Official PyTorch implementation for the manuscript accepted by MIDL (oral):
IST-editing: Infinite spatial transcriptomic editing in a generated gigapixel mouse pup
Jiqing Wu, Ingrid Berg and Viktor H. Koelzer.
Camera-ready version: https://openreview.net/pdf?id=9pBGVsHdzL
If you find this repository helpful for your research, we would appreciate your citation of the following papers.
@inproceedings{wu2024ist,
title={IST-editing: Infinite spatial transcriptomic editing in a generated gigapixel mouse pup},
author={Wu, Jiqing and Berg, Ingrid and Koelzer, Viktor H},
booktitle={International Conference on Medical Imaging with Deep Learning},
year={2024},
organization={PMLR}
}
@article{10.1093/bioinformatics/btae077,
author = {Wu, Jiqing and Koelzer, Viktor H},
title = {SST-editing: in silico spatial transcriptomic editing at single-cell resolution},
journal = {Bioinformatics},
volume = {40},
number = {3},
pages = {btae077},
year = {2024},
month = {02},
issn = {1367-4811},
doi = {10.1093/bioinformatics/btae077},
url = {https://doi.org/10.1093/bioinformatics/btae077},
eprint = {https://academic.oup.com/bioinformatics/article-pdf/40/3/btae077/56850944/btae077.pdf},
}
Conceptual illustrations of the proposed model.
Overview for the generated WSI of the mouse pup.
The code base has been successfully tested under the following configurations:
- Ubuntu 20.04
- Nvidia driver 525.147
- CUDA 11.8
- Python 3.11
- PyTorch 2.1
- Miniconda
Note that lower PyTorch version, E.g., >= 1.13 should also work.
Please check environment.yml for creating the conda environment. In case pyvips package does not work properly, install it via conda commmand instead,
conda install --channel conda-forge pyvips
- Unzip the downloaded mouse.zip to Data/Xenium_mouse folder (create it if not exists).
If you want to process the dataset from scratch, you could first download the raw data via
Please see also prep_inf.py and utils_mouse.py for more implementation details about the data processing pipeline.
Once the mouse dataset is ready, run the train.sh script for training the model on a single GPU (e.g., RTX 2080/3090 Ti)
sh train.sh
In case of multi-gpu training (let's say 2), replace the first line in train.sh with
python -m torch.distributed.launch --nproc_per_node=2 --master_port=4096 train_style2_inf.py \
After complete the training, run the following command to generate the cell-, tissue-and animal-level images including the gigapixel WSI (saved under Path/To/MERG_0_D/*1_53_0_26*.tif)
TASK=all_organs_video-bbox-tif_tile N_ITER=800000 sh infer.sh
For instance, run the following command to conduct direct scaling of gene expressions w.r.t. selected tissue regions
TASK=all_organs_video-scale_tile N_ITER=800000 sh infer.sh
The downloading links to raw and generated WSIs, which can be examined by open-sourced QuPath or commericial software such as HALO
-
Unzip the downloaded brain.zip to Data/MERFISH folder (create it if not exists).
-
To train the model, run
sh train_brn.sh -
To generate the cell-, tissue-and animal-level images including the gigapixel WSI
TASK=all_organs_video-bbox-tif_tile N_ITER=800000 sh infer_brn.sh -
You could examine the generated 73728 X 106496 WSIs for training and test via https://zenodo.org/records/10816781.
This repository is developed upon SST-editing project. Besides, we would like to thank all the researchers contributing to the Xenium mouse pup dataset.
The copyright license of this repository is specified with the LICENSE-IST-editing.