scRNA-seq and the batch effects

Question

[Bogdan]

Dear all,

considering some of the recent methods for BATCH-CORRECTION / DATA INTEGRATION in scRNA-seq :

https://satijalab.org/seurat/vignettes.html#seurat-wrappers

LIGER, MNN, HARMONY, ZINBWAVE, and CONOS ;

i was wondering if you have any preference/ suggestions. thank you !

-- bogdan

[shoujun.gu]

I have tried couple methods. The biggest problem I found is, since you have no idea what's the real data suppose to be, there is no way you can really measure the accuracy of these methods. If I have to pick one to use, I will choose MNN or Scanorama (but most of the time, I prefer not to torture the data).

[jared.andrews07]

Agreed - if you can avoid batch correction (or have no evidence that it's occurring), you should definitely avoid it.

[igor]

I think you already asked this previously: about batch correction in scRNA-seq

[Bogdan]

yes, thank you Igor.

I was wondering what the experience of the people was ... , or if there are any other suggestions.

Answer 1

[Charles Warden]

I am not sure if all batch effects can be corrected, but these are what I can think of:

• total number of actual cells (with or without additional filtering)
• index hopping (from other libraries, which may or may not be scRNA-Seq)

For the 2nd one, I think this is usually not a major issue (unless you look for events that occur at <5% frequency)

[Bogdan]

Thanks Charles. i was wondering if you have any recommendation/preference regarding LIGER, MNN, HARMONY, ZINBWAVE, and CONOS ; thanks !

[Charles Warden]

I don't think I have tried those methods. It has been a little while since I worked on a scRNA-Seq dataset, but those were the main things that I could think of.

Answer 2

[jared.andrews07]

I have used the Seurat wrapper around fastMNN and can recommend it. It is generally less heavy-handed than Seurat's normal integration method and explains the amount of variance lost between each batch, which is a useful check to ensure only a small amount of (presumably technical) variation is lost at each step.

I can't speak to the other methods, as I haven't used them.