A diffusion language model for genome-scale perturbation prediction across diverse cellular contexts
Status: Model weights and inference code coming soon. The Python API, model weights, and tutorials are under active development. Star/watch this repository for release updates.
X-Cell predicts genome-scale transcriptional responses to genetic perturbations across diverse cellular contexts. Trained on X-Atlas/Pisces (25.6M perturbed single cells, 7 CRISPRi Perturb-seq screens), X-Cell integrates multi-modal biological priors through cross-attention and generalizes zero-shot to unseen cell types and perturbations.
- Diffusion LM with iterative inference-time refinement
- Multi-modal priors via cross-attention (ESM-2, STRING, GenePT, DepMap, JUMP-Cell Painting, scGPT)
- X-Cell Mini (55M) — a compact model for single-GPU perturbation prediction
- 5× higher Pearson Δ than the next-best method on held-out perturbations
- Zero-shot generalization to unseen cell types, confirmed on primary human T cells and melanocyte progenitors
pip install xcellimport anndata as ad
from xcell import XCell
# Load pretrained X-Cell Mini
model = XCell.from_pretrained("Xaira-Therapeutics/X-Cell", variant="mini")
# Predict from an AnnData object
adata = ad.read_h5ad("control_cells.h5ad")
predictions = model.predict(adata, perturbation="BRCA1")
# Or from one or more .h5ad paths
predictions = model.predict(
["screen1.h5ad", "screen2.h5ad"],
perturbation="BRCA1",
)See the documentation for full examples.
This project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
| Model | Parameters | HuggingFace |
|---|---|---|
| X-Cell Mini | 55M | Xaira-Therapeutics/X-Cell |
X-Atlas/Pisces is available at Xaira-Therapeutics/X-Atlas-Pisces.
@article{xcell2026,
title = {X-Cell: Scaling Causal Perturbation Prediction Across Diverse
Cellular Contexts via Diffusion Language Models},
year = {2026},
}See LICENSE for details.