combat correction or include batch effect in the model?

Question

[Inbalor]

I am performing differential expression analysis using RNA-seq data. I am using limma package.

I am wondering to know how I can bias my differential expression analysis if previous to differential expression test I use Combat to correct for batches effects although there is no clear evidence of batch effect when I plot data into a PCA/heatmap... However, I know that this batch (library prep) can influence my expression level.

Would it be a better approach to include the batch in the model?

Thanks for your help!

Answer 1

[Irsan]

If you cannot find any apparant batch effects I would not worry about it too much. I question though that PCA analysis is most powerful/powerful enough to identify those batch effects. You would only see batch effects in PCA analysis clearly if the batch effects are the primary source of variation in your samples. I hope that in your case the biological conditions have a way bigger effect. I think you could better use ANOVA to quantify the relative influence of batch relative to the other factors of your study. Then (or anyway), if you are still uncomfortable about it or if reviewers/your PI/anyone force you to correct for batches, then just use batch as a covariate in your analyses. If there really is no batch effect than the differences between including batch as covartiate or not are minimal anyway.

[Inbalor]

Thanks Irsan for your comment. Do you know if I can calculate the relative influence of batch relative to the other factors of my study using lmFit() from limma?

The thing is that when I do the test without batch effect (only consider my biological groups) then I hardly obtained P-values significant after correction by multiple test. However, when I do the batch correction by comBat, then I obtain a lot of significant genes after correction by multiple test.