Exploring New Chemical Space for Antibiotics with Active Learning and Bacterial Phenotypic Profiling

Antimicrobial resistance is a major global health threat, yet no antibiotics with a novel mode of action have been approved in the last 20 years. While machine learning (ML) accelerates drug discovery by optimizing molecules in known chemical spaces, it struggles to explore novel spaces where new mechanisms of action might exist. We use Generative flow networks (GFlowNets), a novel ML architecture, to sample chemical space in proportion to a reward function (e.g., predicted antibiotic activity). In this way, compounds with low antibiotic activity, which are discarded as inactive in traditional screening, still provide information that can point in the direction of new antibiotic activity peaks. This approach uncovers pathways to molecules with novel activity. To enhance training, we employ bacterial cell painting, which uses fluorescent dyes to generate detailed phenotypic profiles of compound effects at high throughput. By linking these microscopy profiles to known antibiotics and whole genome CRISPRi depletion data, ML models can infer mechanisms of action. Using high-throughput microscopy screening and iterative active learning loops, we aim to identify and validate new antibiotic candidates in unexplored chemical spaces.