SpatialFusion

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
• spatial transcriptomics data 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: SpatialFusion: A lightweight multimodal framework for pathway-informed spatial niche mapping.

You can find detailed documentation at https://uhlerlab.github.io/spatialfusion/

---

Prepare SpatialFusion inputs

Before running SpatialFusion, you need to generate unimodal embeddings from two modalities. You can choose which foundation model to use for each:

Modality	Supported models
H&E / whole-slide image	UNI · Virchow
Spatial transcriptomics	scGPT · Nicheformer

We provide Dockstore workflows for generating these embeddings, detailed on our documentation website.

Installation

1. Create virtual environment

mamba create -n spatialfusion python=3.10 -y
mamba activate spatialfusion

2. Install platform-specific libraries

SpatialFusion depends on PyTorch and DGL, which have different builds for CPU and GPU systems.

CPU

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

pip install dgl -f https://data.dgl.ai/wheels/torch-2.4/repo.html

GPU (CUDA 12.4)

pip install "torch==2.4.1" "torchvision==0.19.1" \
  --index-url https://download.pytorch.org/whl/cu124

pip install dgl -f https://data.dgl.ai/wheels/torch-2.4/cu124/repo.html

---

3. Install SpatialFusion package

Basic installation — Recommended for users

pip install spatialfusion

---

Install from source - Recommended for contributors

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

git clone https://github.com/uhlerlab/spatialfusion.git

cd spatialfusion/

pip install -e ".[dev,docs]"

---

4. Verify Installation

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

---

5. Notes

• Default output directory is:

  $HOME/spatialfusion_runs
  

  Override with:

  export SPATIALFUSION_ROOT=/your/path
  

• CPU installations work everywhere but are significantly slower.

---

Tutorials

A complete tutorial notebook to embed / finetune SpatialFusion on a sample is available at:

tutorials/embed-and-finetune-sample.ipynb

To get unimodal embeddings, we provide instructions or a tutorial at:

tutorials/get-unimodal-embeddings-sample.ipynb

To run SpatialFusion on H&E only, we provide a tutorial at:

tutorials/embed-he-only.ipynb

To run SpatialFusion on RNA only, we provide a tutorial at:

tutorials/embed-rna-only.ipynb

Additional packages for the foundation models you choose must be installed manually:

• scGPT: https://github.com/bowang-lab/scGPT
• Nicheformer: https://github.com/theislab/nicheformer
• UNI: https://huggingface.co/MahmoodLab/UNI2-h
• Virchow: https://huggingface.co/paige-ai/Virchow2

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

spatialfusion_env.yml

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

---

Pretrained Weights

SpatialFusion ships pretrained checkpoints for every combination of supported foundation models. Select the pair that matches the embeddings you generated:

H&E model	RNA model	AE checkpoint	GCN checkpoint
UNI	scGPT	spatialfusion-ae-uni-scgpt.pt	spatialfusion-gcn-uni-scgpt.pt
UNI	Nicheformer	spatialfusion-ae-uni-nicheformer.pt	spatialfusion-gcn-uni-nicheformer.pt
Virchow	scGPT	spatialfusion-ae-virchow-scgpt.pt	spatialfusion-gcn-virchow-scgpt.pt
Virchow	Nicheformer	spatialfusion-ae-virchow-nicheformer.pt	spatialfusion-gcn-virchow-nicheformer.pt
UNI	(H&E only)	—	spatialfusion-he-gcn-uni-scgpt.pt
(RNA only)	scGPT	—	spatialfusion-rna-gcn-uni-scgpt.pt

All checkpoints are bundled with the package under src/spatialfusion/data/ and are resolved automatically via spatialfusion.utils.pkg_ckpt.resolve_pkg_ckpt.

---

Usage Example

A minimal example showing how to embed a dataset using the pretrained AE and GCN. Swap in the checkpoint filenames that match your chosen foundation models (see the table above):

from spatialfusion.embed.embed import AEInputs, run_full_embedding
from spatialfusion.utils.pkg_ckpt import resolve_pkg_ckpt
import pandas as pd
import scanpy as sc

# Load H&E embeddings (UNI or Virchow — use whichever you generated)
he_df = pd.read_parquet('UNI.parquet')       # or 'Virchow2.parquet'

# Load RNA embeddings (scGPT or Nicheformer — use whichever you generated)
rna_df = pd.read_parquet('scGPT.parquet')    # or 'nicheformer.parquet'

# Load AnnData object
adata = sc.read_h5ad("object.h5ad")

# Mapping sample_name -> AEInputs
# z_he holds H&E embeddings (UNI or Virchow); z_rna holds RNA embeddings (scGPT or Nicheformer)
sample_name = 'sample1'
ae_inputs_by_sample = {
    sample_name: AEInputs(
        adata=adata,
        z_he=he_df,
        z_rna=rna_df,
    ),
}

# Select the checkpoints matching your FM combination (e.g. UNI + scGPT shown here)
ae_ckpt  = resolve_pkg_ckpt("checkpoint_dir_ae/spatialfusion-ae-uni-scgpt.pt")
gcn_ckpt = resolve_pkg_ckpt("checkpoint_dir_gcn/spatialfusion-gcn-uni-scgpt.pt")

# Run the multimodal embedding pipeline
# spatial_key: key in adata.obsm holding X/Y coordinates — set to match your AnnData
#   (e.g. check with list(adata.obsm.keys()))
# celltype_key: key in adata.obs holding cell type labels — set to match your AnnData
#   (e.g. check with adata.obs.columns.tolist())
emb_df = run_full_embedding(
    ae_inputs_by_sample=ae_inputs_by_sample,
    ae_model_path=ae_ckpt,
    gcn_model_path=gcn_ckpt,
    device="cuda:0",  # or "cpu"
    combine_mode="average",
    spatial_key='spatial_px',
    celltype_key='celltypes',
    save_ae_dir=None,  # optional
)

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

---

Required Inputs

SpatialFusion operates on a single-cell AnnData object paired with an H&E whole-slide image. It also accepts only a WSI with cell coordinates in the H&E only mode, or only an AnnData object with RNA embeddings (no WSI) in the RNA only mode.

AnnData fields

Key	Description
adata.obsm['spatial_px']	X/Y centroid coordinates of each cell/nucleus in WSI pixel space. This is the default key expected by SpatialFusion; pass spatial_key= to run_full_embedding if your AnnData uses a different key (check with list(adata.obsm.keys())).
adata.X	Raw counts (cell × gene). Must be single-cell resolution.
adata.obs['celltypes'] (optional)	Annotated cell types. This is the default key; pass celltype_key= to run_full_embedding if your AnnData uses a different column name (check with adata.obs.columns.tolist()).

Whole-Slide Image (WSI)

A high-resolution H&E image corresponding to the same tissue section used for ST. Used to compute morphology embeddings with your chosen H&E foundation model (UNI or Virchow).

---

Typical Workflow

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

2. Run your RNA foundation model (scGPT or Nicheformer) to compute molecular embeddings

3. Run your H&E foundation model (UNI or Virchow) to compute morphology embeddings

4. Select the matching pretrained checkpoint pair (see Pretrained Weights above)

5. Run SpatialFusion to integrate all modalities into joint embeddings

6. Cluster & visualize

   • Leiden clustering
   • UMAP
   • Spatial niche maps

---

Repository Structure

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

Highlights:

• src/spatialfusion/data/ — packaged 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

---

Citing

If you use SpatialFusion, please cite:

Yates J, Shavakhi M, Choueiri T, Camp SY, Van Allen EM, Uhler C. SpatialFusion: A lightweight multimodal framework for pathway-informed spatial niche mapping. bioRxiv. 2026. doi:10.64898/2026.03.16.712056

---

Version

Version

This release adds RNA-only SpatialFusion support (v0.3.0).

---

License

MIT License. See LICENSE for details.