Chioso: Segmentation-free Annotation of Spatial Transcriptomics (ST) at Sub-cellular Resolution

Key Features of Chioso

• Pixel based annotation at subcellular-resolution
• Does NOT need cell segmentation input
• Scalable to very large dataset (e.g. MOSTA dataset: 20 billion RNA reads, 1 billion locations, < 5 hours wall time)

Installation

pip install git+https://github.com/jiyuuchc/chioso.git

Usage

0. Required Inputs

1. ScRNAseq with cell type annotation in h5ad format
2. Spatial data in space-deliminated text format with four feature columns: gene, x, y, counts.
3. Common genes in both datasets (or a subset genes of interests) as a list of string saved in a JSON file

1. Convert input data to more efficient formats

python -m chioso.pp-ref --data <h5ad file> --genes <gene file> --outdir <outdir>

# repeat if more than one input file
python -m chioso.pp-spatial --data <st text file> --genes <gene file> --outdir <outdir> 

2. Train predictive model based on the reference data

python -m chioso.train-predictor --config <cfg_predictor.py>

Default config files are under the configs/

3. Train generative model on spatial data and reference data

python -m chioso.train-chioso --config <cfg_chioso.py>

Default config files are under the configs/

4. Inference

python -m chioso.inference --config <cfg_chioso.py> --checkpoint <model checkpoint>