Decoding Chemical Predictions: Group Contribution Methods for XAI

Abstract

Graph Neural Networks (GNNs) have recently shown great promise for modeling chemical systems. However, beyond the accuracy and performance of these models, understanding their underlying mechanisms is also crucial. While many general GNN explainers exist, incorporating domain-specific knowledge can enhance the development of explainers tailored to chemical applications. In this study, we developed an approach based on the well-established concept of group contributions, providing additional explanations without compromising model accuracy. Our results indicate that different GNN models may learn distinct patterns from the molecules. Furthermore, by applying a custom loss function, we successfully aligned the learning process of the models with desired group contributions while maintaining the overall model performance.