# Model

## Architecture

X-Cell is a **set-level diffusion transformer** that predicts perturbed gene expression profiles
from control cell populations. Unlike autoregressive single-cell foundation models, X-Cell operates
on *sets* of cells and is trained explicitly on interventional data via distribution-matching objectives.

<p align="center">
  <img src="figs/x-cell-overview.png" alt="X-Cell Architecture" width="100%">
</p>

### Key design choices

**Diffusion-based training.**
Each training sample has a random fraction (25%, 50%, 75%, or 100%) of control gene expression
positions replaced with ground-truth perturbed values. The model learns to predict the full
perturbed profile from this partially revealed input. At inference, predictions are iteratively
refined across 4 steps (coarse-to-fine).

**Multi-modal biological priors via cross-attention.**
At every third self-attention layer, Flamingo-style cross-attention conditions gene representations
on six prior knowledge tokens per perturbation:

| Source | Content | Dimension |
|--------|---------|-----------|
| [ESM-2](https://github.com/facebookresearch/esm) | Protein language model embeddings | 5120 |
| [STRING](https://string-db.org) | Protein-protein interaction network | 512 |
| [GenePT](https://github.com/yiqunchen/GenePT) | LLM gene representations | 3072 |
| [DepMap](https://depmap.org) | Genetic dependency profiles | 1150 |
| [JUMP-Cell Painting](https://jump-cellpainting.broadinstitute.org) | Morphological features | 259 |
| Gene identity | Stop-gradient gene embedding | — |

**Tied output embeddings.**
The output head projects back through the shared gene embedding matrix (PaLM-style 1/√d scaling),
acting as an implicit regularizer against conservative collapse.

---

## X-Cell Mini

| | X-Cell Mini |
|---|---|
| **Parameters** | 55M |
| **Layers** | 12 |
| **Hidden dim** | 512 |
| **Attention heads** | 8 |
| **FFN** | ReLU, 1× |
| **Normalization** | Post-LN (LayerNorm) |
| **Cross-attn layers** | 4 |
| **Init** | scGPT |
| **Training** | Replogle-Nadig |
| **Min GPU VRAM** | 8 GB (1 GPU) |

---

## Scaling

X-Cell follows power-law scaling consistent with large language models. Train loss scales as
L(N) ∝ N⁻⁰·³² (α = 0.32, R² = 0.96) across five model sizes from 83M to 3.1B parameters.

---

## Weights

Model weights are hosted on HuggingFace:

```
Xaira-Therapeutics/X-Cell
└── mini/     # X-Cell Mini (55M)
```

[:hugging: Open on HuggingFace](https://huggingface.co/Xaira-Therapeutics/X-Cell){ .md-button }