Main

Micro-RNAs (miRs) are endogenous ∼22-nt non-coding RNAs, which can alter protein levels in a sequence-specific manner.1, 2 Since their initial identification in nematodes in 1993,3 miRs have been increasingly recognised as key regulators in many biological systems. Current estimates of the number of miRs in the human genome range from ∼450 to 1000.1, 4 These estimates represent 1–4% of all protein-coding genes, rendering them as one of the largest classes of regulatory genes.5 The breadth of their potential relates to their imperfect complementarity, whereby a single miR may bind to as many as 200 target genes, with diverse functions ranging from transcription factors, proteins, and receptors, to transporters.2 Bioinformatic analyses estimate that miRs could regulate the expression of up to one-third of all human mRNAs.6 Approximately half of the known miRs have been mapped to fragile sites in the genome,7 providing the first clue to their role in cancer. Since then, miRs have been associated with almost all varieties of human malignancies including haematological7 and solid cancers,8 such as lung,9 breast,10, 11 prostate,12 pancreas,13 and glioblastomas.14

The global alteration in the miR expression in cancer, compared with that in corresponding normal tissues, could be manifested by either overexpression or underexpression. In silico analyses show that among 18 commonly shared miRs in epithelial malignancies, 44% (115 of 263) of the predicted gene targets were well-known cancer-associated genes, such as Rb, TGFβR2, or Ras.8 Thus, depending on its targets, a miR could function as either a tumour suppressor or as an oncogene. For example, two early reports described the downregulation of miR-15a and mirR-16-1 in chronic lymphocytic leukaemia (CLL),15 which were shown to target Bcl-2.16 The absence of these miRs thus leads to anti-apoptotic signalling. Conversely, miR-21 is often overexpressed in breast cancer and can inhibit pro-apoptotic genes14, 17 and PTEN.18 Hence, the overexpression of miR-21 will result in the suppression of apoptosis and in the upregulation of PTEN targets, such as Akt, with both pathways converging on promoting survival.

Increasing data support the value of miR expression profiling19 in distinguishing one cancer subtype from another,8 such as luminal vs basal classifications for human breast cancer.20 When the classification accuracy of miR profiles was directly compared with that of standard mRNA, miRs were shown to have superior performance for poorly differentiated tumours.8, 21 Several recent reports have correlated miR expression with clinical outcome. Earlier ones have described a 13-miR signature in CLL, which correlated with clinical behaviour.16 Two recent lung studies have identified an 8-miR set for adenocarcinoma,9 and a second 5-miR signature that predicted outcome for patients with non-small-cell lung cancer.22

All these aforementioned studies have used frozen tissues. However, routine histology processing uses formalin fixation to preserve the histological architecture of tissue specimens. Archival collections of formalin-fixed tissues, linked to clinical databases, provide a rich resource from which biological insights could be derived far more expeditiously than the prospective collection of frozen samples. In addition, any biomarker developed from formalin-fixed paraffin-embedded (FFPE) samples could be more readily translated into clinical practice. Unfortunately, RNA is degraded in tissues before, during, and after formalin fixation23, 24 and can continue to deteriorate even during storage, leading to shortened fragments of RNA.25 Interestingly, miRs appear to be better preserved, perhaps because of their intrinsically shorter lengths. Therefore, it should be possible to perform genome-wide screening for miRs using FFPE tissues.20, 26, 27, 28, 29 Indeed, this approach has been successful for colon and breast cancers;20, 29, 30 however, the technical robustness of these platforms has not been thoroughly investigated.

Therefore, in this study, we systematically investigated a comprehensive panel of 365 human miRs on 40 archival breast cancer FFPE samples. We used the Taqman Low Density Array (TLDA), a 384-microfluidic card, designed to perform several hundred quantitative real-time PCR (qRT-PCR) reactions simultaneously. Three criteria were defined to evaluate the robustness of this platform: (a) technical reproducibility; (b) independent validation using an alternate experimental method, such as single-well qRT-PCR; and (c) the data probed biologically relevant processes. On the basis of these a priori defined criteria, the TLDA platform was shown to be technically and biologically robust for the analysis of FFPE samples.

MATERIALS AND METHODS

RNA Purification from FFPE Samples

Thirty-four archival FFPE lumpectomy blocks were evaluated by a breast cancer pathologist (NM). Regions of invasive ductal carcinoma were confirmed, thereby identifying areas containing >70% of malignant epithelial cells for macro-dissection. The clinical descriptors for these 34 breast cancer patients are provided in Table 1. Six reduction mammoplasty specimens were obtained; regions containing predominantly epithelial tissues were also macro-dissected.

Table 1 Clinical descriptors for the 34 breast cancer patients
Full size table

Total RNA enriched for small RNA species was isolated with Recoverall (Ambion, Austin, TX, USA), according to the manufacturer's instructions. Briefly, xylene was added to 4-μm sections of FFPE samples to remove paraffin. The tissues were digested with protease at 50°C and treated with DNase. After washing, total RNA, including the small miR fraction, was eluted with distilled water. Recovered RNA concentrations and quality were measured using the Nanodrop 1000A spectrophotometer (Nanodrop Technologies, Wilmington, DE, USA).

Real-Time Quantification of micro-RNAs

The quality of the 40 FFPE samples was assessed by RT-PCR analysis of the endogenous control RNU44 using Applied Biosystems TaqMan® Micro-RNA Assay. This assay includes a reverse transcription (RT) step using the High-Capacity cDNA Archive Kit (Applied Biosystems, CA, USA), wherein a stem-loop RT primer specifically hybridises with a miR molecule and is then reverse-transcribed with a MultiScribe reverse transcriptase.27 Briefly, the RT mix includes 50-nM stem-loop RT primers, 1 × RT buffer, 0.25 mM each of dNTPs, 3.33 U/μl MultiScribe reverse transcriptase, and 0.25 U/μl RNase inhibitor. The 7.5-μl reaction was then incubated in an Applied Biosystems 7900 Thermocycler for 30 min at 16°C, 30 min at 42°C, 5 min at 85°C, and then held at 4°C. The RT products were subsequently amplified with sequence-specific primers using the Applied Biosystems 7900 HT Real-Time PCR system. The 10-μl PCR mix contains a 0.67-μl RT product, the 1 × TaqMan® Universal PCR Master Mix, a 0.2-μM TaqMan® probe, a 1.5-μM forward primer, and a 0.7-μM reverse primer. The reactions were incubated in a 384-well plate at 95°C for 10 min followed by 40 cycles of 95°C for 15 s and at 60°C for 1 min.

TaqMan Low Density Array

Global profiling of miR expression on 40 FFPE samples was performed using the TaqMan® Array Human Micro-RNA Panel v1.0. (Applied Biosystems, CA, USA), which runs the PCR in 1–2 μl of reaction volume in a 384-well format. The TLDA contains 368 TaqMan® Micro-RNA Assays enabling the simultaneous quantitation of 365 human miRs plus 3 endogenous controls. Similar to the individual TaqMan® Micro-RNA Assay, total RNA was first reverse-transcribed with the Multiplex RT pool set, then quantitated using an Applied Biosystems 7900 HT Real-Time PCR system.

Data Pre-Processing

The TLDA data were pre-processed using the same strategy used earlier for mRNA data,31 with some modifications. For each sample, a normalisation factor was calculated as the mean of the intensity of two endogenous controls, RNU44 and RNU48. A ΔCt was calculated for each miR species as Ct(gene)−Normalisation Factor. The normalised expression level was then calculated as log2∣2−ΔCt∣. No scaling was used, so that the normalised expression values represent the levels of miR species relative to the levels of the two housekeeping control genes. Normalisation was performed in the R statistical environment (v2.6.1).

Assessment of TLDA and PCR Technical Reproducibility in FFPE Samples

To assess the technical reproducibility of the TLDA platform, replication of three samples was performed multiple times. One normal sample was assessed in duplicate (N1R1 and N1R2), another normal sample in triplicate (N2R1, N2R2, N2R3), and a tumour sample in triplicate (T1R1, T1R2, T1T3). We first performed a clustering analysis by calculating Spearman's correlation coefficient between each pair of samples, then subjecting the correlation data to agglomerative hierarchical clustering using Euclidean distance as the distance metric and complete linkage. The results were then visualised using a heatmap. Clustering was implemented in the R statistical environment (v2.6.1), and visualisation was carried out using the lattice (v0.17-2) and latticeExtra (v0.3-1) packages. We also assessed the distributional similarity of the technical replicates using box and scatter plots. We further tested the accuracy of the technical replicates by determining the fraction of tumour–normal differences that were replicable when using different technical replicates at a fold change threshold of two.

Comparison of miRNA Profile in FFPE and Frozen Samples

Corresponding pairs of frozen (F1 and F2) and FFPE (R1 and R2) breast cancer samples were derived from two different patients and subjected to global miR profiling as described above. We performed a clustering analysis by calculating Spearman's correlation coefficient between each pair of samples, then subjecting the correlation data to agglomerative hierarchical clustering using Euclidean distance as the distance metric and complete linkage. The results were then visualised using a heatmap. Clustering was implemented in the R statistical environment (v2.7.1) and visualisation was carried out using the lattice (v0.17-14) and latticeExtra (v0.5-2) packages.

PCR Validation

Single-well PCR analysis was conducted to validate 18 miRs (let-7d, miR-21, miR-24, miR-29c, miR-32, miR-125b, miR-145, miR-155, miR-184, miR-191, miR-194, miR-221, miR-328, miR-497, miR-518b, miR-572, miR-601, and miR-622) in 6 normal breast samples and at least 18 breast cancer samples. The analysis was performed using the TaqMan® Micro-RNA Assay as described above. Pearson's correlation was calculated across all 18 validations to compare the tumour–normal fold changes measured by TLDA with those measured by single-well PCR. Each experiment was conducted in duplicate or triplicate, depending on RNA availability.

Tumour–Normal Comparison

To compare the miR expression profiles of normal and tumour breast tissues, we first used clustering to visualise the profiles. We used the DIANA divisive hierarchical clustering algorithm with Pearson's correlation as the distance metric, as implemented in the cluster package (v1.11.9) for the R statistical environment (v2.6.1) on the pre-processed expression values. Only genes that showed detectable expression in all samples were included. The resulting expression profiles were visualised with a heatmap using the lattice (v0.17-2) and latticeExtra (v0.3-1) packages.

To identify specific miRs that are differentially expressed between tumour and normal samples, we performed gene-wise t-tests with Welch's correction for heteroscedasticity between tumour and normal samples. To account for multiple testing, we performed a false-discovery rate adjustment on the resultant vector of P-values. We also calculated the fold changes (in log2 space) for each gene. Genes detected in fewer than 10 samples were excluded from this analysis. All calculations were performed in the R statistical environment (v2.6.1).

Finally, to determine whether miR expression signatures could predict if a given sample were normal or a tumour, we used a machine-learning approach. We used leave-one-out cross-validation and k-nearest neighbours,32 and set k=5 because the number of normal samples was 6. k-nearest neighbours was implemented in the class package (v7.2-38) of the R statistical environment (v2.6.1).

RESULTS

Global miRNA Profiling Using FFPE Samples is Highly Reproducible

The primary objective of this study was to show the reliability of high-throughput TLDA platform for miR quantitation using archival FFPE samples. To achieve this goal, three criteria were established: technical reproducibility, independent validation, and the generation of biologically relevant results. We evaluated these criteria in sequence, focussing on 34 breast cancer samples obtained from archival FFPE lumpectomy blocks (Table 1) and from six reduction mammoplasty specimens.

To investigate the technical replicability of the TLDA platform, we performed global miR profiling in triplicate for one normal and one breast cancer sample and in duplicate for a second normal breast sample. These eight expression profiles were normalised to a panel of two endogenous controls (RNU44 and RNU48), as established earlier.31 The distributional similarity of the technical replicates was first verified using scatter plots (Figure 1a). Unsupervised machine learning (clustering) was then used to visualise the similarity of the technical replicates (Figure 1b). The two normal samples clustered separately from the single tumour sample, whereas the technical replicates for each biological sample clustered closely together. The expression profiles correlated well between the technical replicates, with Spearman's rho above 0.90 in all cases. In addition, the two normal samples correlated strongly with each other (Spearman's rho above 0.85), but less strongly with the tumour sample (Spearman's rho between 0.75 and 0.85).

Figure 1
figure 1

Replicability of Taqman TLDA analysis. To determine the relationship between technical and biological replicates, we assessed global miR expression in two normal breast samples (N1 and N2) and in one invasive ductal carcinoma sample (T1). Two technical replicates were performed on one normal sample (N1.R1 and N1.R2); three were performed for the other normal sample (N2.R1, N2.R2, and N2.R3), and for the carcinoma sample (T1.R1, T1.R2, and T1.R3). (a) Scatter plot analysis of TLDA analyses showing similar pattern of expression in the replicate experiments. (b) We calculated Spearman's correlation between all pairs of samples and performed agglomerative hierarchical clustering using complete linkage. The resulting heatmap shows that the technical replicates from the two normal samples (blue and green bars) clustered together, separate from the tumour sample (red bars). The correlations range from 0.75 between normal and tumour, to >0.95 for technical replicates of a single sample. (c) Comparison of microRNA expression of two pairs of frozen and FFPE samples from two invasive breast carcinoma patients. Spearman's correlation between the two pairs of samples was calculated and agglomerative hierarchical clustering was performed. The resulting heatmap shows that the paired frozen and FFPE samples of the same patient has correlations of at least 0.94.

Full size image

We also evaluated the technical reproducibility of this data set in terms of the reported tumour–normal differences. The tumour–normal differences were identified in these eight profiles using a two-fold threshold. Predicted tumour–normal differences between technical replicates were again very similar, with a 92.8+1.6% agreement between different technical replicates. Taken together, these data show that the TLDA profiles of miR expression in samples extracted from FFPE blocks were highly reproducible.

Consistent miRNA Profiling Between Frozen and FFPE Samples

To further corroborate the validity of global miR profiling of FFPE samples, the same experiments were repeated, using pairs of frozen and FFPE blocks from the same patients. The miR profiles of these samples were generated and normalised to the two endogenous controls (RNU44 and RNU48). Analysis with unsupervised machine learning (clustering) showed the similarity of the two types of samples (Figure 1c). miRNA profiles from FFPE and frozen samples from the same tumour were highly correlated, with Spearman's rho of at least 0.94. By contrast, miR profiles from different patients were less correlated, with Spearman's rho between 0.86 and 0.92.

PCR Validation

A robust experimental system should not only be technically replicable but should also yield similar results to alternative experimental methodologies. Accordingly, we directly compared the TLDA-derived expression profiles with single-well PCR. Eighteen miRs were selected and validated in six normal and up to 24 tumour samples. These included 11 miRs that were predicted by TLDA to be differentially expressed between tumour and normal samples, and seven miRs predicted not to be differentially expressed. The tumour–normal differential expression identified by these two techniques was concordant in 14/18 miRs (78%), with similar trends observed for the remaining four miRs (Table 2 and Supplementary Figure 1). Furthermore, the fold changes predicted for these 18 miRs by the two methods were highly correlated (R=0.79, P=9.5 × 10−5). Hence, these analyses show that the TLDA results can be confirmed using an independent experimental approach.

Table 2 PCR and TLDA validation
Full size table

Tumour–Normal Comparison

The ultimate test for a novel experimental system is to probe new and biologically important questions. The TLDA system was used to assess the expression levels of 6 normal breast samples and 34 breast cancer samples. This data set was then examined in three different ways: (a) global patterns of miR expression; (b) specific miR species exhibiting differential tumour–normal expression; and (c) the ability to classify samples on the basis of miR levels. We first evaluated the broad patterns of miR expression using unsupervised pattern recognition33 to both visualise the data set and discover unexpected trends. Clustering with a divisive hierarchical algorithm (Figure 2) showed a clear separation between the 6 normal (blue bars on the right side of the figure) and the 34 tumour samples (red bars).

Figure 2
figure 2

Tumour–Normal Heatmap. The miR profiles of normal and malignant breast samples were visualised using agglomerative hierarchical clustering using Spearman's rho as the distance metric. Rows correspond to tissue samples, and are labelled with a box indicating whether the row represents a tumour (red) or a normal (blue) sample. Each column corresponds to an individual miR sequence. This unsupervised and unbiased analysis shows that the 6 normal samples (blue bars) are completely separated from the 34 cancer samples (red bars).

Full size image

Next, standard statistical methodologies were used to assess tumour–normal differential expression in this data set. Among the 208 miR species expressed in at least 10 of the 40 samples, 81 (38.9%) were differentially expressed between tumour and normal tissues at a 5% false-discovery rate. Table 3 lists the 30 most significantly differentially expressed miRs: 26 were overexpressed in tumours, whereas only 4 were underexpressed. The magnitude of this differential expression was considerable, with some miRs reduced to only 5% of their normal tissue levels (eg, miR-125b), whereas others were induced more than 50-fold (eg, miR-196a and miR-622).

Table 3 List of 30 most significantly differentially expressed micro-RNAs between malignant vs normal breast epithelial tissues
Full size table

Finally, supervised machine learning was used to determine if it were possible, as reported earlier,21 to classify tissues on the basis of miR expression profiles. We used the k-nearest neighbours machine learning combined with leave-one-out cross-validation to predict the tumour–normal status for each sample. This machine-learning approach achieved 100% accuracy, verifying that miR profiles could distinguish normal from malignant breast tissue samples.

DISCUSSION

Increasing evidence supports the use of miR profiling to characterise human tumours. To date, six distinct predictive signatures have been reported: for CLL,16 hepatocellular cell,34 esophageal squamous cell,35 colon,36 and two for lung cancer.9, 22 However, all these evaluations have used frozen tissues. In this study, we provide the first characterisation of the technical robustness of high-throughput global miR profiling of 365 human miRs using a qRT-PCR-based assay. We show high sensitivity and specificity in this approach, by systematically evaluating 40 FFPE breast tissue samples. This method facilitates screening for both high and low expressors in a single experiment. In addition, very small quantities of input total RNA are required (∼1–1.5 μg), and no labourious miR fractionations are required, thereby minimizing RNA loss.

There have been extensive evaluations conducted on the inter- and intra-experimental replicabilities for mRNA expression profiling platforms;37, 38, 39, 40 none have been undertaken earlier for this relatively new field of miR profiling. To address this gap, we describe a comprehensive assessment of the viability of coupling sample extraction from archival FFPE blocks to miR expression profiling using the TLDA platform. We established three criteria for evaluating this methodology: technical reproducibility, independent validation, and the generation of biologically relevant data.

We first showed that there is a high degree of inter-experimental reproducibility for TLDA analyses of FFPE samples. Three distinct statistical techniques were used: unsupervised machine learning (clustering), distributional assessments (scatter plots), and evaluation of the reproducibility of tumour–normal differences. In each case, the technical reproducibility was high, with Spearman's correlations above 0.9 and technical replicates clustering together (Figure 1b). These technical replicates also show highly similar expression profiles, with few outliers (Figure 1a). The resulting biological information was also extremely robust, with 92.8% reproducibility of tumour–normal differences between technical replicates. Although these criteria have not been assessed earlier for miR profiling, our results are comparable with similar evaluations of mRNA expression profiling, in which >95% of repeated sample accuracies have been determined for fresh leukaemia cells.41 Therefore, these results indicate that assessing global miR expression using FFPE samples is feasible using relatively simple experimental protocols with small quantities of input RNA, and that technical replicates are likely unnecessary.

Having shown technical reproducibility, we then evaluated the second criterion: independent validation. Eighteen miRs were selected for validation using the traditional single-well qRT-PCR assay. Fourteen of these miRs showed concordant results, whereas consistent trends were observed in the remaining four. Importantly, the fold changes predicted for these 18 miRs by the two methods were significantly correlated between the two methods. These results show that miR profiles generated from FFPE blocks using TLDA can be validated using an independent experimental technique. This level of validation is again similar to that observed for mRNA expression profiling.42 Discrepancies in validation for mRNA experiments have been attributed to inter-experimental variations, differences in hybridisations, and differences in gene sequences used in arrays vs individual qRT-PCR experiments. Although the latter two issues would not apply to these miR experiments, we would attribute the miR discrepancy to inter-experimental variations of these very sensitive qRT-PCR-based assays.

Indeed, we have assessed the reliability of shifting from high-throughput miR profiling to qRT-PCR. The expressions of four genes (RNU6B, miR-328, miR-221, and miR-213) along with two normalisation controls (RNU44 and RNU48) were further evaluated in 4 normal and 12 tumour samples in triplicate (Supplementary Figure 2). A variance analysis was performed to quantify the relative contributions of technical and biological variability. For all the 4 genes across the 16 samples, the technical variability was smaller than the biological variability (Supplementary Figure 3a). We further considered the differential expression between tumour and normal samples identified by TLDA and qRT-PCR. Good concordance was observed for three genes (RNU6B, miR328, and miR213), both in terms of statistical significance (Supplementary Figure 3b) and magnitude of differential expression (Supplementary Figure 3c), further corroborating that the TLDA platform leads to reproducible measurements of miR expression, as confirmed by qRT-PCR.

The capabilities and robustness of global miR profiling using FFPE samples are far superior to that of mRNA profiling, which is not technically reliable because of RNA degradation.25 This issue was specifically addressed in a direct comparison of mRNA profiling of FFPE vs frozen tissues, showing a correlation coefficient of only 0.28.29 The only mRNA assay that could be conducted on FFPE samples is the measurement of already-identified transcripts, such as the 21-gene signature 43 for breast cancer (http://www.genomichealth.com/oncotype), which would not allow for the discovery of novel biologically relevant sequences.

Interesting information has emerged in the evaluations of global miR profilings; in a direct comparison between FFPE and snap frozen samples of a normal human thyroid cell line, this group showed higher miR expression levels observed from FFPE when compared with that from frozen samples. This group hypothesised that this unexpected finding might relate to miRs being resistant to the methylol cross-linking between RNA and proteins, which occurs during tissue fixation.44 Similarly, a systematic comparison of miR profiling between frozen and FFPE normal murine tissues showed an excellent correlation of 0.8929 with remarkable stability even in FFPE blocks that have been in storage for 10 years,29 which is corroborated in yet another recent study reporting similarly high correlations (0.85 to 0.95) using miR microarray analyses.45 Our own data using two corresponding pairs of frozen and FFPE invasive breast cancer samples yielded the same level of outstanding consistency, with Spearman's correlations of 0.94 or greater (Figure 1c).

Even more useful is the ability to conduct in situ hybridisation of miRs in FFPE breast cancer samples, which provides additional valuable information on the cellular localisation of relevant miRs in disease processes.30 Hence, our study contributes to the emerging literature on the utility and robustness of miR profiling of FFPE samples, underscoring opportunities to expeditiously acquire clinically and biologically relevant insight into the role of miRs in human malignancies.

Statistical analysis identified a distinct subset of breast cancer-associated miRs, similar to earlier-reported miR evaluations, including upregulation of miR-21, miR-155, miR-191, miR-196a, miR-200c, miR-210, and miR-213, as well as downregulation of miR-125b and miR-221.10, 12, 30, 46 Micro-RNA-21 (Table 3), which is overexpressed ∼10-fold in our cohort, is among the most-studied miRs in human cancers. Another highly dysregulated miR is miR-196a, which is overexpressed ∼64-fold in this study; this miR has been suggested to target HOXB8 and is significantly differentially expressed between normal and malignant pancreatic tissues.47, 48 Similar to the Sempere et al's30 study, miR-200c was overexpressed (∼16-fold), and miR-221 was downregulated (also by ∼16-fold) in our cohort (Table 3). Micro-RNA-200c can reduce the expression of transcription factor 8 and increase the expression of E-cadherin,30, 46 which might account for an increased propensity for invasion. Micro-RNA-221 targets p2749 and kit,50 and is included in the predictive signature sets for both CLL16 and lung cancer,22 suggesting an important role for this miR in cancer progression. Downregulation of miR-125b has been consistently detected in breast cancers10, 12 and has been shown to suppress both expression and function of ERBB2 and ERBB3.51

This study includes a more up-to-date and comprehensive evaluation of human miRs, enabling the identification of several less well-characterised miRs that are significantly altered in invasive breast cancers. As shown in Table 3, we have identified new breast-related miRs, which might provide additional insight into the complexities of human breast cancer dysregulation. These novel miRs include the less-characterised miRs, such as upregulated miR-423, miR-449, miR-449b, miR-518b, miR-572, miR-601, miR 622, miR-646, and downregulated miR-328 (Table 3).

The overexpression of HER-2/neu and ER expression are two important prognostic factors for human breast cancer. The association of miR with HER-2/neu or ER expression in breast cancer has been investigated earlier,10, 12 noting distinct miR subsets correlating with HER-2/neu, PR, and ER status.10, 12 On the other hand, Sempere et al30 observed no clear association of miR expression with either ER or HER-2/neu. Similarly, in our study, no association of miR with HER-2/neu or ER expression was identified, which might be attributable to the small number (four cases) of HER-2/neu positive tumours in our cohort.

In conclusion, we have successfully conducted global miR profiling for human breast FFPE samples, showing that this platform achieves excellent inter-experimental reproducibility and biological accuracy, rendering technical replicates likely unnecessary. Such high-throughput profiling offers the rich resource of archival clinical FFPE blocks, from which biological insight could now be acquired in an expeditious manner, for the further elucidation of the role of miRs in human malignancies.