PrepSCTFindMarkers skips re-correction after IntegrateLayers(HarmonyIntegration) with SCT v2, requiring recorrect_umi=FALSE - is this expected behavior?

[louhao1129]

Hi Seurat team,
I am encountering a situation where PrepSCTFindMarkers() skips re-correction, but the subsequent FindMarkers() call requires recorrect_umi=FALSE to work. I want to confirm if this is the expected behavior for Seurat v5 with SCTransform v2 and Harmony integration, or if I am missing a step in my workflow.
My Workflow
I am following the Seurat v5 workflow for integrating multiple samples using SCTransform(vst.flavor="v2") followed by HarmonyIntegration:

# Step 1: Merge multiple samples into one object with layers
scRNA<- merge(x = seurat_list[[1]], y = seurat_list[-1], 
                           add.cell.ids = sapply(seurat_list, function(x) unique(x$orig.ident)))

# Step 2: SCTransform v2 (this automatically handles each layer)
scRNA<- SCTransform(scRNA, vars.to.regress = "percent.mt")

# Step 3: Run PCA
scRNA<- RunPCA(scRNA)

# Step 4: Integrate using Harmony via IntegrateLayers
scRNA<- IntegrateLayers(
  object = scRNA, 
  method = HarmonyIntegration,
  normalization.method = "SCT",
  orig.reduction = "pca", 
  new.reduction = "harmony"
)

# Step 5: Clustering and visualization
scRNA<- FindNeighbors(scRNA, reduction = "harmony", dims = 1:30)
scRNA<- FindClusters(scRNA, resolution = seq(0.1, 1, 0.1))
scRNA<- RunUMAP(scRNA, reduction = "harmony", dims = 1:30)

# Step 6: scDblFinder

scRNA= JoinLayers(scRNA, assay = "RNA")
counts_matrix <- LayerData(scRNA, assay = "RNA", layer = "counts")

sce <- SingleCellExperiment(
  assays = list(counts = counts_matrix),
  colData = scRNA@meta.data
)

sce <- scDblFinder(
  sce, 
  samples = "orig.ident",  # 按样本分别处理
)

scRNA$scDblFinder.score <- sce$scDblFinder.score
scRNA$scDblFinder.class <- sce$scDblFinder.class

scRNA_filtered <- subset(scRNA, subset = scDblFinder.class == "singlet")

# Step 6: Prep for DE analysis
Idents(scRNA_filtered) = "SCT_snn_res.0.2"
scRNA_filtered<- PrepSCTFindMarkers(scRNA_filtered)
# Output: "Minimum UMI unchanged. Skipping re-correction."

# Step 7: Find markers
# This works only when recorrect_umi=FALSE:
markers <- FindAllMarkers(scRNA_filtered, recorrect_umi = FALSE)

The Issue/Question
Contradictory messages: PrepSCTFindMarkers says "Skipping re-correction" (implying it's already prepared), but if I omit recorrect_umi=FALSE in FindMarkers, I get the warning:

Warning: Object contains multiple models with unequal library sizes.
Run PrepSCTFindMarkers() before running FindMarkers().
My current solution: Setting recorrect_umi = FALSE in FindMarkers() allows the analysis to run without errors.
My concern: Is using recorrect_umi = FALSE the correct and valid approach here? Specifically:
Does IntegrateLayers(method = HarmonyIntegration) preserve the SCT v2-corrected Pearson residuals such that recorrect_umi = FALSE is the intended usage?
Or should PrepSCTFindMarkers be actually doing something (not skipping) when multiple layers exist post-Harmony?
Is there a risk of incorrect p-values/log2FC when using recorrect_umi = FALSE on a Harmony-integrated object that still contains multiple SCT models (one per original layer)?

R 4.5.1
> packageVersion("Seurat")
[1] ‘5.4.0’
> packageVersion("SeuratObject")
[1] ‘5.3.0’
> packageVersion("sctransform")
[1] ‘0.4.2’
> packageVersion("harmony")
[1] ‘1.2.3’

[igrabski]

Hi, thanks for your question! You should be able to avoid this issue by running PrepSCTFindMarkers prior to subsetting your object (i.e. prior to this step: scRNA_filtered <- subset(scRNA, subset = scDblFinder.class == "singlet")). Please let me know if you still run into problems!