NexuST

Paper: NexuST: A Hierarchical Foundation Model for Spatial Transcriptomics (bioRxiv, 2026).

This repository contains the NexuST checkpoint saved at training step 10,000.

Checkpoint

File	Training step	Size
NexuST-step10000.ckpt	10,000	815,205,507 bytes

The uploaded file is byte-identical to the original training artifact step_step=010000.ckpt; only the repository filename was changed. This is the step-10,000 checkpoint, not the separately saved best-validation-loss checkpoint.

SHA256: 178de8d6ad027c2bb699de9e25ebf0dc76915d63701590752fbd8ed85e647d7e

Download

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="Haiping-UoM/NexuST",
    filename="NexuST-step10000.ckpt",
)

This example downloads the checkpoint. Model construction and inference require the corresponding NexuST code, configuration, and vocabulary; a verified loading example is not yet included in this repository.

Pretraining data

The H5AD release is available in the dataset repository HumanST-46M, with separate train/ and val/ directories.

Release status

This is an initial checkpoint release. Architecture details, preprocessing and inference instructions, evaluation results, and license terms have not yet been documented here.

Citation

If you use NexuST or HumanST-46M in your research, please cite:

@article{liu2026nexust,
  title   = {NexuST: A Hierarchical Foundation Model for Spatial Transcriptomics},
  author  = {Liu, Haiping and Zhao, Qian and Lin, Lijing and Zou, Zhiyong and Cai, Wenhao and Sun, Jingyuan and Zhou, Yuxi and Alvarez, Mauricio A. and Gilmore, Andrew and Rattray, Magnus and Frangi, Alejandro F. and Zhou, Hongpeng},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.09.22.753590}
}