GNNXGroup

This repository contains the code for the Master thesis 'Decoding Chemical Predictions: Group Contribution Methods for XAI'. This branch focuses on the molecule models. For the reaction models, please visit the reaction_models branch.

To reproduce the results obtained in the study, please follow the steps below. Please make sure you have Python, Pytorch, Pytorch geometric, and RDKit installed. It might be necessary to create some folders for the code to work and store the correct files.

Prepare dataset

python 0_preparing_dataset.py --dataset=[dataset]

Collect the groups' information and generate the group contributions ground truth

python 1_linear_regression.py --dataset=[dataset]

Train the models

python 2_training.py \
    --epochs=1000 \
    --model_name=[model] \
    --model_type=[model_type] \
    --target=[target] \
    --root_dir=[root_dir] 
    --dataset=[dataset] \
    --batch_size=32 \
    --weight_decay=0.9e-16

Extract the results from the trained models

Before extracting the results, please prepare a training_data.json file containing the information of the trained models to be processed. Example of the file is given below:

{
    "gnn_models": [
        "qm9_schnet_original_H_2024-05-14_20-45-44",
        "qm9_schnet_group_gap_2024-05-16_06-11-15",
        "qm9_egnn_group_H_2024-05-25_17-26-24_custom_train",
        "alchemy_egnn_original_gap_2024-05-17_02-47-04",
        "alchemy_schnet_group_gap_2024-05-28_12-57-52_custom_train"
    ],
    "regression_models": [
        "ridge_qm9_benson_H_split_scaffold",
        "ridge_qm9_benson_gap_split_scaffold",
        "ridge_alchemy_benson_H_split_scaffold",
        "ridge_alchemy_benson_gap_split_scaffold"
    ],
    "contribution_models": [
        "qm9_schnet_group_H_2024-05-14_20-45-56_contributions.pkl",
        "qm9_egnn_group_gap_2024-05-17_17-05-33_contributions.pkl",
        "qm9_schnet_group_gap_2024-05-25_17-25-56_custom_train_contributions.pkl",
        "alchemy_schnet_group_H_2024-05-17_02-44-02_contributions.pkl",
        "alchemy_egnn_group_H_2024-05-17_05-23-49_contributions.pkl"
    ]
}

python 3_get_results.py

Generate the XAI plots

python 3_get_contribution_plots.py