Reinforcement learning for the design of targeted antimicrobial peptides against resistant pathogens

Abstract

Antimicrobial peptides (AMPs) are considered the most promising alternative to traditional antibiotics, addressing the issue of antibiotic resistance through diverse mechanisms that differ from conventional antibiotics. However, failures in clinical trials of AMPs highlight the importance of balancing low toxicity with high activity.

This study proposes a novel AMP generation model that integrates reinforcement learning, Generative Pretrained Transformer (GPT), and Low-Rank Adaptation (LoRA) to design peptides that account for both activity and non-hemolytic properties for each strain. The goal of this thesis is to contribute to the development of potential AMPs by ensuring activity and simultaneously generating non-hemolytic peptides tailored to individual strain.

To validate the strain-aware generation, the embeddings of peptides generated for each strain were visualized using t-SNE plots. Furthermore, the model demonstrated the ability to generate peptides satisfying both activity and non-hemolytic conditions for unseen strains, highlighting its potential to design AMPs for emerging pathogens.