This report is aimed at understanding the performance characteristics of a newly developed, highly sensitive, high-dimensional miRNA expression platform, the Illumina human miRNA BeadArray. In particular, we wished to rigorously assess reproducibility of the assay and present raw, un-normalized data to better understand the performance of the assay. Overall, the Illumina miRNA array system performed very well. With respect to reproducibility (both within a plate and between plates), this assay was found to be similar to other high-dimensional gene expression platforms. The recommended input of RNA was 200 ng. Using this as the reference, however, we found no evidence for a dilution effect, with comparable results obtained from 25 ng to 800 ng RNA. Therefore, miRNA from small tissue samples can be reliably assayed. Additionally, this platform can test up to 96 samples at a time, providing high throughput capability. A description of the platform has been previously published  and included validation of 12 individual miRNA by qPCR, and on a more global scale, miRNA expression was validated using digital gene expression.
The experimental approach was to look at variability between extractions, between technical replicates (both within and between plates) and between varying inputs of RNA using RNA from five colon cancer specimens and four cell lines. The agreement between extractions appears to be similar to that between technical replicates within an extraction and was similar for all the samples tested. As one would expect, the between-plate agreement was slightly less than within-plate agreement, although most correlation values were >0.90. There were several outliers noted in one SAM. Processed at the same time as another SAM, these may be related to sample handling or perhaps to a few underperforming probes.
We present both scatter plots and MVA plots to evaluate the data; the latter are a useful tool for assessing biases in high-dimensional data. Just as with high-dimensional gene expression microarray platforms [32, 36, 34, 38, 37], the bias curves observed here demonstrate that bias is not constant across all abundance levels for high-dimensional microRNA data indicating a need for nonlinear normalization. There is debate in current literature as to whether normalization algorithms used for high density mRNA microarrays are applicable to data from microRNA microarrays [39–43]. Most commonly used mRNA normalization algorithms assume that only a small portion of probes are differentially expressed, that the distribution of differentially expressed probes is approximately symmetric about identity. In addition, there must be sufficient probes for estimation of bias without over-fitting. These assumptions must be evaluated specific to the experiment at hand. The data we present here suggest that there are sufficient numbers of probes expressed on this microRNA platform for estimation of biases without over-fitting using standard mRNA normalization algorithms in this experiment. In addition, there is no differential expression expected in the replicates analyzed herein. Thus, the assumptions hold for these data and nonlinear normalizations such as quantile [34, 44] and cyclic loess [32, 33] should be directly applicable to this experiment. Indeed, the fastlo normalization method utilized in this report removed the nonlinearity seen in the bias curves.
Finally, we selected 12 miRNA targets to further assess the accuracy of the Illumina array utilizing a second method to quantitate miRNA levels (qPCR). Using RNA from the four cells lines used in this study, the correlation between miRNA expression levels derived from the Illumina platform compared to that from the qPCR analysis was excellent for the 12 targets. These results were essentially the same as those previously reported .
Overall, the Illumina array appears to be quite specific for the mature form of the miRNA. Chen et al. hypothesize that the cDNA synthesis may be more complete for the short mature miRNAs than for the pre-miRNA templates. Also, the stem-loop structure of pre-miRNAs could hinder the cDNA synthesis and annealing of the oligonucleotides, resulting in relatively selective detection of expression of the mature miRNAs .
In summary, we found this high-dimensional miRNA profiling platform to be highly sensitive, providing reproducible data over a wide range of RNA input amounts. The variability between extraction, replicates, and SAMs was found to be acceptable. After comparison with independently produced qPCR data, validation of the absolute and relative quantification of the technology seemed adequate as well.