Skip to content

User guide

This page summarizes common usage patterns and helps you choose the right entry point.
For the conceptual model (masking, rescue, leakage filtering, hierarchical inference), see Concepts.

Typical workflows

Scanpy (clusters → curated markers → LLM labels)

Use this when you already work with AnnData (Scanpy/Scanpy-compatible pipelines).

  1. Start from a clustered AnnData (e.g., adata.obs["leiden"]).
  2. Distill per-cluster marker lists (mask → rescue → leakage filter).
  3. Query an LLM with lineage-aware context to obtain major_type / fine_type labels.

Tip: - See Concepts for what “mask”, “rescue”, and “leakage” mean in practice.

Use this when your primary workflow is Seurat and you want to keep downstream plotting and analysis in R.
LLM-scCurator operates on AnnData, so the recommended pattern is:

Export from Seurat → run the Python workflow → re-import labels into Seurat.

Spatial (Visium / Xenium)

Use the same workflow on spatially resolved matrices; the distillation and hierarchical annotation logic is unchanged.

  • Visium: spot-level annotation with optional pseudo-bulk validation
  • Xenium: cell-level annotation with spatial coordinates preserved

See Tutorials for paper spatial validation notebooks (Xenium/Visium).

Manuscript reproduction (paper/)

For deterministic reproduction of benchmarks, figures, and Source Data, follow: