Seurat: Workflow for Reciprocal PCA integration with SCTransform normalization

[zrcjessica]

Hello,

I am trying to integrate ~20 scRNA-seq samples and want to use the rpca approach after normalizing each sample with SCTransform. However, this isn't explicitly demonstrated in the vignette for rpca, and I have a couple of questions:

1. When running Integrate_Data() after FindIntegrationAnchors(), should I specify normalization.method = "SCT"?
2. Do I need to replace the ScaleData() step with SCTransform()? Do I omit it altogether? Or do I keep it even though the individual samples were normalized with SCTransform?

Here is an example of how I am thinking of doing it, using the data from the Seurat rpca vignette:

bm40k.list <- lapply(X = bm40k.list, FUN = function(x) {
    x <- SCTransform(x)
})

features.list <- SelectIntegrationFeatures(object.list = bm40k.list)

bm40k.list <- lapply(X = bm40k.list, FUN = function(x) {
    x <- PCA(x, features = features)
}

anchors <- FindIntegrationAnchors(object.list = bm40k.list, reference = c(1, 2), reduction = "rpca", 
    dims = 1:30)
bm40k.integrated <- IntegrateData(anchorset = anchors, dims = 1:30, normalization.method = "SCT")

bm40k.integrated <- RunPCA(bm40k.integrated, verbose = FALSE)
bm40k.integrated <- RunUMAP(bm40k.integrated, dims = 1:30)

However, I'm not sure if this is the right way to use rpca with SCT. Could someone please clarify what the proper approach is?

Thanks!

[saisomesh2594]

Hi @zrcjessica ,

I, too, recently, performed the same integration workflow for 16 samples using SCT normalization with Reciprocal PCA integration

I followed the exact same steps as you, and in general, this seems like a proper approach to do so. To specifically answer your points,

1. I think you should mention normalization.method = SCT in the FindIntegrationAnchors() step as well along while doing so in IntegrateData()

2. I do not think it is recommended to redo SCTransform() after integration

I would be happy to get this confirmed by someone else as well.

Cheers,
Somesh

[zrcjessica]

I've found that when I run FindIntegrationAnchors() with normalization.method="SCT" in the code above, I get the following error:

Error in FindIntegrationAnchors(object.list = samples.list, reference = 5, : Please specify the anchor.features to be used. The expected workflow for integratinge assays produced by SCTransform is SelectIntegrationFeatures -> PrepSCTIntegration -> FindIntegrationAnchors.

1. FindIntegrationAnchors(object.list = samples.list, reference = 5, 
 .     normalization.method = "SCT", reduction = "rpca", dims = 1:30)
2. stop("Please specify the anchor.features to be used. The expected ", 
 .     "workflow for integratinge assays produced by SCTransform is ", 
 .     "SelectIntegrationFeatures -> PrepSCTIntegration -> ", "FindIntegrationAnchors.")

[saisomesh2594]

Hi @zrcjessica ,

I remember having the same error. I had to provide the features calculated in the SelectIntegrationFeatures as anchor.features in the FindIntegrationAnchors step. So, the command I used is as follows -

anchors <- FindIntegrationAnchors(object.list = seurat.list,
                                  reference=c(1,2,3,4),
                                  normalization.method="SCT",
                                  anchor.features=integration.features,
                                  verbose=T,
                                  reduction="rpca")

Cheers,
Somesh

[zrcjessica]

Did you also have to run PrepSCTIntegration() prior to running FindIntegrationAnchors()?

[saisomesh2594]

Yes, Precisely.

[jaisonj708]

The user responses are correct. You should specify normalization method in IntegrateData. There is no need to run ScaleData after SCTransform, nor should you re-run SCTransform after integration.