De Novo Design of Membrane-Targeting Antimicrobial Peptides Against Gram-Negative Bacteria Using a Generative Artificial Intelligence Framework.

Antimicrobial resistance (AMR), particularly Gram-negative bacteria, poses significant challenges due to their robust outer membranes limiting antibiotic efficacy. Antimicrobial peptides (AMPs) show promising potential to replace traditional antibiotics. This study proposes a multi-condition constrained directed generation framework guided by the AMP antimicrobial mechanisms for designing membrane-targeting antimicrobial peptides (MTAMPs) against Gram-negative bacteria. By integrating sequence information, physicochemical properties, and spatial structure (PCSS) descriptors related to the outer membrane, a conditional variational autoencoder (GenMTAMP) model was developed for de novo MTAMP design. Then, target PCSS descriptors within a predefined range are input as conditional constraints into the GenMTAMP model to direct generate MTAMPs. Candidate MTAMPs were screened and evaluated through subsequent identification (ClaAMP) and prediction (PreAMP) modules. Experimental validation showed that two top-ranked peptides named MTAMP003 and MTAMP004, exhibit excellent antibacterial activity against Gram-negative bacteria while maintaining low cytotoxicity and haemolytic activity toward mammalian cells. Furthermore, mechanism research indicates MTAMPs can disrupt the Gram-negative bacterial outer membrane barrier, with limited impact on mammalian cell membranes. In summary, the current research establishes a targeted and efficient generative artificial intelligence (AI) framework for de novo MTAMP design, and provides a generalisable framework for the rational design of AMPs with predefined functional properties.