Reactome_GNN is a Python package for creating highly-dimensional representations of human pathway networks, which can then be used in downstream tasks.
The current version uses a multi-layer Graph Convolutional Network (GCN) as a model for obtaining these representations. Several different architectures were trained and compared, but the one with 5 GCN layers and the embedding size of 16 gives the best results. However, hyperparameter optimization is still an open issue in this project and significantly better results could be achieved.
git clone https://github.com/reactome/reactome_gnn.git
cd reactome_gnn
python setup.py install
- Python 3.8
- Pip 20.2.4+
You can find a jupyter notebook with usage instructions in the demo directory. There you can find a step-by-step guide on how to generate networks in several different ways, transform them into DGL graphs, specify the GNN model, and obtain the embeddings.
This project was done as part of Google Summer of Code 2021, under the supervision of Nasim Sanati. The list of my contributions can be found here.
Author: Lovro Vrcek