Recent studies have shown that miR expression profiling may be more accurate for distinguishing disease states than mRNA expression analysis [23
]. Since miRs are better preserved, and can be obtained from FFPE tissue, they may be a better choice for expression studies when using FFPE samples [11
]. miR expression has been examined in FFPE and fresh-frozen tissue specimens. For example, an 89% correlation in miR expression profiles was shown between FFPE and fresh-frozen samples from murine liver, using a locked nucleic acid (LNA)-based miR microarray platform [11
]. Other studies, using human tissue sources (colon, myometrium, lymph node, melanocytes), on various microarray platforms (mir
Vana miRNA Bioarrays, Agilent human miRNA arrays, and Invitrogen Ncode Multi-Species miRNA microarrays), showed high correlation between miR expression profiles in FFPE and fresh-frozen samples [18
]. Technologies such as Luminex's fluorescence labeled beads, when used on FFPE and fresh-frozen breast tissue, also yielded similar results [15
], showing that FFPE tissues are useful for miR expression analysis.
Another platform for miR expression analysis, using high-throughput qRT-PCR, has yielded high correlation coefficients when comparing fresh-frozen and FFPE samples [12
]. Recently, a comparison between FFPE and fresh-frozen tissue using two pairs of breast carcinoma samples was performed on the high-throughput qRT-PCR TaqMan Human MicroRNA Array v1.0 (Early Access), demonstrating a correlation of 0.94 between paired samples [21
]. In our study, we also tested the TaqMan Human MicroRNA Array v1.0 (Early Access) platform to show its utility and reproducibility in determining a miR expression profile using FFPE clinical samples. However, our study differs in several aspects, including the assessment of different extraction methods and demonstration of platform optimization using less than recommended RNA amounts as described in Materials and Methods. These experiments are necessary to warrant the use of this platform in the study of miR expression in FFPE tissues, and have not been previously described. Our study also adds new information to the current literature, showing the importance of stratifying miRs based on abundance when analyzing miR expression data.
Our optimization results showed that the amount of input RNA and dilution of the cDNA can affect miR expression levels detected. These effects are more pronounced if examined from the perspective of miR abundance. Our results show that input RNA from FFPE samples should be maximized and that dilution of cDNA should be kept to a minimum, in order to generate the most reliable and reproducible miR expression profiles using the TaqMan Human MicroRNA Array v1.0 (Early Access) platform. This is apparent when observing the shifts that occur in miR abundance with changes in RNA input concentration. As seen in Figure , with a lower input RNA concentration, a lower number of high abundance miRs are detected. Our results showed a significant shift in the proportion of miRs in each abundance stratum with the increase in input RNA concentration from 10 ng/μL to 200 ng/μL (chi-square test, χ2 = 12.7, df = 2, p = 0.002). The largest change is decreasing frequency of low abundance miRs (163 to 130) and increasing frequency of high abundance miRs (72 to 112), a pattern which is consistent over each increment of input RNA concentration. Of note, the low abundance strata across the input RNA concentrations have a higher variability, compared to the medium and high abundance strata at the same concentrations, as shown in Figure . This demonstrates that expression values obtained for miRs within the medium and high abundance strata are more reliable for analysis and a direct link between increasing input RNA concentration and improved reliability for low and medium abundance miRs.
Well failures are also less common as input RNA concentrations increase or cDNA dilutions decrease. Interestingly, dilution of cDNA can be used in order to compensate for low input RNA amounts, a scenario which can occur should RNA yield be low. This can be observed when comparing an input RNA concentration of 200 ng/μL with a 15× cDNA dilution factor and an input RNA concentration of 66.7 ng/μL with a 5× cDNA dilution factor. Our results show the consistency of Ct values obtained with different cDNA dilution factors (Pearson correlation 0.95). This was also shown to be true when examining correlation coefficients relative to miR abundance. There was also no statistically significant shift in the proportion of miRs in each abundance stratum between equivalent combinations of input RNA concentration and cDNA dilution [chi-square test, (χ2 = 3.3, df = 6, p = 0.77)]. In our laboratory, we have optimized the experiments to use an input RNA concentration of 200 ng/μL with a 15× cDNA dilution factor. This gives us the flexibility to use input RNA concentrations as low as 60 ng/μL with a 5× cDNA dilution factor, which is the lowest cDNA dilution possible according to the manufacturer's protocol, in order to maintain equivalent miR profiles between samples.
Finally, we showed that miR expression was highly correlated in FFPE and fresh-frozen lymphoid tissues, regardless of the extraction method used, despite the fact that raw Ct scores obtained from FFPE samples are significantly different than those from fresh-frozen tissue. This is likely due to reduced yields of RNA from FFPE tissue. Unsupervised hierarchical clustering analysis, using correlation heat-maps, showed that tumour and normal samples clustered in separate groups. The hierarchical clustering performed according to miR abundance demonstrated that the medium abundance miRs appear to be most informative in helping to segregate tumours from normal samples. This may be because many miRs that fall into the low abundance category are potentially not expressed in lymphoid tissue. Our pre-filtering removed miRs with Ct values of 40 in all twelve samples, meaning that if one miR had a Ct value less than 40 but greater than 35 in even one sample, it would be included in the analysis as a low abundance miR. The high abundance miRs may be informative as well, although one tumour sample clusters together with the normal samples. Interestingly, this tumour is a post-treatment sample; the patient having been treated with R-CHOP (rituximab-cyclophosphamide-doxorubicin-vincristine-prednisone) prior to biopsy and sample collection. Rituximab is an anti-CD-20 antibody, which targets B-lymphocytes expressing CD20 (including mantle cell lymphoma), and combined with CHOP chemotherapy, it has been shown to improve overall and complete response rates, as well as time to treatment failure, in mantle cell lymphoma patients [25
Also of note is the shift of miRs from high to medium and medium to low abundance in FFPE compared to fresh-frozen tissue, seen in Table and Figure . The shift in proportion of miRs in the three strata is significantly different by chi-square test between TRIzol and FFPE (χ2
= 17.5, df = 2, p = 0.0002) and between mir
Vana and FFPE (χ2
= 13.5, df = 2, p = 0.001), but not between TRIzol and mir
= 0.32, df = 2, p = 0.85). This decrease in the proportion of high abundance miRs and increase in the proportion of low abundance miRs may be due to known reduction of RNA yields that occur in FFPE tissue [7
]. Reduction in RNA yield is seen due to chemical modification of RNA introduced by formalin fixation. This modification occurs through the addition of mono-methylol (-CH2
OH) groups to RNA, as well as formation of secondary modifications such as methylene bridging, preferentially at adenine residues [8
]. Although miRs may be relatively protected due to association with other proteins (e.g. the RISC complex), some modifications likely occur since the shift in miR abundance from high to medium and from medium to low indicates a reduction in miR yield from FFPE samples. This effect has also been seen by Hoefig et al. [16
] who demonstrated a 1.0-1.5 Ct reduction in miR expression between fresh-frozen tissues and their paired FFPE samples. In our study we observe that this reduction does not affect sample biology. Our PCA results (Fig. ) reveal a complete separation between normals and tumours regardless of sample treatment (FFPE vs
. fresh-frozen). Similar to our results, Hoefig et al. [16
] demonstrated that biologically relevant miR expression profiles between liver and lymph node tissues were unchanged regardless of tissue treatment, and that this biological variance was higher than the technical variance introduced by formalin fixation and paraffin embedding [16
]. In our study, by comparing normal and tumoural tissues from the same source (lymphoid), we were able to demonstrate robust differential expression, which was higher than the differences between FFPE and fresh-frozen tissue, as seen in Figure . Thus, the correlations reported in our study demonstrate that miR levels likely reflect the disease state in patients regardless of tissue source (FFPE vs
. fresh-frozen) used for analysis.
Although our results show that the technical differences introduced by tissue treatment are minimal, we also show that data analysis can be strengthened by examining miR expression according to abundance strata. As seen in Figure , our results indicate that stratification of miRs according to abundance should be considered during data analysis as low abundance miRs (Ct>35) do reveal low correlation coefficients when compared to medium (30≤Ct≤35) and high abundance miRs (Ct<30). Based on our results, we would exclude the low abundance miRs from analysis as the miRs in the medium and high abundance strata show superior correlations and sample clustering. We suggest that investigators stratify their miR expression data and judge whether or not to include miRs from all strata during data analysis on a case-by-case basis. This method of data analysis suggested by us is a potential tool for further optimization of miR expression data derived from FFPE samples and will increase the correlation between these samples and their paired fresh-frozen tissues.