Question about FindVariableFeatures

[Vivianstats]

Hello Seurat developers,

I would like to know for the three options (disp, vst, and mvp) in the FindVariableFeatures function, whether gene expression mean and standard deviation are calculated using raw counts, or normalized and log-transformed counts? This is not clear in the function documentation, so it would be really helpful if you can clarify.

Thank you!

[Dragonmasterx87]

Hi @Vivianstats from the FindVariableFeatures function RDocumentation entry here it says:

"......All mean/variance calculations are not performed in log-space, but the results are reported in log-space - see relevant functions for exact details."

Also please look at Jared's response here.

Taken together, it appears that all calculations are performed on raw counts.

I'm not a Seurat dev, please also wait for a valid response from someone from the satija lab.

🐲

[Vivianstats]

Hi Dragonmasterx87,

Thank you for sharing the information! It seems clear that the calculations are performed before log-transformation, but I still wonder whether they are based on raw counts or library-size normalized ones.

[rschauner]

@Vivanstats, also not a Seurat dev, but thought I could get you your answer. (TL;DR: it's the counts by default)

One helpful thing besides ?FindVaraibleFeatures is Viewing the functions themselves by just running View(FindVariableFeatures), however, this yields slightly unhelpful output:

function (object, ...) 
{
  UseMethod(generic = "FindVariableFeatures", object = object)
}

So then taking a look I figured out the best method to pull will be the method for the Assay class in Seurat. So using View( getS3method("FindVariableFeatures", "Assay")) I found this in the function call:

  if (selection.method == "vst") {
    data <- GetAssayData(object = object, slot = "counts")
    if (IsMatrixEmpty(x = data)) {
      warning("selection.method set to 'vst' but count slot is empty; will use data slot instead")
      data <- GetAssayData(object = object, slot = "data")
    }
  }

So unless counts is empty, FindVariableFeatures are based on counts, not normalized data in the data slot. Hope this helps!

[saketkc]

If you are using the SCTransform based workflow, the variable features are set to ones with the highest pearson residuals. The default FindVariableFeatures on RNA uses the raw counts to choose the variable features (using the mean-variance relationship).