GenOT: Generative optimal transport enables spatiotemporal interpolation and generation in cross-platform spatial transcriptomics

Introduction

Spatial transcriptomics (ST) technologies have revolutionized the analysis of spatial gene expression patterns within tissues. However, existing computational methods still face challenges in integrating spatial information and generating cross-heterogeneous sample data. To address this, we developed GenOT - a spatial data generation framework based on graph self-supervised contrastive learning and optimal transport theory. The framework employs a multimodal feature learning architecture to dynamically identify important genes and hierarchically aggregate spatial neighborhood information, achieving high-precision spatial domain clustering and biologically interpretable feature extraction. The core innovation of GenOT lies in introducing an optimal transport barycenter-based interpolation algorithm, which mathematically models cross-sample spatial distribution differences to reconstruct spatiotemporal continuous gene expression dynamics. Experiments on multiple datasets including human dorsolateral prefrontal cortex (10x visium), mouse embryonic development (Stereo-seq), and olfactory bulb/hippocampal tissues (Slide-seqV2) demonstrate that GenOT significantly outperforms existing methods in spatial domain identification, cross-technology platform integration, and developmental trajectory reconstruction. This provides an innovative tool for tissue structure analysis at single-cell resolution and developmental process modeling.

Repository Structure

GenOT/
├── GenOT/                  # Core Python package
│   ├── __init__.py         # Package entry point; exports Encoder / DualEncoder / Decoder, etc.
│   ├── genot.py            # Main classes: Encoder (single-slice representation), DualEncoder (dual-slice integration), Decoder (embedding → expression reconstruction)
│   ├── model.py            # Network modules: K-hop aggregation (CombUnweighted), MGCN/MGCN2, Discriminator, AvgReadout
│   ├── preprocess.py       # Preprocessing: normalization, KNN graph construction, contrastive labels, PCA features, adjacency normalization
│   ├── OTutils.py          # Optimal transport: FGW barycenter interpolation, EMD transport plans, mapping expression back to target coordinates
│   ├── utils.py            # Clustering (mclust/leiden, etc.), spatial alignment (PASTE2/ICP), marker-gene color alignment
│   └── plotting.py         # Visualization functions
│
├── Tutorial/               # 8 Jupyter tutorials reproducing the paper's experiments
│   ├── Tutorial 1_DLPFC.ipynb                          # Spatial domain identification (10x Visium)
│   ├── Tutorial 2_MOSTA.ipynb                          # Spatial domain identification (Stereo-seq)
│   ├── Tutorial 3_Mouse_Hippocampus.ipynb             # Mouse hippocampus
│   ├── Tutorial 4_Mouse_Olfactory(...).ipynb          # Olfactory bulb (Stereo-seq & Slide-seqV2)
│   ├── Tutorial 5_Mouse_Brain_Merge_...ipynb          # Anterior–posterior brain slice merging
│   ├── Tutorial 6_DLPFC_interpolation.ipynb           # Spatiotemporal interpolation
│   ├── Tutorial 7_MOSTA_integration.ipynb             # Cross-slice integration
│   └── Tutorial 8_Diff_Tech_MOSTA_integration.ipynb   # Cross-platform integration
│
├── Data/                   # Data instructions (no actual data; download links only)
│   └── README.md
│
├── docs/                   # Sphinx documentation source for the ReadTheDocs site
│   ├── source/
│   ├── Figures/
│   └── ...
│
├── somde/                  # Bundled SOMDE dependency (spatially variable gene detection)
│
├── GenOT.jpg               # Model overview figure (referenced in README)
├── requirements.txt        # Python dependencies
├── smoke_test.py           # End-to-end integration smoke test on synthetic data
├── readthedocs.yaml        # ReadTheDocs build configuration
├── LICENSE                 # MIT license
└── README.md

Environment

First, clone and navigate to the repository.

git clone https://github.com/wrab12/GenOT
cd GenOT                                          

Create and activate a virtual environment using python 3.9 with conda:

conda create -n GenOT python=3.9
conda activate GenOT

Install R + mclust first — rpy2's build step needs R headers, so this must happen before pip install. The conda-forge route works on Linux / macOS / Windows without sudo:

conda install -c conda-forge r-base r-mclust

(If you prefer system R from CRAN, install it now and then run install.packages("mclust") inside R.)

Install somoclu from conda-forge — the pip wheel may ship without the compiled library:

conda install -c conda-forge somoclu

Install GenOT and its Python dependencies:

pip install -e .

This process usually takes around 5 minutes.

Next, install torch and torch_geometric. The installation depends on your CUDA version. Please follow these steps:

1. Run nvidia-smi to check your CUDA version.
2. Install the appropriate version of torch (see the PyTorch installation guide).
3. After installing torch, install torch_geometric (see the PyG installation guide).

A known-working combination on CUDA 12.1 is:

pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 \
    --index-url https://download.pytorch.org/whl/cu121
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv \
    -f https://data.pyg.org/whl/torch-2.4.0+cu121.html
pip install torch_geometric   

Replace 2.4.0 / cu121 with your own torch + CUDA combination if needed.

Testing

After setting up the environment, verify your installation with the included smoke test. It runs both main GenOT pipelines (spatial domain identification and spatiotemporal interpolation) end-to-end on a small synthetic dataset in about a minute, and exits non-zero if anything is broken.

cd GenOT          # repo root (the folder containing the GenOT/ package)
python smoke_test.py

A successful run prints [smoke] ALL PASS and exits with code 0.

Datasets

• Download all datasets from this Google Drive link.

Documentation

See detailed documentation and examples at https://GenOT.readthedocs.io/en/latest/index.html.

Acknowledgements

We thank the developers of GraphST (GitHub), Somde (GitHub), Paste2 (GitHub), and POT (GitHub) for their valuable tools and resources.

License

This project is licensed under the MIT License. See the LICENSE file for details.

References

Wang, R., Liu, X., Zhuo, L. et al. GenOT: generative optimal transport enables spatiotemporal interpolation and generation in cross-platform spatial transcriptomics. Genome Biology (2026). https://doi.org/10.1186/s13059-026-04166-z