scGPT-spatial: Continual Pretraining of Single-Cell Foundation Model for Spatial Transcriptomics

 

🟩 TL,DR Highlights 🟩

✨ Spatial-omic foundation model ✨ ✨ Continual pretraining of scGPT on 30 million cells/spots ✨

✨ Novel MoE (Mixture of Experts) decoders ✨ ✨ Spatially-aware sampling ✨ ✨ Neighborhood-based reconstruction objective ✨

✨ Curation of SpatialHuman30M corpus ✨ ✨ Visium, Visium HD, Xenium, MERFISH ✨

✨ Multi-modal and multi-slide integration ✨ ✨ Cell-type deconvolution ✨ ✨ Missing gene imputation ✨

🟧 Model Weights 🟧

scGPT-spatial V1 weights on figshare.

🟫 SpatialHuman30M 🟫

Pretraining dataset names, slide metadata, and access links are summarized in data source table. Processed data will be available upon publication given permission under license of the original data source.

🟦 Setup and Tutorials 🟦

To start, clone the current repo:

git clone https://github.com/bowang-lab/scGPT-spatial

Special acknowledgement to the scGPT codebase - for environment setup please follow instructions there.

Check out our zero-shot inference tutorial on github! More code coming soon.

🟪 Preprint and Citation 🟪

Check out our preprint! https://www.biorxiv.org/content/10.1101/2025.02.05.636714v1

@article{wang2025scgpt,
  title={scGPT-spatial: Continual Pretraining of Single-Cell Foundation Model for Spatial Transcriptomics},
  author={Wang, Chloe Xueqi and Cui, Haotian and Zhang, Andrew Hanzhuo and Xie, Ronald and Goodarzi, Hani and Wang, Bo},
  journal={bioRxiv},
  pages={2025--02},
  year={2025},
  publisher={Cold Spring Harbor Laboratory}
}