Painting Peptides With Antimicrobial Potency Through Deep Reinforcement Learning

Abstract In the post‐antibiotic era, antimicrobial peptides (AMPs) are considered ideal drug candidates because of their lower likelihood of inducing resistance. Computational models provide an efficient way to design novel AMPs. However, current optimization and generation approaches are tailored for specific application scenarios, which hinders the ease of use. To address this challenge, a novel AMP design model named AMPainter is proposed. Based on deep reinforcement learning, AMPainter integrates optimization and generation tasks in a unified framework. AMPainter is applied to three types of peptides, including known AMPs, signal peptides (SPs), and random sequences. AMPainter outperforms ten related models in enhancing the activity of known AMPs on the predicted antimicrobial potency and diversity. Several AMPs demonstrate a 128‐fold decrease in their actual minimal inhibitory concentrations (MICs). AMPainter evolves effective AMPs from membrane‐active SPs with an experimental success rate of 80%. In terms of generation, de novo designed AMP from an inactive random sequence achieves an average MIC of 2.88 µM against four bacteria. In vitro MICs of peptides along the virtual evolutionary path match the predicted scores. Therefore, AMPainter can significantly improve the antimicrobial potency of various peptides, expand the AMP sequence space, and discover novel antimicrobial agents.