Implementation of P-Net as a flexible deep learning tool to generate insights from genetic features.
Current pytorch implementation in revision.
Pnet uses the Reactome hierarchical graph as underlying structure to reduce the number of connections in a fully connected feed forward neural network. The sparse layers connect only known pathways. This limits the number of parameters to be learnt in a meaningful way and facilitate learning via gradient descent and leads to more generalizable models.
- Clone the github repository and navigate into it.
- Create the conda environment with
conda env create -f pnet.yml, activate it withconda activate pnet - Run
pip install -e .to install the package locally. - To check successful installation run
python test/test_data_loading.pywhich will verify basic import and file structure.
For further functional testing see the testing notebook
Detailed sepcific usage examples are provided in the notebooks. Generally the network structure expects gene level data for each sample (e.g. read counts, CNA indication etc.). different data modalities can be concatenated as a dictonary and passed to the pnet_loader object. A good starting place to familiarize yourself with the usage of pnet is this example notebook