// Multi-resolution typed consensus over sc-clustering. Fans out leiden / louvain at several resolutions in parallel, scores members by silhouette + cross-method NMI, runs kmode / weighted / LCA consensus on the surviving base clusterings, and emits a verified report carrying the mandatory A-path banner per ADR 0010.
LLM-grounded biological interpretation of a verified typed consensus run. Reads the typed run dir + the original adata, runs inline per-cluster DE, looks up markers in a bundled tissue-keyed marker DB, and asks the chair LLM to (γ) name each cluster's likely cell type with mandatory marker citations and (β) recommend top-3 next-step skills with mandatory evidence_refs. Output banner [A+I: Interpreted on verified consensus]. Failure-mode contract per ADR 0012.
Multi-method consensus over spatial-domains. Fans out 5 methods in parallel, computes a SACCELERATOR-style base-clustering ranking, runs typed consensus (kmode / weighted / LCA), and emits a verified consensus report with the mandatory A-path banner per ADR 0010.
Load when scaffolding a NEW OmicsClaw skill from a natural-language request — generates the skill directory layout (SKILL.md, parameters.yaml, references/, tests/) under the chosen domain. Skip when modifying an existing skill (edit its files directly) or when only routing a query (use `orchestrator`).
Load when routing a natural-language omics query to the correct domain skill across spatial / singlecell / genomics / proteomics / metabolomics / bulkrna domains via keyword / LLM / hybrid matching. Skip when the target skill is already known — invoke that skill directly.
Load when copying this directory to bootstrap a new OmicsClaw v2 skill (rename, fill in, then `git add`). Skip when an existing skill already covers the request.
Load when extracting gene programs (NMF / cNMF factorisation) and per-cell program usage scores from a non-negative scRNA AnnData. Skip when ranking marker genes per cluster (use sc-markers) or for inferring TF → target regulons (use sc-grn).
| name | sc-consensus-clustering |
| description | Multi-resolution typed consensus over sc-clustering. Fans out leiden / louvain at several resolutions in parallel, scores members by silhouette + cross-method NMI, runs kmode / weighted / LCA consensus on the surviving base clusterings, and emits a verified report carrying the mandatory A-path banner per ADR 0010. |
| version | 0.1.0 |
| author | OmicsClaw |
| license | Apache-2.0 |
| tags | ["singlecell","consensus","typed-consensus","multi-resolution","silhouette","bc-ranking","kmode","lca","weighted"] |
| requires | ["anndata","scanpy","numpy","pandas","scipy","scikit-learn","pyyaml"] |
The user has a preprocessed scRNA AnnData (PCA + neighborhood graph
already computed via sc-preprocessing) and wants robust cell-cluster
assignments insensitive to the chosen resolution. Single-resolution
Leiden/Louvain results are notoriously resolution-sensitive — at
r=0.4 you get 6 broad types, at r=1.5 you get 22 sub-states. This
skill runs a SACCELERATOR-style consensus across a resolution sweep
(and optionally across leiden vs louvain) and reports the stable
core of the labels.
It does NOT replace sc-clustering; it wraps it.
| Input | Format | Required |
|---|---|---|
| Preprocessed AnnData | --input <preprocessed.h5ad> (with obsm["X_pca"]) | yes |
| Output directory | --output <dir> | yes |
| Resolutions to sweep | --resolutions 0.4,0.8,1.0,1.4,2.0 | no (default 0.5,0.8,1.0,1.4,2.0) |
| Cluster methods to use | --cluster-methods leiden,louvain | no (default leiden) |
| Explicit member list | --members leiden:resolution=0.5,louvain:resolution=1.0 | no (overrides the sweep) |
| Fan-out everything | --all | no (sweeps both methods × all default resolutions) |
| Operator | --operator {kmode,weighted,lca} | no (default kmode) |
| Score weights | --alpha 0.6 --beta 0.4 | no (ADR 0011 defaults) |
| Class-imbalance cap | --max-class-frac 0.8 | no |
| Pre-run plan confirm | --confirm-plan | no |
| Non-interactive | --non-interactive | no |
| Seed | --seed 0 | no |
| Output | Path | Notes |
|---|---|---|
| Verified consensus labels | consensus_labels.tsv | columns cell_id,consensus_<operator> |
| Per-member labels | member_<name>/figure_data/embedding_points.csv | from sc-clustering |
| Cross-method NMI matrix | cross_method_nmi.csv | square per member |
| Composite scores | member_scores.csv | ADR 0011 schema |
| Markdown report | report.md | starts with [A: Verified consensus] (non-configurable) |
| Audit | plan.json | resolution sweep + chosen operator + filtered members |
--members / --all) or
derived from the resolution sweep × cluster-methods combinations.sc-clustering once per member.silhouette_score from each member's
clustering_summary.csv is the intrinsic-quality signal; cross-method
NMI is computed across members.--cluster-methods defaults to leiden ONLY, not both, because
louvain and leiden agree to within 1–2% on most datasets and the
consensus signal comes mostly from the resolution sweep.runtime/consensus/dispatch.output_banner. Do NOT strip it.requires_preprocessed: true — run sc-preprocessing first.# Default sweep (leiden at 5 resolutions)
oc run sc-consensus-clustering --input preprocessed.h5ad --output out/
# Both methods × 5 resolutions = 10 members; SACCELERATOR-style benchmark
oc run sc-consensus-clustering --input preprocessed.h5ad --output out/ \
--cluster-methods leiden,louvain --resolutions 0.5,0.8,1.0,1.4,2.0
# Explicit
oc run sc-consensus-clustering --input preprocessed.h5ad --output out/ \
--members leiden:resolution=0.5,leiden:resolution=1.0,louvain:resolution=1.0
skills/spatial/consensus-domains/ — sibling spatial-side skill