Strategy for reclustering an Integrated object

[saeedfc]

Dear Seurat Team,

After integration, I can either subset and run the UMAP/tSNE and Findneighbours and Findclusters functions with integrated assay.(So Seurat will use the subset of the integrated matrix it created with all the cells).

SO <- subset(Int.SO, idents = c(1,2,3,5,6,8,13,14,15))
DefaultAssay(SO) <- "integrated"
SO <- RunPCA(SO, npcs = 100, verbose = FALSE, assay = "integrated")
SO <- RunUMAP(SO, dims = 1:20, assay = "integrated", umap.method = 'umap-learn', metric = 'correlation', n.neighbors = 10L, min.dist = 0.5)
SO <- FindNeighbors(SO, dims = 1:40, prune.SNN = 1/15,k.param = 20, force.recalc = T)
SO <- FindClusters(SO, resolution = 0.65)
UMAPPlot(SO)

DefaultAssay(SO) <- "RNA"
SO <- NormalizeData(SO)

Alternatively, I can select the cell names that I need for reclustering from the original integrated object and then start fresh by creating new seurat objects with only these selected cell names and integrate freshly.(To create a new integrated matrix based on the cells present).And then do UMAP/tsne and clustering.

cell_list <- WhichCells(Int.SO, idents = c(1,2,3,5,6,8,13,14,15))
SO1 <- CreateSeuratObject(counts = X1, project = "WT", min.cells = 3, min.features = 200)
SO1 <- subset(SO1, cells = intersect(Cells(SO1), cell_list))
SO2 <- CreateSeuratObject(counts = X1, project = "KO", min.cells = 3, min.features = 200)
SO2 <- subset(SO2, cells = intersect(Cells(SO2), cell_list))
POI.sub <- list(SO1,SO2)

for (i in 1:length(POI.sub)) {
    POI.sub[[i]] <- SCTransform(POI.sub[[i]], vars.to.regress = "percent.mt",verbose = FALSE)
}

POI.features <- SelectIntegrationFeatures(object.list = POI.sub, nfeatures = 2000)

POI.sub <- PrepSCTIntegration(object.list = POI.sub, anchor.features = POI.features, 
    verbose = FALSE)
POI.anchors <- FindIntegrationAnchors(object.list = POI.sub, normalization.method = "SCT", 
    anchor.features = POI.features, verbose = FALSE)
SO <- IntegrateData(anchorset = POI.anchors, normalization.method = "SCT", 
    verbose = FALSE)

DefaultAssay(SO) <- "integrated"
SO <- RunPCA(SO, npcs = 100, verbose = FALSE, assay = "integrated")
SO <- FindNeighbors(SO, dims = 1:30)
SO <- FindClusters(SO, resolution = 0.3,graph.name = 'integrated_snn')

SO <-  RunUMAP(SO, min.dist = 0.5, assay = "integrated", umap.method = 'umap-learn', metric = 'correlation', n.neighbors = 20L,spread = 0.7, graph = 'integrated_snn')

DefaultAssay(SO) <- "RNA"
SO <- NormalizeData(SO)

Is either of these strategies 'wrong'?

[timoast]

Neither of these are "wrong", but I don't think there is any advantage to the second approach