APPLICATION OF DEEP REINFORCEMENT LEARNING FOR THE RATIONAL DESIGN OF MAGAININ ANALOGUES WITH ANTIMICROBIAL POTENTIAL

Antimicrobial resistance (AMR) constitutes a growing global health threat, driven by the inevitable evolution of microorganisms under the selective pressure imposed by antimicrobial use. The rapid dissemination of resistant organisms compromises the effectiveness of available treatments and contributes to increased morbidity and mortality. It is estimated that antimicrobial-resistant infections are associated with approximately 700,000 deaths annually, underscoring the urgent need for novel therapeutic strategies. Antimicrobial peptides (AMPs), widely distributed among plants, animals, and microorganisms, are important components of the innate immune system and exhibit activity against a range of pathogens, including bacteria, fungi, and viruses such as HIV-1, influenza viruses, and coronaviruses. Among these molecules, magainin-2, originally isolated from Xenopus laevis, is one of the most studied AMPs. However, its clinical application is limited by insufficient selectivity at relevant concentrations. In this context, the rational design of magainin-derived analogues represents a strategy for enhancing therapeutic selectivity while preserving antimicrobial efficacy. In this work, peptides derived from the magainin sequence were designed and selected using a Deep Reinforcement Learning (DRL)-based generative approach. The model was employed to predict sequence modifications aimed at enhancing antimicrobial activity through cycles of sequence generation and in silico evaluation. The 10 promising candidates were subjected to structural modeling using AlphaFold3 and molecular dynamics simulations with GROMACS in a bacteria-mimetic membrane. Peptides displaying α-helical and amphipathic conformations, with distinct polar and apolar faces, and favorable hydrophobicity indices and hydrophobic moments, were prioritized for analysis. Based on the integrated computational assessment, the ten promising peptide candidates were analyzed according to their predicted antimicrobial potency, low hemolytic activity, and structural and biological robustness. Antimicrobial activity and hemolytic potential were predicted using the CAMPR4 and HemoPI2 servers, respectively. This enabled the identification of peptide sequences with predicted antibacterial activity, while reducing hemolytic potential. The selected peptides will be synthesized by solid-phase peptide synthesis (SPPS) and evaluated in vitro against bacteria belonging to the ESKAPE group. This computational and experimental strategy is expected to provide a rational framework for the development of safer and more effective magainin-derived AMPs and to expand their potential applications in antimicrobial therapy and biotechnology.