omicverse-single-cell-foundation-model
Single-cell foundation model — ov.llm.SCLLMManager
When to use this skill
Pick this skill when the user wants a transformer-style per-cell representation that is structurally richer than PCA: e.g. cell embedding, zero-shot or fine-tuned cell-type annotation, batch integration on the foundation-model latent, or (model permitting) perturbation prediction. The five fully-supported backends are scGPT, Geneformer, scFoundation, UCE, CellPLM; pick by data type, hardware, and gene-ID convention.
This skill does not cover marker-rule annotators (CellTypist / SCSA / gpt4celltype) or reference-mapping annotators (popV / scmap / SingleR). For those, prefer the single-cell-annotation skill — both can coexist; the foundation embedding is often used as the input space for downstream annotation/integration.
Backend selection cheatsheet
| Model | Tasks supported | Species | Gene IDs | Min VRAM | CPU? | Strengths |
|---|---|---|---|---|---|---|
| scGPT | embed, integrate, fine-tune→annotate | human, mouse | symbol | 8 GB | yes | General RNA, longest-running, multi-modal extensions |
| Geneformer | embed, integrate, fine-tune→annotate | human | ENSEMBL | 4 GB | yes | Ensembl-id pipelines, low-VRAM |
| scFoundation | embed, integrate | human | symbol | 16 GB | no | xTrimoGene architecture; perturbation work upstream |
| UCE | embed, integrate | 7 species (cross-species) | symbol | 16 GB | no | Zebrafish / macaque / pig / frog / lemur transfer |
| CellPLM | embed, integrate, annotate (zero-shot) | human | symbol | 8 GB | yes | Fastest inference; cell-centric pretraining |
Common pitfalls the agent should avoid: