5. Differentially-Expressed miRNA Detection
In miRNA microarray data analysis, the main interests are usually focused on whether a specific miRNA or a set of miRNAs are differentially expressed in the two cell populations. The concept of “differentially-expressed” is not well-defined, which makes it challenging to detect the differentially expressed miRNAs. Without loss of generality, we assume an miRNA has expression, x, in the control sample, and y, in the test sample. Usually, we will judge whether the miRNA is differentially expressed by checking the fold-change (FC), which is the ratio of y/x. If FC=1, the miRNA is not differentially expressed. If FC is much larger than 1.0, we say the miRNA is upregulated; otherwise, if FC is much smaller than 1.0, we say the miRNA is downregulated. An miRNA that is either downregulated or upregulated is said to be differentially expressed. However, it is not clear how different that can be treated as differentially expressed. We need to take both the fold-change and the basal level of the miRNA into consideration. For example, if the miRNA is highly expressed, a two-fold change might be practically meaningful, and the miRNA can be considered as differentially expressed. Meanwhile, if the miRNA is weakly expressed, a large FC value might be practically meaningless to the researchers. In practice, we do not know the actual value of FC; instead, we estimate it using FC^=g1(X)/g2(Y). Here, g1(.) and g2(.) are two functions that are expected to be able to calibrate the measurement errors and will approximate the true values of x and y, respectively. As shown in the measurement quality and background correction section, the majority of miRNAs have low SNRs, and thus, the measurement errors could play very important roles in the intensity measures, Y and X and, hence, could severely influence the estimate of the fold-change. In other words, measurement errors pose a higher degree of uncertainty in identifying the differentially expressed miRNAs. We need to use the FC with caution in differentially-expressed miRNA detection. The fold-change criteria should be used only to miRNAs that are turned on (“expressed”) in both samples. In more detail, if an miRNA is expressed in one sample, but not expressed in the other sample, the FC value could be zero or infinity, in theory. In practice, due to the existence of background signal, the observed FC value could be very small or very large. In such cases, it is relatively easy to detect the differentially-expressed miRNAs, though such miRNAs should be dealt with separately, because they will make the distribution of FC values of all miRNAs very skewed. While in another scenario, if an miRNA is not expressed in both samples, the FC value is dominated by background noise, and such miRNAs should be excluded from the study.
In miRNA discovery studies, when a large number of replicates are available, statistical tests, such as t-test, ANOVA or other omnibus tests can be used for differentially expressed miRNA detection. It is worth noting that the non-linear normalization methods, including the loess method and quantile normalization method, are preferred, due to their capability of dealing with the nonlinear changes of miRNAs with different expression levels. Fan et al. proposed a model with a parametric component and a non-parametric component to test the differentially expressed genes by taking the treatment effect, block (position) effect and the nonlinear relationship into consideration [54].
In practice, many laboratories do not repeat array experiments. As a result, there is no biological replicates for each miRNA. For some microarray platforms, there might be a few technical replicates available on each array. The lack of sufficient replicates makes it a big challenge to identify the differentially expressed miRNAs. For instance, in the 48 LNAv9 arrays, the ten osterosarcoma xenograft specimens are treated with three chemo-drugs and saline, respectively. Most of the samples are tested once, and some are repeated two times. If research interests are focused on the drug-resistance of different patients to the three chemo drugs, respectively, there is only one or no biological replicate. For the LNAv7 and LNAv9 arrays, there are four technical replicates on each array for every miRNA. However, for the bead arrays, there is neither a biological replicate nor technical replicates. In [40], the authors proposed to identify the differentially expressed miRNAs by regressing the expressions from the test sample on the expressions from the control sample using an errors-in-variables non-parametric regression model. This method can be applied to detect the differentially expressed miRNAs based on two arrays without replicates. When replicated samples or probes are available, data from various arrays can be integrated and, hence, improve the overall performance of the regression model to detect differentially expressed miRNAs, which is also validated using the qRT-PCR results [40].
The measurement error model-based tests can make good use of the global information from each of the expression profiles, and the required number of samples per gene/miRNA can be greatly reduced [18,19]. However, when the overall signal quality of an array is fair or poor, the outcomes could be quite questionable if no biological replicate is available to use. It is crucial to have multiple biological replicates (at least two) to improve the sensitivity and specificity in differentially-expressed miRNA detection.