This repository contains the code for the work on protein-ligand interaction with GNNs, in which we investigate what GNNs learn by using explainable AI (XAI).

This code relies on the usage of EdgeSHAPer as an XAI method. However, there is no need to install it since all the files needed are provided in the src folder.

We suggest to run the code in a conda environment. We provide an environment.yml file that can be used to install the needed packages:

conda env create -f environment.yml

Note: this file was created from a conda working environment under Windows 11, so some packages may not be available in different versions/OS. We suggest to manually install them in this case. Edit the file with your system conda envs folder. We also provide the environment_no_builds.yml file for a higher cross-platform compatibility (minor manual edits to the file may be needed).

If you encounter problems during the installation of the PyTorch Geometric dependencies, manually install them in the conda environment using:

pip install torch_sparse==0.6.15 torch_cluster==1.6.0 torch_spline_conv==1.2.1 torch_geometric==2.1.0.post1 -f https://data.pyg.org/whl/torch-1.12.0+cu116.html

Note that the versions above are the ones found in the environment.yml file. We suggest to install such version for reproducibility, but you are encouraged to try different versions to check for compatibility!

For compatibility with EdgeSHAPer code, this additional module must be installed.

Finally, download the data we used in our experiments from here (copy and paste in a browser the link address if the download does not start) and nzipped them into the data.

The config file parameters.yml contains settable parameters that are loaded by the scripts provided in the repo.

We provide the trainer_script.py file to train custom GNN models (check parameters.yml for details).

To train your model simply run:

python trainer_script.py

Alternatively, to reproduce the experiments, you can use the pretrained models available in models/pretrained_models folder.

Using one of the pretrained models, or any custom model trained with trainer_script.py, you can use explainer_script.py to run EdgeSHAPer to generate explanations in terms of important edges. Among other parameters found in parameters.yml, AFFINITY_SET defines the affinity set for which to run the explanations (low, medium or high). To explain the predictions of the model defined by the MODEL_PATH parameter for the selected affinity set, simply run:

python explainer_script.py

Files contatining edge importance as Shapley value estimates by EdgeSHAPer will be saved into the folder specified by the SAVE_FOLDER parameters.

Once the explanations for all the affinity sets are obtained, it is possible to run:

python top_k_computation.py

to generate statistics and graphics for the top-k important edges of the samples explained. Additional parameters accepted can be found in the config file parameters.yml.

A possible top-k edges (k=25) output image may look as:

The folder additional_scripts contains scripts for additional experiments and computations. A README file is provided in the folder with instructions.

For any questions, feel free to drop an email.

If you use our work or results, please cite our publication in Nature Machine Intelligence 🧠

Mastropietro, A., Pasculli, G. & Bajorath, J. Learning characteristics of graph neural networks predicting protein–ligand affinities. Nat Mach Intell (2023). https://doi.org/10.1038/s42256-023-00756-9