HEIST

HEIST

Pre-trained checkpoint for HEIST: Hierarchical Embeddings for Integrated Spatial Transcriptomics (ICLR 2026).

Model code, training scripts, and tutorials live on GitHub: https://github.com/Graph-and-Geometric-Learning/HEIST

Models

v2.0 was pretrained on rebuilt data that fixes a preprocessing bug in v1, where every cell of a cell type shared one gene-expression vector. v2 models embed each gene by its index in v2.0/gene_vocab.json (6,495 human gene symbols). Genes outside this vocabulary are ignored, so datasets with partly unknown panels still work. Mouse symbols need mapping to human orthologs first.

Usage

Clone the repo and install dependencies:

git clone https://github.com/Graph-and-Geometric-Learning/HEIST
cd HEIST
pip install -e .

Embed a dataset (raw counts in .X, coordinates in .obsm['spatial']; needs a CUDA GPU):

python embed.py --adata sample.h5ad --out sample_heist.h5ad --model v2.0/marker_all

or from Python:

from utils.inference import load_heist, embed_adata

model, gene_vocab = load_heist("v2.0/marker_all")   # "v1" loads the original model
adata.obsm["X_heist"] = embed_adata(adata, model, gene_vocab, save_root="data/preprocessed", name="sample")

embed_adata runs the same preprocessing as the v2 pretraining data: spatial tiling of large tissues, top-200 HVGs, Leiden clusters, MAGIC, a Voronoi cell graph, and GPU KSG mutual-information gene networks. For a walkthrough with PHATE visualization, see cell_embeddings.ipynb.

Citation

@inproceedings{madhu2026heist,
  title={{HEIST}: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data},
  author={Madhu, Hiren and Rocha, Jo{\~a}o Felipe and Huang, Tinglin and Viswanath, Siddharth and Krishnaswamy, Smita and Ying, Rex},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=lK82jpa8jr}
}