sunweiwei/ai4sci-virtual-cell-zeroshot-data

ai4sci virtual-cell-zeroshot — release v1

Prepared data for the `virtual-cell-zeroshot` task: predict single-cell CRISPRi knockdown responses in cellular contexts a model has never seen perturbed (the Arc Virtual Cell Challenge 2026 zero-shot setting), rebuilt from public data.

dev/        what the agent sees (mount read-only at /workspace/data)
  train/{k562,jurkat,hct116,hek293t}/   cells.h5ad, pseudobulk.h5ad, se_embeddings.npy, dev_split/
  test/{ctx_near,ctx_mid,ctx_far}/public/   basal.h5ad, requests.jsonl, context.json
  priors/     ESM-2 gene embeddings, GO annotations, DepMap gene effects (training lines)
  manifest.json, splits.json, gene_panel.csv, gene_panel_intersection.json
private/    grader only: responses of the test contexts, request subsets/strata, cached no-effect + DE tables
manifest.json

dev/se600m/ is not included: fetch SE-600M (se600m_epoch16.ckpt, config.yaml, protein_embeddings.pt, licences) from arcinstitute/SE-600M at revision 5a9a80f into dev/se600m/ (the task's materialize.sh / fetch_release.sh do this).

Schemas, split design, sampling caps and QC: the task's environment/data/README.md. Every source file, size and SHA-256: manifest.json.

Sources and licences

Derived (harmonised, subsampled, re-split) from: Replogle et al. 2022 (Figshare 20029387, CC BY 4.0); Nadig et al. 2025 (GEO GSE264667); X-Atlas/Orion, Xaira Therapeutics (HF Xaira-Therapeutics/X-Atlas-Orion @ 53a5bc98, CC BY-NC-SA 4.0 — the hct116/hek293t trees inherit that licence: non-commercial, share-alike, attribution); Arc Institute Virtual Cell Challenge 2025 H1 hESC data (Arc VCC data terms); DepMap 24Q4 (CC BY 4.0); Gene Ontology annotations (CC BY 4.0). Please cite the original datasets.