Interpreting Graph Neural Networks with Myerson Values for Cheminformatics Approaches


Here we introduce a novel method to interpret the predictions of graph neural networks (GNNs) based on Myerson values from cooperative game theory. Myerson values are closely related to Shapley values and thus provide an interpretability approach similar to the SHAP values. We developed the technique for applications in drug discovery, but it can be used with any graph. Using the GNN as a coalition game and the interpreted graph as the cooperation structure, the Myerson values determine the worth of each node of the graph. The worth of all nodes of the graph adds up to the predicted value of the model, allowing for a simple and intuitive interpretation of the prediction. To interpret predictions on molecular graphs we show visual explanations on molecular structures using two molecular datasets.