Framework for evaluating explainable AI in antimicrobial drug discovery

Explainable artificial intelligence (XAI) methods for molecular property prediction lack standardized evaluation criteria, preventing widespread deployment in drug development and hit optimisation, where proper understanding of structure-activity relationship is essential. We developed an evaluation framework for XAI using fragment-based explainability tests to compare XAI with different molecular representation and challenge the different XAI approaches for proper explanation of activity cliffs. The evaluation methods include essential scaffold recognition, scaffold sensitivity and substructure specificity for explaining activity cliff, and technical evaluation on model robustness and consistency. Using a curated dataset of antibiotic molecules we established three XAI models with fundamentally different molecular representation: Random Forest on chemical features using SHAP, CNN on sequence-based SMILES using token occlusion, and RGCN on molecular graphs with substructure masking. Together with detailed case study, we evaluated the explainability behaviours and quality of the different XAI approaches and highlighted their limitations. While all XAI approaches displayed good predictive and scaffold recognition capabilities, and comparable robustness and consistency, they displayed quite different explainability behaviour for activity cliffs, revealing their different utility for medicinal chemistry.