Workflow for Multi-sample SCTransformed Sketch-based Integration

[TBThien25]

Hi Seurat team and community, thank you all for your contributions in science.

I’m currently integrating multiple samples using a sketch-based approach on SCTransformed data, because running full-dataset integration per sample is exceeding our computational capacity. The standard sketch-integration tutorial (1 million cell dataset from Parse Biosciences) does something like:

object <- NormalizeData(object)
object[["RNA"]] <- split(object[["RNA"]], f = object$sample)
object <- FindVariableFeatures(object, verbose = FALSE)
object <- SketchData(object = object, ncells = 5000, method = "LeverageScore", sketched.assay = "sketch")
DefaultAssay(object) <- "sketch"
object <- FindVariableFeatures(object, verbose = F)
object <- ScaleData(object, verbose = F)
object <- RunPCA(object, verbose = F)
object <- IntegrateLayers(object, ...)

However, when I try to adapt this to SCTransform, I can't re-apply variable-feature selection or re-scaling on the downsampled SCT assay. To work around this, I've devised two approaches:

Approach 1: Sketch after SCTransform, then skip re-selection & scaling

# Full SCTransform on each sample
object[["RNA"]] <- split(object[["RNA"]], f = object$sample)
object <- SCTransform(object)

# Sketch on the SCT assay
object <- SketchData(object = object, assay = "SCT", sketched.assay = "sketch", ncells = 3000*length(unique(object$SampleID)), method = "LeverageScore")

# Proceed with PCA and integration
DefaultAssay(object) <- "sketch"
object <- RunPCA(object, reduction.name = "sketch_PCA", features = VariableFeatures(object))
object <- IntegrateLayers(object, orig.reduction = "sketch_PCA", new.reduction = "sketch_RPCA", method = RPCAIntegration, normalization.method = "SCT")

# Project back to full SCT assay
object <- ProjectIntegration(object = object, sketched.assay = "sketch", reduction = "sketch_RPCA", assay = "SCT", reduction.name = "SCT_RPCA")

Approach 2: Re-SCTransform on the downsampled RNA assay

object[["RNA"]] <- split(object[["RNA"]], f = object$sample)
object <- SCTransform(object)

# Sketch on SCT assay
object <- SketchData(object = object, assay = "SCT", sketched.assay = "sketch", ncells = 3000*length(unique(object$SampleID)), method = "LeverageScore")

# Subset the original RNA assay to the sketched cells
object_sub <- subset(object, cells = colnames(object@assays$sketch))

# Re-apply SCTransform on RNA to create a new "sketch" assay
object_sub <- SCTransform(object_sub, assayIn = "RNA", assayNew = "sketch")

# Replace the sketch assay in the main object
object@assays$sketch <- object_sub@assays$sketch

# Proceed with PCA and integration as above
DefaultAssay(object) <- "sketch"
object <- RunPCA(object, reduction.name = "sketch_PCA", features = VariableFeatures(object))
object <- IntegrateLayers(object, orig.reduction = "sketch_PCA", new.reduction = "sketch_RPCA", method = RPCAIntegration, normalization.method = "SCT")
object <- ProjectIntegration(object = object, sketched.assay = "sketch", reduction = "sketch_RPCA", assay = "SCT", reduction.name = "SCT_RPCA")

Question:
Which approach best follows Seurat's intended workflow for sketch-based integration on SCTransformed data? Or is there a more canonical method I'm missing?

Thank you for your guidance!
Thien