SUICA: Learning Super-high Dimensional Sparse Implicit Neural Representations for Spatial Transcriptomics
Qingtian Zhu1*, Yumin Zheng2,3*, Yuling Sang4, Yifan Zhan1, Ziyan Zhu5, Jun Ding2,3,6, Yinqiang Zheng1
1The University of Tokyo, 2McGill University, 3MUHC Research Institute, 4Duke-NUS Medical School,
5Carnegie Mellon University, 6Mila-Quebec AI Institute
ICML 2025
SUICA pipeline
conda create -n SUICA python=3.9 -y && conda activate SUICA
pip install -r requirements.txtThe typical data structure is as follows:
|-- custom_root_path
|-- configs #configuration file for GAE training and GAE-INR joint training
|-- ST
|--embedder_gae.yaml
|--inr_embd.yaml
|-- data
|--preprocessed_data
|--your_data.h5ad
|logs
|--networks
|--scripts
|--systems
|--datasets.py
|--train.py # The main file
|--utils.py
Train the Graph AutoEncoder (GAE)
python train.py --mode embedder --conf ./configs/ST/embedder_gae.yaml
Train the GAE-INR
python train.py --mode inr --conf ./configs/ST/inr_embd.yaml
@article{zhu2024suica,
title={SUICA: Learning Super-high Dimensional Sparse Implicit Neural Representations for Spatial Transcriptomics},
author={Zhu, Qingtian and Zheng, Yumin and Sang, Yuling and Zhan, Yifan and Zhu, Ziyan and Ding, Jun and Zheng, Yinqiang},
journal={arXiv preprint arXiv:2412.01124},
year={2024}
}