.519 to 0.704; P = 0.018).

In multivariate Cox regression analysis, including lymph node status, age, tumor size, and histological grade, only the combined B-Cell/IL8-metagene score showed strong independent prognostic value in both the discovery cohort (HR 0.38, 95% CI 0.22 to 0.67, P = 0.001) and in the second, larger validation cohort-C, (HR 0.21, 95% CI 0.07 to 0.62, P = 0.005). The only other variable with borderline statistical significance (HR 0.40; 95% CI 0.17 to 0.99, P = 0.046) was lymph node status in validation cohort-C (Table 4). However, even in univariate analyses the remaining clinical variables did not show a significant prognostic value in the analyzed cohorts. This might be attributed to the fact that most TNBC are usually highly proliferating and grading is not as important for prognosis in this subtype as it is in ER positive disease; in addition, the power of our analysis may be limited to detecting the modest effect of age and tumor size on prognosis within this sample set. The inclusion of a term for chemotherapeutic treatment in the multivariate analysis further reduced the sample size to 213 patients in cohort-A (no treatment information was available for patients from validation cohort-B). Of these 213 patients only 37 were treated with chemotherapy. The combined B-Cell/IL8-metagene score remained significant (P = 0.001) in the corresponding multivariate analysis (Additional file 2, Supplementary Table S9A). Unexpectedly, chemotherapy treatment was associated with a worse prognosis probably due to chance or some form of selection bias to include higher risk patients in these public data sets (Additional file 2, Supplementary Table S9A). This selection bias is consistent with a significant higher portion of node positive patients in the chemotherapy group (P = 0.001) and a trend for a higher histological grade (P = 0.074; Additional file 2, Supplementary Table S9B).

The correlation of several published prognostic gene signatures to the metagenes discovered within the pure TNBC cohort was analyzed by hierarchical clustering using the gene expression data from cohort-A (Additional file 4, Supplementary Methods Section 13). As shown in Additional file 1, Supplementary Figure S11, the "recurrence score" [36], "genomic grading index" (GGI) [37], and the "wound response signature" [38] display high correlation to the proliferation metagene. On the other hand the "7-gene immune response (IR) signature" [39], the "stroma derived prognostic predictor" (SDPP) [40], and the "368 gene medullary breast cancer signature" [16] were all highly correlated to immune cell metagenes. The magnitude of the correlation (R² = 0.4 to approximately 0.7) between the different immune metagenes and the related signatures is at the same high level as the correlation between genes within other metagene clusters (R² = 0.5 to approximately 0.7; Table 2). We demonstrated previously [22] that even if the different immune metagenes can discriminate between distinct types of immune cells, the actual infiltration of tumors generally represents a mixture of these different immune cells. In most cases, the differences in the proportions in this mixture are smaller than the global differences in lymphocyte infiltration between individual tumors. Therefore, different immune signatures often carry redundant prognostic information and can replace each other. In contrast to the immune cell metagenes no correlation between the IL8 metagene and other signatures were observed.