foundation-model-analysis
Installation
SKILL.md
Foundation Model Analysis
Use this skill when a user wants to generate cell embeddings, annotate cell types, integrate batches, or predict perturbation effects using single-cell foundation models. The ov.fm module provides a unified 6-step API that works identically across all 22 supported models.
Model Selection Guide
Pick a model based on your task, species, and hardware. The 5 skill-ready models have full adapter support:
| Model | Tasks | Species | Gene IDs | Min VRAM | CPU? | Best when |
|---|---|---|---|---|---|---|
| scGPT | embed, integrate | human, mouse | symbol | 8 GB | Yes | General RNA, multi-modal (RNA+ATAC+Spatial) |
| Geneformer | embed, integrate | human | ensembl | 4 GB | Yes | Ensembl IDs, CPU-only environments, network biology |
| UCE | embed, integrate | 7 species | symbol | 16 GB | No | Cross-species (zebrafish, macaque, pig, frog, lemur) |
| scFoundation | embed, integrate | human | custom | 16 GB | No | xTrimoGene architecture, perturbation tasks |
| CellPLM | embed, integrate | human | symbol | 8 GB | Yes | Fastest inference (batch_size=128), cell-centric |
12 additional partial models (scBERT, GeneCompass, Nicheformer, scMulan, tGPT, CellFM, scCello, scPrint, AiDocell, Pulsar, Atacformer, scPlantLLM) and 5+ experimental models are also registered.