Xiang-zx-zx/xenium-senescence-data

# Xenium Senescence Dataset

Cell morphology images (64x64 grayscale) from Xenium spatial transcriptomics, with senescence scores.

## Structure

```
data_ccdm/
xenium_{tissue}_64x64_aligned.h5 # Images + senescence scores (aligned version)
checkpoints/
latents_{tissue}.npy # VAE-encoded latents (128-dim)
ot_pairs_{tissue}_typed.npy # Within-cell-type OT pairs (cosine cost)
celltype_ids_{tissue}.npy # Integer cell type IDs
celltype_map_{tissue}.npy # CT name → ID mapping
sig_metric_map_lung_full.json # Per-CT primary metric + direction
```

## Tissues

| Tissue | Cells | H5 Size |
|--------|------:|--------:|
| lung | 41,702 | 28M |
| brain | 9,960 | 6.2M |
| cervical | 96,340 | 76M |
| prostate | 29,280 | 19M |
| skin | 17,396 | 11M |
| ovary_cancer | 23,880 | 20M |

## Usage

```python
from huggingface_hub import hf_hub_download
import numpy as np, h5py

# Download Lung data
h5_path = hf_hub_download("Xiang-zx-zx/xenium-senescence-data", "data_ccdm/xenium_lung_64x64_aligned.h5", repo_type="dataset")
latents = np.load(hf_hub_download("Xiang-zx-zx/xenium-senescence-data", "checkpoints/latents_lung.npy", repo_type="dataset"))
ot_pairs = np.load(hf_hub_download("Xiang-zx-zx/xenium-senescence-data", "checkpoints/ot_pairs_lung_typed.npy", repo_type="dataset"))
ct_ids = np.load(hf_hub_download("Xiang-zx-zx/xenium-senescence-data", "checkpoints/celltype_ids_lung.npy", repo_type="dataset"))

with h5py.File(h5_path) as f:
images = f['images'][:] # (N, 1, 64, 64) uint8
labels = f['labels_norm'][:] # (N,) float [0,1]
```

## H5 File Format

Each aligned H5 contains:
- `images`: (N, 1, 64, 64) uint8 — grayscale cell patches
- `labels_raw`: (N,) float — raw senescence scores
- `labels_norm`: (N,) float [0, 1] — normalized scores

## OT Pairs

`ot_pairs_{tissue}_typed.npy` shape (M, 2): pairs of cell indices where:
- Both cells are same cell type (typed/within-CT OT)
- Cosine-similarity cost in latent space
- Direction-filtered: primary metric moves in expected direction