To integrate or not to integrate?

[beletskayo]

Hi hi,

I am trying to run an analysis on 5 publicly available single-cell datasets, spanning 3 different conditions (I have at least 2 different datasets for each condition). Cell type and condition are already annotated in each dataset. Because of this, and since ultimately, my goal is to run differential expression analyses using a pre-defined list of genes (which shouldn't be done on the integrated assay anyway), I am wondering if there is any use in running the integration workflow at all in this case? To reduce the batch effect, I used SCTransform on the merged object, here’s my rough workflow:

1. merge all 5 objects into merged.seurat object
2. SCTransform + standard pre-processing:

merged.seurat <- JoinLayers(merged.seurat)
merged.seurat.sct <- SCTransform(merged.seurat, conserve.memory = TRUE,  vst.flavor = "v2",
                                 vars.to.regress = c("percent.mito", "nFeature_RNA", "nCount_RNA",  "author"))
DefaultAssay(merged.seurat.sct) <- "SCT"  
merged.seurat.sct <- FindVariableFeatures(merged.seurat.sct)
merged.seurat.sct <- RunPCA(merged.seurat.sct)
merged.seurat.sct <- RunUMAP(merged.seurat.sct, dims = 1:30, reduction = "pca", reduction.name = "umap.sct")

3. Transfer cell type labels using one of the datasets as reference with FindTransferAnchors() and `TransferData()
4. DE analysis with FindMarkers() using the SCT assay a) across the entire dataset, b) in each cell type (subset Seurat Object + FindMarkers())

Does this workflow make sense, should any steps be re-ordered, and do I need to include integration at any point?

Thank you!

Sera