Multi-Modal Foundation · Xaira-Therapeutics
X-Cell
A diffusion model predicting single-cell transcriptomic responses to CRISPRi perturbations across cellular contexts.
What X-Cell is
X-Cell is a set-level diffusion language model developed by Xaira Therapeutics for predicting genome-scale transcriptomic responses to genetic perturbations at single-cell resolution. It is designed to model the downstream effects of CRISPR interference (CRISPRi) gene knockdowns across diverse cellular settings.
The model addresses the challenge of predicting cellular responses across unseen cell types and perturbations. By integrating multi-modal biological priors through cross-attention, it enables zero-shot causal perturbation prediction, capturing complex transcriptional shifts without requiring prior experimental screens for every target condition.
How it works
X-Cell is structured as a set-level diffusion transformer that operates on sets of cells rather than individual cells. Predictions are generated iteratively using a masked diffusion process with a four-step coarse-to-fine refinement routine during inference. The architecture incorporates tied output embeddings scaled by 1/sqrt(d) following a PaLM-style formulation, and the X-Cell Mini variant is initialised from scGPT. It also integrates multi-modal biological priors via Flamingo-style cross-attention mechanisms, drawing from ESM-2, STRING, GenePT, DepMap, JUMP-Cell Painting, and scGPT.
The model is trained on the X-Atlas/Pisces dataset, a compendium comprising 25.6 million perturbed single cells across seven CRISPRi Perturb-seq screens. These screens span diverse cellular contexts: colorectal cancer (HCT116, 3.4M cells, 18,924 perturbations), kidney epithelial cells (HEK293T, 4.5M cells, 18,312 perturbations), hepatocellular carcinoma (HepG2, 2.6M cells, 9,735 perturbations), induced pluripotent stem cells (iPSC, 4.2M cells, 10,095 perturbations), resting Jurkat T lymphoblastic leukaemia cells (2.8M cells, 10,872 perturbations), CD3/CD28-stimulated active Jurkat cells (2.8M cells, 10,878 perturbations), and an iPSC multi-lineage differentiation screen (5.1M cells, 12,175 perturbations).
Key facts
- Model architecture
- Set-level diffusion transformer
- Variant parameters
- 55M (X-Cell Mini)
- Initialisation
- scGPT (for X-Cell Mini)
- Training dataset
- X-Atlas/Pisces (25.6M cells across 7 CRISPRi Perturb-seq screens)
- Inference process
- 4-step coarse-to-fine masked diffusion refinement
What you can use it for
- Transcriptomic perturbation prediction
- Forecast genome-scale single-cell gene expression changes resulting from CRISPRi gene knockdowns using control cell transcriptomic data.
- Zero-shot cell type generalization
- Predict perturbation outcomes in cell contexts outside the training set, including primary human CD4+ T cells and melanocyte progenitors.
- Immune regulation modeling
- Model T-cell inactivation zero-shot, identifying CD3 complex inactivators and novel regulators such as LRBA and APPL2.
Run it here
We list this model but cannot run it yet: unsupported task: other.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Achieves a 5x higher Pearson delta compared to the next-best method on held-out iPSC perturbations.
- Demonstrates zero-shot generalisation to unseen cell types and genetic perturbations.
- Integrates multi-modal biological priors using cross-attention over protein sequences, interaction networks, and functional assays.
- Displays LLM-class scaling behaviour with training loss scaling as L(N) ~ N^-0.32 (R^2 = 0.96).
Watch out for
- Designed specifically for CRISPRi gene knockdowns rather than other perturbation modalities like CRISPR activation or small molecules.
- Model weights, Python API, and inference code are designated as under active development and coming soon.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · cc-by-nc-sa-4.0
The cc-by-nc-sa-4.0 licence carries terms worth reading before you depend on it.
Questions
Can I use X-Cell commercially?
Check the licence before commercial use. The cc-by-nc-sa-4.0 licence carries terms worth reading before you depend on it.
What kind of perturbations does X-Cell predict?
X-Cell is designed to predict genome-scale transcriptional responses to CRISPRi (CRISPR interference) gene knockdowns.
What is the size of the X-Cell Mini variant?
X-Cell Mini has 55 million parameters and is initialised from scGPT for fast inference.
How does X-Cell process single-cell input data?
Rather than evaluating cells individually, X-Cell is a set-level transformer that operates over sets of cells and iteratively refines predictions via a four-step masked diffusion process.
Which biological priors are incorporated into the model?
It integrates priors across multiple modalities via Flamingo-style cross-attention, including ESM-2, STRING, GenePT, DepMap, JUMP-Cell Painting, and scGPT.
How to cite it
@article{xcell2026, title = {X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models}, year = {2026}}
Written from X-Cell's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.