Interpreting Graph Neural Networks for Antibacterial Drug Repurposing against Acinetobacter baumannii

Antimicrobial resistance in Acinetobacter baumannii, whose carbapenem-resistant strains are classified by the World Health Organization as a critical priority, has motivated machine-learning screens for new antibacterials, but translating predictions into chemical hypotheses requires evaluating the molecular features on which a model relies. A graph neural network ensemble achieves competitive predictive performance against a descriptor-augmented reference model on the same antibacterial screen while using only atom features and connectivity, allowing interpretation to focus on the molecular graph without a separate input of handcrafted descriptors. The atom-level explanations identify features supporting the model's activity predictions under perturbation and restoration, while comparisons with equally sized perturbations within each molecule distinguish piperazine as the most consistently supported recurring contributor to predicted activity. Prospective screening of the Drug Repurposing Hub followed by developability and safety assessment yields a ranked shortlist for experimental follow-up and leads to testing of CE3F4, which reduced A. baumannii growth by 57% at 0.25 mM in a preliminary assay with a non-monotonic concentration--response profile.