What is the most appropriate method for correlation analysis of mRNA-target correlations? And why?

Question

[BioNovice247]

Hi all,

I am struggling to find the most appropriate method for miRNA-mRNA correlation analysis. Based on the published papers, the Pearson correlation seems to be the most common method used. But is there a specific reason behind this? I have also (rarely) seen papers that use Spearman or Kendall correlation for this analysis. So I'm wondering, which one is the best method to choose for miRNA-target correlation analysis and what are the upsides and downsides of each method. Thanks in advance for your response

Answer 1

[Kevin Blighe]

Pearson is for parametric data; Spearman and Kendall are for non-parametric.

For more information on each, I may suggest to search via the search engine of your choice. There is much documentation on these methods on statistics-oriented websites.

Kevin

[BioNovice247]

Thank you for your response. I am aware of the assumptions and general characteristics of these methods. However, what I'm seeking is context-specific guidance regarding the utilization of these methods for gene expression data and specifically, mRNA-miRNA correlation analysis. The closest thing I could find was a paper from Briefings in bioinformatics (https://doi.org/10.1093/bib/bbt051) but that study does not exactly differentiate between these three methods since all of them are only appropriate for the extraction of functional and monotonic dependencies and as it appears, all these methods can be successfully implemented for expression data. So my question concerns whether there is any specific reason for the prominent use of the Pearson correlation coefficient for mRNA-target correlation analysis. And since they are all applicable to gene expression data, is there any difference between the reliability of their results with the biology of the subject in consideration? Thanks again for your time and attention

[Kevin Blighe]

And since they are all applicable to gene expression data, is there any difference between the reliability of their results with the biology of the subject in consideration?

I guess that this is the crux of the matter because they are not all applicable to gene expression data, depending on the study design and nature of the data being used for the correlation. One would not use Pearson for a study on n=6, for example, or where one is correlating FPKMs or normalised counts. This, in turn --yes-- implies that many published results are implementing incorrect statistical tests.

If in doubt, use Spearman. For a powered study design and using expression data that follows a normal distribution, use Pearson.

I attended a 1 hour presentation on Pearson correlation one day in the Longwood Medical Area in Boston, and I wish that I could put you in touch with that presenter. They had grown up in an era when even Pearson correlation was done manually on paper.

[BioNovice247]

Thank you very much for your response Kevin. This is really interesting.

One would not use Pearson for a study on n=6, for example, or where one is correlating FPKMs or normalised counts.

For the sake of knowledge, can I ask what type of normalization you have in mind and kindly put me in the right direction as to why Pearson correlation is not appropriate in these situations? As far as I understand, normalization of RNA-seq data (using an approach such as DESeq2's VST or edgeR's TMM), in order to make appropriate sample-to-sample comparisons possible, is a prerequisite of correlation analysis. And considering that even in high-powered datasets, there are usually a significant number of genes that do not follow normal distribution even after such normalizations, overall these suggest that Pearson correlation is in fact useless when dealing with RNA expression data.