Discovery of naturally inspired antimicrobial peptides using deep learning.

Non-ribosomal peptides (NRPs) are promising lead compounds for novel antibiotics. Bioinformatic mining of silent microbial NRPS gene clusters provide crucial insights for the discovery and de novo design of bioactive peptides. Here, we describe the efficient discovery and antibacterial evaluation of novel peptides inspired by metabolite scaffolds encoded by NRPS gene clusters from 216,408 bacterial genomes. In total, 335,024 NRPS gene clusters were identified and dereplicated, yielding 328 unique peptide scaffolds. Using deep learning-based scoring, five antimicrobial peptide candidates (P1-P5) were synthesized via solid-phase chemical synthesis. Among them, peptide P2 exhibited potent antibacterial activity with MIC values of 1-2 μM against two pathogenic strains. Subsequent amino acid optimization guided by deep learning algorithms produced P2.2, a derivative with significantly enhanced antibacterial activity. Mechanistic studies revealed that P2.2 disrupts bacterial membranes and increases permeability by modulating proteins involved in the type VI and III secretion systems. Furthermore, P2.2 demonstrated synergistic effects when combined with conventional antibiotics and exhibited reduced hemolytic activity, improving its therapeutic potential. These findings underscore the immense potential of deep learning to accelerate the discovery of naturally inspired antimicrobial peptides from silent biosynthetic gene clusters.