Coming from Scanpy or Seurat

Scarf keeps the familiar biological sequence of quality control, feature selection, normalization, reduction, neighbourhood graphs, embeddings, clustering, and marker discovery. The main change is how the work is executed. A Scarf DataStore opens a Zarr store on local disk or object storage, streams bounded blocks, and writes each completed result back as an immutable artifact.

Choose the guide for the ecosystem you already use:

• Scarf for Scanpy users maps Scanpy stages and explains H5AD transfer.

• Scarf for Seurat users maps Seurat stages, RDS import, and multimodal WNN analysis.

If you want to run Scarf before comparing APIs, start with the Quick start.

Where Scarf fits

Scarf is most useful when:

• the count matrix is too large for convenient in-memory analysis

• counts should stay on local disk or object storage while the analysis runs

• completed steps should persist so they can be inspected or reused

• RNA, ATAC, or CITE-seq assays should live in one analysis store

You do not need to move an entire project to Scarf. A common pattern is to run the large, graph-based part of an analysis in Scarf, then export the required data for a method in another single-cell ecosystem. Scarf does not include scVI, Scanorama, RNA velocity, or replicate-aware differential expression.

The mental-model change

Familiar concept	Scarf
An AnnData or SeuratObject holds counts, metadata, and results	A DataStore opens a Zarr store containing counts, metadata, and results
Analysis changes an object in the current session	Each supported analysis step persists an immutable artifact
Filtering subsets an object or creates a view	Filtering returns an immutable selection artifact without deleting cells
Graphs and embeddings occupy named object slots	Graphs and embeddings are passed by exact artifact reference
One object exposes an active result set	A PipelineRun exposes one durable frozen result set without replacing other runs

Feature selection is artifact-only. A call such as features = ds.select_hvgs(cells, ...) returns an immutable reference that normalization and downstream methods consume. Related analysis branches can therefore coexist in one datastore without an implicit active result replacing an earlier one.

See Why Scarf records provenance for the result model and Import and export for the full format contracts.