scRNA-seq preprocessing and clustering
Use this skill after input inspection when a real AnnData .h5ad dataset must be normalized, reduced, and clustered.
Scientific contract
- Treat the source file as read-only and record its SHA-256 before and after execution.
- Accept non-negative raw count matrices, or matrices explicitly marked as log1p-transformed in
adata.uns["log1p"]. - Refuse ambiguous continuous matrices rather than guessing whether TPM, normalized values, or log values need another transformation.
- For raw counts, filter cells with fewer than 200 detected genes and genes detected in fewer than 3 cells, retain filtered counts in
layers["counts"], normalize to 10,000 counts per cell, and applylog1p. - Select up to 2,000 highly variable genes, then run PCA, a neighbor graph, Leiden clustering, and UMAP with the supplied seed.
- Preserve retained cell order. Never overwrite the input or claim that clusters are biological cell types.
- Record every selected branch, threshold, package version, cell disposition, and cluster size in machine-readable artifacts.
Outputs
processed.h5ad: normalized data, retained counts, HVG flags, PCA, neighbor graph, Leiden clusters, and UMAP.preprocessing_report.json: input identity, detected state, parameters, filtering, dimensions, cluster sizes, warnings, and source-integrity evidence.cell_clusters.csv: one row per retained cell in input order with Leiden cluster and UMAP coordinates.cell_disposition.csv: one row per original cell indicating whether it passed the explicit cell filter.
