Four miRNAs associated with aggressiveness of lymph node-negative, estrogen receptor-positive human breast cancer

Abstract

In this study, we quantified 249 mature micro-RNA (miRNA) transcripts in estrogen receptor-positive (ER⁺) primary breast tumors of patients with lymph node-negative (LNN) disease to identify miRNAs associated with metastatic capability. In addition, the prognostic value of the candidate miRNAs was determined in ER⁻/LNN breast cancer. Unsupervised analysis in a prescreening set of 38 patients identified three subgroups predominantly driven by three miRNA signatures: an ER-driven luminal B-associated miRNA signature, a stromal miRNA signature, and an overexpressed miRNA cluster located on chromosome 19q23, but these intrinsic miRNA signatures were not associated with tumor aggressiveness. Supervised analysis in the initial subset and subsequent analysis in additional tumors significantly linked four miRNAs (miR-7, miR-128a, miR-210, and miR-516–3p) to ER⁺/LNN breast cancer aggressiveness (n = 147) and one miRNA (miR-210) to metastatic capability in ER⁻/LNN breast cancer (n = 114) and in the clinically important triple-negative subgroup (n = 69) (all P < 0.05). Bioinformatic analysis coupled miR-210 to hypoxia/VEGF signaling, miR-7 and miR-516–3p to cell cycle progression and chromosomal instability, and miR-128a to cytokine signaling. In conclusion, our work connects four miRNAs to breast cancer progression and to several distinct biological processes involved therein.

Various prognostic mRNA expression signatures have been revealed (1–3), but the forces driving breast cancer aggressiveness are poorly understood. Micro-RNAs (miRNAs) are small nonprotein-coding RNAs (4, 5) that use the endogenous RNA interference pathway to modulate gene expression, thereby contributing to various normal cellular and developmental processes and to malignant transformation and tumor progression (6, 7). Currently, >500 human miRNA genes have been identified (http://microrna.sanger.ac.uk/sequences), and biochemical and in silico analyses have suggested that each of these miRNAs targets a plethora of mRNAs. Furthermore, it is likely that miRNAs can act as tumor suppressor genes and oncogenes (5, 7, 8) because they are located in genomic regions that show copy number alterations in a variety of cancer types (9). To make the picture more complete, next to miRNAs, a new class of nonprotein-coding RNAs encoded by transcribed ultra-conserved regions exists that might interact with miRNAs and are also located in genomic regions that show gain or loss in human cancers (8). A current model proposes that coding and noncoding RNAs cooperate in the initiation and progression of malignancy (8). Focusing on breast cancer, levels of specific miRNAs vary between normal and malignant breast tissue, tumors of different grade, molecular subtype, and nodal and steroid hormone-receptor status (10, 11), and a few miRNAs have been causally linked to breast cancer proliferation and invasion (12–14). However, convincing data in extensive series of well-defined clinical specimens are currently lacking. In this study, we explored whether miRNAs are associated with prognosis in ER⁺/LNN breast cancer. In addition, biological pathways were connected to the identified prognostic miRNAs using global gene expression data also available for the studied specimen. Finally, the prognostic significance of selected miRNAs was also studied in ER⁻/LNN breast cancer and in triple-negative breast cancer (i.e., those tumors that are negative for both steroid receptors and do not show HER-2 amplication).

Results

To identify clinically relevant miRNAs associated with time to distant metastasis (TDM) (i.e., tumor aggressiveness) in ER⁺/LNN primary breast cancer, we analyzed the miRNA expression levels by real-time RT-PCR in an extensive cohort of breast cancers. None of these patients had received any (neo)adjuvant systemic therapy. We performed a prescreen on 38 samples to identify prognostic candidates among the 249 miRNAs quantitated by real-time PCR. Unsupervised Pearson correlation clustering using the 75% most variable miRNAs subdivided these 38 samples into three groups (Fig. 1A). Careful examination of the cluster arms showed that three “intrinsic” miRNA signatures drove this subdivision [supporting information (SI); Fig. S1a]. To further explore these intrinsic miRNA signatures, we re-clustered the samples using these three miRNA signatures separately and associated clinical and biological information with the generated subgroups. This revealed that miRNA signature 1 was dominated by a single co-expressed cluster of miRNAs [i.e., miR-515–522, located on chromosome 19q23 (Fig. 1B)]. The second miRNA signature (Fig. 1C) consisted of stroma-specific miRNAs because these miRNAs were predominantly expressed in stromal fibroblast and endothelial cells and marginally in the epithelial breast cancer cell lines (Fig. 1D). Furthermore, their expression correlated with the stromal content of the specimens (Fig. 1E; Mann–Whitney test, P = 0.04). The third intrinsic miRNA signature (Fig. 1F) associated with ER mRNA expression (Fig. 1G; Mann–Whitney test, P = 0.0006) separated luminal B from other intrinsic subtypes of breast cancer (Fig. 1G; Fisher's exact test, P = 0.0002) and specifically identified ER⁺ breast cancer cell lines (Fig. 1H) among the epithelial breast cancer cell lines and stromal endothelial and fibroblast cells studied (see Fig. S1b for complete clustering diagram). Several miRNAs in this latter signature have previously been linked to ER and PgR status in breast cancer (10). This indicates that also within ER⁺ breast tumors, miRNA expression is influenced by steroid hormones or their receptors, as is the case for many ordinary coding messengers (15). Although we found specific biological features connected to these three intrinsic miRNA signatures, they were not linked to breast cancer aggressiveness.

Fig. 1.

Exploring the miRNA expression profile of 38 ER⁺/LNN breast cancer specimens screened for the expression of 249 miRNAs. (A) Pearson correlation plot of the specimens based on the most variable miRNAs representing the top 75% of those measurable, revealed three main clusters marked a, b, and c. Red indicates a strong positive correlation, whereas green indicates a weak positive correlation between samples. Three miRNA signatures (1, 2, and 3) were identified driving the unsupervised hierarchical clustering (see Fig. S1a; colors of the subgroups a, b, and c match between Fig. 1A and Fig. S1a). (B, C, F) Hierarchical clustering of specimens based on the three dominant miRNA signatures separately. (D) MiRNAs specifically expressed in the stromal cell lines. (E) Tumor content of tumor samples with the sample groups after clustering using miRNA signature 2. (F) Cluster arm of c is presented transposed, with (Right) the tumor content (relative to stromal content) observed in the hematoxylin-eosin–stained sections of the samples. Shown is a part of the hierarchical clustering plot of breast cancer, endothelial and fibroblast cell lines that represent the miRNAs specifically expressed in the stromal cell lines. The complete clustering plot of the cell lines mentioned is shown in Fig. S1b. (G) Biological features (intrinsic subtypes and mRNA expression of ER, PgR, and Her-2) associated with the sample groups after clustering using miRNA signature 3. The cluster arm of F is presented transposed, with the subtype of the samples (right) and the quantitative mRNA expression levels of ER, PgR, and Her-2. For the subtypes, dark blue represents luminal B, light blue luminal A, red Her-2, green normal-like, and orange basal tumors. (H) MiRNAs specifically expressed in the ER⁺ breast cancer cell lines. Shown is a part of hierarchical clustering plot that represents the miRNAs more abundantly expressed in the ER⁺ cell lines. For the complete clustering plot, see Fig. S1b. Note the significant overlap (<60%) between the miRNAs present in C and D and in F and H. For B, C, D, F, and H, red color indicates high expression level and green color indicates low expression level of the indicated miRNA in a particular specimen.

Supervised analysis in this prescreening set of 38 ER⁺/LNN breast tumors, on the other hand, suggested that several miRNAs were differentially expressed between patients with a short TDM versus those with a long TDM. Twelve candidate miRNAs most differentially expressed between the two groups, miR-7, miR-22, miR-34b, miR-128a, miR-145, miR-151, miR-193b, miR-205, miR-210, miR-449, miR-489, and miR-516–3p, were further explored by qRT-PCR in the remaining 147 ER⁺ cases and in 114 available ER⁻/LNN primary breast cancer cases. Patient and tumor characteristics in the complete cohort, including the 38 initial cases, and their associations with the levels of the 12 candidate miRNAs, are summarized in Table S1. Of note are the positive association of miR-7, miR-34b, and miR-151 with tumor size (P < 0.05); the association of miR-7, miR-210, miR-489, and miR-516–3p with pathological grade (P < 0.01); and the positive association of miR-34b, miR-145, miR-193b, miR-449, and miR-489 with each other (n = 298, Spearman rank correlation, P < 0.001; not shown) and with steroid hormone receptor levels (P < 0.001).

In the independent 147 ER⁺/LNN samples, only levels of miR-7, miR-128a, miR-210, and miR-516–3p were significantly associated with a shorter TDM in univariate and multivariate Cox regression analysis corrected for traditional prognostic factors (all P < 0.05 using a false discovery rate of 10%; see Table S2 and Kaplan-Meier survival curves in Fig. S2 a–d). Combining the information of these four prognostic miRNAs either by using hierarchical clustering (Fig. 2A; P = 0.002) or by building a Cox model (Fig. 2B; P < 0.0001) showed that this panel of miRNAs significantly identified subgroups with a very different TDM.

Fig. 2.

Kaplan–Meier TDM curves of primary breast cancer patients based on miRNA expression levels. Kaplan–Meier TDM curves for 147 ER⁺/LNN cases as a function of four subgroups (clusters 1–4) revealed by hierarchical clustering of the samples based on the levels of miR-7, miR-128a, miR-210, and miR-516-3p (clustering plot is shown above the Kaplan–Meier curve) (A) and as a function of three equally sized subgroups (quantiles 1–3) based on a score of the coefficients of the multivariable Cox regression model of miR-7, miR-128a, miR-210, and miR-516-3p as continuous variables (B). (C) Kaplan–Meier TDM curves for 69 triple-negative (ER⁻/PgR⁻/Her-2⁻) breast cancer cases as a function of subgroups dichotomized at the median miR-210 expression level. Patients at risk at the indicated time intervals are shown at the bottom of each graph.

In the 114 ER⁻/LNN breast cancer cases, of the 12 candidate prognostic miRNAs identified in ER⁺ disease, only higher miR-210 levels were in univariate analysis significantly related to shorter TDM (hazard ratio = 1.72; 95% confidence interval: 1.12–2.64; P = 0.01). Equally sized subgroups based on the miR-210 expression levels showed a significant different prognosis in the whole ER⁻ subgroup (see Fig. S2e; P = 0.0026) and in the triple-negative (ER⁻/PgR⁻/Her-2⁻) subgroup (Fig. 2C; P = 0.05).

By using global testing (16), the availability of transcriptome information on these same specimens allowed us to associate biological pathways with the four prognostic miRNAs (see Table S3 for all top ranking pathways). We restricted the analysis to the ER⁺ tumors, the primary focus of our study. MiR-210 was most significantly coupled to “VEGF, hypoxia and angiogenesis” (comparative P = 0.015), in which HRAS, PTK2, SHC1, and HIF1A were overexpressed and KDR was under-expressed in tumors having high versus those having low miR-210 expression (Fig. 3A). Furthermore, based on two published hypoxia gene expression signatures (17, 18), hypoxic tumors had higher miR-210 expression levels than normoxic tumors (both P < 0.001; see Fig. S3). MiR-7 and miR-516–3p were most significantly associated with cell cycle-related events (miR-7 to “G2-to-M cell cycle checkpoint” [Fig. 3B, comparative P = 0.012]; miR-516–3p to “cyclins and cell cycle regulation” [Fig. 3C, comparative P = 0.002]). Cell cycle checkpoint deregulation in cancer has a link to chromosomal instability. Indeed, samples that had predicted chromosomal instability based on a published chromosomal instability gene signature (19) had high miR-7 (P < 0.001) and high miR-516–3p (P = 0.008) expression (see Fig. S3). Finally, miR-128a was inversely related to various IL (IL-17 and IL-4), lymphocytic and inflammation signaling pathways (Fig. 3D; only the IL-17 pathway is shown; comparative P < 0.05 in all cases).

Fig. 3.

Pathways associated with clinically relevant miRNAs. Pathway analysis plots of Biocarta pathways most strongly associated with miR-210 (A), miR-7 (B), miR-516–3p (C), and miR-128a (D). A pathway plot [generated by the Global Test program (16)] shows the genes annotated to the indicated pathway in Biocarta with their association with tumors showing a low versus high expression of the indicated miRNA. Each bar represents a gene in the pathway. The height of the bar indicates the contribution (influence) of each individual gene to the significance of the pathway. Horizontal markers in a bar indicate 1 SD away from the reference point, and two or more horizontal lines in a bar indicate a statistically significant association of the corresponding gene with the subgroups identified by a specific miRNA. Red and green colors indicate positive and negative associations with the tumor subgroup having high versus low expression of a particular miRNA.

Discussion

miRNAs are an emerging group of ribonucleotides, and their clinical value in breast cancer is still largely unknown. The implications of our findings in the current study are twofold. Unsupervised analysis of miRNA expression data in 38 ER⁺ clinical specimens screened for the entire panel of 249 miRNAs identified three miRNA signatures that we arbitrarily labeled intrinsic miRNA signatures. Interestingly, one signature was co-expressed in luminal B intrinsic subtype cancers, indicating, in line with suggestions of others (11), that intrinsic molecular subtypes based on mRNA expression (20, 21) have specific miRNAs co-expressed. The second miRNA signature was abundantly expressed in stromal cell lines, and is thus likely to be expressed in tumor stroma. These miRNAs might thus prove to be useful tumor-specific stromal markers. The third signature identified a single cluster of miRNAs located on chromosome 19q23, a region that does not show any amplification in the breast cancer specimen in our study cohort (unpublished data), suggesting these miRNAs are expressed in a co-regulated manner. Although the observations above are interesting, their clinical significance remains to be determined and requires additional validation.

The supervised approach, however, identified among 249 miRNAs several candidate prognostic miRNAs. These were further studied in a significant number of well characterized clinical specimens, leading to the identification of four miRNAs (miR-7, miR-128a, miR-210, and miR-516–3p) associated with tumor aggressiveness in ER⁺/LNN and one miRNA (miR-210) linked with early relapse in ER⁻/LNN breast cancer. MiRNA-210 was also significantly associated with poor outcome in the 69 triple-negative breast cancers present among the specimens studied, and this miRNA may prove to be a valuable marker, because strong prognostic markers in this clinically important breast cancer subgroup are sparsely available.

Noteworthy is that our large-scale screen in clinical specimens did not single out miRNAs miR-10b, miR-126, miR-335, miR-373, and miR-520c, which were recently functionally linked to breast cancer migration and invasion in animal models and subsequently linked to progression of clinical breast cancer (12–14). Even though these miRNAs were detectable by our method, some of them may be more relevant in ER⁻ or other specific subtypes of breast cancer. Alternatively, they could have remained unrevealed in our study because of the notorious heterogeneous nature of breast cancer. On the other hand, our study did confirm the clinical significance of miR-210, which was just recently connected to breast cancer outcome in a heterogeneous cohort of 219 breast cancer patients (22), which, unlike ours, also included patients with nodal involvement and those having received various types of adjuvant treatment. That previous study does not necessarily associate miR-210 with pure prognosis of breast cancer as was established in our current study, because the observed effects may have been confounded by the treatments given.

The potential biological significance of our identified miRNAs in breast cancer was revealed by studying co-expressed messenger RNAs. This way, miR-210 was connected to hypoxia in agreement with the fact that miR-210 expression is induced under hypoxic conditions in breast and ovarian cancer cell lines and is under direct transcriptional control of the transcription factor HIF-1α (22–24). The independent conformation by several groups that miR-210 has a link with hypoxia confirms that global testing of co-expressed messengers grouped in biological pathways is a feasible approach. In a similar way, miR-7 and miR-516–3p, which both were associated with pathological grade, were linked to cell cycle deregulation, the pathway currently most frequently linked to prognosis, particularly in ER⁺ breast cancer (1, 3, 25, 26). Finally, miR-128a was inversely associated with genes usually expressed in lymphocytes, suggestive of a role of this miRNA in lymphocytic infiltration or immune response in breast cancer metastasis (26).

In conclusion, this study provides a leap forward because it has connected expression levels of four miRNAs and associated pathways with breast tumor aggressiveness. Only miR-210 was also associated with clinical outcome in ER⁻/LNN breast cancers, including the clinically important triple-negative subgroup. Finally, if the miRNAs identified here, which are all over-expressed in tumors that relapse early, are causally linked to breast cancer progression, silencing them has great therapeutic potential and may guide the biotechnical development of a new type of targeted drugs (27, 28).

Materials and Methods

Patients' Samples.

The institutional medical ethics committee approved the current study (medical ethics committee number 02·953), which was performed in accordance with the Code of Conduct of the Federation of Medical Scientific Societies in The Netherlands (http://www.fmwv.nl/) and which is reported according to the REMARK guidelines (29). The 344-LNN patient cohort from which our 299 samples were selected because of the availability of sufficient good-quality miRNA and Affymetrix U133A mRNA expression data has been described in detail previously (2, 26). None of the patients had received any systemic (neo)adjuvant therapy. From 185 ER⁺ and 114 ER⁻ tumors, sufficient total RNA was present to perform the miRNA expression profiling analysis. In the prescreen, 38 ER⁺ samples were analyzed, leaving 147 ER⁺ cases in which the miRNA candidates from the prescreen were subsequently studied. For one prescreen sample, we ran out of RNA during the study, leaving 184 ER⁺ cases for the final candidate analysis. For these 184 ER⁺ and 114 ER⁻ cases (298 in total), the patient and tumor characteristics are presented in detail in Table S1. The median age of the patients in the study at surgery was 52 years (range: 26–83 years). Two hundred seventeen patients (72%) had undergone breast-conserving surgery, and the remaining patients were treated with radical mastectomy. Time to distant metastasis was defined as the time between surgery and the development of distant metastasis. Median follow-up time of patients alive was 99 months (range: 20–186 months). One hundred one (34%) patients developed a distant metastasis, which counted as an event in the analysis for TDM. Patients who died without evidence of disease were censored at last follow-up in the analysis of TDM.

Tissue Processing, RNA Isolation, Multiplex and Uniplex cDNA Synthesis, and Quantification of miRNA Species.

Tissue processing, total RNA isolation, and total RNA quality control checks have been described elsewhere (30, 31). The TaqMan Human MicroRNA Assay Set (release date March 2006; see Table S4; Applied Biosystems [ABI]), consisting of 259 unique assays to quantify 249 miRNAs and 10 controls (Z30 and nine different SNORs/RNUs), was used to screen a set of 38 ER⁺ breast cancer specimens for differentially expressed miRNAs. The specimens were selected from the ER⁺ cases in the cohort at random with the restriction that 19 were from patients who developed a distant metastasis within 3 years and 19 were from patients who did not develop a distant metastasis within 7 years. The use of a specific primer with a hairpin structure during cDNA synthesis and mature miRNA specific detection probes precluded the detection of precursor miRNAs (ABI). For analysis of the remaining cohort with the 12 selected miRNAs, the individual TaqMan human MicroRNA Assay kits from ABI were used. In brief, up to 48 different RT primers (250 nM each) were pooled according to the recommendations of ABI, concentrated for 30 min in a speed vacuum centrifuge at 50°C and re-suspended in nuclease-free ddH₂O to a final concentration of 50 nM each. Twenty-five to 50 ng of total RNA sample aliquots were reverse-transcribed in a final volume of 20 μl with a final concentration of 12.5 nM for each RT primer using the TaqMan MicroRNA Reverse Transcription Kit (ABI) according to the manufacturer's instructions. To verify that the multiplex RT approach did not affect the quantification of specific miRNAs, some RT primers were included in different RT pools and expression levels of a number of the more lowly expressed miRNAs were validated in a uniplex RT reaction. These experiments showed that especially for the lower expressed miRNAs, some gain in sensitivity could be achieved in a uniplex RT reaction compared with a multiplex RT reaction. We therefore decided to assess expression levels of all the low expressed miRNAs (requiring >30 cycles before passing the threshold level) in a uniplex RT format.

Before performing a duplicate real-time PCR for each of the miRNAs separately, RT samples were diluted in nuclease-free ddH₂O and amplified in a final volume of 20 μl containing cDNA synthesized from 0.25 ng of total RNA, 6 μl of TaqMan universal PCR master mix without UNG, and 6 μl of the individual TaqMan MicroRNA primer and probe assays. Real-time PCR was performed in an Mx3000P Real-Time PCR System (Stratagene) with cycling conditions according to the manufacturer's instructions. A pool consisting of RNA from all 38 human breast tissues included in the prescreening set was included in each cDNA synthesis run, and the resulting data were used to normalize for in-between experimental variations. In addition, all cDNA synthesis runs incorporated a minus RT reaction, which proved to be negative for all assays in this study. PCR efficiency, linearity, and the upper and lower detection limits of each of the individual miRNA assays were validated with a standard curve constructed from a simultaneously run serially diluted pool of cDNA from all samples included in the particular run. To check for RNA input, integrity, and yield of amplifiable cDNA, the reverse primer (HPRT1-R) of the medium abundant reference gene hypoxanthine-guanine phosphoribosyltransferase (HPRT1) was used in a uniplex cDNA synthesis reaction, followed by amplification of the HPRT1 gene: HPRT1-F, 5′-TTC CTT GGT CAG GCA GTA TAA TCC-3; HPRT1-R, 5′-GGT CCT TTT CAC CAG CAA GCT-3; HPRT1 probe, FAM-[CTG GCT TCA CCA TCG]-BHQ1, TAMRA (Sigma Genosys). Because appropriate reference molecules for small RNAs are still unknown for clinical breast cancer, the most stable miRNA in our prescreening set of 38 ER⁺ breast cancer specimens was calculated with geNorm (32) and NormFinder (33) with the Datan Framework GenEx Pro package version 4.3.2 and was shown to be miR-132. In addition, geNorm and NormFinder identified miR-374 as second best. Reassuring is that we found no correlation between these two miRNAs and ER-protein levels (Spearman r_s = −0.03 for miR-374 and 0.11 for miR-132, respectively) which is the most important confounding factor in clinical breast cancer. Thus, we selected miR-132 (median Ct: 26.71) and miR-374 (median Ct: 25.24) to normalize the raw expression data of the 213 different miRNAs that could be reliably assessed in our prescreening breast tumor cohort (median Ct: 27.94, range: 20.22–39.55). Finally, miR-374 and miR-132 turned out to be the least variable miRNAs in the extended series that were analyzed for the 12 candidate miRNAs. Concentrations of the miRNAs, expressed relative to our miRNA reference set (miR-132 and miR-374), were quantified as follows: miRNA target = 2^{(average Ct miR-132 + miR-374 − mean Ct target miR)}.

Statistics and Bioinformatics.

Differences in expression levels between groups were assessed with the Mann–Whitney U test or Kruskal–Wallis test. In these tests, patient and tumor characteristics were used as grouping variables. The Fisher's exact test was used to test for associations between categorized variables. The associations between continuous variables were tested using Spearman rank correlation (r_s). In the subset of 38 ER⁺/LNN primary breast cancers, the Student's t test was used to identify differentially expressed miRNAs between the patients with a short or long TDM. Because this prescreening aimed at identifying candidate miRNAs for confirmation in independent samples, a P of 0.1 without correcting for multiple comparisons was used. In the remaining 147 ER⁺ and 114 ER⁻ cases, the prognostic value of the clinical and biological variables, with TDM as the endpoint in the univariate and multivariable analyses, was investigated with the Cox proportional hazards model, and the hazard ratio and its 95% confidence interval were estimated (34). To adjust for multiple comparisons, univariate P-values were considered significant using a false discovery rate of 10% (35). For this regression analysis, miRNA expression levels were either log- or Box-Cox–transformed to reduce skewness. To combine the information of multiple miRNAs, we calculated an index using the estimates from the Cox model. Kaplan-Meier survival plots, log-rank tests, and log-rank tests for trend were used to assess the differences in TDM of the different predicted risk groups of patients. Triple-negative breast cancer (i.e., ER⁻, PgR⁻, non-HER2 amplified), was defined with cutpoints for ER and PgR protein as described, respectively (2, 36). Her-2 mRNA levels, as determined by real-time PCR (37), displayed a clear bimodal distribution; therefore, the cutpoint for being Her-2 amplified or not was determined as described (38) and defined at 18.01355. All P-values are two-sided and, except for the prescreen, a P ≤ 0.05 was considered statistically significant. Regression analysis and association studies were performed using the statistical package STATA, release 10 (Stata).

Pathway Analysis.

Affymetrix microarray gene expression data (HG-U133A chips) previously deposited in the NCBI/GEO database (entries GSE2034 and GSE5327) were available from all 184 ER⁺ patients and 114 ER⁻ patients. As done previously (2, 26), gene expression signals were calculated using Affymetrix GeneChip analysis software MAS 5.0. Global scaling was performed to bring the average signal intensity of a chip to a target of 600 before data analysis. Samples were split according to their ESR1 mRNA expression level (39). Within each sample set, the intensities for each probe set were thresholded at 30 and were then expressed relative to the geometric mean of that probe set and 2-log transformed. A sample was labeled “high” or “low” for an miRNA if it belonged to the top 25% or bottom 25% of all samples according to the expression levels of that particular miRNA. Thus, the 25% (n = 46) ER⁺ samples with the highest miR-7 expression levels were compared with the 46 ER⁺ samples with the lowest miR-7 expression levels, and likewise for each miRNA of interest. The Global Test program (16) was used (version 4.2.0) to relate Biocarta pathways (http://www.biocarta.com/) to samples expressing high or low levels of a particular miRNA, irrespective of their association with TDM. All P-values were corrected for multiple testing and checked by re-sampling at 1000 times to ensure that an equally sized and randomly chosen group of genes would not also be significant. Pathways were considered of interest if the P value of the Global Test after correcting for multiple testing and the re-sampling P value were both below 0.05. Pathway P-values were two-sided and corrected for multiple testing. The contribution of individual genes in a pathway was evaluated using z-scores calculated by the Global Test program. Genes with z-scores >1.96 were considered significant contributors to the pathway. R version 2.4.1 (http://www.cran.org) was used to run the Global Test package.