BATCH EFFECT REMOVAL

Question

[Rishabh Jha]

Hello everyone, I have been working on expression analysis past few months. Lately, I hadn't been asked to address batch effects from my data. I have two datasets on which I had performed a meta-analysis. I had certain questions regarding batch effects and removal of batch effects removal from my data.

Q - 1. I had read that the batch effect removal algorithm is required when your approach is "Data Merging" and not "Meta-analysis". I hadn't merged my data - Just performed limma analysis and then check for common genes and finally used the pval combination method. Finally, find out DEGS. I hadn't found papers that address batch effect removal in such a strategy. Please help.

Q. 2 - Secondly if I performed sva on the individual dataset, perform limma analysis and then use the pVal combination method to combine the dataset. Is it possible?

Q -3.How to create a model and null matrix and finally how to get the batch effect adjusted expression matrix.

Kindly help. Thanks in Advance.

Answer 1

[Kevin Blighe]

Q - 1. I had read that the batch effect removal algorithm is required when your approach is "Data Merging" and not "Meta-analysis". I hadn't merged my data - Just performed limma analysis and then check for common genes and finally used the pval combination method. Finally, find out DEGS. I hadn't found papers that address batch effect removal in such a strategy. Please help.

In this case, if you are just deriving differentially expressed genes, just include batch in the design formula for limma. In this way, the effect of batch will be 'adjusted for' when deriving test statistics for your main outcome / condition of interest.

Q. 2 - Secondly if I performed sva on the individual dataset, perform limma analysis and then use the pVal combination method to combine the dataset. Is it possible?

You can use SVA to identify 'extraneous' (unknown) batch effects in the data and then include these surrogate variables in the design formula, similar to my comment above. Be careful about 'wiping out' the actual effect of your main outcome / condition, though. Some people are going crazy about batch effect removal and I feel that they end up doing more harm than good [to their data].

Q -3.How to create a model and null matrix and finally how to get the batch effect adjusted expression matrix.

The design would be of form: ~ condition + batch or ~ condition + sv1 + sv2 + ... svn

To actually remove the batch effect for downstream analyses like PCA, clustering, etc., use limma::removeBatchEffect()

Also, if it helps: Batch effects : ComBat or removebatcheffects (limma package) ?

---------------

You will have more questions, but please do read more answers / threads in order to learn more.

Kevin