DESeq2/RNA-Seq: Cross-study comparisons with large control-variance?

Question

[Felix]

I have created a cross-study (20, >250 samples in total) RNA-seq expression matrix of the same cell type (different conditions) and planned to do a combined DE analysis - in terms of checking for similarities of DE across different comparisons. By 'pooling' all controls, defining condition + series as design to account for the batch effect, I have then plotted the PCA to get an idea of the data.

I then removed the batch effect using limma for visualization purposes (as suggested here: https://support.bioconductor.org/p/76099/#93176) Here, the DESeq2 transformed PCA without and with removal of the 'study'/batch effect (color=study):

Since the variation of 'control' samples across the different studies is still very large after removing the 'series' batch effect, I would like to do WT vs. condition comparisons for every study separately - is that a reasonable approach and doable with DESeq2? Most of the studies do have replicates.

Thank you very much for your input!

Answer 1

[Kevin Blighe]

I would proceed [with caution] by including batch as a covariate in your design formula. Then, when you derive test statistics for your conditions of interest, an adjustment will be made for the effect of batch. So, no modification or use of removeBatchEffects() for this particular part. If you later want to use the data for other downstream functions, like heatmaps, etc., then use removeBatchEffects() on the vst expression levels.

Kevin