I think batch correction didn't working in my dataset(ft. RunHarmony)

Question

[kayah]

Hi all, now I'm analyzing scRNA seq. The problem is that I don't think batch correction work well. I want to see the aging difference between young and old group so I want to do batch correction in sex. but, I don;t think it work well. (I don't have any repeats in my dataset) Thank you so muuch!!

WAT <- merge(WAT_M_Y, y = c(WAT_M_O, WAT_F_Y, WAT_F_O), 
                 project = "WAT")
    [e-mail]$type <- c(rep("Young", ncol(WAT_M_Y)),
                            rep("Old", ncol(WAT_M_O)),
                            rep("Young", ncol(WAT_F_Y)),
                            rep("Old", ncol(WAT_F_O)))
    [e-mail]$sex <- c(rep("Male", ncol(WAT_M_Y)),
                           rep("Male", ncol(WAT_M_O)),
                           rep("Female", ncol(WAT_F_Y)),
                           rep("Female", ncol(WAT_F_O)))

    WAT$lowQC <- ifelse(WAT$nFeature_RNA>500 & 
                          WAT$percent.mt <10, "PASS", "FAIL")
    WAT<- subset(WAT, subset = lowQC == "PASS")
    WAT <- NormalizeData(WAT)
    WAT_variable  <- FindVariableFeatures(WAT, selection.method = "vst", nfeatures = 2000) 
    WAT_variable <- ScaleData(WAT_variable, vars.to.regress = c("percent.mt"))
    VariableFeaturePlot(object= WAT_variable)
    WAT_variable <- RunPCA(WAT_variable, features = VariableFeatures(WAT_variable), verbose = T) 
    #batch correction 
    library(harmony)
    library(Rcpp)
    WAT_variable2 <- RunHarmony(WAT_variable, group.by.vars = "sex")

Answer 1

[OmnibusX]

They do look aligned together. The difference might come from the difference in cell type composition between samples. One quick method I often use to assess batch correction efficiency is coloring your cell clusters by cell type labels and checking whether the same cell types are clustered together. If they are, it suggests that the batch correction has been effective in aligning the data appropriately across your batches. This visualization helps confirm that the batch effect isn't dominating the biological signal, which is crucial for your analysis of age-related differences. If you continue to see the same cell types split into distinct clusters, it may indicate that the batch correction hasn't fully addressed the underlying batch effects.

[kayah]

Based on the plots generated using your guidance, it seems that batch correction might not be necessary, or it has been effectively applied. Thank you for your kind and detailed response.

Answer 2

[BioinfGuru]

Using RLE plots alongside PCA plots is a more helpful assessment of the presence/removal of technical variation.

Great youtube explanation

Paper

Workflow