batch effects of RNAseq data from public databases - need to correct or not ?

Question

[CrazyB]

CCLE, TCGA, & GTEx have provided RNAseq data but there was very little information available regarding whether batch effects correction is needed when data from different datasets are combined or compared. I found only one publication that addresses the batch effect, as if batch effects per se is not a big issue. www.nature.com/articles/sdata201861

Q1. Across TCGA tumor types, is there a need to normalize all data before comparing the expression of gene of interest?

Q2. How big an issue is the batch effect between different databases, especially as they seem to be processed by slightly different pipelines (in addition to other technical factors during sample procurement and handling).

Any suggestions or references regarding the issue (or non-issue) of batch effects among large datasets contributed by multiple institutes

Answer 1

[Charles Warden]

You need to consider the possibilty of batch effects, but I would not expect any method to be able to completely correct confounding variables (without any over-correction, for example). So, I would recommend some sort of visualization before and after your adjustment.

Perhaps this answer may be of interest to you?

A: Are all RPKM/FPKMs in RNAseq "equal" ?

Answer 2

[cfos4698]

To be confident that any potential batch effect isn't affecting your work, you could always obtain the reads/raw count data and run a differential expression analysis yourself. You could model for any potential batch effect by including the batch in the differential expression model, e.g. design <- model.matrix(~batch+treatment). You could also detect, and model for, any batch effects using sva (see "Removing hidden batch effects" at https://f1000research.com/articles/4-1070/v2).