Explainable Graph Convolutional Neural Network for a Mixture of Molecules

Abstract

Use of the Graph Convolutional Neural Network, e.g. Message Passing Neural Network (MPNN), is growing and the needs for improved interpretability and expansion of the scope of application is gaining. For the purpose of extending the scope of MPNN to the mixture of molecules, I propose a framework to predict the property of the mixture of molecules using an explainable model for MPNN based on gradient-weighted Class Activation Mapping and the novel explainable neural network architecture which can input a mixture of molecules. I evaluated the explanation quality of the molecule property prediction of MPNN and the mixture property prediction of the proposed neural network. As a result, I confirmed the consistency of the predicted explanation and the domain knowledge of the molecular property. I also confirmed that by using the proposed framework, one can obtain a graph based explanation for mixture property, which is consistent with domain knowledge.