We previously reported that a comparison of Agilent microarray profiling and SYBR green-based quantitative RT-PCR (qPCR) of ten miRNAs in seven different human tissues found the two measurements correlated quite well 26. To perform a more extensive comparison, we analyzed the expression of 61 human miRNAs in nine different tissues (brain, breast, heart, liver, placenta, testes, ovary, skeletal muscle, thymus), using both Agilent miRNA microarrays and TaqMan stem-loop qRT-PCR 15. Aliquots of the same RNA samples were used for both the microarray and qPCR measurements. We chose these particular 61 miRNAs for several reasons. First, they represent a wide range of expression levels, as determined in an initial array analysis of some of the tissues. Second, they have wide differences in GC content, ranging from 23% (miR-190) to 68% (miR-328). Third, we chose several miRNAs which had potentially problematic sequences or exhibited atypical behavior during the development of the Agilent microarray platform: two of these did not show as good a linear titration curve as other miRNAs tested in a previous study (miR-126*, miR-296) 26, and two other miRNAs were previously reported not to be labeled by enzymatic methods similar (but not identical) to that used with the Agilent microarray assay (miR-208, miR-219) 33.

Of the 61 miRNAs examined, only miR-637 was not detected by either method in any of the tissues. The rest of the miRNAs assayed were detected in most or all of the tissues by both methods, with two exceptions: miR-208, expressed only in the heart 34 and at very low levels in skeletal muscle, and miR-138, expressed in the brain, and at lower levels in placenta and thymus (all data is shown in Additional File 1).

The qPCR and microarray results were compared by plotting the qPCR cycle threshold (Ct) value versus the log₂ of the array signal for each miRNA in all nine tissues (representative plots are shown in Figure 1, with the remaining plots shown in Additional Files 2, 3, 4, 5, 6, 7). These two values should be directly comparable, since both the qPCR Ct value and the log₂ of the microarray signal change by a value of 1 for every 2-fold change in miRNA concentration. If the qPCR and microarray measurements are equivalent, the plots will show a linear correlation (R = -1) with a slope of -1. Figure 2 shows the slopes and the correlation values for each of the 60 miRNAs. 56 of the 60 miRNAs show correlation values (R) between -0.8 and -1.0, and 50/60 plots have slopes between -1.2 and -0.8. Of the four miRNAs which were selected as potentially problematic in the microarray measurements, only miR-296 did not correlate between the microarray and qPCR assays; miR-208, miR-219, and miR-126* all gave excellent correlations.

To examine the results for all 60 miRNAs on one plot, we cannot simply plot qPCR Ct values versus microarray signals for all miRNAs in all tissues, because both the qPCR and microarray assays have differential sensitivities to different miRNAs. Thus, instead of looking at absolute expression levels, we must look at relative ratios of miRNA expression between two different tissues. To judge the consistency of fold-changes measured by microarray and qPCR platforms, we plotted the ratios of miRNA expression between all 36 possible pairs of tissues as measured by qPCR (Ct(tissue1)-Ct(tissue2)) and by microarrays (log₂(signal in tissue1)-log₂(signal in tissue2)). Four such plots are shown in Figure 3 (the other 32 plots are shown in Additional Files 8, 9, 10, 11, 12, 13, 14, 15), while Figure 4 shows the slopes and R values for each of the plots for the 36 tissue pairs. The plots all show very good correlation between the qPCR and array ratios, with R values between -0.984 and -0.821. The slopes of the 36 plots vary between -1.05 and -0.793. The intercepts of these fold-change plots (shown in Figures 3 and Additional Files 8, 9, 10, 11, 12, 13, 14, 15) indicate the consistency between fold-changes measured by the two methods. The mean of the intercepts of the line fits for the 36 tissue pairs was 0.00 +/- 0.23 (1 SD) (data not shown). This level of variability (17%) is comparable to that seen among independent measurements using the same technique.

miR-494, miR-296, and miR-99a are the three miRNAs that exhibit the most discrepant correlation values and slopes between the qPCR and microarray assays (Fig. 2); however, if the measurement of miR-99a in placenta is omitted, the slope for this miRNA becomes -0.844 with R = -0.929 (see plot in Fig. 1). For miR-494 and miR-296, if one platform were measuring levels of these miRNAs accurately, while the other platform were not, then we might expect a significant divergence from linearity to be observed between the two measurements when adding increasing amounts of synthetic miR-494 or miR-296 RNA into a total RNA sample. To test this, we added 1 zmol to 10 fmols of synthetic miR-296 and miR-494 RNAs to 100 ng of total RNA from liver or placenta, and measured the qPCR and microarray responses (Figure 5). For both miRNAs, in both tissues, the relation between qPCR measurement and array measurement is linear above a threshold spike-in concentration. The R values of the linear regions are very close to -1, with slopes between -0.842 and -0.935, indicating that both the qPCR and the microarrays are producing sample-responsive and internally consistent measurements of miR-296 and miR-494 at these concentration levels. Below the threshold spike-in levels, the qPCR Ct values and microarray signals are unchanged for miR-494, while for miR-296 the Ct values increase slightly, but the array measurements are unchanged. We conclude that the difference between the two platforms is not due to different sensitivity, since both the microarray and qPCR measurements are capable of measuring miR-296 and miR-494 accurately above a spike-in concentration threshold. Presumably some type of interference confounds the measurement of endogenous expression levels in the complex sample, on one or both of the platforms.

A question which arises when comparing the miRNA profiling results reported in different studies is whether the methods used to isolate RNA from tissue or cell line samples systematically affect the miRNA profiles. To examine whether the miRNA profile of a sample is affected by the type of total RNA prep method used, we prepared one large frozen cell pellet from each of two different human cell lines, HeLa (a cervical carcinoma line) and ZR-75-1 (a breast carcinoma line). We then subdivided these pellets into equal aliquots, and performed total RNA isolation on the aliquots using three different techniques: phenol/guanidinium (TRIzol, Invitrogen) followed by isopropanol precipitation, and two column-based techniques, miRNeasy (QIAGEN) and mirVana (ABI). Four to eleven replicate preps were performed on each cell type with each method.

Mean RNA yields, as measured by absorbance at 260 nm, and quality metrics for each prep type are shown in Table 1. The RNA integrities of the preps were analyzed on the Agilent 2100 Bioanalyzer, and all the preps had high quality RNA according to the RIN number 3536. This indicates that most of the RNAs in the various preps were intact, with minimal breakdown. However, RIN values do not provide information about non-RNA contaminants, such as organic reagents and DNA. The TRIzol preps showed the lowest mean 260/230 ratios, possibly indicating the presence of some remaining TRIzol reagent in the final product.

Since the absorbance at 260 nm is used to quantitate the amount of RNA for use in the measurement assays, and since the 260:230 ratios can only serve as a crude guideline to possible contaminants, it is important to examine the absorption spectra in more detail (Figure 6). Some of the spectra clearly show the presence of additional peaks between 220–230 nm, indicating the presence of contaminant(s). While this is most consistently seen in the TRIzol preps, it is also sometimes seen in the mirVana preps and in one of the miRNeasy preps. In some of the samples where no distinct peak is observed between 220–230 nm, the spectra show significant absorbance in the wavelengths immediately below 220 nm, with a shoulder tailing from 220 nm to 240 nm. At 240 nm the absorption increases again to peak at around 260 nm, which is the wavelength of maximum absorption for nucleic acids of mixed oligonucleotide composition. For many of these spectra, this absorption pattern suggests that the absorbance of the contaminant(s) whose peak is below 240 nm may overlap with the nucleic acid absorbance peak at 260 nm, which would result in the overestimation of nucleic acid quantity as determined by the absorption at 260 nm. Also, a couple of the absorbance spectra show a slight shoulder in the 260–270 nm range, indicating a contaminant which could also affect RNA quantitation. Thus, careful examination of sample spectra can be important for identifying samples where measured miRNA levels might be compromised by absorbance-based RNA quantitation artifacts.

For miRNA microarray profiling analysis we took three of the replicate total RNA preps of each different prep method in each cell type and hybridized them to Agilent miRNA microarrays. A total of 42 hybridizations were done, with all but two samples hybridized at least twice (Additional File 16).

A good measure of the reproducibility of replicate measurements is the Root Mean Square (RMS) deviation of the natural logs of all signals that are well above background levels. The RMS deviation is approximately equal to the coefficient of variation of signals, and is an estimate of the proportional error of the measurement. For example, if the RMS deviation is 0.15 (15%), then a measured fold change between samples of 1.15 is a difference of one standard deviation.

We compared the RMS deviations between pairs of hybridizations performed using the same total RNA prep (hybridization replicates) with those performed using different total RNA preps done with the same method (prep replicates), and also with those using different total RNA prep methods. We also compared hybridizations performed on the same or different days, in order to take into account any day-to-day variability in the results. Figure 7a shows box plots of the RMS deviations between all possible replicate hybridization pairs, categorized by same or different prep method, same or different prep replicate, same or different hybridization replicate, and same or different hybridization day. The box plots for the six different categories are shown in decreasing order of variability. Hybridizations using preps from different methods show the most variability, with same day hybs of the different prep methods showing slightly less variability than different day hybs. Replicate preparations using the same method are the next lowest in variability, again with same day hybs being less variable than different day hybs. Finally, hybs done with aliquots of the same preparation have the lowest variability, again with same day hybs being less variable than different day hybs. The three sources of variability can thus be put in order of their magnitude: different prep methods > different preps using the same method > different hybridization day.

We also plotted the RMS deviations for hybridizations involving prep replicate pairs from each of the three different prep methods (regardless of the hybridization day), in order to examine whether the different prep methods showed different amounts of variation between prep replicates (Figure 7b). The TRIzol prep replicates showed less variability than the other two prep methods. A Student's t-test between pair-wise comparisons of the three prep methods confirmed that this difference is statistically significant (data not shown). There was no significant difference in variability between the mirVana and miRNeasy prep replicates.

To examine whether there are systematic differences among the miRNA profiles observed for RNA isolated by different methods, we first looked at the overall signal levels of the hybridizations. The grand means of the mean total gene signal for all hybridizations of the same prep type for each cell line are shown in Table 2. For both cell types, the miRNeasy preps gave about 25% higher overall signals than the other two prep methods. The overall signal differences between the TRIzol and mirVana preps of the same cell type were minimal. It should be noted that the miRNeasy preps had the highest 260:230 ratios (Table 1), and their spectra generally showed lower absorbance in the 220–230 nm range compared to the other two prep types. It is possible that some of the material which is absorbing at 220–230 nm in RNA extracted using the other two methods is contributing to the 260 nm peak, and it is also possible that DNA contaminants are present. The presence of either or both types of contaminant can lead to an overestimation of the amount of RNA present in these preps.

We next examined whether specific miRNAs systematically differ among the different prep methods (Figures 8 and 9). We calculated an average expression profile for each prep method for the two cell lines, by first averaging the total gene signals for each miRNA from all hybridizations of the same RNA prep, and then averaging together these individual prep averages for all preps of the same prep and cell type. Since there were differences seen in the overall signal levels between the different preps (Table 2), we normalized each pair-wise comparison to the 75^th percentile of one of the pairs. The expression profile of most miRNAs in HeLa cells does not depend on the RNA prep method (Figure 8). However, there is a small subset of miRNAs that consistently report different relative expression levels depending on the prep method. The miRNAs that are labeled in the figure show expression levels that differ by at least 2-fold in different prep methods. Additional File 17 lists the miRNAs which are 1.5x and 2x higher in one prep type compared to another. Three miRNAs are found at consistently lower levels in TRIzol preps than in the other two preps: miR-29b, miR-33, and miR-219. mirVana preps show consistently higher levels of four miRNAs when compared to the other two HeLa preps: miR-149, miR-328, miR-574, and miR-766. Figure 9 (and Additional File 17) shows the results from the ZR-75-1 breast cell line. While fewer miRNAs show different profiles among the three different prep methods in this cell line compared to HeLa, four out of the five that are observed to be discrepant in ZR-75-1 cells are the same as those seen in HeLa cells (miR-29b, miR-33, miR-219, and miR-328).

The finding that a small number of miRNAs report different microarray signals when prepared by different methods raises the question of whether these differences reflect real differences in the concentrations of these miRNAs in the different sample preps. To examine this, we assayed individual samples prepared with the three methods by qPCR, using primers for three of the miRNAs showing differences between the prep methods. We then compared these results to those obtained from microarray analysis of the same preps (Figure 10). The good agreement of the qPCR results with the array results for these miRNAs strongly suggests that the differences in the miRNA levels observed between the sample prep methods reflect true differences in the miRNA content of the extracted RNA, and are not artifacts of the measurement assay. At present, we have no explanation for why these particular miRNAs are found at different levels when using different extraction techniques.

Total RNA samples from normal human tissues were from Ambion (Austin, TX). Frozen HeLa cell pellets were from Cell Trends (Middletown, MD), and frozen ZR-75-1 cell pellets were from BioProcessing Inc. (Portland, ME).

miRNA microarrays were manufactured by Agilent Technologies (Santa Clara, CA)., and contain 20–40 features targeting each of 470 human miRNAs (Agilent design IDs 015508 (sample prep studies) and 016436 (nine tissue comparison studies)) 37. Sequences of the 470 miRNAs were obtained from the Sanger miRBase, release 9.1 101112. Labeling and hybridization of total RNA samples were performed according to the manufacturer's protocol. 100 ng total RNA was used as input into the labeling reaction, and the entire reaction was hybridized to the array for 20 hours at 55°C. For the microarray versus qRT-PCR comparisons, the labeling and hybridizations of the nine human tissues were done 4–5 times, and the mean and standard deviation for each miRNA were calculated.

Microarray results were extracted using Agilent Feature Extraction software (v9.5.3.1) and analyzed using GeneSpring GX 7.3.1 software (Agilent Technologies) and Spotfire DecisionSite 8.1 software (TIBCO Software, Palo Alto, CA). Box plots were calculated using JMP 5.1 software (SAS, Cary, NC). Original microarray data is deposited in the Gene Expression Omnibus 38 (Series GSE11879).

All scatter plots of miRNA microarray data use the total gene signal, which is proportional to the total number of targets bound by the probes targeting each miRNA 3132. For comparison of two hybridizations, the natural logs of the total gene signals for all genes expressing above 10x the background noise in both samples were regressed against each other, and the standard deviation of the residuals from the regression line were reported as the RMS deviation. For most pairs of samples prepared by the same method, residuals were normally distributed, so that the RMS deviation describes true random variation in the assay. In pairs of samples prepared by different methods, residuals of most of the miRNAs were also normally distributed, with systematic exceptions of some miRNAs as discussed in the text. No normalization was performed for either microarray or qPCR data, except for an overall intensity normalization applied to the average signals from different prep methods, as described in the text (Figures 8 and 9). For this comparison, the 75^th percentile of the total gene signal for all the miRNAs on the array was calculated by sorting the total gene signals for 470 miRNAs on the array in ascending order, and the signals from the three methods were normalized to the signal from the 353rd miRNA.

miRNA qRT-PCR analysis was performed using Taqman miRNA assays (Applied Biosystems, Foster City, CA), according to the manufacturer's protocol. 5 ng total RNA was input into each reverse transcription reaction (RT) for each miRNA. Four replicates were done for each miRNA, consisting of two replicate PCR reactions from each of the two replicate RT reactions, and the results were averaged. PCR reactions were run on a 7500 Real Time PCR machine (Applied Biosystems) and analyzed using 7500 System SDS software (v1.4).

Synthetic miRNAs were manufactured by TriLink BioTechnologies (San Diego, CA) and spiked into human liver and placenta total RNA (Ambion). 100 ng of these RNA mixes were then used for labeling and hybridization onto the microarrays, while 5 ng were used as input into the reverse transcriptase reaction for qPCR. Two replicate microarray hybridizations and four replicate qPCR reactions were done for each dilution in each tissue.

Frozen cell pellets were resuspended in phosphate buffered saline and divided into equal aliquots of 5 × 10⁶ (HeLa) or 1 × 10⁷ (breast) cells and refrozen. Individual aliquots were subsequently thawed just before use.

TRIzol preps were performed according to the manufacturer's protocol (Invitrogen, Carlsbad, CA) using an isopropanol precipitation. Briefly, 1 ml of TRIzol reagent was added to the cell pellet and cells were lysed by repetitive pipetting, and then incubated at room temperature for 5 minutes. 200 μl of chloroform were added, followed by vigorous shaking and incubation for 2–3 minutes at room temperature. Samples were centrifuged 15 minutes at 12000 × g at 4°C. The aqueous layer was transferred to a new tube, and the RNA was precipitated by adding 0.5 ml isopropanol, incubating 10 minutes at room temperature, and spinning for 10 minutes (12000 × g at 4°C). Pellets were washed with 80% ethanol and resuspended in nuclease-free dH₂O (Ambion).

miRNeasy total RNA preps (QIAGEN, Valencia, CA) were performed according to the manufacturer's protocol. The mirVana miRNA Isolation kit (Applied Biosystems) was used according to the manufacturer's protocol for total RNA isolation.

All total RNA preps were analyzed using the 2100 Bioanalyzer (Agilent Technologies), RNA 6000 Nano LabChip kits, and 2100 expert software (version B.02.05.SI360). Absorption spectra were measured on an ND-1000 spectrophotometer (NanoDrop Technologies, Wilmington, DE).