Purpose: To investigate potential associations between gene modules representing key biologic processes and response to aromatase inhibitors (AI) in estrogen receptor–positive (ER+) breast cancer.

Patients and Methods: Paired gene expression and Ki67 protein expression were available from 69 postmenopausal women with ER+ early breast cancer, at baseline and 2 weeks post-anastrozole treatment, in the presurgical setting. Functional gene modules (n = 26) were retrieved from published studies and their module scores were computed before and after elimination of proliferation-associated genes (PAG). Ki67 and module scores were assessed at baseline and 2 weeks post-anastrozole. Unsupervised clustering was used to assess associations between modules and Ki67.

Results: Proliferation-based modules were highly correlated with Ki67 expression both pretreatment and on-treatment. At baseline with and without PAGs, Ki67 expression was significantly inversely correlated with ERG, ESR1.2, SET, and PIK3CA modules. Modules measuring estrogen signaling strongly predicted antiproliferative response to therapy with and without PAGs. Baseline expression of insulin-like growth factor-1 (IGF-I) module predicted a poor change in Ki67-implicating genes within the module as involved in de novo resistance to AIs. High expression of Immune.2.STAT1 module pretreatment predicted poor antiproliferative response to therapy. A significant association between estrogen-regulated genes modules (ESR1, ESR1-2, SET, and ERG) was evident post AI.

Conclusions: Multiple processes and pathways are affected by AI treatment in ER+ breast cancer. Modules closely associated with ESR1 expression were predictive of good antiproliferative response to AIs, but modules representing immune activity and IGF-I/MAPK were predictive of poor Ki67 response, supporting their therapeutic targeting in combination with AIs. Clin Cancer Res; 20(9); 2485–94. ©2014 AACR.

Translational Relevance

Response and resistance to endocrine therapy in estrogen receptor–positive (ER+) breast cancer is poorly understood. Using gene expression and Ki67 data from patients treated with the aromatase inhibitor (AI) anastrozole in the presurgical setting, we interrogated key biologic processes represented by gene modules (n = 26). Proliferation (CIN70, GGI, AURKA, and Gene70) and estrogen-responsive pathways (ESR1.1, ESR1.2, and SET) were profoundly reduced by anastrozole treatment, as were signaling pathways PTEN, CASP3, E2F3, MYC, AKT/mTOR, and IGF-I. In contrast, PIK3CA, MAPK, Stroma.2-PLAU, Stroma.1, and Immune.1 expression module scores were significantly increased by treatment, suggesting that the pathways relevant to resistance are upregulated at this early time point. A significant association between estrogen-regulated modules (ESR1, ESR1-2, SET, and ERG) was evident post AI and may be indicators of late relapse. These observations reveal that different processes and pathways are affected differently with AI treatment between patients with ER+ breast cancer, emphasizing the importance of tumor heterogeneity.

About 75% to 80% of patients with breast cancer at primary diagnosis present with breast tumors that are hormone receptor positive disease (1). The standard of care for these patients is endocrine treatment, such as tamoxifen, which competes with estrogen for the estrogen receptor (ER), or aromatase inhibitors (AI) that block the conversion of androgen to estrogen. The clinical efficacy of AIs is superior to that of tamoxifen for the treatment of postmenopausal women with ER-positive (ER+) breast cancer in both the adjuvant and metastatic setting (2). Neoadjuvant endocrine therapies, in particular AIs, are also effective at downstaging breast tumors and facilitating breast conservation in many postmenopausal patients (3).

Gene expression profiling has revealed the high degree of heterogeneity in breast cancer and the impact that particular molecular attributes can have on clinical outcome (4, 5). In particular, analyses of global gene expression profiles have revealed that breast cancer may be characterized as a collection of biologically distinct subgroups defined by “intrinsic” molecular markers. The principal subgroups include (i) luminal cancers, which are mainly ERα+, (ii) tumors characterized by amplification/overexpression of the HER2 gene, and (iii) basal-like cancers in which hormone receptors and HER2 are generally absent but basal cytokeratins are expressed. In addition to subtype-associated differences in prognosis and response to treatment, molecular differences within subtypes are increasingly being considered as contributors to the diversity of treatment response. The luminal group is generally divided into two subgroups: luminal A tumors characterized by greater expression of estrogen-regulated genes and markers associated with luminal breast epithelial cells; and luminal B in which these features are less apparent and there is greater expression of proliferation-associated genes (PAG).

Despite the efficacy of endocrine therapy in patients with ER+ breast cancer, approximately 25% of patients recur within 10 years after receiving 5 years of tamoxifen (6). Although the molecular events underpinning the efficacy of endocrine therapies are not fully understood (7, 8), it is clear that downregulation of proliferation plays a major role. Uncontrolled proliferation is a hallmark feature of tumor progression and in breast cancer this is most frequently assessed by immunohistochemical assessment of Ki67 (9). This marker has been used to assess prognosis, predict sensitivity or resistance to endocrine therapy, and serve as a dynamic biomarker of treatment efficacy in samples taken before, during, and after neoadjuvant endocrine therapy (10). Ki67 is suppressed to a variable extent by AIs in the large majority of breast cancer and overall by a mean of approximately 75% after 2 weeks, increasing to >80% after 12 to 16 weeks (11); the degree of Ki67 suppression by endocrine agents has been found to predict their efficacy as adjuvant therapy. Despite this focus of investigation on proliferation, it is known that many genes not associated with proliferation are also affected by estrogen deprivation (12). It seems likely that these other effects may contribute to the overall impact of AIs on clinical outcome. Preoperative neoadjuvant therapy provides a unique setting for characterizing such biologic changes in tumors and the molecular features that affect the changes.

To summate the activity of biologic processes/pathways relevant to cancer, a number of gene modules have been derived, including those that represent Wnt, SRC, Ras, Myc, E2F3, PTEN loss (13, 14), activation of insulin-like growth factor 1 (IGF-I; ref. 15), mitogen-activated protein kinase (MAPK; ref. 16), AKT/mTOR pathways (17), and PIK3CA mutations (18). Key biologic processes in breast cancer have also formed the basis of functional gene signatures (19), including tumor invasion (Stroma-PLAU), immune response (Immune-STAT1), angiogenesis (VEGF), apoptosis (CASP3), proliferation (AURKA), ER signaling (ESR1), and HER2 signaling (ERBB2). Prognostic gene modules such as genomic grade index (GGI; ref. 5), Gene70 (20), and a signature that has been reported to recapitulate deficiency in DNA-repair mechanisms and chromosomal instability (CIN70; ref. 21) have also been described. These gene modules provide an opportunity to characterize the functional heterogeneity of ER+ tumors and to study the potential impact of AI therapy on key biologic processes that may contribute to clinical benefit.

The primary objective of this study was to determine systematically whether baseline expression of 26 published gene modules predicted for antiproliferative response in ER+ tumors with matched global gene expression data and Ki67 measurements, before and after 2 weeks of neoadjuvant anastrozole. Given that many of these gene modules contain PAGs, we first assessed the specific contribution of the PAGs to the correlations of the gene signatures with changes in Ki67. Second, we correlated the pretreatment expression of the individual gene modules with the change in Ki67 after 2 weeks of anastrozole therapy. Third, the effect of AI treatment on the gene modules was assessed and finally, the associations of the gene modules with Ki67 were determined before and after treatment, as well as their changes resulting from treatment.

Presurgical study of anastrozole

Pre- and on-treatment (week 2) core biopsies were available from postmenopausal women recruited into the anastrozole (1 mg/day)-only arm of a multicenter, randomized, double-blind, phase II neoadjuvant trial of anastrozole alone or with gefitinib in early breast cancer (ClinicalTrials.gov identifier: NCT00255463). This anastrozole-treated subgroup constitutes the Functional Aromatase Inhibitor Molecular Study (FAIMoS). Sample acquisition, storage, and RNA extraction was done as previously described (12). Following exclusions, paired samples were available from 81 patients treated for 2 weeks, of which 69 patients had corresponding Ki67 protein expression pre- and post treatment. A detailed CONSORT diagram is shown in Supplementary Fig. S1. These patients were used in the following analyses. Patient demographics are shown in Supplementary Table S2.

Gene expression data and probe annotation

Gene expression profiles were generated on the Illumina HumanWG-6 v2 Expression BeadChip. Illumina raw data were extracted using the BeadStudio software and were transformed and normalized using variance-stabilizing transformation and robust spline normalization method in the Lumi package in R (http://www.bioconductor.org). Probes were discarded if they were not detected in any of the 69 paired samples (detection P > 0.01), resulting in 53.7% (26,163/48,701) of the probes remaining for the downstream data analysis. A corresponding annotation file, HumanWG-6_V2_0_R4_11223189_A, was retrieved from the manufacturer's website (http://www.switchtoi.com/annotationfiles.ilmn). Gene expression data from this study were deposited by the authors of reference (12) and are available on http://synapse.sagebase.org/#Dataset:16243 (22).

Gene modules

Twenty-six publicly available gene modules representing potentially important biologic processes in breast cancer were assessed: (i) ER signaling and estrogen-regulated modules: ESR1.1, ESR1.2, sensitivity to endocrine therapy index (SET; refs. 19, 23, 24), and ERG (12); (ii) growth factor receptors and cytoplasmic signaling–related modules: IGF-I, obesity, MAPK (15, 16, 25), SRC, betacatenin, RAS (13), ERBB2, VEGF (19), PIK3CA (PIK3CA-GS; ref. 18), PTEN loss (14), and AKT/mTOR (17); (iii) cell cycle–related module: MYC (13); (iv) apoptosis and cell survival signaling related module: CASP3 (19); (v) proliferation-based modules: CIN70 (21), GGI (5), AURKA (19), Gene70 (20), and E2F3 (13); (vi) tumor invasion: Stroma.1 (26) and Stroma.2-PLAU (Stroma.2; ref. 19); (vii) immune response: Immune.1 (27) and Immune.2-STAT1 (Immune.2; ref. 19).

Composition and mapping of the gene modules

EntrezGeneIDs were used as gene identifiers in all selected modules with the exception of the ESR1.1 module (23), in which the UniGeneID was used. In this instance, we converted to EntrezGeneID using the ROCK database (www.rock.icr.ac.uk; ref. 28). The HumanWG-6_V2 annotation file was used to map the EntrezGeneIDs to the corresponding Illumina probe IDs. When multiple probes mapped to the same EntrezGeneID, the probe with the highest variance across pre- and posttreatment samples was selected to represent the gene. Genes were discarded from further analysis if they were not mapped to either the annotation file or the expression data.

PAGs

As Ki67 was used as the comparator with each of the gene modules, we hypothesized that PAGs, that occur in and dominate the behavior of many gene modules, may influence the representation of the key features in the respective process. To address this, three gene sets were identified, which were associated with proliferation: (i) genes functionally associated with cell-cycle progression and cell division according to Gene Ontology annotations (29), (ii) cycling genes (genes showing cell cycle stage-specific expression; ref. 29, 30), and (iii) tumor-based “proliferation cluster” genes (belonging to a ‘proliferation cluster’ defined in human breast cancer expression datasets; refs. 19, 29, 31). This process identified a total of 1,178 PAGs (Supplementary Table S3).

Computation of module scores

For each sample, a gene module score was calculated according to the expression of the relevant genes in the FAIMoS sample, as previously described (19, 32, 33).

formula

where n is the number of genes in a module, Xi represents the normalized gene expression in the test sample from the FAIMoS study, and gene-specific weights Wi are equal to +1 or −1 according to the direction of their association with the phenotype in the original publication. Data retrieved from publicly available datasets and the gene-set lists used in the study are detailed in Supplementary Table S4.

Data analysis

All analyses were performed using R version 2.14.1 (http://www.r-project.org/). P values less than 0.05 were considered statistically significant. Reported P values are two-sided. Unadjusted P values are shown but because large numbers of statistical analyses have been performed, we have additionally shown adjusted P values to aid in their cautious interpretation by using R function p.adjust.methods (i.e., False discovery rate, FDR).

The effect of the 2-week AI treatment on log2 Ki67 and gene modules was tested using Wilcoxon signed rank test based on the changes in expression. Geometric mean of intensity (G) was calculated as:

formula

and G is the anti of |$\log _2^{\rm G}$|⁠, where Xi represents the gene expression intensity or Ki67 protein level. The %Δ of geometric mean of intensity in a module and in Ki67 was defined as

formula

To study the associations of Ki67 and the twenty-six gene modules together with the impact of removing the PAGs, we performed (i) Spearman rank correlation between Ki67 and gene modules and (ii) pvclust (34) to visualize the degree of association between Ki67 and gene modules.

Both approaches were performed using log2 Ki67 protein level and the expression module scores (i) at the baseline, (ii) in their response to estrogen deprivation, and (iii) posttreatment with and without PAGs. Here, |$\log _2^{({\rm Ki}67{\rm pre})}$| and |$\log _2^{({\rm Ki}67{\rm post})}$| were used for Ki67 pretreatment and posttreatment protein level respectively, and the 2-week protein level change in Ki67 was defined as |$\Delta \log _2^{{\rm Ki}67} \left({\log _2^{({\rm Ki}67{\rm post})} - \log _2^{({\rm Ki}67{\rm pre})}} \right)$|⁠.

The R package pvclust, which calculates P values for each cluster using bootstrap resampling was used to assess the stability of the clusters. Two types of P values are available from this package: approximately unbiased (AU) P value and bootstrap probability (BP) value. The AU P value uses the multiscale bootstrap resampling methodology for calculating its P value and is superior to the BP P value calculated by the ordinary bootstrap resampling method. Hence, the AU method was used for these analyses and P values were reported as percentage (%), cluster with P value greater than 95% representing significantly associated modules.

To test the presence of clusters, we used the Ward method and correlation-based dissimilarity matrix with the parameters set to 10,000 bootstrap replicates, and the relative sample sizes set from 0.5 to 1.4 (default setting), with incremental steps of 0.1 to determine AU P value. In the data matrices, rows correspond to patients/samples, columns correspond to |$\log _2^{{\rm Ki}67}$|/gene module scores, and the values were centered and scaled via R function scale.

Effect of removal of PAGs on gene modules

As proliferation has been strongly associated with the predictive and prognostic utility of gene modules (35), we took the strategy of analyzing the endocrine responsiveness of the gene modules with and without PAGs (with the exception of the AURKA module given that its primary purpose was to represent proliferation; Supplementary Table S5). Analysis of the PAGs showed that they not only changed significantly with AI, but their change was significantly concordant with changes in Ki67 (Supplementary Table S6).

Pretreatment associations of gene modules and Ki67 with and without the PAGs

Associations between the gene modules and Ki67 at baseline were assessed using pvclust. GGI, which mainly quantifies tumor proliferation, was in a significantly correlated cluster with AURKA, CIN70, PTEN loss, and Gene70 modules (AU P = 98%; Fig. 1A). Removal of the PAGs from these five gene modules resulted in the loss of the significance of the association with Gene70 and PTEN (Fig. 1B). Gene modules measuring estrogen signaling (ESR1.1, ESR1.2, and SET) and obesity with IGF-I remained significantly correlated in the absence or presence of PAGs. Another notable significant clustering was between AKT/mTOR, ERBB2, and VEGF, but this was disrupted by removal of the PAGs. In general, functionally related biologic processes were in highly correlated clusters.

Figure 1.

Cluster dendrogram pretreatment includes PAGs (A), excluding PAGs (B); cluster dendrogram of changes in expression after treatment includes PAGs (C), excluding PAGs (D); cluster dendrogram of posttreatment includes PAGs (E), excluding PAGs (F). The cluster stability analysis of the unsupervised hierarchical clustering uses pvclust to calculate the AU P value. The AU P values are reported as percentage (%) in red. Clusters representing significantly associated modules (i.e., those clusters with greater than 95% significance) are shown by the red rectangle. Gene module scores and Ki67 protein expression at pretreatment, posttreatment, and change in the treatment were used for this analysis.

Figure 1.

Cluster dendrogram pretreatment includes PAGs (A), excluding PAGs (B); cluster dendrogram of changes in expression after treatment includes PAGs (C), excluding PAGs (D); cluster dendrogram of posttreatment includes PAGs (E), excluding PAGs (F). The cluster stability analysis of the unsupervised hierarchical clustering uses pvclust to calculate the AU P value. The AU P values are reported as percentage (%) in red. Clusters representing significantly associated modules (i.e., those clusters with greater than 95% significance) are shown by the red rectangle. Gene module scores and Ki67 protein expression at pretreatment, posttreatment, and change in the treatment were used for this analysis.

Close modal

At baseline, with and without PAGs, baseline Ki67 protein expression was significantly positively correlated with the following modules: GGI, CIN70, AURKA, PTEN, Gene70, IGF-I, obesity, ERBB2, and RAS, and inversely correlated with ERG, ESR1.2, SET, and PIK3CA gene modules (Supplementary Table S7A and S7B). The significance of these associations with the proliferation-based modules, GGI, CIN70, AURKA, PTEN, and Gene70, became weaker after eliminating PAGs, whereas the relationship of Ki67 with the remaining modules, IGF-I, obesity, ERBB2, RAS, ERG, ESR1.2, SET, and PIK3CA was unaffected.

Correlation of pretreatment gene modules with change in Ki67

To assess the effect of the gene modules at predicting antiproliferative response to AI therapy, we compared baseline expression of each gene module versus the change (Δ) in Ki67 as a marker of response (Table 1), in which a negative ρ indicates a positive relationship of the module or metagene with Ki67 response to AI therapy, because the response is a reduction in Ki67. We found that the ER signaling module scores (ESR1.1, SET, and ESR1.2) were significantly associated with the reduction in Ki67 (unadjusted P <0.005, 0.04, and <0.05, respectively). In contrast, Immune.2.STAT1, Gene70, and IGF-I baseline module scores were significantly negatively associated with change in Ki67, indicating that the higher the module score, the smaller the antiproliferative response. This inverse correlation became stronger when PAGs were removed from the Gene70 and GGI modules. The ESR1.1 and Immune.2.STAT1 modules remained significant (adjusted P < 0.05) after adjustment for multiple testing. ρ values for the relationship of obesity, Stroma.2.PLAU and MAPK modules with change in Ki67 were all positive, but with 0.05 < unadjusted P < 0.10.

Table 1.

Spearman correlation between individual pretreatment module score and change in Ki67 protein expression, |$\Delta \log _2^{{\rm Ki}67} \left({\rm\log _2^{({\rm Ki}67{\rm post})} - \log_2^{({\rm Ki}67{\rm pre})}} \right)$|

All matchedPAGs excluded
Module nameρUnadjusted PAdjusted PρUnadjusted PAdjusted PReference
ESR1.1 −0.337 0.0046 0.0469 −0.339 0.0044 0.0469 Mackay et al. (23) 
SET −0.248 0.0396 0.1803 −0.252 0.0370 0.1803 Symmans et al. (24) 
ESR1.2 −0.241 0.0465 0.1803 −0.244 0.0431 0.1803 Desmedt et al. (19) 
ERG −0.226 0.0619 0.1860 −0.206 0.0890 0.2188 Dunbier et al. (12) 
MYC −0.160 0.1889 0.3705 −0.147 0.2282 0.4157 Bild et al. (13) 
VEGF −0.065 0.5955 0.7820 −0.065 0.5955 0.7820 Desmedt et al. (19) 
E2F3 −0.041 0.7374 0.8646 −0.052 0.6729 0.8579 Bild et al. (13) 
Betacatenin 0.015 0.9016 0.9384 −0.002 0.9888 0.9888 Bild et al. (13) 
AKT/mTOR 0.026 0.8294 0.9000 −0.006 0.9587 0.9779 Majumder et al. (17) 
ERBB2 0.042 0.7313 0.8646 0.037 0.7629 0.8646 Desmedt et al. (19) 
SRC 0.048 0.6942 0.8635 0.033 0.7890 0.8748 Bild et al. (13) 
PIK3CA 0.065 0.5980 0.7820 0.071 0.5638 0.7820 Loi et al. (18) 
PTEN 0.072 0.5545 0.7820 0.038 0.7536 0.8646 Saal et al. (14) 
Stroma.1 0.122 0.3196 0.5046 0.127 0.2991 0.4921 Farmer et al. (26) 
CASP3 0.127 0.2971 0.4921 0.092 0.4520 0.6780 Desmedt et al. (19) 
RAS 0.132 0.2808 0.4921 0.120 0.3265 0.5046 Bild et al. (13) 
CIN70 0.152 0.2123 0.4010 0.017 0.8891 0.9384 Carter et al. (21) 
Immune.1 0.161 0.1866 0.3705 0.161 0.1866 0.3705 Teschendorff et al. (27) 
AURKA 0.175 0.1507 0.3342 N/A N/A N/A Desmedt et al. (19) 
Obesity 0.206 0.0901 0.2188 0.196 0.1064 0.2467 Creighton et al. 
GGI 0.217 0.0731 0.1962 0.277 0.0213 0.1552 Sotiriou et al. (5) 
Stroma.2.PLAU 0.221 0.0679 0.1924 0.226 0.0620 0.1860 Desmedt et al. (19) 
MAPK 0.230 0.0576 0.1860 0.237 0.0495 0.1803 Creighton et al. (16) 
IGF-I 0.250 0.0385 0.1803 0.244 0.0433 0.1803 Creighton et al. (15) 
Gene70 0.299 0.0127 0.1080 0.357 0.0026 0.0442 van't Veer et al. (20) 
Immune.2.STAT1 0.381 0.0013 0.0332 0.381 0.0013 0.0332 Desmedt et al. (19) 
All matchedPAGs excluded
Module nameρUnadjusted PAdjusted PρUnadjusted PAdjusted PReference
ESR1.1 −0.337 0.0046 0.0469 −0.339 0.0044 0.0469 Mackay et al. (23) 
SET −0.248 0.0396 0.1803 −0.252 0.0370 0.1803 Symmans et al. (24) 
ESR1.2 −0.241 0.0465 0.1803 −0.244 0.0431 0.1803 Desmedt et al. (19) 
ERG −0.226 0.0619 0.1860 −0.206 0.0890 0.2188 Dunbier et al. (12) 
MYC −0.160 0.1889 0.3705 −0.147 0.2282 0.4157 Bild et al. (13) 
VEGF −0.065 0.5955 0.7820 −0.065 0.5955 0.7820 Desmedt et al. (19) 
E2F3 −0.041 0.7374 0.8646 −0.052 0.6729 0.8579 Bild et al. (13) 
Betacatenin 0.015 0.9016 0.9384 −0.002 0.9888 0.9888 Bild et al. (13) 
AKT/mTOR 0.026 0.8294 0.9000 −0.006 0.9587 0.9779 Majumder et al. (17) 
ERBB2 0.042 0.7313 0.8646 0.037 0.7629 0.8646 Desmedt et al. (19) 
SRC 0.048 0.6942 0.8635 0.033 0.7890 0.8748 Bild et al. (13) 
PIK3CA 0.065 0.5980 0.7820 0.071 0.5638 0.7820 Loi et al. (18) 
PTEN 0.072 0.5545 0.7820 0.038 0.7536 0.8646 Saal et al. (14) 
Stroma.1 0.122 0.3196 0.5046 0.127 0.2991 0.4921 Farmer et al. (26) 
CASP3 0.127 0.2971 0.4921 0.092 0.4520 0.6780 Desmedt et al. (19) 
RAS 0.132 0.2808 0.4921 0.120 0.3265 0.5046 Bild et al. (13) 
CIN70 0.152 0.2123 0.4010 0.017 0.8891 0.9384 Carter et al. (21) 
Immune.1 0.161 0.1866 0.3705 0.161 0.1866 0.3705 Teschendorff et al. (27) 
AURKA 0.175 0.1507 0.3342 N/A N/A N/A Desmedt et al. (19) 
Obesity 0.206 0.0901 0.2188 0.196 0.1064 0.2467 Creighton et al. 
GGI 0.217 0.0731 0.1962 0.277 0.0213 0.1552 Sotiriou et al. (5) 
Stroma.2.PLAU 0.221 0.0679 0.1924 0.226 0.0620 0.1860 Desmedt et al. (19) 
MAPK 0.230 0.0576 0.1860 0.237 0.0495 0.1803 Creighton et al. (16) 
IGF-I 0.250 0.0385 0.1803 0.244 0.0433 0.1803 Creighton et al. (15) 
Gene70 0.299 0.0127 0.1080 0.357 0.0026 0.0442 van't Veer et al. (20) 
Immune.2.STAT1 0.381 0.0013 0.0332 0.381 0.0013 0.0332 Desmedt et al. (19) 

NOTE: The bolded letters and figures represent the gene modules that are significantly altered.

Dynamic expression changes in gene modules and Ki67 in response to 2 weeks of AI treatment

Fourteen gene module scores (including PAGs) were significantly decreased after 2 weeks of anastrozole therapy, although none to the same magnitude as the single immunohistochemistry marker Ki67 (Table 2). The ERG module showed the most significant change in response to therapy. CIN70, GGI, AURKA, and the Gene70 module, which mainly quantify tumor proliferation, were each profoundly reduced by anastrozole. For CIN70, GGI, AURKA, and ERG, the P values for the change were all relatively similar to that for Ki67 at c.10−12. Modules representing SET, PTEN, CASP3, E2F3, MYC, ESR1.1, ESR1.2, AKT/mTOR, and IGF-I were also all significantly suppressed but at a lesser magnitude by estrogen deprivation. In contrast, MAPK, Stroma.2.PLAU, Stroma.1, and Immune.1 expression module scores were significantly increased by anastrozole treatment. When the PAGs were removed from the gene modules, the change in expression was significantly reduced in each of the proliferation-based modules. In particular, Gene70 was profoundly affected and was no longer significantly suppressed by AI treatment. However, removal of the PAGs from the remaining gene modules namely ERG, ESR1.1 ESR1.2, SET, MYC, MAPK, Stroma.2.PLAU, and Stroma1 had no significant impact.

Table 2.

Dynamic changes in expression of Gene modules and Ki67 in response to the 2 weeks' AI treatment

All matchedPAGs excluded
Module namea of Geometric mean of intensities of a module of pre- and posttreatmentTwo-sided unadjusted P value of Wilcoxon testAdjusted P%Δ of Geometric mean of intensities of pre- and posttreatmentTwo-sided unadjusted P value of Wilcoxon testAdjusted P
Ki67 −75.23 4.75E−12 4.77E−11 N/A N/A N/A 
ERG −28.36 1.40E−12 3.64E−11 −28.51 1.20E−12 3.64E−11 
CIN70 −24.25 4.20E−12 4.77E−11 −11.34 1.60E−08 7.56E−08 
GGI −22.32 5.00E−12 4.77E−11 −7.82 3.90E−09 2.54E−08 
AURKA −12.34 5.50E−12 4.77E−11 N/A N/A N/A 
Gene70 −8.37 8.60E−09 4.47E−08 −1.11 0.3400 0.3930 
SET −7.99 1.00E−05 3.25E−05 −8.20 6.70E−06 2.32E−05 
PTEN −7.22 2.80E−10 2.08E−09 −2.61 0.0002 3.85E−04 
CASP3 −5.19 0.0114 0.0180 −4.45 0.0420 0.0642 
E2F3 −4.40 4.40E−09 2.54E−08 −2.63 8.10E−05 1.91E−04 
MYC −4.32 1.80E−06 7.20E−06 −4.74 3.90E−07 1.69E−06 
ESR1.2 −4.39 2.00E−04 3.85E−04 −4.53 2.00E−04 3.85E−04 
ESR1.1 −4.18 0.0015 0.0027 −4.56 5.00E−04 9.29E−04 
AKT/mTOR −2.75 0.0052 0.0085 −2.07 0.0590 0.0877 
IGF-I −2.44 3.80E−05 9.88E−05 −1.94 2.00E−04 3.85E−04 
VEGF −1.80 0.1300 1.73E01 −1.80 0.1300 0.1730 
RAS −1.14 0.1100 1.55E01 −1.00 0.2500 0.3020 
SRC 0.52 1.0000 1.0000 0.34 0.7700 0.7850 
PIK3CA 0.47 0.3400 0.3930 0.28 0.4700 0.5200 
Obesity 0.74 0.7100 0.7380 0.67 0.6100 0.6470 
Betacatenin 1.01 0.3800 0.4300 0.73 0.4900 0.5310 
ERBB2 1.62 0.1700 0.2100 2.12 0.0770 0.1110 
Immune.2.STAT1 3.62 0.1400 0.1780 3.69 0.1400 0.1780 
MAPK 5.03 4.80E−06 1.78E−05 4.50 2.20E−05 6.73E−05 
Stroma.2.PLAU 7.57 0.0038 0.0066 7.72 0.0040 0.0067 
Immune.1 13.45 3.20E−05 8.76E−05 13.45 3.20E−05 8.76E−05 
Stroma.1 20.75 4.60E−05 1.14E−04 20.05 8.60E−05 1.94E−04 
All matchedPAGs excluded
Module namea of Geometric mean of intensities of a module of pre- and posttreatmentTwo-sided unadjusted P value of Wilcoxon testAdjusted P%Δ of Geometric mean of intensities of pre- and posttreatmentTwo-sided unadjusted P value of Wilcoxon testAdjusted P
Ki67 −75.23 4.75E−12 4.77E−11 N/A N/A N/A 
ERG −28.36 1.40E−12 3.64E−11 −28.51 1.20E−12 3.64E−11 
CIN70 −24.25 4.20E−12 4.77E−11 −11.34 1.60E−08 7.56E−08 
GGI −22.32 5.00E−12 4.77E−11 −7.82 3.90E−09 2.54E−08 
AURKA −12.34 5.50E−12 4.77E−11 N/A N/A N/A 
Gene70 −8.37 8.60E−09 4.47E−08 −1.11 0.3400 0.3930 
SET −7.99 1.00E−05 3.25E−05 −8.20 6.70E−06 2.32E−05 
PTEN −7.22 2.80E−10 2.08E−09 −2.61 0.0002 3.85E−04 
CASP3 −5.19 0.0114 0.0180 −4.45 0.0420 0.0642 
E2F3 −4.40 4.40E−09 2.54E−08 −2.63 8.10E−05 1.91E−04 
MYC −4.32 1.80E−06 7.20E−06 −4.74 3.90E−07 1.69E−06 
ESR1.2 −4.39 2.00E−04 3.85E−04 −4.53 2.00E−04 3.85E−04 
ESR1.1 −4.18 0.0015 0.0027 −4.56 5.00E−04 9.29E−04 
AKT/mTOR −2.75 0.0052 0.0085 −2.07 0.0590 0.0877 
IGF-I −2.44 3.80E−05 9.88E−05 −1.94 2.00E−04 3.85E−04 
VEGF −1.80 0.1300 1.73E01 −1.80 0.1300 0.1730 
RAS −1.14 0.1100 1.55E01 −1.00 0.2500 0.3020 
SRC 0.52 1.0000 1.0000 0.34 0.7700 0.7850 
PIK3CA 0.47 0.3400 0.3930 0.28 0.4700 0.5200 
Obesity 0.74 0.7100 0.7380 0.67 0.6100 0.6470 
Betacatenin 1.01 0.3800 0.4300 0.73 0.4900 0.5310 
ERBB2 1.62 0.1700 0.2100 2.12 0.0770 0.1110 
Immune.2.STAT1 3.62 0.1400 0.1780 3.69 0.1400 0.1780 
MAPK 5.03 4.80E−06 1.78E−05 4.50 2.20E−05 6.73E−05 
Stroma.2.PLAU 7.57 0.0038 0.0066 7.72 0.0040 0.0067 
Immune.1 13.45 3.20E−05 8.76E−05 13.45 3.20E−05 8.76E−05 
Stroma.1 20.75 4.60E−05 1.14E−04 20.05 8.60E−05 1.94E−04 

NOTE: The bolded letters and figures represent the gene modules that are significantly altered.

a%Δ of Geometric mean of intensities of pre and post treatment = [(post-geometric mean intensity − pre-geometric mean intensity)/pre geometric mean intensity] × 100.

Associations of changes in gene modules and changes in Ki67 with and without PAGs

Assessment of the interconnectivity of the changes in the modules during treatment was strikingly different. ERBB2 with VEGF, Strom.1 with Strom.2.PLAU, and ESR1.1 with ESR1.2 and SET remained in their significantly associated cluster irrespective of the presence of PAGs (Fig. 1C & D). However, removal of the PAGs resulted in a strong association between IGF-I, obesity, MAPK, Src, and betacatenin, also, between changes in PTEN loss, E2F3, CASP3, AKT/mTOR, and CIN70. Notably, this was not a result of shared genes between the modules (Supplementary Table S8).

The change in expression of the ERG module was significantly correlated with the change in Ki67 protein level. Similar correlations with Ki67 were evident for the modules CIN70, GGI, AURKA, AKT/mTOR, PTEN, Gene70, and E2F3. However, these associations became weaker or even non-existent after removal of the PAGs from these modules, indicating that the previous correlations were partially dependent on the colinearity provided by the PAGs within those modules (Supplementary Table S9A and S9B).

Posttreatment associations of gene modules and Ki67 with and without the PAGs

After 2 weeks of treatment, the gene modules were clustered mainly into five pathway-associated clusters as follows (Fig. 1E & F): (i) ER signaling cluster consisting of ESR1-1, ESR1-2, SET, and ERG was evident (AU P ≥ 97%); of note, these modules shared only two genes (NAT1, TFF1); (ii) cell proliferation related cluster; (iii) growth factor receptors and cytoplasmic signaling–related cluster, which contained IGF-I, obesity, MAPK, RAS, ERBB2, VEGF, betacatenin, and SRC gene modules; (iv) tumor invasion cluster (Strom.1, Strom.2.PLAU); and (v) a significantly correlated immune response cluster (Immune.1, Immune.2.STAT1).

Interestingly, in these on-treatment samples, removal of the PAGs had no significant impact on the clustering. This is likely due to the suppression of proliferation in most samples by the AI. Of note, with and without PAGs, the PIK3CA module was not highly associated with any of the clusters, despite its module score being significantly positively correlated with ERG module score (unadjusted P = 0.007 and 0.004 respectively), weak but significantly inversely correlated with obesity module score (unadjusted P = 0.013 and 0.024, respectively; Supplementary Table S10A and S10B).

Posttreatment Ki67 values were significantly positively associated with the proliferation-based modules (Gene70, AURKA, CIN70, and GGI), PTEN, IGF-I, AKT/mTOR, RAS, and Immune.2.STAT1 with and without PAGs, despite the decreased proliferation. However, weak but significant negative correlations were evident between 2-week Ki67 values with expression module scores ESR1.1 (with and without PAGs, unadjusted P = 0.027 and 0.036, respectively). Two-week Ki67 was significantly positively associated with CASP3 and E2F3 only when PAGs were present in modules (unadjusted P = 0.014 and 0.025, respectively), and the statistical significance was lost in both cases when PAGs were removed from the modules (unadjusted P = 0.062 and 0.317, respectively; Supplementary Table S10A and S10B).

The goal to predict those patients most likely to gain benefit from or be resistant to endocrine therapy remains a major clinical challenge. Investigations in the neoadjuvant setting are favored for study of mechanisms. In the current study, we used the change in Ki67 as the index of response. Although pathologic complete response (pCR) is a validated endpoint for chemotherapy in ER or HER2+ disease, it is not a valuable endpoint in ER+ disease, whether this is treated with endocrine therapy or chemotherapy (36, 37). Consistent with this, in the clinical trial from which the data discussed were derived, no patient showed pCR on any arm of the trial (11). In contrast, the change in Ki67 is a widely accepted biomarker of endocrine treatment efficacy (9, 38).

Over the last decade, a vast amount of gene expression data has been generated and used to develop gene modules or metagenes that represent numerous individual pathways or processes. Some are either prognostic or predictive of response to therapy. In the current study, for the first time, we assessed the ability of 26 publicly available gene expression modules encompassing a number of processes associated with tumor biology, to predict Ki67 response to an AI. We also assessed the impact of AI treatment on their expression.

As expected, gene modules based around proliferation (GGI, Gene70, CIN70, and AURKA) correlated strongly with the proliferation marker Ki67. The differences in the correlation of many of the modules with one another, when PAGs were removed, emphasize the influence of proliferation features in many of them and also the importance of their removal, to consider the biologic importance of the modules and represented processes in relation to highly antiproliferative agents, such as AIs. The observation that removal of PAGs had little impact on clustering in the on-treatment samples in which proliferation was markedly reduced suggests the effectiveness of this subtractive approach.

At baseline, expression of the ERG module, consisting of estrogen-regulated genes, showed a trend toward predicting a change in Ki67 as a marker of response to therapy. In addition, the ERG module scores were the most significantly reduced after the 2 weeks' AI treatment. This latter observation supports the constituent genes of the ERG module as being highly estrogen dependent and their baseline expression, as somewhat indicative of endocrine responsiveness. However, it should be noted that this gene signature was originally derived from this cohort of patients (8).

The ESR1.1, ESR1.2, and the SET modules were generated using an association-based approach in which all genes highly correlated with and including ESR1 expression were included (19, 23, 24). In this study, we found that high baseline module scores of these gene signatures predicted good Ki67 response to AI therapy more strongly than the ERG module, indicating that ER-associated rather than estrogen-regulated genes are more important for predicting response to AIs.

The IGF, together with its receptor (IGF-IR), has been strongly implicated in endocrine resistance via hyperactivation of both the PI3K and MAPK pathways, leading to the ability of tumors to circumvent the need for steroid hormones (15, 39, 40). This is consistent with high IGF modules scores being associated with poor response to anastrozole. Furthermore, the MAPK module provided similar trend when PAGs were removed from the module. Not surprisingly, association studies between the gene modules showed that IGF-I was correlated with both the obesity and MAPK gene modules as well as with the obesity module itself, showing a trend to association with poor Ki67 response. This last observation is particularly notable given the recent evidence for obesity being associated with reduced benefit from AIs relative to that from tamoxifen (25).

The PIK3CA-GS (18) showed that at baseline the module was significantly inversely correlated with Ki67, IGF-I, MYC, E2F3, PTEN loss, and gene modules indicative of proliferation. This suggests the lower the PIK3CA-GS signature score the greater the pathway activity, but does not predict change in Ki67, consistent with our recent data on PIK3CA mutations, having no impact on Ki67 response in these same samples (41). One explanation for this observation is that the PIK3CA-GS was derived from a PIK3CA exon 20 mutation. As such, the gene signature may have greatest relevance in selecting patients most likely to respond to PI3K/mTOR inhibition. A recent retrospective study (42) of patients receiving the mTOR inhibitor everolimus in combination with letrozole in the neoadjuvant setting showed, surprisingly, that PIK3CA mutations did not predict for response to the combination or either agent but the PI3KCA-GS was able to identify those patients with ER+ breast cancer who gained benefit from the combination.

One of the most striking observations was that Immune.2.STAT1 predicted poor Ki67 response to anastrozole. This is consistent with our recent report showing that immune-related genes are highly predictive of poor antiproliferative response to AIs (12). This is further supported by the recently published article by Tsang and colleagues (43), which showed that higher lymphocytic infiltration correlated with poor prognosis in ER+ patients. The data suggest that patients showing evidence of high immune-related gene expression and/or lymphocytic infiltration may be candidates for treatment with immune-modulators in combination with an AI. It is, however, in stark contrast with that seen for response to chemotherapy in which high expression of these immune modules was associated with a significantly better outcome in two independent studies (19, 27). A dynamic network of positive and negative regulatory cytokines, which influence not only the type of response but also the amplitude of that response, modulates the immune system. For instance, the immune system has conflicting roles in tumor progression, on the one hand suppressing tumor growth, whereas on the other, providing a proproliferative environment by generating a source of growth factor cues. Hence, one explanation for this result could be that chemotherapy destroys not only the tumor, but also immune-associated cells, whereas endocrine therapy impacts more on cross-talk between the tumor and the immune system. This hypothesis is supported to some extent by recent data, which suggest that letrozole can decrease recruitment of regulatory T cells, which are known to dampen the immune response (44).

Despite these distinct correlations between the E2F3 signature and other modules, no correlation was found between E2F3 and Ki67 protein expression at the baseline, in the presence or absence of the PAGs. This contrasts with a previous study (45), which suggested that an E2F3 module, consisting of 24 non-cell-cycle genes, was significantly associated with Ki67 protein expression at baseline. Furthermore, the authors suggested the expression signature of E2F3 activation in ER+ tumors maybe linked with resistance to AI-induced estrogen deprivation. One potential explanation for these conflicting observations is that 13 of the 24 genes within the E2F3 signature (45) are known PAGs, accounting for the strong correlation with Ki67 protein levels (45).

The biology of breast cancer does not seem conducive to the identification of robust predictive signatures. The molecular mechanisms underpinning resistance to targeted therapeutics may be confined to the functional alteration of key genes, which may not necessarily influence global gene expression profiles. Another important caveat is that resistance mechanisms, such as gene mutations and epigenetic modifications, may be beyond the perception of gene expression analysis, limiting the identification of potentially important predictive alterations (46–48). The heterogeneity of breast cancer also raises the possibility that resistance may emerge through selection of nonmodal subpopulations, which are unable to be discerned by the composite nature of microarray-based analyses (49, 50). Technological advances have the potential to address some of these shortfalls; in particular, the advent of next-generation sequencing technology provides the opportunity for DNA and RNA aberrations to be integrated with transcriptional profiling.

In conclusion, the molecular response to AI treatment varies greatly between patients, consistent with the variable clinical benefit, as a result of tumor heterogeneity. Multiple processes and pathways are affected by AI treatment in ER+ breast cancer; however, reduced proliferation seems to be the dominant process. As expected, those modules most closely associated with ESR1 expression were predictive of good antiproliferative response to AIs. Of note, modules representing immune activity and IGF-I/MAPK were predictive of poor Ki67 response to therapy, supporting our earlier study (8) and highlighting their potential utility as biomarkers.

M. Dowsett has received commercial research grants and speakers bureau honoraria from AstraZeneca. No potential conflicts of interest were disclosed by the other authors.

Conception and design: Q. Gao, L.-A. Martin, M. Dowsett

Development of methodology: Q. Gao, N. Patani, Z. Ghazoui

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): N. Patani, A.K. Dunbier, M. Dowsett

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Q. Gao, N. Patani, A.K. Dunbier, Z. Ghazoui, M. Zvelebil, L.-A. Martin, M. Dowsett

Writing, review, and/or revision of the manuscript: Q. Gao, N. Patani, A.K. Dunbier, Z. Ghazoui, L.-A. Martin, M. Dowsett

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Q. Gao, Z. Ghazoui

Study supervision: L.-A. Martin, M. Dowsett

This work was supported by the Breakthrough Breast Cancer Research Centre and The Mary-Jean Mitchell Green Foundation, and The National Health Service provided funding to the NIHR Biomedical Research Centre at the Royal Marsden Hospital.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
ASCO
. 
American Society of Clinical Oncology
.
Available from:
http://www.cancer.net/cancer-types/breast-cancer/diagnosis.
2.
Smith
IE
,
Dowsett
M
. 
Aromatase inhibitors in breast cancer
.
N Engl J Med
2003
;
348
:
2431
42
.
3.
Chia
YH
,
Ellis
MJ
,
Ma
CX
. 
Neoadjuvant endocrine therapy in primary breast cancer: indications and use as a research tool
.
Br J Cancer
2010
;
103
:
759
64
.
4.
Sorlie
T
,
Perou
CM
,
Tibshirani
R
,
Aas
T
,
Geisler
S
,
Johnsen
H
, et al
Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications
.
Proc Natl Acad Sci U S A
2001
;
98
:
10869
74
.
5.
Sotiriou
C
,
Wirapati
P
,
Loi
S
,
Harris
A
,
Fox
S
,
Smeds
J
, et al
Gene expression profiling in breast cancer: understanding the molecular basis of histologic grade to improve prognosis
.
J Natl Cancer Inst
2006
;
98
:
262
72
.
6.
Davies
C
,
Godwin
J
,
Gray
R
,
Clarke
M
,
Cutter
D
,
Darby
S
, et al
Relevance of breast cancer hormone receptors and other factors to the efficacy of adjuvant tamoxifen: patient-level meta-analysis of randomised trials
.
Lancet
2011
;
378
:
771
84
.
7.
Ali
S
,
Coombes
RC
. 
Endocrine-responsive breast cancer and strategies for combating resistance
.
Nat Rev Cancer
2002
;
2
:
101
12
.
8.
Musgrove
EA
,
Sutherland
RL
. 
Biological determinants of endocrine resistance in breast cancer
.
Nat Rev Cancer
2009
;
9
:
631
43
.
9.
Dowsett
M
,
Nielsen
TO
,
A'Hern
R
,
Bartlett
J
,
Coombes
RC
,
Cuzick
J
, et al
Assessment of Ki67 in breast cancer: recommendations from the International Ki67 in Breast Cancer working group
.
J Natl Cancer Inst
2011
;
103
:
1656
64
.
10.
Dowsett
M
,
Smith
IE
,
Ebbs
SR
,
Dixon
JM
,
Skene
A
,
A'Hern
R
, et al
Prognostic value of Ki67 expression after short-term presurgical endocrine therapy for primary breast cancer
.
J Natl Cancer Inst
2007
;
99
:
167
70
.
11.
Smith
IE
,
Walsh
G
,
Skene
A
,
Llombart
A
,
Mayordomo
JI
,
Detre
S
, et al
A phase II placebo-controlled trial of neoadjuvant anastrozole alone or with gefitinib in early breast cancer
.
J Clin Oncol
2007
;
25
:
3816
22
.
12.
Dunbier
AK
,
Ghazoui
Z
,
Anderson
H
,
Salter
J
,
Nerurkar
A
,
Osin
P
, et al
Molecular profiling of aromatase inhibitor-treated post-menopausal breast tumors identifies immune-related correlates of resistance
.
Clin Cancer Res
2013
;
19
:
2775
86
.
13.
Bild
AH
,
Yao
G
,
Chang
JT
,
Wang
Q
,
Potti
A
,
Chasse
D
, et al
Oncogenic pathway signatures in human cancers as a guide to targeted therapies
.
Nature
2006
;
439
:
353
7
.
14.
Saal
LH
,
Johansson
P
,
Holm
K
,
Gruvberger-Saal
SK
,
She
QB
,
Maurer
M
, et al
Poor prognosis in carcinoma is associated with a gene expression signature of aberrant PTEN tumor suppressor pathway activity
.
Proc Natl Acad Sci U S A
2007
;
104
:
7564
9
.
15.
Creighton
CJ
,
Casa
A
,
Lazard
Z
,
Huang
S
,
Tsimelzon
A
,
Hilsenbeck
SG
, et al
Insulin-like growth factor-I activates gene transcription programs strongly associated with poor breast cancer prognosis
.
J Clin Oncol
2008
;
26
:
4078
85
.
16.
Creighton
CJ
,
Hilger
AM
,
Murthy
S
,
Rae
JM
,
Chinnaiyan
AM
,
El-Ashry
D
. 
Activation of mitogen-activated protein kinase in estrogen receptor alpha-positive breast cancer cells in vitro induces an in vivo molecular phenotype of estrogen receptor alpha-negative human breast tumors
.
Cancer Res
2006
;
66
:
3903
11
.
17.
Majumder
PK
,
Febbo
PG
,
Bikoff
R
,
Berger
R
,
Xue
Q
,
McMahon
LM
, et al
mTOR inhibition reverses Akt-dependent prostate intraepithelial neoplasia through regulation of apoptotic and HIF-1-dependent pathways
.
Nat Med
2004
;
10
:
594
601
.
18.
Loi
S
,
Haibe-Kains
B
,
Majjaj
S
,
Lallemand
F
,
Durbecq
V
,
Larsimont
D
, et al
PIK3CA mutations associated with gene signature of low mTORC1 signaling and better outcomes in estrogen receptor-positive breast cancer
.
Proc Natl Acad Sci U S A
2010
;
107
:
10208
13
.
19.
Desmedt
C
,
Haibe-Kains
B
,
Wirapati
P
,
Buyse
M
,
Larsimont
D
,
Bontempi
G
, et al
Biological processes associated with breast cancer clinical outcome depend on the molecular subtypes
.
Clin Cancer Res
2008
;
14
:
5158
65
.
20.
van ‘t Veer
LJ
,
Dai
H
,
van de Vijver
MJ
,
He
YD
,
Hart
AA
,
Mao
M
, et al
Gene expression profiling predicts clinical outcome of breast cancer
.
Nature
2002
;
415
:
530
6
.
21.
Carter
SL
,
Eklund
AC
,
Kohane
IS
,
Harris
LN
,
Szallasi
Z
. 
A signature of chromosomal instability inferred from gene expression profiles predicts clinical outcome in multiple human cancers
.
Nat Genet
2006
;
38
:
1043
8
.
22.
Dunbier
A
.
Available from
: http://synapse.sagebase.org/#Dataset:16243.
23.
Mackay
A
,
Urruticoechea
A
,
Dixon
JM
,
Dexter
T
,
Fenwick
K
,
Ashworth
A
, et al
Molecular response to aromatase inhibitor treatment in primary breast cancer
.
Breast Cancer Res
2007
;
9
:
R37
.
24.
Symmans
WF
,
Hatzis
C
,
Sotiriou
C
,
Andre
F
,
Peintinger
F
,
Regitnig
P
, et al
Genomic index of sensitivity to endocrine therapy for breast cancer
.
J Clin Oncol
2010
;
28
:
4111
9
.
25.
Creighton
CJ
,
Sada
YH
,
Zhang
Y
,
Tsimelzon
A
,
Wong
H
,
Dave
B
, et al
A gene transcription signature of obesity in breast cancer
.
Breast Cancer Res Treat
2012
;
132
:
993
1000
.
26.
Farmer
P
,
Bonnefoi
H
,
Anderle
P
,
Cameron
D
,
Wirapati
P
,
Becette
V
, et al
A stroma-related gene signature predicts resistance to neoadjuvant chemotherapy in breast cancer
.
Nat Med
2009
;
15
:
68
74
.
27.
Teschendorff
AE
,
Miremadi
A
,
Pinder
SE
,
Ellis
IO
,
Caldas
C
. 
An immune response gene expression module identifies a good prognosis subtype in estrogen receptor negative breast cancer
.
Genome biol
2007
;
8
:
R157
.
28.
Ur-Rehman
S
,
Gao
Q
,
Mitsopoulos
C
,
Zvelebil
M
. 
ROCK: a resource for integrative breast cancer data analysis
.
Breast Cancer Res Treat
2013
;
139
:
907
21
.
29.
Ben-Porath
I
,
Thomson
MW
,
Carey
VJ
,
Ge
R
,
Bell
GW
,
Regev
A
, et al
An embryonic stem cell-like gene expression signature in poorly differentiated aggressive human tumors
.
Nat Genet
2008
;
40
:
499
507
.
30.
Whitfield
ML
,
Sherlock
G
,
Saldanha
AJ
,
Murray
JI
,
Ball
CA
,
Alexander
KE
, et al
Identification of genes periodically expressed in the human cell cycle and their expression in tumors
.
Mol Biol Cell
2002
;
13
:
1977
2000
.
31.
Hu
Z
,
Fan
C
,
Oh
DS
,
Marron
JS
,
He
X
,
Qaqish
BF
, et al
The molecular portraits of breast tumors are conserved across microarray platforms
.
BMC genomics
2006
;
7
:
96
.
32.
Ignatiadis
M
,
Singhal
SK
,
Desmedt
C
,
Haibe-Kains
B
,
Criscitiello
C
,
Andre
F
, et al
Gene modules and response to neoadjuvant chemotherapy in breast cancer subtypes: a pooled analysis
.
J Clin Oncol
2012
;
30
:
1996
2004
.
33.
Wirapati
P
,
Sotiriou
C
,
Kunkel
S
,
Farmer
P
,
Pradervand
S
,
Haibe-Kains
B
, et al
Meta-analysis of gene expression profiles in breast cancer: toward a unified understanding of breast cancer subtyping and prognosis signatures
.
Breast Cancer Res
2008
;
10
:
R65
.
34.
Suzuki
R
,
Shimodaira
H
. 
Pvclust: an R package for assessing the uncertainty in hierarchical clustering
.
Bioinformatics
2006
;
22
:
1540
2
.
35.
Venet
D
,
Dumont
JE
,
Detours
V
. 
Most random gene expression signatures are significantly associated with breast cancer outcome
.
PLoS Comput Biol
2011
;
7
:
e1002240
.
36.
Kaufmann
M
,
von Minckwitz
G
,
Mamounas
EP
,
Cameron
D
,
Carey
LA
,
Cristofanilli
M
, et al
Recommendations from an international consensus conference on the current status and future of neoadjuvant systemic therapy in primary breast cancer
.
Ann Surg Oncol
2012
;
19
:
1508
16
.
37.
von Minckwitz
G
,
Untch
M
,
Blohmer
JU
,
Costa
SD
,
Eidtmann
H
,
Fasching
PA
, et al
Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes
.
J Clin Oncol
2012
;
30
:
1796
804
.
38.
Ellis
MJ
,
Suman
VJ
,
Hoog
J
,
Lin
L
,
Snider
J
,
Prat
A
, et al
Randomized phase II neoadjuvant comparison between letrozole, anastrozole, and exemestane for postmenopausal women with estrogen receptor-rich stage 2 to 3 breast cancer: clinical and biomarker outcomes and predictive value of the baseline PAM50-based intrinsic subtype–ACOSOG Z1031
.
J Clin Oncol
2011
;
29
:
2342
9
.
39.
Drury
SC
,
Detre
S
,
Leary
A
,
Salter
J
,
Reis-Filho
J
,
Barbashina
V
, et al
Changes in breast cancer biomarkers in the IGF1R/PI3K pathway in recurrent breast cancer after tamoxifen treatment
.
Endocr Relat Cancer
2011
;
18
:
565
77
.
40.
Johnston
SR
,
Dowsett
M
. 
Aromatase inhibitors for breast cancer: lessons from the laboratory
.
Nat Rev Cancer
2003
;
3
:
821
31
.
41.
Segal
C
,
Lopez-Knowles
E
,
Patel
V
,
Garcia-Murillas
I
,
Turner
N
,
Martin
L
, et al
Relationship of PIK3CA mutation and PI3K pathway activity with antiproliferative response to anastrazole in ER positive breast cancer
.
Annals of Oncology
2013
;
24
Suppl 3
:
iii
30
.
42.
Loi
S
,
Michiels
S
,
Baselga
J
,
Bartlett
JM
,
Singhal
SK
,
Sabine
VS
, et al
PIK3CA genotype and a PIK3CA mutation-related gene signature and response to everolimus and letrozole in estrogen receptor positive breast cancer
.
PLoS ONE
2013
;
8
:
e53292
.
43.
Tsang
JY
,
Hui
SW
,
Ni
YB
,
Chan
SK
,
Yamaguchi
R
,
Kwong
A
, et al
Lymphocytic infiltrate is associated with favorable biomarkers profile in HER2-overexpressing breast cancers and adverse biomarker profile in ER-positive breast cancers
.
Breast Cancer Res Treat
2014
;
143
:
1
9
.
44.
Generali
D
,
Bates
G
,
Berruti
A
,
Brizzi
MP
,
Campo
L
,
Bonardi
S
, et al
Immunomodulation of FOXP3 +regulatory T cells by the aromatase inhibitor letrozole in breast cancer patients
.
Clin Cancer Res
2009
;
15
:
1046
51
.
45.
Miller
TW
,
Balko
JM
,
Fox
EM
,
Ghazoui
Z
,
Dunbier
A
,
Anderson
H
, et al
ERalpha-dependent E2F transcription can mediate resistance to estrogen deprivation in human breast cancer
.
Cancer Discov
2011
;
1
:
338
51
.
46.
Ding
L
,
Ellis
MJ
,
Li
S
,
Larson
DE
,
Chen
K
,
Wallis
JW
, et al
Genome remodelling in a basal-like breast cancer metastasis and xenograft
.
Nature
2010
;
464
:
999
1005
.
47.
Weigelt
B
,
Pusztai
L
,
Ashworth
A
,
Reis-Filho
JS
. 
Challenges translating breast cancer gene signatures into the clinic
.
Nat Rev Clin Oncol
2012
;
9
:
58
64
.
48.
Creighton
CJ
,
Fu
X
,
Hennessy
BT
,
Casa
AJ
,
Zhang
Y
,
Gonzalez-Angulo
AM
, et al
Proteomic and transcriptomic profiling reveals a link between the PI3K pathway and lower estrogen-receptor (ER) levels and activity in ER+ breast cancer
.
Breast Cancer Res
2010
;
12
:
R40
.
49.
Borst
P
,
Wessels
L
. 
Do predictive signatures really predict response to cancer chemotherapy?
Cell Cycle
2010
;
9
:
4836
40
.
50.
Colombo
PE
,
Milanezi
F
,
Weigelt
B
,
Reis-Filho
JS
. 
Microarrays in the 2010s: the contribution of microarray-based gene expression profiling to breast cancer classification, prognostication and prediction
.
Breast Cancer Res
2011
;
13
:
212
.

Supplementary data