Contrastive explanations for machine learning predictions in chemistry

The concept of contrastive explanations originating from human reasoning is used in explainable artificial intelligence. In machine learning, contrastive explanations relate alternative prediction outcomes to each other involving the identification of features leading to opposing model decisions. We introduce a methodological framework for deriving contrastive explanations for machine learning models in chemistry to systematically generate intuitive explanations of predictions in high-dimensional feature spaces. The molecular contrastive explanations (MolCE) methodology explores alternative model decisions by generating virtual analogues of test compounds through replacements of molecular building blocks and quantifies the degree of "contrastive shifts" resulting from changes in model probability distributions. In a proof-of-concept study, MolCE was applied to explain selectivity predictions of ligands of D2-like dopamine receptor isoforms.