# scFATE — NeurIPS 2026 datasets

Paper-canonical splits and biological-descriptor multiview tensors for the four perturbation benchmarks evaluated in
[`Angione-Lab/scFATE`](https://huggingface.co/Angione-Lab/scFATE).

## Files

| Path | Source | Use |
|---|---|---|
| `CRISPRa-norman/norman2019_gears_split.h5ad`     | Norman et al. 2019, GEARS split | CRISPRa (Norman) |
| `replogle_k562/replogle_k562.h5ad`               | Replogle et al. 2022 | K562 CRISPRi |
| `replogle_rpe1/replogle_rpe1.h5ad`               | Replogle et al. 2022 | RPE1 CRISPRi |
| `sciplex3/sciplex3.h5ad`                          | Srivatsan et al. 2020 | SciPlex3 chemical |
| `gene_embeddings/{norman,replogle_k562,replogle_rpe1}_multiview.pt` | precomputed | gene-descriptor multiview |
| `drug_embeddings/sciplex3_multiview_v2.pt`        | precomputed (Morgan + RDKit + ChemBERTa) | drug-descriptor multiview |
| `splits/sciplex3_split_v2b.json`                  | mechanism-of-action stratified | SciPlex3 25-drug OOD split |

## Usage

```python
from huggingface_hub import snapshot_download
data_dir = snapshot_download("Angione-Lab/scFATE-datasets", repo_type="dataset")
# Then in scFATE training code, point --dataset_h5ad at e.g.
# {data_dir}/replogle_k562/replogle_k562.h5ad
```