The search for antibiotics is urgent because of the global antibiotic resistance problem. While a variety of strategies are actively sought, interest in antimicrobial peptides persists due to high potency and low chance of resistance development. The establishment of the antimicrobial peptide database laid the foundation for peptide prediction and design. Both artificial intelligence and non-AI approaches have been demonstrated. Since AI remains a black box and does not teach us how to design peptide antibiotics, this study took a database-guided approach. Our peptide design benefited from the recent classification of peptides into hemolytic and nonhemolytic groups in the APD6. Our designed peptides rapidly killed Gram-negative bacteria Escherichia coli and Acinetobacter baumannii, but not Gram-positive methicillin-resistant Staphylococcus aureus, Staphylococcus epidermidis, and Bacillus subtilis. In addition, our peptide inhibited bacterial attachment, biofilm formation, and disrupted preformed biofilms. Remarkably, YZ200, designed based on the nonhemolytic group, showed no sign of hemolysis even at 400 μM, whereas YZ201, designed based on the hemolytic group, displayed toxicity. Our analysis uncovered a higher hydrophobic ratio for the hemolytic group. Mechanistic studies revealed that the peptide permeabilized and depolarized bacterial membranes. The predicted membrane-bound structure of YZ200 contains a longer amphipathic helix, explaining its higher potency than YZ201. By comparing our experimental results for the designed peptides with AI-predicted activity and toxicity outcomes, it becomes evident that great progress has been made for AI prediction of antimicrobial peptides and such predictions will be improved in the future by including good data as illustrated herein.
Mechesso et al. (2026) studied this question.