Why using unnormalized data for FindVariableFeatures ?

[nebetbastet]

Hi,

I switched quite recently (few months ago) on Seurat3 as I was used to Seurat2.

I use a "personnal" normalization (scran) that I integrated in a Seurat object. By mistake, the normalized data were added in two slots:
object@assays$RNA@data AND object@assays$RNA@counts
In other words, I lost raw data.

Until yesterday, I thought it was OK as I thought raw counts were not used again in the worflow.
It's only yesterday, when I decided to have a closer look on your paper on integration that I noticed that the "FindVariableFeatures" procedure was performed on raw data (at least with the default method "vst").

Unfortunately, the results are very different when I have raw counts or normalized counts in the "counts" slot : only 1/3 of genes are consistent...

I have two questions:

• Why this choice of un-normalized data to identify variable genes ? Maybe I am mistaken, but I think this is not explained in the paper or in your vignette
• If by mistake, you use normalized data in this procedure (as I did), which bias can you expect? In other words, why would it be "better" to use unnormalized data?

Thank you in advance for your answers!

[timoast]

This is explained in the section "Feature selection for individual datasets" in Stuart and Butler et al. 2019

The section shown here:

It is not valid to run the same procedure (selection.method="vst") on normalized count data

[duocang]

@timoast

So unnormalized data is for FindVariableFeatures. We assume it is under RNA assay.

Can I run SCTransform first and FindVariableFeatures later?

I get different variable feature if SCTransform done. The unnormalized data from SCT assay is the corrected read counts, does the corrected read counts work better than counts from RNA?