7. Conclusions
As a well-established discovery tool in biological and biomedical research, microarray has been successfully migrated to miRNA studies. Different microarray platforms have their own strengths and weaknesses. In ideal situations, we can make good uses of various microarray platforms to borrow the strengths from each other. However, the reality is that most researchers tried to stay with as few platforms as possible, due to the low reproducibility among arrays across platforms and/or across laboratories. The diversity of miRNA microarray platforms and lack of reliable analytical methods actually have made cross-platform miRNA microarray data comparison and integration challenging.
Measurement errors exist in all microarray platforms. In miRNA microarrays, the expressions of numerous miRNAs might be dominated by measurement errors, and the overall signal quality of the miRNA microarrays is usually deficient. Thus, it is especially important that the miRNA microarray experiments will be well designed and well conducted. In some literature, it was emphasized to ensure the measurement quality via controlling the experimental factors, and it was suggested that the background signal subtracted signals should be used directly without any normalization for miRNA microarray data [1,67,68]. However, we need to be aware that measurement errors are inevitable in microarray experiments. Conclusions based on microarray data without proper normalization might be quite misleading.
Before we apply any statistical method to normalize microarray data for further analysis or to calibrate measurement errors to detect differentially expressed miRNAs, it is always good to screen first the bad microarrays off—if evidence shows that an array completely failed, it should be excluded from further analyses. To our knowledge, there is no existing method to evaluate systematically the quality of different slides. However, we can assess the quality of various arrays via the following different ways. First, if the experiment is conducted well, a reasonably large portion of strong signals are expected from each array. One of the LNAv9 array contains only 3.89% of spots that are not flagged. This is a strong indication that the whole slide might not be usable. Automatic flagging can be applied to check the signal quality. However, we should not simply discard the measures from those probes that are flagged as weak. On the other hand, if a probe is found to be contaminated, the corresponding measure should not be used. Second, we can compute the Spearman’s correlation coefficient between any pair of arrays. The Spearman’s correlation coefficient is rank based, so it won’t be affected by the distributions of the miRNA expressions in the same array. For example, arrays tested with the same platform are supposed to have reasonably high reproducibility. A very small Spearman’s correlation coefficient usually can be used as an indicator of something wrong in the array data. Third, we can compare the density distributions of expressions of all miRNAs being tested in every array. For miRNA microarrays, due to the small total number of miRNA and also the violation of the assumption that the numbers of upregulated and downregulated miRNAs are approximately the same, it is hard to judge the signal quality of an array when it has a different expression distribution.
Normalization is an essential matter for microarray data analysis. A well-developed normalization method can efficiently calibrate the measurement errors and can offer a powerful tool for cross-platform and cross-laboratory microarray data integration. Most existing normalization methods for miRNA microarray data are adopted from mRNA/cDNA microarrays with or without modifications. However, the unique signature of miRNA has reduced the enthusiasm of such adoption. In our previous studies and research by others, it has been found that the two-component measurement error model and the generalized logarithm transformation work well for microarray data. Based on the measurement error model, several normalization methods and differentially-expressed miRNA detection algorithms have been developed and achieved good results. We should keep in mind that the measurement errors coupled with data from different platforms have different characteristics, and hence, it is not realistic to develop one or a few methods that can deal with data from all platforms. For example, expression data obtained from bead arrays have stronger signals, but without technical replicates. As a result, we have to heavily count on the global information from the profile for bead array when there is no replicated array for the same specimen. In addition, most of the existing normalization methods based on the global information are not applicable to normalize the sub-profiles, due to the extreme small number of miRNAs tested in each pool. When dealing with the traditional glass-based arrays, which have weaker signals compared with the bead arrays, but usually have several technical replicates, special attention needs to be paid to the background correction and how to find robust estimates from several replicates on each array. It is also worth noting that borrowing the strengths of some reliable analytical platforms, such as the TaqMan Array Human MicroRNA Panel, might be a good approach for miRNA microarray data normalization. Meanwhile, we also need to keep an eye on the quality of qRT-PCR results as well—measurement errors also exist in qRT-PCR platforms. The qRT-PCR results should be used as “gold standards” with caution.
We recommend to keep the measurements from the weakly expressed miRNAs in the analysis. However, in detecting the differentially expressed miRNAs, miRNAs expressed at different levels should be tested (viewed) differently. According to their expressions in the control sample, the miRNAs can be grouped into three groups: group 1, for those with expression significantly stronger than the background noise; group 2, for those with expression close to the background noise; and group 3, for all others. For miRNAs in group 3, it is meaningless to test whether any of them are downregulated, and attention should be paid more to those in the other two groups, especially in group 1.
In summary, it is challenging, but necessary, to develop some novel adaptive statistical methods to efficiently calibrate the measurement errors for normalization and for differentially-expressed miRNA detection. In that way, we can reuse the miRNA microarray data saved in a variety of databases and to integrate data from similar studies contributed by different laboratories using various platforms. Even in an era in which the next-generation deep sequencing technologies have been widely used, microarray is still very valuable as a reliable and affordable profiling tool. In addition, measurement errors and bias exist in small RNA sequencing data, too. Some of the normalization methods for miRNA, including the quantile normalization, smoothing-based normalization methods and measurement error model-based normalization methods, can also be applied to small RNA sequencing data with or without modifications.