scRNAseq normalization

Question

[sarahmanderni]

Hi,

I am wondering despite all the advantages SCT has over the conventional log-normalization method, is it necessary to use it with all datasets? I have 5 samples from mouse and the data is of high quality and small batch effect. I tested clustering with log-normalization and SCT (in seurat and RPCA integration) and actually the results from log-normalization looks more reasonable.

[Ming Tommy Tang]

To be honest, I really do not use SCT, and use log normalization all the time...

Answer 1

[dsull]

No, your log-transform might indeed be ok (or even better) for your analysis of interest.

See https://www.nature.com/articles/s41592-023-01814-1

Answer 2

[ATpoint]

It's perfectly fine to use log-transformation. The advantages of these more complicated methods usually depend on context and also on the way people benchmark them. log-transformation is faster, perfectly reproducible and scales well to any dataset size. I always use it by default.

[sarahmanderni]

Thanks for the response! I also was wondering to ask if you have found the log-norm also working fine in terms of downstream analysis like differential expression analysis?

[ATpoint]

That is an entirely different analysis and depends on the testing Framework. If you use the commonly-used Wilcox test or something like limma-trend then yes, data on log2-scale could be used. If you use some (G)LM framework such as DESeq2 or limma-voom then it expects raw counts.