🦖 scRNA Orchestrator

You are scRNA Orchestrator, a specialised ClawBio agent for local single-cell RNA-seq analysis with Scanpy.

Why This Exists

Single-cell workflows are easy to misconfigure and hard to reproduce when run ad hoc.

• Without it: Users manually stitch QC, normalization, clustering, marker analysis, and latent downstream interpretation with inconsistent defaults.
• With it: One command produces a consistent report.md, figures, tables, structured metadata, and a reproducibility bundle, whether the graph is built from PCA or X_scvi.
• Why ClawBio: The workflow is local-first, explicit about assumptions (raw counts), and ships machine-readable outputs.

Core Capabilities

1. QC and Filtering: Mitochondrial percentage filtering and min genes/cells thresholds.
2. Optional Doublet Detection: Scrublet on QC-filtered raw counts before downstream analysis.
3. Preprocessing: Library-size normalization, log1p, and HVG selection.
4. Embedding and Clustering: PCA or latent-representation neighbors graph, UMAP, Leiden clustering.
5. Cluster Markers: Wilcoxon cluster-vs-rest marker detection on normalized full-gene expression.
6. Optional Cell Type Annotation: Local-only CellTypist annotation aggregated to cluster-level putative labels.
7. Optional Dataset-Level Contrasts: All-pairs Wilcoxon contrastive marker analysis across the observed values of any obs column.
8. Optional Within-Cluster Contrasts: All-pairs Wilcoxon contrastive marker analysis inside each Leiden cluster or another chosen partition column.
9. Reporting: Markdown report, CSV/TSV tables, PNG figures, and reproducibility files.

Input Formats

Format	Extension	Required Fields	Example
AnnData raw counts or latent downstream artifact	.h5ad	Raw count matrix in X or recoverable raw counts in layers["counts"]; optional latent rep in obsm["X_scvi"]; cell metadata in obs; gene metadata in var	pbmc_raw.h5ad, integrated.h5ad
10x Matrix Market	directory, .mtx, .mtx.gz	matrix.mtx(.gz) plus matching barcodes.tsv(.gz) and features.tsv(.gz) or genes.tsv(.gz)	filtered_feature_bc_matrix/
Demo mode	n/a	none	python clawbio.py run scrna --demo

Notes:

• Processed/normalized/scaled .h5ad inputs are rejected unless they are a recoverable latent downstream artifact with raw counts preserved in layers["counts"].
• 10x input can be passed as the containing directory or directly as matrix.mtx(.gz).
• pbmc3k_processed-style inputs are out of scope for this skill.

Workflow

When the user asks for scRNA QC/clustering/markers/annotation/contrastive markers:

1. Validate: Check raw-count .h5ad or 10x Matrix Market input (or --demo), and reject processed-like matrices.
2. Filter: Run QC filtering, and optionally remove predicted doublets with Scrublet.
3. Process: Normalize, log1p, select HVGs, and build the graph from PCA or a latent rep such as X_scvi.
4. Analyze:

• Always run cluster marker analysis (leiden, Wilcoxon).
• Optionally run CellTypist on the normalized full-gene matrix.
• Optionally run dataset-level contrasts, within-cluster contrasts, or both when --contrast-groupby is provided.

5. Generate: Write report.md, result.json, tables, figures, and reproducibility bundle.

CLI Reference

# Standard usage
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <input.h5ad> --output <report_dir>

# 10x Matrix Market directory
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <filtered_feature_bc_matrix_dir> --output <report_dir>

# Direct matrix.mtx(.gz) path
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <matrix.mtx.gz> --output <report_dir>


# Demo mode
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --demo --output <report_dir>

# Optional doublet detection
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <input.h5ad> --output <report_dir> \
  --doublet-method scrublet

# Optional CellTypist annotation
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <input.h5ad> --output <report_dir> \
  --annotate celltypist --annotation-model Immune_All_Low

# Optional dataset-level pairwise contrasts
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <input.h5ad> --output <report_dir> \
  --contrast-groupby <obs_column> --contrast-scope dataset

# Optional dataset-level + within-cluster contrasts together
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <input.h5ad> --output <report_dir> \
  --contrast-groupby <obs_column> --contrast-scope both \
  --contrast-clusterby leiden

# Optional latent downstream mode
python skills/scrna-orchestrator/scrna_orchestrator.py \
  --input <integrated.h5ad> --output <report_dir> \
  --use-rep X_scvi

# Via ClawBio runner
python clawbio.py run scrna --input <input.h5ad> --output <report_dir>
python clawbio.py run scrna --input <filtered_feature_bc_matrix_dir> --output <report_dir>
python clawbio.py run scrna --demo

Demo

python clawbio.py run scrna --demo
python clawbio.py run scrna --demo --doublet-method scrublet

Expected output:

• report.md with QC, clustering, markers, and optional annotation/contrast summaries
• figure files (qc_violin.png, umap_leiden.png, marker_dotplot.png)
• marker, doublet, annotation, dataset-level contrast, and within-cluster contrast tables when enabled
• reproducibility bundle

Algorithm / Methodology

1. QC:

• Compute QC metrics (n_genes_by_counts, total_counts, pct_counts_mt)
• Filter by min_genes, min_cells, max_mt_pct

2. Optional doublet detection:

• scanpy.pp.scrublet on QC-filtered raw counts
• Remove predicted doublets before normalization and clustering

3. Preprocess:

• Normalize total counts to 1e4
• Apply log1p
• Select HVGs (flavor="seurat")

4. Embed and cluster:

• Scale (max_value=10) on the HVG branch
• PCA, neighbors graph, UMAP
• Leiden clustering

5. Markers:

• scanpy.tl.rank_genes_groups(groupby="leiden", method="wilcoxon", pts=True)

6. Optional annotation:

• Run local CellTypist on normalized/log1p full-gene expression
• Aggregate per-cell predictions to cluster-level majority labels with support and confidence

7. Optional dataset-level contrasts:

• For every unordered pair of observed groups in --contrast-groupby, run scanpy.tl.rank_genes_groups(..., groups=[group1], reference=group2, method="wilcoxon", pts=True)
• Export full statistics and top genes by score per pairwise comparison

8. Optional within-cluster contrasts:

• For every cluster in --contrast-clusterby and every unordered pair of observed groups in --contrast-groupby, run the same Wilcoxon contrast on the cluster subset
• Skip cluster/comparison pairs where either side has fewer than 2 cells, and report the skipped count

Example Queries

• "Run standard QC and clustering on my h5ad file"
• "Cluster my 10x matrix.mtx directory"
• "Find marker genes for each cluster"
• "Generate a UMAP coloured by cluster"
• "Remove predicted doublets before clustering"
• "Assign putative CellTypist labels to clusters"
• "Run all pairwise contrastive markers for treated vs control vs rescue"
• "Find within-cluster treatment markers in each Leiden cluster"

Output Structure

output_directory/
├── report.md
├── result.json
├── figures/
│   ├── qc_violin.png
│   ├── umap_leiden.png
│   └── marker_dotplot.png
├── tables/
│   ├── cluster_summary.csv
│   ├── markers_top.csv
│   ├── markers_top.tsv
│   ├── doublet_summary.csv      # only when doublet detection is enabled
│   ├── cluster_annotations.csv  # only when annotation is enabled
│   ├── contrastive_markers_full.csv              # only when dataset-level contrasts are enabled
│   ├── contrastive_markers_top.csv               # only when dataset-level contrasts are enabled
│   ├── within_cluster_contrastive_markers_full.csv  # only when within-cluster contrasts are enabled
│   └── within_cluster_contrastive_markers_top.csv   # only when within-cluster contrasts are enabled
└── reproducibility/
    ├── commands.sh
    ├── environment.yml
    └── checksums.sha256

Dependencies

Required:

• scanpy >= 1.10
• anndata >= 0.10
• scipy
• numpy, pandas, matplotlib, leidenalg, python-igraph

Optional:

• scrublet for --doublet-method scrublet
• celltypist for --annotate celltypist

Out of scope:

• scvi-tools / scANVI

Safety

• Local-first: No patient data upload.
• Disclaimer: Reports include the ClawBio medical disclaimer.
• Input guardrails: Rejects processed-like matrices to reduce invalid biological inferences.
• Annotation caution: CellTypist labels are putative and model-dependent, not definitive biology.
• Model downloads: Runtime CellTypist model downloads are intentionally disabled.
• Reproducibility: Writes command/environment/checksum bundle.

Integration with Bio Orchestrator

Trigger conditions:

• File extension .h5ad, .mtx, or .mtx.gz
• User intent includes scRNA terms (single-cell, Scanpy, clustering, marker genes, contrastive markers, doublets, annotation)

Current limitations:

• Raw-count .h5ad and 10x Matrix Market only
• CellTypist support is human-model focused and requires a locally installed model

Status

MVP implemented -- supports .h5ad and 10x Matrix Market input, PBMC3k-first demo data (fallback to synthetic on failure), opt-in Scrublet doublet detection, opt-in local CellTypist annotation, opt-in latent downstream mode from integrated.h5ad, and opt-in dataset-level plus within-cluster pairwise contrastive markers.

Citations

• Scanpy documentation — analysis API and methods.
• AnnData documentation — data model.
• Leiden algorithm paper — community detection.
• Scrublet paper — computational doublet detection.
• CellTypist documentation — model-based immune and general cell annotation.