---
license: mit
tags:
- spatial-transcriptomics
- computational-pathology
- histology
- gene-expression-prediction
- uncertainty-quantification
pipeline_tag: image-to-image
---

# DELPHI

Pre-trained weights for the DELPHI model described in:

> *DELPHI: Dual-scale calibrated confidence enables trustworthy spatial gene
> expression prediction from histology.* Under review.

## Model

| File | Size | Description |
|------|------|-------------|
| `delphi_her2st.pt` | 34 MB | HER2ST full-dataset model (seed 42, epoch 22) |

## Usage

```python
from src.model import DELPHI
import torch

model = DELPHI(uni2h_dim=1536, hidden_dim=384, num_genes=785,
               gh=12, gw=12, knn_k=8, n_swin_blocks=4)
model.load_state_dict(torch.load("delphi_her2st.pt", map_location="cpu"))
model.eval()

# Inference: x [N, 1536] UNI2-h features, pos [N, 2] coordinates
with torch.no_grad():
    mu, log_phi, pi, sigma_al, sigma_ep = model(x, pos)
```

## Architecture

Frozen UNI2-h ViT encoder → 12×12 Swin grid (K=8) → Bayesian Last Layer → Hurdle-Gaussian head

## Citation

```bibtex
@article{delphi2025,
  title={DELPHI: Dual-scale calibrated confidence enables trustworthy
         spatial gene expression prediction from histology},
  author={...},
  journal={Under review},
  year={2025}
}
```