claude-skill-registry

"

Universal Single-Cell Annotator

This skill wraps multiple cell type annotation strategies into a single Python class. It allows agents to flexibly choose between rule-based (markers), data-driven (CellTypist), or reasoning-based (LLM) approaches depending on the context.

When to Use This Skill

• Initial Analysis: When processing raw AnnData objects.
• Validation: When cross-referencing automated labels with known markers.
• Discovery: When identifying rare cell types using LLM reasoning on marker lists.

Core Capabilities

1. Marker-Based Scoring: Scores cells based on provided gene lists (e.g., "T-cell": ["CD3D", "CD3E"]).
2. Deep Learning Reference: Wraps celltypist to transfer labels from massive atlases.
3. LLM Reasoning: Extracts top markers per cluster and constructs prompts for LLM interpretation.

Workflow

1. Load Data: Ensure data is in AnnData format (standard for Scanpy).
2. Choose Strategy:
3. Use Markers if you have a known gene panel.
4. Use CellTypist for broad immune/tissue profiling.
5. Use LLM for novel clusters.
6. Annotate: Run the corresponding method.
7. Inspect: Check adata.obs for the new annotation columns.

Example Usage

User: "Annotate this dataset looking for T-cells and B-cells."

Agent Action:

from universal_annotator import UniversalAnnotator
import scanpy as sc

adata = sc.read_h5ad('data.h5ad')
annotator = UniversalAnnotator(adata)

markers = {
    'T-cell': ['CD3D', 'CD3E', 'CD8A'],
    'B-cell': ['CD79A', 'MS4A1']
}

annotator.annotate_marker_based(markers)
# Results in adata.obs['predicted_cell_type']