Sc annotation

Skill mianaz/sc-skills/skills/sc-annotation

Use when assigning cell-type identities to single-cell clusters — generating candidate labels via reference-atlas label transfer, pretrained models (SCimilarity, Azimuth), SingleR/celldex, or LLM marker-prompting (GPT-4o/Claude), then confirming every label against canonical (human or mouse) markers with citations using a dotplot, plus AddModuleScore signature scoring (TLS, chemokine, MHC, Ig). Coarse compartments before fine subtypes.From its SKILL.md

Install

npx -y skills add mianaz/sc-skills --skill sc-annotation

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

2.7 KB, as published. Nobody here has run it

sc-annotation

Overview

Automation proposes, markers dispose. Candidate labels come from label transfer / pretrained models; canonical-marker dotplot verification with citations is the mandatory final gate. Two-level: coarse compartment → fine subtype.

Flow

Cluster on the integrated embedding.
Candidate labels (not final) — use one or more, then cross-check agreement:
- label transfer (references/label-transfer.md)
- pretrained models (references/pretrained-models.md) — incl. Garnett marker-file classifier (literature-driven, portable, independent of your clustering)
- SingleR + celldex reference (references/singler.md)
- LLM marker-prompting, GPT-4o / Claude (references/llm-annotation.md)
- When several references/methods disagree, quantify agreement with a Cohen's κ concordance matrix + spot outlier references → references/label-transfer-concordance.md
MANDATORY: canonical-marker dotplot verification (references/marker-verification.md) using references/markers.md (human) or references/markers-mouse-immune.md (mouse), both cited.
Assign identities; log supporting markers + citations + which automated methods agreed in methods.md.
Optional — signature/state scores: AddModuleScore for published programs (TLS, 12-chemokine, MHC-I/II, Ig), or a signature derived from a bulk DE table projected onto cells (bulk→sc bridge, with pseudobulk reverse-validation). See references/module-scores.md. Compare across groups with a test + exact p-value.
Optional — marker-panel QC: to test whether a small gene panel actually separates the annotated types (e.g. for a flow panel), measure discriminatory power with a classifier benchmarked against ceiling/floor. See references/panel-discriminatory-power.md.
Optional — cluster-specificity QC: to defend annotation granularity or make keep-vs-merge calls, score whether each cluster is separable on the whole transcriptome (SVM cross-validation F1 vs a permuted-label null) — clusters near the null floor are over-clustering. See references/cluster-specificity.md.

When NOT to use

Embeddings/clustering inputs → sc-integration. Figures → scientific-plotting.