On The Interpretability of Graph Neural Networks in QSPR Modeling

Artificial intelligence-based models (AI) and in particular Graph Neural Networks (GNN) are considered a more promising approach for modeling molecular properties compared to the use of traditional descriptor-based models due to their enhanced ability to express the structural information and their ability to better generalize to unseen data. However, their ‘black box’ nature and the lack of transparency and interpretability could hinder their wider acceptance and usage. In this work, we combine knowledge-based molecular descriptors and two AI-based concepts, the junction tree model and the attention mechanism to produce an interpretable model. The model is trained on the enthalpy of formation of organic compounds and the insights gained are highlighted and compared to the insights gained from models with a higher level of interpretability such as the group-contribution models. The results obtained show consistency with the insight gained in the form of the relative importance of the molecular sub-structures to the overall property

Recommended citation: Fan, F., Aouichaoui, A. R., & Sin, G. (2022). On The Interpretability of Graph Neural Networks in QSPR Modeling. In Computer Aided Chemical Engineering (Vol. 51, pp. 1393-1398). Elsevier. https://doi.org/10.1016/B978-0-323-95879-0.50233-2