A generative artificial intelligence approach for the discovery of antimicrobial peptides against multidrug-resistant bacteria.

Abstract

The discovery of novel antimicrobial peptides (AMPs) against clinical superbugs is urgently needed to address the ongoing antibiotic resistance crisis. AMPs are promising candidates due to their broad-spectrum activity, rapid bactericidal mechanisms and reduced likelihood of inducing resistance compared with conventional antibiotics. Here, a pre-trained protein large language model (LLM), ProteoGPT, was established and further developed into multiple specialized subLLMs to assemble a sequential pipeline. This pipeline enables rapid screening across hundreds of millions of peptide sequences, ensuring potent antimicrobial activity and minimizing cytotoxic risks. Through transfer learning, we endowed the LLMs with different domain-specific knowledge to achieve high-throughput mining and generation of AMPs within a unified methodological framework. Notably, both mined and generated AMPs exhibited reduced susceptibility to resistance development in ICU-derived carbapenem-resistant Acinetobacter baumannii (CRAB) and methicillin-resistant Staphylococcus aureus (MRSA) in vitro. The AMPs also showed comparable or superior therapeutic efficacy in in vivo thigh infection mouse models compared with clinical antibiotics, without causing organ damage and disrupting gut microbiota. The mechanisms of action of these AMPs involve disruption of the cytoplasmic membrane and membrane depolarization. Overall, this study presents a generative artificial intelligence approach for the discovery of novel antimicrobials against multidrug-resistant bacteria, enabling efficient and extensive exploration of AMP space.