Fragment-Based Explainable AI for Pathogen-Selective Antimicrobial Drug Design

Abstract

The rise and extent of antimicrobial resistance demand computational tools that go beyond simple predictions of antimicrobial activity to deliver applicable medicinal chemistry insights for an accelerated and more efficient development of novel antimicrobials. Here, we present a fragment-based explainable artificial intelligence (XAI) framework, based on Relational Graph Convolutional Network (R-GCN) models trained with over 127,000 compounds from the Community for Open Antimicrobial Drug Discovery (CO-ADD) database and ChEMBL, targeting Staphylococcus aureus (Gram-positive), Escherichia coli (Gram-negative), and Candida albicans (fungal). External validation on 100,000 high-throughput screening compounds from the European Chemical Biology Database (ECBD) demonstrated the predictive performance of the three models and their utility in selecting compounds from a structurally divergent library, with enrichment factors up to 19-fold. Using substructure mask explanation (SME), we further decompose each prediction into fragment-based contribution scores and map them onto the chemical structures, allowing medicinal chemists to identify which scaffolds and substituents have a positive or negative effect on the activity against specific pathogens. From correctly predicted compounds, we further extracted pathogen-selective XAI fragments representing the core scaffolds responsible for activity, allowing a comparative analysis of these selective fragments between the different microbial classes and providing an evaluation tool for the antimicrobial potential of novel compound libraries. The developed XAI framework, together with the pathogen-specific predictive models, the pathogen-selective fragment lists, and their mapping onto the chemical structures, is able to provide a highly valuable and practical guide to the rational design of new antimicrobial agents.