bioresearch-foundation-embeddingslisted

# BioResearch Agent — Foundation Model Embeddings Skill ## Capability Generates cell-level embeddings from three single-cell foundation models and evaluates them on a standardized cell-type recovery benchmark: 1. **scGPT** (Cui et al. 2024, Nature Methods) — 512-dim, pretrained on 33M human cells, CPU-compatible. Input: normalized + log1p AnnData. Live API: `scgpt.tasks.embed_data(adata, model_dir, gene_col)`. 2. **UCE** (Rosen et al. 2023, bioRxiv) — 1280-dim, pretrained on 36M multi-species cells, GPU required. Input: raw counts h5ad. Live API: `AnndataProcessor` from `evaluate.py`. 3. **scFoundation** (Hao et al. 2024, Nature Methods) — 512-dim, pretrained on 50M+ human cells, GPU required. Input: CSV aligned to 19,264-gene list. Live API: `get_embedding.py --input_type singlecell`. Evaluation metrics (all custom implementations, no sklearn dependency for core metrics): - **Silhouette score** — cluster cohesion vs separation - **ARI** (Adjusted Rand Index) — agreement with ground-truth labels - **NMI** (Normalized Mutual Information) — information-theoretic agreement - **Cross-model kNN overlap** — neighborhood consistency between model pairs - **Robustness sweep** — metrics across noise levels [0.1–0.6] Returns per-model metrics CSV, cross-model consistency table, robustness sweep results, UMAP/PCA visualization, clustering heatmap, summary report with evidence grade, and a JSON evidence package. ## Run ```bash bioresearch run foundation-embedding