single-cell-analysis

Files (4)

• SKILL.md
• openai.yaml · agents/
• handler.py · scripts/
• skill.json

SKILL.md (first lines)

# Single-cell Analysis

Use `single_cell_preflight` before proposing analysis. It accepts cell-by-gene CSV/TSV, AnnData `.h5ad`, 10x H5, or standard 10x MTX files. Confirm that the selected `X` or AnnData layer contains non-negative integer raw counts, cell IDs match optional metadata exactly, QC thresholds retain a useful dataset, and the local Scanpy runtime is available.

Use `single_cell_run_analysis` only through a researcher-approved workflow. The bounded pipeline performs count QC, optional Scrublet scoring, CP10k normalization, log1p transformation, highly variable gene selection, PCA, nearest-neighbor graph construction, UMAP, Leiden clustering, and descriptive Wilcoxon marker ranking with a fixed random seed and persisted provenance. Scrublet is off by default; predicted cells remain in the analysis unless `exclude_predicted_doublets` is separately approved.

Never call a Leiden cluster a cell type without external annotation evidence. Treat UMAP geometry, Scrublet predictions, and cell-level marker p-values as exploratory. The workflow does not perform ambient-RNA correction, batch integration, trajectory inference, or donor-aware differential expression. Preserve the input format, selected count layer, batch key, automatic threshold, prediction count, and whether exclusions were applied.

From phenixace/Molemo_WorkBench, MIT.

More from phenixace/Molemo_WorkBench