SpaCeNet

This package implements SpaCeNet. It is an extension of the traditional Gaussian Graphical Model which allows modelling of spatially distributed observations and the associations of their variables.

Setup

You can either install the package directly from github via

pip install git+https://github.com/sschrod/SpaCeNet.git

or use the provided Docker image

docker build -t spacenet -f DOCKERFILE .

SpaCeNet is implemented using PyTorch Tensors to facilitate efficient matrix computations on the GPU. Hence, we suggest setting up PyTorch with CUDA.

Get started

For a quick introduction we prepared a jupyter notebook demonstrating the inference of spatial relationships on simulated data with SpaCeNet.

The simplest way to analyse your own data using SpaCeNet is with SpaCeNet_main.py. Simply save the preprocessed data in a compressed numpy format

np.savez(<file path>, X_mat=<Gene matrix (Sxnxp)>, coord_mat=<associated coordinates (Nxp)>, GeneNames=<List of Gene names (optional)>)

and call SpaCeNet_main.py with --exp_name, --data_path, --preprocessed_data, --results_folder. Further, set -gs True to run a grid-search, -sr True to run a single SpaCeNet model or -ar True for some general analysis. For the full list of argparse arguments refer to SpaCeNet_main.py

To run a hyper-parameter search and analyse the findings on the simulated data run

python3 SpaCeNet_main.py --exp_name Simulation --data_path example --preprocessed_data simulated_data.pickle --results_folder sim_results/ -ss 1e-6 -e 1e-5 -l 3 -gs True -ar True

To reproduce the results on the MOSTA data, download the data (instructions in \data), run preprocess_MOSTA.py to save the preprocessed data and call

python3 SpaCeNet_main.py --exp_name MOSTA30 --data_path data --preprocessed_data MouseBrainAdult_30Percent.npz --results_folder results/ -nr 2 -st True -gs True -ar True```

References

Lück, N., Lohmayer, R., Solbrig, S., Schrod, S., Wipfler, T., Shutta, K.H., Guebila, M.B., Schäfer, A., Beißbarth, T., Zacharias, H.U. and Oefner, P.J., 2022. SpaCeNet: Spatial Cellular Networks from omics data. bioRxiv, pp.2022-09.