Data vs Counts-level Variable Features for Spatial Datasets

[CalebYuDot]

Hi all — I’m new to Seurat working through the Visium vignette and noticed FindVariableFeatures() computes variable genes using the counts layer by default. This is even though downstream steps like normalization and dimensionality reduction operate more on the data layer.

Why does Seurat default to counts for variable feature selection, instead of the normalized data layer? Just curious about the reasoning behind the design choice.

[YoukaiFromAccounting]

Hello @CalebYuDot; thank you for your message. By default, FindVariableFeatures() is designed to operate on the raw counts layers, which should capture more biologically meaningful variability prior to normalization or transformation. Further downstream functions like FindTransferAnchors will operate on the normalized data layer.

LeverageScore() was recently updated to use the counts layer variable features found by FindVariableFeatures() without needing users to manually specify to use those features in SketchData. Alternatively, you can identify your own set of variable features and pass them into SketchData if you want to use a custom set that you previously identified.

Thank you for your continued interest in the Seurat project!