The number of infections caused by antibiotic-resistant pathogens is increasing alarmingly every year and will continue to grow. Antimicrobial peptides (AMPs) are considered new therapeutic agents for the effective combat against infectious diseases. We report the design of antimicrobial peptides derived from a database of leech metagenome proteins using a machine learning approach. Peptides with antimicrobial activity and reduced toxicity were identified by the CatBoost algorithm. Among them, Hm-AMP2 possesses the most promising application in practice, including antibacterial activity against both Gram-positive and Gram-negative bacteria, low cytotoxic and hemolytic effects. Hm-AMP2 kills various bacteria at low concentrations (4.6–18.5 μM) by the disruption of bacterial membranes. According to nuclear magnetic resonance analysis, the peptide adopts an α-helical structure in a membrane environment. Hm-AMP2 interacts with lipopolysaccharides of different bacteria according to microscale thermophoresis and CD spectroscopy analysis. This effect can play a role in the first defense against the organism’s bacterial invasion. The computational approach developed for the identification of AMP can be useful for the rational design of effective non-toxic peptide antibiotics.