README.md

---
pretty_name: GHIST+ Data and Model Bundle
license: other
tags:
- spatial-transcriptomics
- computational-pathology
- xenium
---

# GHIST+ data and model bundle

This bundle contains the model artifacts, predictions, evaluation inputs,
comparison outputs, and plot-ready tables released with
[GHIST+](https://github.com/SydneyBioX/GHIST_plus). Source code is provided in
that repository.

## Download

    hf download SydneyBioX/GHIST-Plus-bundle \
      --repo-type dataset \
      --local-dir bundle

The bundle is approximately 38 GB. Individual files can also be downloaded
from this page.

## Use with the figure notebooks

These notebooks regenerate figures from released expression predictions and
analysis tables. Training and checkpoint reconstruction are not required.

Complete the GitHub [installation](https://github.com/SydneyBioX/GHIST_plus#installation)
and [notebook setup](https://github.com/SydneyBioX/GHIST_plus#reproduce-figures-35),
then run:

```bash
conda activate model_env
cd GHIST_plus
export GHIST_BUNDLE_ROOT="/path/to/bundle"
python -m jupyterlab
```

Set `GHIST_BUNDLE_ROOT` to the downloaded bundle directory containing
`GHIST_plus/`, `evaluation_data/`, `figure_data/`, and `other_models/`.

Open the desired `Figure3.ipynb`–`Figure5.ipynb`, select **GHIST+ (model_env)**,
and choose **Restart Kernel and Run All**. Figures appear inside the notebook;
save the notebook to keep its outputs.

## Model checkpoints

The four released GHIST+ checkpoint files are:

- GHIST_plus/models/breast_multi/ghist_plus_breast_multi_checkpoint.pth
- GHIST_plus/models/breast_single/ghist_plus_breast_single_checkpoint.pth
- GHIST_plus/models/imputation/ghist_plus_gene_imputation_checkpoint.pth
- GHIST_plus/models/pancancer/ghist_plus_pancancer_checkpoint.pth

They exclude the frozen third-party UNI2-H encoder weights. Users must obtain
UNI2-H directly from
[MahmoodLab/UNI2-h](https://huggingface.co/MahmoodLab/UNI2-h), accept its
terms, and follow the **Pretrained Checkpoints** instructions in the
[GHIST+ README](https://github.com/SydneyBioX/GHIST_plus#pretrained-checkpoints).
That procedure reconstructs the complete checkpoints locally at the same
filenames, so the existing inference commands and configs remain unchanged.

The bundle does not redistribute UNI2-H weights.

## Contents

- **Figure 3:** evaluation data, GHIST+ predictions, and comparison-model
  predictions under `evaluation_data/`, `GHIST_plus/predictions/`, and
  `other_models/`.
- **Figure 4:** imputation inputs and predictions, including the bundled
  VQ/composition ablation predictions.
- **Figure 5:** paired PCC and coverage tables under `figure_data/figure5/`.

    bundle/
    ├── GHIST_plus/
    │   ├── models/
    │   └── predictions/
    ├── evaluation_data/
    ├── figure_data/
    └── other_models/

The paths and lightweight input schemas were checked against the released
Figure 3–5 notebooks.

## Data and third-party terms

The bundle combines author-generated artifacts, processed public source data,
and outputs from comparison methods. It therefore has no single blanket
license; the Hugging Face license field is other. See
THIRD_PARTY_NOTICES.md for component-specific sources, versions, attributions,
and terms. Upstream terms remain applicable.

## Tutorial data

tutorial.ipynb does not use this bundle as its DATA_ROOT. The tutorial requires
a separately prepared GHIST data directory containing aligned H&E images,
segmentation masks, nuclei metadata, and inputs for the selected training mode.