NOTE: THIS CODE WAS ORIGINALLY FORKED FROM KEVIN MICHALEWICZ

GCN-Bf

1. Create a conda environment, install all the dependencies, and then:

git clone https://github.com/kevinmicha/GCN-Bf
cd GCN-Bf
pip install .

List of dependencies:

• logomaker
• matplotlib
• numpy
• pandas
• requests
• scikit-learn
• scipy
• torch
• torch_geometric
• transformers

2. Request three folders of data from the authors:

• adjacencies_sparse (go to gcn_bf/config.py and update ADJACENCIES_DIR accordingly)
• contact_maps (go to gcn_bf/config.py and update CM_DIR accordingly)
• chothia_ext (go to gcn_bf/config.py and update STRUCTURE_DIR accordingly)

3. Go to scripts and run in the following order (graphs can be gnm or bagpype):

• python generate_epitope.py [--thr THR]
• python generate_input.py
• python generate_graph_arrays.py [--graphs GRAPHS] [--noise NOISE]
• python training_script_epitope.py [--graphs GRAPHS] [--lr LR] [--hidden_channels HIDDEN_CHANNELS] [--epochs EPOCHS] [--seq_only] [--noise NOISE]

4. For evaluation and interpretability, once models and data are saved in the checkpoints folder:

• python evaluation_epitope.py [--graphs GRAPHS] [--lr LR] [--hidden_channels HIDDEN_CHANNELS] [--epochs EPOCHS] [--noise NOISE]
• python evaluation_test.py [--graphs GRAPHS] [--lr LR] [--hidden_channels HIDDEN_CHANNELS] [--epochs EPOCHS] [--noise NOISE] [--weighted]

5. Further insights can be found in the notebook Interpretability.