Accelerating the Discovery of Antimicrobial Peptides against Foodborne Pathogens via a Stacked Ensemble Learning Framework with Multimodal Feature Filtering

Recurrent outbreaks of foodborne pathogens substantially threaten public health, leading to the overuse of antibiotics and hence the selection of multidrug-resistant strains. In response to the urgent demand for novel antimicrobial agents, we herein report the development of a computational-experimental closed-loop framework for the accelerated discovery of antimicrobial peptides. Utilizing a multi-modal feature-stacked ensemble learning model (AUC: 0.9951), we analyzed 11,822 Streptomyces genomes and predicted an antimicrobial peptide, Stre5, from 54,163 candidate sequences. Stre5 showed potent bactericidal activity against E. coli with a minimum inhibitory concentration (MIC) of 32 µg/mL, alongside low cytotoxicity and negligible hemolytic activity at its effective dose. Furthermore, stability assays demonstrated that Stre5 maintains robust antimicrobial efficacy across a wide range of temperatures (up to 80 °C), pH levels (3-8), and in the presence of various metal ions. Scanning electron microscopy, propidium iodide staining, and molecular dynamics simulations all showed that Stre5 disrupts bacterial membrane integrity through a carpet model mechanism. This combined in silico and in vitro strategy provides a powerful tool for combating antibiotic resistance and offers a promising candidate for novel food bio-preservatives.