Batch effect correction

Question

[jtull]

Hi,

I am interested in human validation of a set of genes identified as differentially expressed between 2 groups in mice and for that, I am using some control samples from a publication. I've read that to remove batch effects with DESeq2, introducing the batch in the design of the experiment as design = ~ batch + condition is recommended. However, in my case, the 2 designs are going to be exactly the same. What can I do? Do you suggest using sva (surrogate variable analysis)? Many thanks!

Best,
jtull

[swbarnes2]

Just to be clear...are all your controls from one batch?

[jtull]

Yes, they are from the same batch.

Answer 1

[swbarnes2]

You cannot magically remove batch effect while keeping condition effects if they are totally confounded.

[jtull]

I see, thanks. Any suggestions though, e.g. sva package? Also, when I use design = ~ condition, I still get some of the genes from the initial list as differentially expressed. Do you think the analysis still has some meaning to it or we cannot do the DE analysis at all?

[ATpoint]

As said, perfect confounding cannot be magically removed. sva and any other package obeys the rules of linear algebra. Answer is unfortunately no. I mean you can perform DE without batch correction and take the results for hypothesis generation, but be aware that this is not reliable. If you then find certain e.g. pathways enriched that are promising see whether you can confirm it with other datasets or experiments. Treat results with care, try to confirm findings and do not make strong claims on the results unless properly validated. The data are not fully useless as long as you treat them with the necessary care. Not more you can do I would say. Without validation it could well be that you chase ghosts though.

[jtull]

Thank you very much.