De novo design of lipopeptides with antimicrobial potency enabled by a chain-as-prompt deep learning framework

Antimicrobial lipopeptides (AMLPs) represent a promising solution to global antibiotic crisis, yet their precision design remains challenging due to the inability of current computational models to account for N-terminal fatty acid modification. Here we develop a deep learning-driven AMLP model for de novo design of AMLPs by integrating a chain-as-prompt mechanism with domain-adaptive transfer learning, effectively bridging the representation gap between peptide sequence and lipid-induced bioactivity. Using the AMLP model, 2657 evolutionarily selected RiPP-derived core peptide sequences from 96,911 bacterial genomes are inspiringly designed for virtual lipidation with seven N-terminal fatty acid chains (C6–C18) followed by predictive screening, generating 6197 candidate sequences. Notably, we find that both high KRH rate ( ≥ 0.3) and common fatty acid tails (C12 and C16) are also critical for further selection of candidate AMLPs, resulting in 289 sequences. Then, 46 are randomly synthesized, with 44 showing antimicrobial activities (a positive rate of 95.7%). The lead antibiotic LP630 exhibits broad-spectrum activities against typical Gram-positive, Gram-negative and mycobacterial strains by targeting multiple bacterial cell membrane components, and demonstrates potent therapeutic efficacy in murine wound infection models. Our study provides a streamlined and universal paradigm for deep learning-enabled high-throughput discovery of novel lipopeptide therapeutics. Authors present their chain-as-prompt deep learning framework, which identifies antimicrobial lipopeptides that show efficacy against drug-resistant bacterial pathogens in laboratory and animal model.