Moremi Bio Agent: Leveraging Agentic Large Language Model for the Discovery of Broad-Spectrum Antibiotics for Enterobacteriaceae

Abstract

Antimicrobial resistance (AMR) is a pressing global health crisis, exacerbated by a stagnating antibiotic discovery pipeline and the emergence of multidrug-resistant pathogens such as Klebsiella pneumoniae. We hypothesize that dual-target strategies may offer a more robust means to overcome AMR by reducing the likelihood of resistance development compared to single-target approaches. In response, we leveraged Moremi Bio Agent, an agentic large language model (LLM) for the autonomous design, in silico validation, and prioritization of broad-spectrum antibiotics targeting Enterobacteriaceae. Using our proposed dual-target strategy, we generated and evaluated 1,002 candidate molecules predicted to simultaneously inhibit the FabI enzyme and the AcrAB-TolC efflux pump--two key resistance mechanisms in Gram-negative bacteria. Our fully autonomous pipeline integrated compound generation, molecular docking, pharmacodynamics/pharmacokinetic predictions & toxicity profiling, and molecule-ranking based on ADMET and drug-likeness properties. Out of 1,002 molecules generated, 774 passed preliminary ADMET benchmarks, with majority of the 60 top-performing candidates (score[≥] 0.8) showing favorable drug-likeness, minimal toxicity. 391 of the compounds exhibited moderate binding interaction to both targets. This study demonstrates the feasibility of AI-driven antibiotic discovery and lays the foundation for future experimental validation to address AMR.