Comparing microRNA expression at different tumour stages and identifying trends in miRNA expression I'm currently working on a university project using publicly available data. I'm collecting genome data on cancer patients at different pathological stages (Stage I, II, III and IV).
I'm aiming to compare microRNA expression at each of these stages in order to identify potential biomarkers for disease states, and I hope to identify any trends in the upregulation or downregulation of certain microRNA.
What would be the best way to go about this - would an ANOVA test work in order to analyse the expression across the groups? And for trends, would linear regression be the way to go about this?
Thank you so much
 A: In principle, ANOVA-type comparisons are OK. In practice with gene-expression data, you typically don't care about the overall significance of differences among groups. Usually there are specific comparisons that are of most interest. You can focus your attention on those specific comparisons without a preliminary ANOVA. For example, in your case you might want only to evaluate which miRNA expression levels increase with each successive increase in Stage (II vs I, III vs II, IV vs III).
If you are looking at trends over only 4 Stage values, linear regression probably isn't best. That assumes an equal increment in expression for each additional Stage, which isn't likely to hold. Treat Stage as a 4-level factor.
With RNA expression values you need to take special care. There are several issues to address, such as the variance of expression measures differing as a function of expression level and the multiple comparisons problem when you do hundreds or thousands of comparisons across Stages on individual RNA species.
Happily, there are well-established tools that help do all this. I'd recommend the limma package in Bioconductor. This article nicely outlines how to proceed. You have to set up design and contrast matrices that represent the structure of your study and the particular comparisons you want to evaluate, as explained here. If all you are modeling is the association of miRNA expression with Stage, that will be straightforward. The design matrix will be based on the four levels of Stage, and the contrast matrices will represent the particular comparisons between Stage levels that are of interest.
