interpretable_yield_prediction

Pytorch implementation of the method described in the paper Interpretation of chemical reaction yields with graph neural additive network

Overview

• This study introduces an interpretable chemical reaction yield prediction method, which models the overall yield as a summation of component-wise contributions from each reaction participant. To enable interpretability, we propose a Graph Neural Additive Network (GNAN) architecture, which processes individual reaction components through shared neural networks and computes their contributions using a reaction-level context embedding. The final yield is predicted by aggregating these contributions. The model is trained with a tailored objective that emphasizes the impact of key components while reducing the influence of less significant ones.
• This GitHub repository provides running examples of the proposed method on the Buchwald-Hartwig and Suzuki-Miyaura datasets.

Components

• data/split/* - original data files
• data/* - processed data files
• data/get_data.py - script for preprocessing the data files
• model/* - trained model files
• result/* - interpretation results
• dataset.py - data structure & functions
• model.py - model architecture & training/inference functions
• util.py - functions used across scripts.
• run.py - script for model training & performance evaluation
• visualization.ipynb - Jupyter Notebook example for visualization of interpretatations

Usage Example

Data processing

• Train/test splits for the Buchwald-Hartwig and Suzuki-Miyaura datasets are located in the ./data/split/ directory.
• To preprocess the data, run the following command:

python ./data/get_data.py

• The processed data files are stored in the ./data/ directory.

Training an interpretation model

• To train the model on the Buchwald-Hartwig dataset, run the following command (data_id: 1 (Buchwald-Hartwig), split_id: 0, train_size: 2767):

python run.py -d 1 -s 0 -t 2767

• To train the model on the Suzuki-Miyaura dataset, run the following command (data_id: 2 (Suzuki-Miyaura), split_id: 0, train_size: 4032):

python run.py -d 2 -s 0 -t 4032

• The trained models are stored in the ./model/ directory.
• The predictions and interpretations are stored in the ./result/ directory.

Dependencies

• Python
• Pytorch
• DGL
• RDKit

Citation

@Article{Kwon2025,
  title={Interpretation of chemical reaction yields with graph neural additive network},
  author={Kwon, Youngchun and Jung, Yongsik and Choi, Youn-Suk and Kang, Seokho},
  journal={Machine Learning: Science and Technology},
  volume={6},
  pages={025054},
  year={2025},
  doi={10.1088/2632-2153/addfaa}
}