soGEN

soGEN generates realistic omics readouts (RNA, surface protein, chromatin accessibility, ...) from spatial niche context — e.g. the multi-scale neighborhood composition around each cell or spot — using zero-inflated / count / continuous likelihood models. It supports both single-modal (one omics readout) and multi-modal (two jointly modeled readouts) generation.

Full documentation: https://baolab-fudan.github.io/soGEN/ · Demo notebooks: single-modal · multi-modal

Installation

pip install sogen-bio

Requires Python >= 3.9 and PyTorch (installed automatically as a dependency).

Quickstart

from sogen import soGEN

# niches: list of niche/context feature matrices (DataFrame or array), one row per spot/cell
# x: target omics matrix (DataFrame or array), one row per spot/cell, one column per feature
model = soGEN(niches, x, mode="ZINB", device="cpu")
model.fit(niches, x, epochs=300, lr=1e-4, batch_size=256)

generated = model.generate(niches_new)   # DataFrame, same columns as x
model.save_model("model.pt")

For two jointly modeled omics readouts (e.g. RNA + protein):

from sogen import soGEN_multimodal

model = soGEN_multimodal(niches, x1, x2, mode="ZINBZIPoisson", device="cpu")
model.fit(niches, x1, x2, epochs=300, lr=5e-4, batch_size=256)

x1_gen, x2_gen = model.generate(niches_new)

See examples/demo_1_single_modal.ipynb (Visium spatial transcriptomics) and examples/demo_2_multimodal.ipynb (spatial CITE-seq, RNA + protein) for complete, runnable walkthroughs on real data, including how to build niche features from spatial coordinates and cluster labels.

Choosing a distribution mode

Development

git clone https://github.com/BaoLab-fudan/soGEN.git
cd soGEN
pip install -e ".[dev,examples]"
python -m pytest tests/

Citation

If you use soGEN in your research, please cite the accompanying manuscript (Bao Lab, Fudan University). Citation details will be added upon publication.

License

MIT — see LICENSE.