Hi all,

I am working on a large dataset consisting of multiple different RNAseq and Microarray studies from different labs and times. While we have a (functional) pipeline setup for this, I recently saw a post which mentioned using voomLmFit to counter the issues which might stem from having excess zeroes in the data.

Since we are combining RNAseq data and microarray data for a combined analysis, we see some of these data-sparsity issues; a number of the genes in the RNAseq dataset are simply not found in the microarray datasets, and not all of our microarray datasets share a full geneset either. Ideally, we would like not to simply remove the partially sparse genes from the dataset, since doing that would drastically reduce the amount of genes available for further analyses.

My question is therefore whether a voomLmFit pipeline could be used for both the RNAseq and microarray data? I.e. is voom transformation of microarray data harmful, and if so is there another way to account for these data-sparsity issues without having to cut down our genesets drastically?

Thanks, Adam

edit: For some more context, we are not combining samples across studies into any single groups. Rather, we want to perform a group-wise comparison (with the original groups from each study), contrasting the changes between conditions. We have no repeats of condition comparisons across studies.

No, that's not a good idea at all. You should be doing a meta-analysis with these data rather than trying to combine like that. GeneMeta is one package you could use, but there are others that you can search for on the BioC website, like metapod.

The latter is a bit tricky if you use Stouffer's method because it's supposed to be based on one tailed p-values, and there's no obvious way to implement that in a package, so you have to know that, and use your own code to accommodate. In other words, you have to convert all your p-values to one tailed, do the test, and then convert back.

I agree with James, although I tend to combine analyses using geneset tests rather than by meta analysis tools like GeneMeta.

My question is therefore whether a voomLmFit pipeline could be used for both the RNAseq and microarray data?

No, voomLmFit() is only for count data such as RNA-seq. If you want to undertake a voom-like analysis of microarray data, then limma provides the function voomaLmFit() for microarrays. Generally, however, observation level weights are not so important for microarray as for RNA-seq, and you would usually just use the standard limma pipelines for microarray data.

You should be processing each technology (RNA-seq and each different type of microarray) according to their own characteristics. You cannot combine different technologies into a combined limma analysis, because the different technologies have different probe sets, different mean-variance relationships and different biases.

is voom transformation of microarray data harmful

Yes it is. A voom analysis of microarray data is basically nonsense because it is assuming characteristics that don't exist for microarrays. Just to mention one thing, voom is assuming that counts need to be normalized by library size, but microarrays don't have library sizes.

is there another way to account for these data-sparsity issues without having to cut down our genesets drastically?

Even if you could run voomLmFit on microarray data, it would do nothing to change the fact that different technologies measure different genes and isoforms. Software normalization cannot make a platform appear to measure something it doesn't.

My practice has been to analyse each technology dataset separately keeping all possible probes in the analysis, and then correlate results between datasets using geneset tests or similar.