OmicVerse Single-Cell — Cell Ontology Mapping (CellMatch)

Goal

Take an annotated single-cell AnnData whose obs[<cell_name_col>] carries free-text cell-type labels (often author-specific shorthand) and map every label to a canonical Cell Ontology (CL) term via NLP. Output is adata.obs['cell_ontology'] (matched CL term name), adata.obs['cell_ontology_cl_id'] (e.g. CL:0000084), and adata.obs['cell_ontology_score'] (cosine similarity).

Three escalating modes:

1. Plain mapping — sentence-transformer cosine-similarity match. Fastest; works on clean labels.
2. LLM abbreviation expansion — calls an LLM to expand 'TIL-1' → 'tissue-resident memory CD8+ T cell' before matching. Required when labels are short / acronym-heavy.
3. Cell Taxonomy enhancement — adds species + tissue context via the Jin 2023 Cell Taxonomy resource; matches against species-specific taxonomies as well as CL.

This skill addresses the gap noted in coverage analysis: single-cell-annotation skill handles CellTypist / SCSA / gpt4celltype annotators, but doesn't cover the post-annotation problem of mapping author-text to a standard ontology.

Quick Workflow

Mode 1 — Plain mapping (cleanest workflow when labels are full text)