With all due respec to Gordon Smyth and other authors of limma which is a great package, in my experience the empirical Bayes moderated regression in ComBat (package sva) simply works better than the simple linear model implemented in limma (at least last time I checked limma).

Hi Michael,

In our data, I see a strong batch effect when I just do a straight up rlog transformation. However, remove the batch effects as you suggested (removeBatchEffect()) seems to work. Now, if I want to still look at differentially expressed features, should I just account for the batch effect by adding it as another factor in the design or should I somehow use the batch effect removed values? Personally, I think it will make sense to add the batch effect as a linear factor (design = ~Sample+Batch) and adjust the results for it instead of subtracting the batch effect. But honestly I am not sure.

I know you guys don't suggest using rlog/vst transformed values for differential expression analysis, but if I have to use the batch effect removed data (rlog transformed), then how could I do that?

I hope my question makes sense and isn't that stupid.

Thanks,

Praful

Hi, Michael --

Can you explain how what you are suggesting here relates to the use of the 'blind' argument in the rlog and varianceStabilizingTransformation functions?  Specifically, and supposing I have included a batch term in my DESeq2 model, if I do rlog(blind=T) followed by removeBatchEffect() to remove mean shifts, how would that differ from just doing rlog(blind=F)?  Would not the latter also remove mean shifts due to the batch effect?

Thanks for your help.

Bob