I am integrating 3 samples. Each of the samples have been normalized with SCTransform and they all look good. I am having some issues when I am integrating the samples. I have tried 2 different approaches:
When I look at the clustering and UMAP with both approaches they are very different. The SCT UMAP has a lot of small clusters and the integration without SCT looks more realistic. My main question is, is it ok to integrate without SCT when the samples have been normalized individually using SCTransform?
saketkc
commented
2 years ago
Hello,
I am integrating 3 samples. Each of the samples have been normalized with SCTransform and they all look good. I am having some issues when I am integrating the samples. I have tried 2 different approaches:
1) Integration with SCT
features <- SelectIntegrationFeatures(object.list = ntd.list, nfeatures = 3000) NTD.integrated <- PrepSCTIntegration(object.list = ntd.list, anchor.features = features) anchors <- FindIntegrationAnchors( object.list = NTD.integrated, object.list = ntd.list, normalization.method = "SCT", anchor.features = features ) NTD.integrated <- IntegrateData(anchorset = anchors, normalization.method = "SCT" )
NTD.integrated <- RunPCA(NTD.integrated) NTD.integrated <- RunUMAP(NTD.integrated, reduction = "pca", dims = 1:30) NTD.integrated <- FindNeighbors(NTD.integrated, reduction = "pca", dims = 1:30) NTD.integrated <- FindClusters(NTD.integrated, resolution = c(0.5))
2) Integration without SCT
features <- SelectIntegrationFeatures(object.list = ntd.list, nfeatures = 2000) anchors <- FindIntegrationAnchors( object.list = NTD.integrated, object.list = ntd.list, anchor.features = features ) NTD.integrated <- IntegrateData(anchorset = anchors )
NTD.integrated <- ScaleData(NTD.integrated, verbose = FALSE) NTD.integrated <- RunPCA(NTD.integrated) NTD.integrated <- RunUMAP(NTD.integrated, reduction = "pca", dims = 1:30) NTD.integrated <- FindNeighbors(NTD.integrated, reduction = "pca", dims = 1:30) NTD.integrated <- FindClusters(NTD.integrated, resolution = c(0.5))
When I look at the clustering and UMAP with both approaches they are very different. The SCT UMAP has a lot of small clusters and the integration without SCT looks more realistic. My main question is, is it ok to integrate without SCT when the samples have been normalized individually using SCTransform?
I have attached the UMAPs below for reference.
Thanks in advance!
Carly
MoA-UMAP-NTD-Normalized.pdf MoA-UMAP-NTD-SCTInt.pdf