Batch effect correction of RNAseq data using scRNAseq batch effect correction tools (e.g., Harmony, Liger, Seurat3)

Question

[mickael.leclercq]

Hi,

I have several standard RNA-seq datasets where I observed strong batch effect. So I tried various new batch effect correction tools, but they are developed for single cell RNA-seq, such as Harmony, Liger and Seurat3. Do you see any objection to that approach? I mean, algorithmically speaking, I wouldn't see a problem, but a reviewer rised this question and I wonder how the community feels about that.

Thank you!

Answer 1

[rpolicastro]

Counts from scRNA-seq integration methods are only supposed to be used for dimension reduction and clustering of scRNA-seq data. The corrected counts themselves are invalid for e.g. differential expression or machine learning. To correct for batch effect in bulk RNA-seq you should include them in your regression model if possible, or a software tailored to bulk RNA-seq such as ComBat-seq like Nicolas mentioned.

Answer 2

[Nicolas Rosewick]

It would be preferable to use methods designed for bulk RNA-Seq as ComBat : https://github.com/zhangyuqing/ComBat-seq

I never tried to use scRNA-Seq methods for bulk rna-seq. Did you plot a PCA bi-plot to evaluate your batch effect ?

[mickael.leclercq]

I tried combat, compared to Harmony, it is hard to tell how well it is better than another. I integrated 4 databases, same experiments (we applied various conditions to a cell line), same protocole, same instrument. The only variation is the date of acquisition and the operator who did the experiment. As you can see in the picture (PCA and UMAP), there is no correction, combat correction, and harmony correction. For some conditions, I observed a good integration, but for other, not at all... Sometimes it worked better with harmony, sometimes better with Combat. It is such a mess to get something generalizable !!

[ATpoint]

What is the aim of all this?

[Radu Tanasa]

I believe you're correct. There is no "best" method. For single cell, a comparative analysis yielded scanorama better in some cases, while scVI was better in others.