Intratumoral Heterogeneity of MicroRNA Expression in Rectal Cancer

Introduction An increasing number of studies have investigated microRNAs (miRNAs) as potential markers of diagnosis, treatment and prognosis. So far, agreement between studies has been minimal, which may in part be explained by intratumoral heterogeneity of miRNA expression. The aim of the present study was to assess the heterogeneity of a panel of selected miRNAs in rectal cancer, using two different technical approaches. Materials and Methods The expression of the investigated miRNAs was analysed by real-time quantitative polymerase chain reaction (RT-qPCR) and in situ hybridization (ISH) in tumour specimens from 27 patients with T3-4 rectal cancer. From each tumour, tissue from three different luminal localisations was examined. Inter- and intra-patient variability was assessed by calculating intraclass correlation coefficients (ICCs). Correlations between RT-qPCR and ISH were evaluated using Spearman’s correlation. Results ICCsingle (one sample from each patient) was higher than 50% for miRNA-21 and miRNA-31. For miRNA-125b, miRNA-145, and miRNA-630, ICCsingle was lower than 50%. The ICCmean (mean of three samples from each patient) was higher than 50% for miRNA-21(RT-qPCR and ISH), miRNA-125b (RT-qPCR and ISH), miRNA-145 (ISH), miRNA-630 (RT-qPCR), and miRNA-31 (RT-qPCR). For miRNA-145 (RT-qPCR) and miRNA-630 (ISH), ICCmean was lower than 50%. Spearman correlation coefficients, comparing results obtained by RT-qPCR and ISH, respectively, ranged from 0.084 to 0.325 for the mean value from each patient, and from -0.085 to 0.515 in the section including the deepest part of the tumour. Conclusion Intratumoral heterogeneity may influence the measurement of miRNA expression and consequently the number of samples needed for representative estimates. Our findings with two different methods suggest that one sample is sufficient for adequate assessment of miRNA-21 and miRNA-31, whereas more samples would improve the assessment of miRNA-125b, miRNA-145, and miRNA-630. Interestingly, we found a poor correlation between the expression estimates obtained by RT-qPCR and ISH, respectively.


Materials and Methods
The expression of the investigated miRNAs was analysed by real-time quantitative polymerase chain reaction (RT-qPCR) and in situ hybridization (ISH) in tumour specimens from 27 patients with T3-4 rectal cancer. From each tumour, tissue from three different luminal localisations was examined. Inter-and intra-patient variability was assessed by calculating intraclass correlation coefficients (ICCs). Correlations between RT-qPCR and ISH were evaluated using Spearman's correlation.

Conclusion
Intratumoral heterogeneity may influence the measurement of miRNA expression and consequently the number of samples needed for representative estimates. Our findings with Introduction MicroRNAs (miRNAs) are short non-coding RNA molecules acting as negative gene regulators at the post-transcriptional level. It is well known that the same miRNA may affect several genes, but also that the same gene may be downregulated by different miRNAs. Thus, miRNAs represent a complex regulatory system of central biological importance. The system is dysregulated during malignant transformation, and it is generally anticipated that change of miRNA activity plays a central role in carcinogenesis [1][2][3][4]. Consequently, a considerable and increasing number of studies have investigated miRNAs as potential markers of diagnosis, treatment, and prognosis. The clinical implications, however, have so far been minimal [5]. One reason may be different methodological approaches, which hampers comparison of studies. Another issue is the different composition and quality of patient materials investigated. Moreover, the literature is dominated by many small studies without subsequent validation of relevant findings [5][6][7][8][9][10].
Another important obstacle is normalisation of real-time quantitative polymerase chain reaction (RT-qPCR) data, when measuring miRNA expression, as appropriate normalisation is important to ensure reliable and reproducible results. Despite an increasing number of miRNA expression studies, the literature is sparse when it comes to reliable normaliser candidates [11][12][13][14].
Intratumoral heterogeneity is a general characteristic of malignant tumours [15][16][17] and represents a major obstacle in the study of tumour biology, miRNAs being no exception. To date, the issue has been almost neglected despite its obvious importance; the activities of miR-NAs are likely to vary in different parts of the tumour with potential major impact on the clinical application. Only few publications have addressed the problem of intratumoral heterogeneity affecting the measurement of miRNA expression. One study is dealing with the issue in breast cancer [18], another one in hepatocellular carcinoma [19].
These challenges are especially relevant in patients with locally advanced rectal cancer, where treatment decisions are often based on single biopsies and the later interpretation of the tumour is influenced by the pre-operative radiotherapy [20,21].
The aim of the present study was to analyse the expression heterogeneity of a panel of selected miRNAs in rectal cancer specimens by two different technical approaches, and to elucidate their mutual concordance. RT-qPCR was chosen, since it is widely accepted as the gold standard for miRNA expression analyses. In situ hybridization (ISH) was chosen because of its ability to visualise the cellular localisation of the investigated miRNAs in addition to the expression level.

Patients and tissue samples
Twenty-seven patients were randomly selected from a cohort of 239 patients who had undergone resection for rectal cancer at Vejle Hospital, Denmark from 1999 to 2008. Inclusion was made only if the histopathological diagnosis was rectal adenocarcinoma pT3 or pT4, if no preoperative chemo-radiotherapy was given, and if formalin-fixed paraffin-embedded (FFPE) rectal cancer tissue was available from three different localisations of the tumour, including the luminal aspect.
According to The Regional Scientific Ethical Committees for Southern Denmark, ethics approval and written informed consent were not required, since the purpose of the study was to develop new methodology and no clinical data were further analysed (Act on Research Ethics Review of Health Research Projects, cf. § 14, section 1). The samples were collected during standard rectal resection, and the project was carried out without knowledge of the clinical data. The study was registered with The Danish Data Protection Agency, and The Danish Registry of Human Tissue Utilisation was consulted before any tissue samples were used.
Available histological sections stained with hematoxylin and eosin (H&E) were examined by a pathologist in order to select three localisations from each tumour. The sections had to be cut from tissue blocks sampled from three different localisations, and all had to include the luminal aspects of the tumour mirroring the locations where diagnostic, endoscopic biopsies would have been taken. According to this selection, adjacent sections were cut from each of the FFPE tissue blocks for the heterogeneity analyses as follows: (a) one section to be stained with H&E for identification of Region of Interest (ROI), (b) four sections for ISH (5 μm), (c) two sections to be stained with hematoxylin for RT-qPCR (8 μm), (d) one look-up section to be stained with H&E. On a print of the H&E stained section, tumour and tumour microenvironment were identified by the pathologist and encircled as regions of interest (ROI). Subsequently, the marking was transferred to the digital whole slide image of the ISH-slide, and to the H&E stained section and the hematoxylin stained section for RT-qPCR. For the RT-qPCRanalysis, the marked area was removed using a scalpel and collected in 1.5 ml RNase-free PCR tubes containing a drop of ethanol (99%) (S1 and S2 Figs).
Addressing the question of normalisation we recently performed a study on miRNA expression profiling to identify and validate reference genes for relative quantification of miRNAs. MicroRNA-193a-5p, miRNA-27a, and let-7g were identified as the most stably expressed miR-NAs in our cohort of rectal cancer and were consequently used as normalisers in the present study [24]. 001563 were used according to the manufacturer's standard instructions for low sample input (LSI).
Custom MicroRNA RT Primer Pool and Custom MicroRNA PreAmp Primer Pool for analysing the above mentioned miRNAs were created according to Protocol for Creating Custom RT and Preamplification Pools using TaqMan1 MicroRNA Assays (Publication Part Number 4465407, Life Technologies).
RNA was reverse-transcribed using the Custom MicroRNA RT Primer Pool. Each RT reaction contained 3 μl total RNA and 12 μl RT reaction mix from TaqMan1 MicroRNA Reverse Transcription Kit.
Preamplification (12 cycles The RT-qPCR analyses were carried out on Applied Biosystems 7900 HT Real-Time System using 1 μl diluted PreAmp product, TaqMan1 MicroRNA Assays and TaqMan1 Universal Master Mix II NoAmpErase1 UNG in a total reaction volume of 20 μl. All reactions were performed in triplicate. Data analysis was performed using SDS software ver. 2.2.2 (Life Technologies).
Water was used as negative control. The no template control was included in the entire process (RT, preamplification and qPCR) and analysed together with samples.
Relative quantification. Average values of triplicate Cq values were used for further analysis. Cq is defined as the PCR cycle number at which the fluorescence meets the threshold in the amplification plot. The arithmetic mean of Cq values for miRNA-193a-5p, miRNA-27a, and let-7g was used for normalisation as mentioned above.
ΔCq values were converted to linear values for statistical analyses by the equation linear value (ΔCq) = 2 -ΔCq assuming that the amplification efficiency of the assays was close to 100%.

Image analysis and quantification
Digital whole slide images were obtained with an x20 objective using an Axio Scan Z1 bright field scanner (Carl Zeiss, Germany). Image analysis of the digital slides was performed using VisiomorphDP software (Visiopharm, Denmark).
The areas of ROI encircled on the H&E stained sections were outlined on the ISH digital whole slide images, and staining artefacts as well as tissue artefacts were cleared. Some slides were in a poor condition caused by folded or lost tissue, which limited the size of the area included in the ROI. We found, not surprisingly, that images with a small ROI (less than 2 mm 2 ) contributed with quite dramatic variation and they were therefore omitted from the statistical analyses. For miRNA-21 five patients were excluded; for miRNA-125b and miRNA-630 three patients were excluded; and for miRNA-145 four patients were excluded.
A pixel classifier was trained to discriminate the blue ISH signal from the red counterstain, the unstained and weakly stained tissue, and the tissue-free areas. The following parameters were obtained during image processing from each ROI: area of intense blue (the ISH signal), area of weak blue (considered background ISH signal), area of purple blue (blue located over nuclear red), and the total area of the individual ROIs. The relative area fractions (areas of blue colour of interest divided by the total ROI area) were considered representative of the relative expression of the miRNA of interest.

Data analysis
All data were summarised using standard descriptive statistics. Inter-and intra-patient variability was assessed by calculation of intraclass correlation coefficients (ICC), using a one-way random-effects model. Specifically, the ICC may in this setting be regarded as the percentage of the total variance in the samples accounted for by differences between the patients examined. If the majority of variation in the sample measurements is from inter-patient variation, the ICC is high (ICC ! 1). If the majority of variation is from intra-patient variation, the ICC is low (ICC ! 0). We considered ICC both for single samples (ICC single ) and for the mean of samples within patients (ICC mean ). The ICC single provides an estimate of the reliability of a given sample as to rendering information about the specific patient (compared to all other patients). Correspondingly, the ICC mean estimates the reliability of the information rendered by taking the mean of three samples from a patient.
The correlation between RT-qPCR and ISH-fractions for each miRNA was estimated for individual samples and for mean values for individual patients using the Spearman correlation coefficient.
In order to illustrate the inter-and intra-patient variation as well as the correlation between RT-qPCR and ISH values, all data were normalised to allow for plotting on a common scale. For each miRNA, the mean and the standard deviation for all RT-qPCR measurements of each patient were calculated, and any individual sample measurement was normalised to a standard scale by extracting the mean and dividing it by the standard deviation. This was also done for all ISH measurements, for each miRNA. The resulting values for both RT-qPCR and ISH were plotted separately for each miRNA.

In situ hybridization
Relative expression estimates for the target miRNAs (miRNA-21, miRNA-125b, miRNA-145, and miRNA-630) were obtained by image analysis of the ISH-stained slides ( Table 1).The miRNA-21 ISH resulted in a strong signal in the stromal tissue surrounding the epithelial tumour cells as also described previously [25,26]. A strong signal in enteric neurons in the stromal tissue and a weaker signal in other stromal cells were detected for miRNA-125b. The ISH for miRNA-145 resulted in a strong and a weak signal mainly in the smooth muscle cells and myofibroblastic cells surrounding the tumour cells [26,27].

Intraclass correlation coefficient (ICC)
The ICC of each miRNA was calculated for both RT-qPCR and ISH as shown in Table 2. For miRNA-21, miRNA-145, and miRNA-630, the TBp-fraction (i.e. all blue colours) was used, but for miRNA-125b we did not include the areas with intense blue staining representing the enteric neurons.

Correlation between qPCR and ISH
The Spearman Correlation Coefficient was calculated to compare the results obtained by RT-qPCR and ISH, respectively (Table 3). For ISH, when comparing with the RT-qPCR, the TBpfraction (all blue colours) was used for all four miRNAs, since RT-qPCR measures the total expression of the miRNA of interest, regardless of localisation. In general, there was a poor correlation between the estimates obtained by the two techniques with correlation coefficients ranging from 0.084 to 0.325 for the mean value for each patient, and ranging from -0.085 to 0.515 in the section including the deepest part of each tumour. The intra-and inter-patient variation and the correlation between RT-qPCR and ISH for miRNA-21, miRNA-125b, miRNA-145, and miRNA-630 are illustrated in Fig 2.

Discussion
An increasing number of studies focus on miRNAs as potential biomarkers in cancer diagnosis and treatment, but the divergent and contradicting results in the literature are a major obstacle to clinical application [5,21]. The situation calls for a critical approach to methodological aspects among which heterogeneity ranks high.
Tissue heterogeneity is a general characteristic of malignant tumours. It is well known that the micro-environment varies throughout the tumour with effect on growth, proliferation, and metastatic potential. The behaviour of tumour mutations is inconsistent, too [16,17].  [19] performed a miRNA profiling in hepatocellular carcinoma comparing different compartments of liver tumours and found that the miRNA dysregulation varies within tumour compartments. The two studies cannot without reservation be compared with our results, as none of them report the ICC. On the other hand, they both report significant intratumoral heterogeneity in accordance with the present study.
To the best of our knowledge, this is the first study to assess the intratumoral heterogeneity of miRNAs in rectal cancer. Moreover, we are not aware of any other studies applying the two techniques of RT-qPCR and ISH on adjacent sections of tissue to elucidate their mutual concordance. RT-qPCR was chosen, since it is widely accepted as the gold standard for miRNA expression analyses. ISH was chosen because of its ability to visualise the cellular localisation of the investigated miRNA in addition to the expression level. We used tumour specimens from The ICC single was higher than 50% for miRNA-21 (RT-qPCR: 65%, ISH: 63%) and miRNA-31 (RT-qPCR: 78%), and for miRNA-125b, miRNA-145, and miRNA-630, it was lower than 50%.
doi:10.1371/journal.pone.0156919.t002 patients with rectal cancer to analyse the intratumoral heterogeneity of a panel of miRNAs with potential clinical relevance [4, 6-9, 22, 23]. We found the overall extent of heterogeneity, as expressed by ICC, to be different for the five miRNAs. The CV is not as useful a parameter in this setting, as it does not provide a direct measure of the intra-patient variation relative to the overall variation in the measurements. The ICC, however, allows for comparison of the intra-and inter-patient variation in the datasets from the two different methods, thus providing an estimate of the precision of the individual measurements referring to a specific patient.
The inter-patient variation was higher than that of the intra-patient variation with respect to miRNA-21 and miRNA-31. This is an indication that the two miRNAs have potential as biomarkers, when analysing one sample (e.g. one diagnostic biopsy). Here we found that mean ICC values for RT-qPCR and ISH were almost identical for miRNA-21, 0.848 and 0.838, Fig 2. Inter-and intra-patient variation and correlation between RT-qPCR and ISH measurements for miRNA-21, miRNA-125b, miRNA-145, and miRNA-630. All data were normalised to be plotted on a common scale: see the main text for details. Each vertical line represents the range of the three measures obtained in the individual patient with different colours to represent RT-qPCR and ISH measurements, respectively. For each miRNA, patients were sorted according to their mean RT-qPCR measurement for plotting. respectively. These values are in good concordance with Nielsen et al. (2011), who reported an ICC value for miRNA-21 ISH of 84% in CRC tissue [25].
On the other hand, for miRNA-125b, miRNA-145, and miRNA-630, the ICC single was below 50%, requiring more cautiousness when interpreting results from a single sample. For miRNA-125b and miRNA-630 the ISH analysis showed lower ICC values (single and mean) than the RT-qPCR analysis, indicating a higher intra-patient variation for the ISH analysis. For miRNA-145 it was an opposite scenario with the ISH analysis showing higher ICC values (single and mean) than the RT-qPCR analysis, indicating a higher intra-patient variation for the RT-qPCR analysis. Therefore, relying only on the ICCs, none of the two techniques are superior to the other.
Interestingly, we found a poor correlation between the results from RT-qPCR and ISH for all the miRNAs; so even though ICC (single and mean) is equal for RT-PCR and ISH for miRNA-21, the results are not comparable. The main explanation is likely to be inherent differences in the methodology of RT-qPCR and ISH. E.g. different from the RT-qPCR technique, the ISH technique detects both precursor and mature miRNAs. However, it cannot be excluded that part of an explanation for the poor correlation could be small displacements of ROI when copying it from the primary plotting to the ISH digital whole slide image and to the slide for RT-qPCR. This would mean a difference in the analysed area, especially in the stroma, where all the miRNAs are represented.
In general, RT-qPCR has a high reproducibility and is a relatively low cost method applicable in most laboratories. FFPE tissue is well suited for miRNA expression analysis performed on a small amount of tissue with only the mature miRNAs being detected. However, RT-qPCR includes several preparatory steps (RNA extraction, reverse transcription, preamplification) in which technical variation can be introduced, and in addition proper normalisation is essential to ensure reproducible results.
The ISH technique presents several advantages, including the ability to visualise the cellular localisation of the investigated miRNAs in addition to the expression level. The quantitative ISH method requires standardised section thickness, because the quantitative estimates are not normalised against a reference standard. Other preparatory steps that may introduce some technical variation include proteolytic pre-treatment as well as tissue folds and detachment. In addition, ISH is a more laborious method not suitable for all laboratories, and data interpretation requires specialised staff In the present study it was possible to visualise the cellular localisation of the investigated miRNAs, and the corresponding ISH signal allowed for focus on the relevant cells in the ISH data analysis. For miRNA-125b we could leave out the strong ISH signal form the enteric neurons when calculating the ICC. For RT-qPCR, the same option is not available, and miRNA-125b from the enteric neurons is a part of the measured expression level. Since the enteric neurons are present regardless of cancer, leaving them out of the analysis would be the most accurate thing to do. In that case, RT-qPCR including miRNA-125b from the enteric neurons would give a biased estimate, and ISH would probably give a more correct result. This has to be balanced with the loss of cases in the ISH analyses due to folded or lost tissue. Thus, several cases were actually lost for all the analysed miRNAs, probably due to erroneous fixation of the tumours.
For all miRNAs ICC mean was higher than ICC single . A low ICC value could be explained as a high measurement noise, whereas a high ICC value indicates less measurement noise. The fact that ICC mean was higher than ICC single suggests that pooling all three samples reduces the measurement noise, but only one available biopsy is often the scenario in clinical practice.
Relying on the ICCs calculated here, it was not possible to conclude which of the two techniques, RT-qPCR or ISH, provides the most representative expression estimates and thereby comprise the tumour heterogeneity. However, for clinical application of a method, practical factors such as number of biopsies needed for examination, accessibility and cost of the method, and interpretation of the results are indeed important factors In conclusion, intratumoral heterogeneity should be considered in the analysis of miRNAs as biomarkers, and miRNA analysis should hence be performed on all available samples. One sample may be sufficient in some settings as indicated for miRNA-21 and miRNA-31 in the current study. The poor correlation between RT-qPCR and ISH suggests that results obtained by the two techniques should be compared with caution. In a clinical setting, RT-qPCR seems preferable as an easy method of getting reproducible results, however, it cannot be excluded that the ISH method may provide more clinically relevant expression estimates. . Data were normalised for plotting on a common scale. For any given measurement, the mean and the standard deviation for that measurement for all samples were calculated, and any individual data point was normalised to a standard scale by extracting the mean and dividing it by the standard deviation. The resulting values for RT-qPCR were plotted for each patient. (EPS) S1