README.md

---
license: mit
library_name: expert-model
pipeline_tag: other
tags:
  - single-cell
  - perturb-seq
  - scRNA-seq
  - perturbation-prediction
  - biology
---

# ExPert paper checkpoints

The additional checkpoints behind the ExPert paper's benchmark and ablation figures. The production model that the
paper's application figures use, and that you should use for your own data, is
[GoyalLab/ExPert](https://huggingface.co/GoyalLab/ExPert).

Code and the scripts that use these checkpoints: <https://github.com/GoyalLab/ExPert> (`paper/`)

| folder | what | used in |
|---|---|---|
| `fold0` … `fold11` | the 12 cross-validation models: each holds out one context (cell line / donor / condition) from training | benchmark and out-of-distribution figures |
| `EFFGATE3` | the ablation arm with the gated prior pull | prior-pull ablation figures |
| `banks/v0.1.0_fullbank_oldfill.npz` | v0.1.0 prototypes with the earlier fill for genes never perturbed in training (trained rows identical to the model's) | ALD and PMD readouts |

Each checkpoint folder holds `model.pt` and `gene_embeddings.npz` and loads like the production model:

```python
from huggingface_hub import snapshot_download
from expert import ExPert

path = snapshot_download("GoyalLab/ExPert-paper-models", revision="v0.1.0", allow_patterns="fold4/*")
model = ExPert.load(f"{path}/fold4")
```

The bank is a NumPy archive: `P` (perturbations × latent dims), `labels`, `trained`.

## Citation

_Preprint reference to be added._