Single-Cell Scaling

Purpose

Scaling standardizes gene expression values before PCA and other downstream steps.

In Seurat, ScaleData() centers and scales expression values for each gene. After scaling, genes are more comparable for dimensionality reduction.

Typical position:

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

Scale Variable Features

By default, the standard workflow scales variable features before PCA.

seu <- Seurat::ScaleData(
  object = seu,                         # Seurat object
  features = Seurat::VariableFeatures(seu), # use variable features
  vars.to.regress = NULL,               # no regression; can use c("S.Score", "G2M.Score")
  scale.max = 10,                       # maximum scaled value
  do.scale = TRUE,                      # scale each feature
  do.center = TRUE,                     # center each feature
  verbose = TRUE                        # show progress messages
)

What this does:

• centers each feature
• scales each feature
• stores the result in scale.data

Scale All Features

Sometimes all genes are scaled, for example when downstream plotting or heatmaps need genes outside the variable feature set.

all_genes <- rownames(seu)

seu <- Seurat::ScaleData(
  object = seu,
  features = all_genes,
  verbose = TRUE
)

Scaling all genes may use more memory.

Regress Variables

ScaleData() can also regress out unwanted sources of variation.

Common examples:

• nCount_RNA
• percent.mt
• S.Score
• G2M.Score

Example:

seu <- Seurat::ScaleData(
  object = seu,
  features = Seurat::VariableFeatures(seu),
  vars.to.regress = c("nCount_RNA", "percent.mt"),
  verbose = TRUE
)

Cell cycle regression:

seu <- Seurat::ScaleData(
  object = seu,
  features = Seurat::VariableFeatures(seu),
  vars.to.regress = c("S.Score", "G2M.Score"),
  verbose = TRUE
)

Do not regress variables blindly. Regression can remove biological signal if the variable is part of the question.

Check Scaled Data

Extract scaled data:

scaled_data <- Seurat::GetAssayData(
  object = seu,
  assay = "RNA",
  slot = "scale.data"
)

Check dimensions:

dim(scaled_data)
scaled_data[1:5, 1:5]

If only variable features were scaled, scale.data may contain fewer rows than the full count matrix.

Note

Scaling is mainly preparation for PCA and related dimensionality reduction.

For a simple first workflow, scale variable features first. Add regression only when there is a clear reason.