Chemically intuitive explainable machine learning for molecular modelling

Predictions made by machine learning models are often difficult to understand, hence they are referred to as “black box” models. This lack of transparency restricts the acceptance of model predictions for interdisciplinary environments such as drug discovery. As a result, there has been growing interest in novel methods from the field of explainable artificial intelligence (XAI) that help to rationalize model decisions. Several XAI approaches have been applied to models used for compound property prediction tasks such as compound activity or potency prediction. For instance, local feature attribution methods, including Shapley additive explanations (SHAP), have been extensively explored to quantify the importance of individual chemical features for predictions of test compounds. Other XAI approaches that are applicable for explaining predictions include, for example, the concept of counterfactuals or anchors, which have thus far only been little investigated in chemistry. However, if applied to molecules, these concepts can produce explanations that are chemically intuitive and accessible to non-experts. Herein, counterfactuals and anchors are introduced and their adaptation for explaining compound activity predictions is presented.