To generate a comprehensive set of miRNA expression profiles for primary human breast cancer we selected 99 primary human tumors, 5 normal breast samples and 33 breast cancer cell lines for miRNA expression profiling. Tumor samples were fresh-frozen and collected from Nottingham City Hospital Tumor Bank and are representative with regard to tumor subtypes and clinical parameters 808182. For miRNA profiling we chose a bead-based flow-cytometric miRNA expression platform, which has recently been developed and was found to have several advantages over glass-slide microarray profiling, including increased specificity 56. We developed this platform further to include 333 probes for 309 unique human miRNAs based on the miRNA repository miRBase 8.1 1718. miRNA labeling included RNA size selection using native polyacrylamide gels, ensuring that only mature miRNAs were assayed.

Using this miRNA expression platform we analyzed a total of 137 samples in 168 assays. Assays for 119 of these 137 samples (87%) passed our quality control, including 93 primary tumor samples, 5 normal breast samples and 21 cell lines (Additional data file 1). We detected the expression of 137 miRNAs in this sample set, 133 of which we detected in normal breast or breast tumors. We included a number of replicate probes and technical replicate samples and found that results were reproducible (Additional data files 5 and 6). For a subset of miRNAs and a subset of samples we also performed quantitative RT-PCR to independently assess miRNA expression (Additional data file 7). While there is generally good correlation between miRNA expression on both platforms, we do observe probe-specific differences. Sample quantity did not permit validation of miRNA expression using northern blotting; however, the bead-based flow-cytometric miRNA expression platform had been validated using northern blotting previously 56.

Unsupervised hierarchical clustering of miRNA expression clearly separated cell lines from both normal breast and tumor samples and suggested that miRNA expression in cell lines is largely deregulated (Figure 1a). We did not observe a perfect separation of normal and tumor samples, as has been described before for primary human tumors 56. However, as our study was focused on tumor subtypes, we profiled only a small number of normal breast samples. As we found major differences in miRNA expression between primary human tissue and cell lines, we excluded cell lines from subsequent analyses. Unsupervised clustering of the tumor samples revealed striking differences in miRNA expression between ER- and ER+ tumors (Figure 1b).

Next we tested if miRNAs are differentially expressed among breast cancer subtypes. To identify the molecular subtypes of our tumor samples we used a single sample predictor (SSP), which classifies breast tumors into five subtypes: luminal A, luminal B, basal-like, HER2+ and normal-like 65697076. In addition to differences in mRNA expression profiles, these tumor subtypes also display distinct clinicopathological characteristics, including different survival rates (Additional data files 8 and 9). For example, the basal-like and HER2+ tumors are less differentiated and more aggressive and the luminal A and luminal B tumors are mostly ER+ with good and poor clinical outcome, respectively. Based on Agilent and Illumina mRNA expression data for 86 of our tumor samples 83 (unpublished results) we were able to classify 51 of the 93 tumor samples as 16 basal-like, 15 luminal A, 9 luminal B, 5 HER2+ and 6 normal-like tumors (Additional data file 1). miRNAs that were found to be differentially expressed in the tumor subtypes are shown in Figure 2a,b. miRNAs that are part of the same family show highly correlated expression. For example, the nine miRNAs that were found to be differentially expressed between luminal A and luminal B tumors represent seven miRNA families (Figure 2b).

Given the large number of miRNAs differentially expressed between molecular subtypes, we investigated the predictive potential of miRNAs in an independent test set. Using all 137 expressed miRNAs, we performed a model-based discriminant analysis 84 for the 16 basal-like and 15 luminal A tumors, the two largest subtype groups in our study (Additional data file 1). As we aimed to distinguish between molecular subtypes, we required a test set of samples with both miRNA and mRNA expression data available. The bead-based miRNA expression data in Lu et al. 56 included 11 breast tumor samples with corresponding Affymetrix gene expression data published in 5685. To our knowledge, no other studies with miRNA and mRNA data on breast tumor samples have been published. Based on the gene expression profiles and the SSP, six tumor samples could be assigned to molecular subtypes, three of which were classified as basal-like, two as luminal A and one as HER2+ (Additional data files 1, 14 and 19). Using the discriminator derived from our miRNA data, all three basal-like and two luminal A tumors in the independent miRNA data set were classified in concordance with their SSP molecular subtype classification.

We next assessed associations between individual miRNAs, molecular tumor subtypes and clinicopathological factors (Figure 3 and Additional data file 18). We tested for statistically significant associations with tumor characteristics such as molecular subtype, grade, stage, vascular invasion, ER status, Nottingham Prognostic Index (NPI) as well as TP53 status as determined by mutation screening and HER2 status assessed by immunohistochemistry (unpublished results). Figure 3 summarizes data for those 31 miRNAs and clinical factors for which there are significant associations at an adjusted p value less than 0.01 (Materials and methods). These 31 miRNAs represent 20 distinct miRNA families. Most of these miRNAs are expressed in the less aggressive, grade 1, ER+ samples. However, some miRNAs are expressed in the more aggressive grade 3/ER- tumors. We did not find any strong associations with stage, vascular invasion, NPI, TP53 or HER2 status.

Given the changes in miRNA expression we observed, it is important to ask how these changes come about. We first tested if the changes in miRNA expression are likely due to chromosomal loss, gain or amplification as inferred from array comparative genomic hybridization (CGH) data. For 82 of the 93 tumor samples we analyzed for miRNA expression, we performed array CGH analysis based on gene centric oligonucleotide microarrays 8687. For each miRNA locus that was identified as altered in any of the samples, we performed separate non-parametric Wilcoxon rank sum tests to assess differences in expression between samples with loss, gain or amplification compared to samples without changes.

We found that in many cases expression differences could not be explained by any genome alterations detected by our array CGH data (Figure 4). The expression of 17 out of 129 mature miRNAs transcribed from genomic regions with an observed aberration correlated with genomic changes at 15 distinct chromosomal loci (p < 0.01). For miR-33 and miR-320, we found strong associations between miRNA expression and genomic alterations (p < 0.001), suggesting chromosomal change is a possible mechanism for mis-expression of these genes in primary human breast cancers. We also identified miRNA clusters whose changes in expression were correlated with copy number, for example, for miR-30b and miR-30d at C8q24.22 (p < 0.001) and miR-15b and miR-16-2 at C3q26.1 (p < 0.05).

We noticed that miRNA clusters are often expressed coordinately in our sample set. For example, miR-106b, miR-93 and miR-25 situated on C7q22.1 are highly expressed in high-grade tumors. To further examine this phenomenon, we calculated the pairwise Pearson correlation of expression between miRNAs on the same chromosome and strand. We observed an abrupt drop in correlation of miRNA expression for pairs of miRNAs that were more than 50 kb apart (Additional data file 12). These observations agree well with what has been observed earlier in human tissue samples 88. We therefore used a distance of 50 kb as a cut-off to identify 56 intergenic or gene-resident spatial clusters, 44 of which are represented in the set of 137 miRNAs detected in our sample set. Interestingly, 26 of 31 clusters for which expression data from multiple stem-loop regions were available show correlated expression with r > 0.4 (Figure 5 and Additional data file 13). For example, the miR-15 and miR-16 family are expressed from two clusters at chromosomes 3q and 13q, which are both highly correlated (r > 0.8). In many cases these correlations are likely due to shared regulatory elements or polycistronic expression of several miRNAs from a single primary transcript 88.

Since only 17 of the 137 miRNAs expressed in our samples showed changes in their expression associated with detected chromosomal abnormalities, changes in miRNA expression may be due to changes in transcription of primary miRNA transcripts. We showed above that miRNA clusters are expressed coordinately. We therefore asked if expression levels of miRNAs that are intronic are correlated with the expression of their host gene, as this suggests changes in primary transcription rates. To test this hypothesis, we compared miRNA expression data with Illumina mRNA expression data for our tumor sample set (unpublished results; Additional data file 13). We only detected correlations for seven miRNA host gene pairs (r > 0.4), suggesting that changes in miRNA expression in our tumor sample set are not generally linked to host gene expression (Table 1). These seven miRNA host gene pairs were miR-30e-5p/NFYC, miR-149/GPC1, miR-25/93/106b/MCM7, miR-342/EVL and miR-99a/C21orf34.

For miRNA genes that are intergenic, we performed a similar comparison using the most proximal probes (within 50 kb) from the Illumina platform as these probes might correspond to primary miRNA transcripts (Additional data file 13). Only 23 out of 243 miRNA/proximal probe pairs at 11 distinct loci correlated in expression (r > 0.4; Table 1). Some of these miRNAs have proximal probes that are very close and likely represent primary miRNA transcripts. For example, miR-205 expression is highly correlated with the proximal probe for transcript NPC-A-5 (r > 0.75). One striking example of correlated expression of miRNAs and proximal probes was miR-10a, which is part of the HOXB cluster (C17q21.32), where Illumina probe data suggest the co-regulation of a region from HOXB2 to HOXB6 including miR-10a (Table 1).

As genomic changes and transcriptional regulation of miRNA expression do not explain the changes in miRNA expression we observed in human breast cancers, post-transcriptional regulation of miRNA expression has to be considered. Indeed, recent studies suggested that primary miRNA processing might be deregulated in human cancer 648990. Therefore, we tested whether genes required for miRNA biogenesis are differentially expressed in our breast cancer samples. As we found many changes in miRNA expression across the five clinical tumor subtypes we had defined above (Figure 2), we asked whether DICER1, DROSHA, DGCR8, AGO1, AGO2, AGO3 or AGO4 expression differs among these subgroups. We found significant changes in the expression of DICER1 (p < 0.001), which was low in the more aggressive basal-like, HER2+ and luminal B type tumors, and AGO2, which was high in basal-like, HER2+ and luminal B type tumors (Figure 6). We did not find significant changes in the expression of DROSHA, DGCR8, or any of the other AGO genes (Figure 6 and Additional data file 10). We also observed significant changes in AGO2, DICER1 and DROSHA expression in relation to ER status, with AGO2 and DROSHA being higher and DICER1 lower in ER- tumor samples (Figure 6).

The observed deregulation of genes required for miRNA biogenesis may be expected to lead to global changes in miRNA expression. To further investigate this possibility, we utilized an alternative approach to between-sample normalization. For the analyses described previously, sample median centering proved advantageous in removing technical variation between samples without changing trends in differential expression (Additional data files 1 and 4). However, this method necessarily removed any global changes in miRNA expression. Using an alternative normalization based on spike-in controls, similar to the method described in 56, we detected small differences in mean miRNA levels according to ER status with lower mean miRNA expression in ER- tumors (Figure 6d).

Primary breast tumor specimens were obtained with appropriate ethical approval from the Nottingham Tenovus Primary Breast Cancer Series (Nottingham City Hospital Tumor Bank, Nottingham, UK). All cases were primary operable invasive breast carcinomas collected from 1990 to 1996. Clinical information, including therapy, has been published previously 8081828387.

RNA was extracted from primary tumors and cell lines using a standard Trizol (Invitrogen, Carlsbad, CA, USA) protocol, modified by washing the final RNA pellet with 80% EtOH. Frozen tumors were sectioned on a cryostat prior to homogenization in Trizol. RNA quantity and quality were assessed by Nanodrop (Nanodrop Technologies, Wilmington, DE, USA) and Agilent 2100 bioanalyzer (Agilent Technologies, Santa Clara, CA, USA), respectively.

miRNAs were extracted from 5 μg of sample total RNA using denaturing PAGE. Briefly, samples were spiked with three synthetic pre-labeling control RNAs (5'-pCAGUCAGUCAGUCAGUCAGUCAG-3', 5'-pGACCUCCAUGUAAACGUACAA-3', 5'-pUUGCAGAUAACUGGUACAAG-3'; Dharmacon, Lafayette, CO, USA) to control for target preparation efficiency, at 3 fmoles/sample. After purification of 18-26 bp RNAs, adaptors were ligated at the 3' and 5' ends using T4 RNA ligase (Fermentas, Burlington, OT, CA), a RNA-DNA hybrid 5'-pUUUaaccgcgaattccagt-idT-3' (Dharmacon; X = RNA, x = DNA, p = phosphate, idT = inverted [3'-3' bond] deoxythymidine) was ligated to the 3' end and 5'-acggaattcctcactAAA-3' (Dharmacon) was ligated to the 5' end using T4 RNA ligase. These bi-ligated products underwent reverse transcription using an adaptor specific primer (M37, 5'-pTACTGGAATTCGCGGTTA-3') and then amplified and labeled using PCR (M37 and M33, 5'-biotin-CAACGGAATTCCTCACTAAA-3'). Amplification was performed on an Eppendorf thermal cycler at 95°C for 30 s, 50°C for 30 s and 72°C for 40 s for 18 cycles. PCR products were precipitated without glycogen and redissolved in 66 μl 1 × TE buffer containing 1 μl of three biotinylated post labeling controls (100 fmols each, FVR506, PTG20210, MRC677).

Oligonucleotide probes were coupled to color-coded polystyrene beads, allowing the simultaneous detection of about 90 different target oligonucleotides. To obtain expression profiles for 309 miRNAs, we created four distinct sets of bead-coupled miRNA probes. Each sample was hybridized to the four bead sets to generate a complete miRNA profile. Oligos were 5'-amino modified with a 6-carbon linker and conjugated to carboxylated xMAP beads (Luminex, Austin, TX, USA) in 96-well formats following the standard manufacturer's protocol. To generate bead set pools, 3 μl of each oligo-bead conjugate was mixed into 1 ml 1.5× TMAC buffer (4.5 M tetramethylammonium chloride, 0.15% sarkosyl, 75 mM Tris-HCl pH 8.0, 6 mM EDTA). Samples were hybridized in a 96-well format with two water-only blanks and at least three bead blanks containing water instead of the labeled sample for use as a background control. We included replicate probes and technical replicate samples across bead sets and sample plates, respectively, to aid quality control and data preprocessing. Hybridization was carried out overnight at 50°C with 33 μl of the bead pool and 15 μl of labeled sample.

Unbound sample was removed from beads by washing with 1 × TE and re-suspending in 1× TMAC buffer. SAPE, streptavidin-phycoerythrin, premium grade (Invitrogen) was added to the beads (1:100 dilution) and incubated for 10 minutes at 50°C to bind to biotin moieties on the cDNA. Samples were processed on a Luminex 100 machine and median fluorescence intensity values acquired using the StarStation software (ACS, Sheffield, UK).

Each array in the preprocessed Illumina and Agilent 83 gene expression data set was normalized to have zero mean and standard deviation one, and each probe was centered to have zero median. An SSP annotation for Agilent probes was provided in 76. Detailed information on the SSP annotation for Illumina probes can be found in Additional data files 1 and 15. Multiple probes for the same UniGene cluster ID in either data set were summarized by their median profile. Samples were then assigned to the nearest subtype centroid as determined by Spearman correlation, requiring a minimum correlation of 0.3. Samples that could be assigned to subtypes based on both Agilent and Illumina expression profiles were classified according to the Illumina assignment (Additional data file 1).