SCTransform

Purpose

SCTransform() is an alternative preprocessing workflow in Seurat.

It is not an extra step after the classic workflow. It replaces the usual sequence of normalization, variable feature selection, and scaling.

Classic workflow:

NormalizeData() -> FindVariableFeatures() -> ScaleData() -> RunPCA()

SCT workflow:

SCTransform() -> RunPCA()

What SCTransform Does

SCTransform performs model-based normalization and variance stabilization.

In practice, it handles:

• normalization
• variance stabilization
• variable feature selection
• scaled residuals for downstream PCA
• optional regression of unwanted variables

It creates a new assay, usually named SCT.

Run SCTransform

Basic SCT workflow:

seed.use <- 1234

seu <- Seurat::SCTransform(
  object = seu,              # Seurat object
  assay = "RNA",             # input assay
  new.assay.name = "SCT",    # output assay
  vst.flavor = "v2",          # SCT v2 regularization
  method = "glmGamPoi",       # faster model fitting; requires glmGamPoi
  ncells = 5000,              # cells sampled for model fitting
  seed.use = seed.use,        # reproducible subsampling
  vars.to.regress = NULL,    # no regression by default
  variable.features.n = 3000, # number of variable features
  verbose = TRUE             # show progress messages
)

method = "glmGamPoi" is commonly used with vst.flavor = "v2" to speed up SCTransform. It requires the glmGamPoi package.

ncells = 5000 is the default number of cells sampled to fit the SCT model. seed.use makes this subsampling reproducible.

If mitochondrial signal should be regressed:

seu <- Seurat::SCTransform(
  object = seu,
  assay = "RNA",
  new.assay.name = "SCT",
  vst.flavor = "v2",
  method = "glmGamPoi",
  ncells = 5000,
  seed.use = seed.use,
  vars.to.regress = "percent.mt",
  variable.features.n = 3000,
  verbose = TRUE
)

Do not regress variables blindly. percent.mt, cell cycle scores, or other metadata columns may reflect biological differences in some datasets.

Multi-Sample SCT

For formal multi-sample analysis, run SCTransform() separately for each sample or library.

Do not treat a merged multi-sample object as the default input for one shared SCT model.

SCT estimates a model of technical variation related to sequencing depth. Different samples or libraries can have different depth, capture efficiency, chemistry, batch structure, or cell composition. Fitting one model across all samples assumes those technical structures are shared, which is often not a good assumption.

Recommended pattern:

seed.use <- 1234

seu.list <- Seurat::SplitObject(
  object = seu,
  split.by = "sample"
)

seu.list <- lapply(
  X = seu.list,
  FUN = function(x) {
    Seurat::SCTransform(
      object = x,
      assay = "RNA",
      new.assay.name = "SCT",
      vst.flavor = "v2",
      method = "glmGamPoi",
      ncells = min(5000, ncol(x)),
      seed.use = seed.use,
      variable.features.n = 3000,
      verbose = FALSE
    )
  }
)

Here ncells is computed inside each sample object. It means each sample uses up to 5000 cells for model fitting.

After this step, seu.list is still a list of Seurat objects. Each object has its own SCT assay and its own SCT model.

After Per-Sample SCT

Per-sample SCT does not directly return one combined Seurat object.

The result is:

class(seu.list)
length(seu.list)
Seurat::Assays(seu.list[[1]])

To continue downstream analysis, choose one route explicitly.

SCT Integration

If batch correction or sample integration is needed, keep seu.list as a list and run the SCT integration workflow.

The common anchor-based route uses CCA by default or explicitly with reduction = "cca":

features <- Seurat::SelectIntegrationFeatures(
  object.list = seu.list,
  nfeatures = 3000
)

seu.list <- Seurat::PrepSCTIntegration(
  object.list = seu.list,
  anchor.features = features
)

anchors <- Seurat::FindIntegrationAnchors(
  object.list = seu.list,
  normalization.method = "SCT",
  anchor.features = features,
  reduction = "cca",
  dims = 1:30,
  k.anchor = 5,
  k.filter = 200,
  k.score = 30
)

seu <- Seurat::IntegrateData(
  anchorset = anchors,
  normalization.method = "SCT",
  dims = 1:30
)

Seurat::DefaultAssay(seu) <- "integrated"

seu <- Seurat::RunPCA(
  object = seu,
  assay = "integrated",
  npcs = 50,
  seed.use = 42,
  verbose = TRUE
)

k.anchor, k.filter, and k.score control anchor selection, filtering, and scoring. The values above are Seurat defaults.

After IntegrateData(), seu is a single Seurat object again. Set the default assay to integrated and run PCA on the integrated assay before graph construction, clustering, and UMAP.

Harmony Integration

Harmony is another common route after per-sample SCT.

Unlike anchor-based SCT integration, Harmony works on a single merged Seurat object and corrects the PCA embedding. It returns an integrated reduction, usually named harmony, rather than an integrated assay.

If there are multiple samples, first choose shared variable features and merge the SCT-normalized objects:

features <- Seurat::SelectIntegrationFeatures(
  object.list = seu.list,
  nfeatures = 3000,
  assay = rep("SCT", length(seu.list))
)

seu <- merge(
  x = seu.list[[1]],
  y = seu.list[-1],
  merge.data = TRUE
)

Seurat::DefaultAssay(seu) <- "SCT"
Seurat::VariableFeatures(seu, assay = "SCT") <- features

Then run PCA on the merged SCT assay and correct the PCA space with Harmony:

seu <- Seurat::RunPCA(
  object = seu,
  assay = "SCT",
  npcs = 50,
  verbose = TRUE
)

seu <- harmony::RunHarmony(
  object = seu,
  group.by.vars = "sample",
  theta = NULL,
  sigma = 0.1,
  dims.use = 1:30,
  verbose = TRUE
)

Downstream neighbors, clustering, and UMAP should then use the Harmony reduction:

seu <- Seurat::FindNeighbors(
  object = seu,
  reduction = "harmony",
  dims = 1:30
)

seu <- Seurat::FindClusters(seu)

seu <- Seurat::RunUMAP(
  object = seu,
  reduction = "harmony",
  dims = 1:30
)

This keeps the per-sample SCT models in the merged object. For marker testing after merging SCT-normalized objects, check whether PrepSCTFindMarkers() is needed.

Check SCT Assay

For a single SCT-normalized object, or for the merged object used before Harmony, check the SCT assay directly.

Check assays:

Seurat::Assays(seu)

Set SCT as the active assay:

Seurat::DefaultAssay(seu) <- "SCT"

Check variable features:

variable_features <- Seurat::VariableFeatures(seu)

length(variable_features)
head(variable_features, 10)

Downstream PCA

After SCTransform() on a single object, run PCA on the SCT assay:

seu <- Seurat::RunPCA(
  object = seu,
  assay = "SCT",
  features = Seurat::VariableFeatures(seu),
  npcs = 50,
  verbose = TRUE
)

Then continue with neighbor graph, clustering, and UMAP using selected PCs.

After anchor-based SCT integration, downstream PCA usually uses the integrated assay returned by IntegrateData(). After Harmony integration, downstream steps should use reduction = "harmony" instead of rerunning this SCT PCA section.

Classic Or SCT

Use one preprocessing route clearly.

Route	Main Steps	Notes
Classic	NormalizeData() -> FindVariableFeatures() -> ScaleData()	simple and widely used
SCT	SCTransform()	model-based normalization and variance stabilization

Do not mix the two routes casually. If using SCT, downstream PCA and clustering should usually use the SCT assay.

Note

For single-sample or simple exploratory analysis, either classic normalization or SCT can be used.

For multi-sample analysis, the default SCT principle is to split by sample or library first, then run SCTransform() on each object separately. SCT integration should be treated as a separate workflow after per-sample SCT.