SpatialFusion is a Python package for deep learning–based analysis of spatial omics data. It provides a lightweight framework that integrates spatial transcriptomics (ST) with H&E histopathology to learn joint multimodal embeddings of cellular neighborhoods and group them into spatial niches.

The method operates at single-cell resolution, and can be applied to:

• paired ST + H&E datasets
• H&E whole-slide images alone

By combining molecular and morphological features, SpatialFusion captures coordinated patterns of tissue architecture and gene expression. A key design principle is a biologically informed definition of niches: not simply spatial neighborhoods, but reproducible microenvironments characterized by pathway-level activation signatures and functional coherence across tissues. To reflect this prior, the latent space of the model is trained to encode biologically meaningful pathway activations, enabling robust discovery of integrated niches.

The method is described in the paper: XXX (citation forthcoming).

We provide pretrained weights for the multimodal autoencoder (AE) and graph convolutional masked autoencoder (GCN) under data/.

SpatialFusion depends on PyTorch and DGL, which have different builds for CPU and GPU systems. You can install it using pip or inside a conda/mamba environment.

mamba create -n spatialfusion python=3.10 -y
mamba activate spatialfusion
# Then install GPU or CPU version below

pip install "torch==2.4.1" "torchvision==0.19.1" \
  --index-url https://download.pytorch.org/whl/cu124
conda install -c dglteam/label/th24_cu124 dgl

Note: TorchText issues exist for this version: pytorch/text#2272 — this may affect scGPT.

pip install torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 \
  --index-url https://download.pytorch.org/whl/cu121
conda install -c dglteam/label/th21_cu121 dgl

# Optional: embeddings used by scGPT
pip install --no-cache-dir torchtext==0.18.0 torchdata==0.9.0

# Optional: UNI (H&E embedding model)
pip install timm

pip install "torch==2.4.1" "torchvision==0.19.1" \
  --index-url https://download.pytorch.org/whl/cpu
conda install -c dglteam -c conda-forge dgl

# Optional, used for scGPT
pip install --no-cache-dir torchtext==0.18.0 torchdata==0.9.0

# Optional, used for UNI
pip install timm

💡 Replace cu124 with the CUDA version matching your system (e.g., cu121).

cd spatialfusion/
pip install -e .

Includes: pytest, black, ruff, sphinx, matplotlib, seaborn.

cd spatialfusion/
pip install -e ".[dev,docs]"

python - <<'PY'
import torch, dgl, spatialfusion
print("Torch:", torch.__version__, "CUDA available:", torch.cuda.is_available())
print("DGL:", dgl.__version__)
print("SpatialFusion OK")
PY

• Default output directory is:

  $HOME/spatialfusion_runs
  

  Override with:

  export SPATIALFUSION_ROOT=/your/path
  

• CPU installations work everywhere but are significantly slower.

A minimal example showing how to embed a dataset using the pretrained AE and GCN:

from spatialfusion.embed.embed import AEInputs, run_full_embedding
import pandas as pd
import pathlib as pl

# Load external embeddings (UNI + scGPT)
uni_df = pd.read_parquet('UNI.parquet')
scgpt_df = pd.read_parquet('scGPT.parquet')

# Paths to pretrained models
ae_model_dir = pl.Path('../data/checkpoint_dir_ae/')
gcn_model_dir = pl.Path('../data/checkpoint_dir_gcn/')

# Mapping sample_name -> AEInputs
ae_inputs_by_sample = {
    sample_name: AEInputs(
        adata=adata,
        z_uni=uni_df,
        z_scgpt=scgpt_df,
    ),
}

# Run the multimodal embedding pipeline
emb_df = run_full_embedding(
    ae_inputs_by_sample=ae_inputs_by_sample,
    ae_model_path=ae_model_dir / "spatialfusion-multimodal-ae.pt",
    gcn_model_path=gcn_model_dir / "spatialfusion-full-gcn.pt",
    device="cuda:0",
    combine_mode="average",
    spatial_key='spatial',
    celltype_key='major_celltype',
    save_ae_dir=None,  # optional
)

This produces a DataFrame containing the final integrated embedding for all cells/nuclei.

SpatialFusion operates on a single-cell AnnData object paired with an H&E whole-slide image.

A high-resolution H&E image corresponding to the same tissue section used for ST. Used to compute morphology embeddings such as UNI.

1. Prepare ST AnnData and the matched H&E WSI

2. Run scGPT to compute molecular embeddings

3. Run UNI to compute morphology embeddings

4. Run SpatialFusion to integrate all modalities into joint embeddings

5. Cluster & visualize

   • Leiden clustering
   • UMAP
   • Spatial niche maps

A complete tutorial notebook is available at:

tutorials/embed-and-finetune-sample.ipynb

Additional required packages (scGPT, UNI dependencies) must be installed manually. Follow the instructions at: https://github.com/bowang-lab/scGPT

We also provide a ready-to-use environment file:

spatialfusion_env.yml

Tutorial data is available on Zenodo: https://zenodo.org/records/17594071

.
├── data
│   ├── checkpoint_dir_ae
│   │   └── spatialfusion-multimodal-ae.pt
│   └── checkpoint_dir_gcn
│       ├── spatialfusion-full-gcn.pt
│       └── spatialfusion-he-gcn.pt
├── LICENSE
├── pyproject.toml
├── README.md
├── src
│   └── spatialfusion
│       ├── embed/
│       ├── finetune/
│       ├── models/
│       └── utils/
├── tests
│   ├── test_basic.py
│   ├── test_finetune.py
│   └── test_imports.py
└── tutorials
    ├── data
    └── embed-and-finetune-sample.ipynb

Highlights:

• data/ — pretrained AE and GCN checkpoints

• src/spatialfusion/ — main library modules

  • embed/ — embedding utilities & pipeline
  • finetune/ — niche-level finetuning
  • models/ — neural network architectures
  • utils/ — loaders, graph utilities, checkpoint code

• tests/ — basic test suite

• tutorials/ — practical examples and sample data

If you use SpatialFusion, please cite:

Broad Institute Spatial Foundation, SpatialFusion (2025). https://github.com/broadinstitute/spatialfusion

Full manuscript citation will be added when available.

This is the initial public release (v0.1.0).

MIT License. See LICENSE for details.