EdgeSHAPer: Bond-centric Shapley value-based explanation method for graph neural networks

Protocol to explain graph neural network predictions using an edge-centric Shapley value-based approach

“…An exemplary file is shown in Figure 2 (having same format for training, validation, and test sets). Note: The repository contains data from the original publication, 1 which can be used to test the algorithm. …”

“…Note: The repository contains data from the original publication, 1 which can be used to test the algorithm.…”

“… Note: At the time of writing, the default values found in the parameters.yml file are the ones used in Mastropietro et al. 1 Note: If training, validation, and test data are not provided, the complete dataset will be divided into training, validation, and test sets according to an 80%:10%:10% ratio. In this case, it might be useful to specify the parameter SAVE_FOLDER_DATA_SPLIT in order to save the resulting data split.…”

“…The method was implemented using Python code with the aid of several deep learning libraries such as PyTorch 4 and PyTorch Geometric. 5 Below we illustrate the installation and workflow with compound data from the original EdgeSHAPer publication 1 as well as with custom data. The installation and execution times reported are based on a machine with the capabilities listed in the materials and equipment section.…”

“…3 https://github.com/AndMastro/EdgeSHAPer/tree/main/experiments/data Mendeley Data: https://doi.org/10.17632/bs6myg75tr.2 Software and algorithms RDKit 2021.09.4 Zenodo Zenodo: https://doi.org/10.5281/zenodo.6605135 Lilly-Medchem-Rules GitHub https://github.com/IanAWatson/Lilly-Medchem-Rules EdgeSHAPer This paper Mastropietro et al. 1 Zenodo 2 https://github.com/AndMastro/EdgeSHAPer Zenodo: https://doi.org/10.5281/zenodo.7267068 scikit-learn 1.0.2 GitHub https://github.com/scikit-learn/scikit-learn PyTorch 1.10.1 or PyTorch 1.10.1_cuda10.2 or PyTorch 1.12.1 or PyTorch 1.12.1_cuda11.6 (use CUDA versions if a GPU is available) GitHub https://github.com/pytorch/pytorch rdkit-heatmap 0.1 GitHub https://github.com/c-feldmann/rdkit_heatmaps matplotlib 3.5.1 GitHub https://github.com/matplotlib/matplotlib networkx 2.6.3 GitHub https://github.com/networkx/networkx numpy 1.22.0 GitHub https://github.com/numpy/numpy tqdm 4.62.3 GitHub https://github.com/tqdm/tqdm pyg 2.0.3 GitHub https://github.com/pyg-team/pytorch_geometric torchdrug 0.1.2 GitHub https://github.com/DeepGraphLearning/torchdrug …”